paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1704.04199 | 2 | 1704 | 2017-04-19T00:53:16 | Evolution and Analysis of Embodied Spiking Neural Networks Reveals Task-Specific Clusters of Effective Networks | [
"q-bio.NC",
"cs.NE"
] | Elucidating principles that underlie computation in neural networks is currently a major research topic of interest in neuroscience. Transfer Entropy (TE) is increasingly used as a tool to bridge the gap between network structure, function, and behavior in fMRI studies. Computational models allow us to bridge the gap even further by directly associating individual neuron activity with behavior. However, most computational models that have analyzed embodied behaviors have employed non-spiking neurons. On the other hand, computational models that employ spiking neural networks tend to be restricted to disembodied tasks. We show for the first time the artificial evolution and TE-analysis of embodied spiking neural networks to perform a cognitively-interesting behavior. Specifically, we evolved an agent controlled by an Izhikevich neural network to perform a visual categorization task. The smallest networks capable of performing the task were found by repeating evolutionary runs with different network sizes. Informational analysis of the best solution revealed task-specific TE-network clusters, suggesting that within-task homogeneity and across-task heterogeneity were key to behavioral success. Moreover, analysis of the ensemble of solutions revealed that task-specificity of TE-network clusters correlated with fitness. This provides an empirically testable hypothesis that links network structure to behavior. | q-bio.NC | q-bio | Evolution and Analysis of Embodied Spiking Neural Networks
Reveals Task-Specific Clusters of E(cid:128)ective Networks
Madhavun Candadai Vasu
Indiana University
Bloomington, Indiana 47405
[email protected]
Eduardo J. Izquierdo
Indiana University
Bloomington, Indiana 47405
[email protected]
7
1
0
2
r
p
A
9
1
]
.
C
N
o
i
b
-
q
[
2
v
9
9
1
4
0
.
4
0
7
1
:
v
i
X
r
a
ABSTRACT
Elucidating principles that underlie computation in neural networks
is currently a major research topic of interest in neuroscience. Trans-
fer Entropy (TE) is increasingly used as a tool to bridge the gap
between network structure, function, and behavior in fMRI stud-
ies. Computational models allow us to bridge the gap even further
by directly associating individual neuron activity with behavior.
However, most computational models that have analyzed embodied
behaviors have employed non-spiking neurons. On the other hand,
computational models that employ spiking neural networks tend to
be restricted to disembodied tasks. We show for the (cid:128)rst time the
arti(cid:128)cial evolution and TE-analysis of embodied spiking neural net-
works to perform a cognitively-interesting behavior. Speci(cid:128)cally,
we evolved an agent controlled by an Izhikevich neural network to
perform a visual categorization task. (cid:140)e smallest networks capable
of performing the task were found by repeating evolutionary runs
with di(cid:130)erent network sizes. Informational analysis of the best so-
lution revealed task-speci(cid:128)c TE-network clusters, suggesting that
within-task homogeneity and across-task heterogeneity were key to
behavioral success. Moreover, analysis of the ensemble of solutions
revealed that task-speci(cid:128)city of TE-network clusters correlated
with (cid:128)tness. (cid:140)is provides an empirically testable hypothesis that
links network structure to behavior.
CCS CONCEPTS
•Computing methodologies → Cognitive science; Arti(cid:128)cial
life; Evolutionary robotics; Model development and analy-
sis; Optimization algorithms; Cognitive robotics; •Applied com-
puting → Biological networks; •(cid:135)eory of computation → Evo-
lutionary algorithms;
KEYWORDS
Spiking Neural networks, Evolutionary algorithms, Transfer En-
tropy, Information (cid:140)eory, Evolutionary Robotics
Permission to make digital or hard copies of all or part of this work for personal or
classroom use is granted without fee provided that copies are not made or distributed
for pro(cid:128)t or commercial advantage and that copies bear this notice and the full citation
on the (cid:128)rst page. Copyrights for components of this work owned by others than ACM
must be honored. Abstracting with credit is permi(cid:138)ed. To copy otherwise, or republish,
to post on servers or to redistribute to lists, requires prior speci(cid:128)c permission and/or a
fee. Request permissions from [email protected].
GECCO '17, Berlin, Germany
© 2017 ACM. 978-1-4503-4920-8/17/07...$15.00
DOI: h(cid:138)p://dx.doi.org/10.1145/3071178.3071336
ACM Reference format:
Madhavun Candadai Vasu and Eduardo J. Izquierdo. 2017. Evolution and
Analysis of Embodied Spiking Neural Networks Reveals Task-Speci(cid:128)c Clus-
ters of E(cid:130)ective Networks. In Proceedings of GECCO '17, Berlin, Germany,
July 15-19, 2017, 8 pages.
DOI: h(cid:138)p://dx.doi.org/10.1145/3071178.3071336
1 INTRODUCTION
(cid:140)ere is an interest like never before to understand how nervous
systems produce behavior. Although many aspects of the environ-
ment and behavior are largely described with continuous variables,
including for example the continuous motion of our bodies and
other objects in the environment, the neurons that process this
information in many organisms operate through discrete-like ac-
tion potentials. One of the grand challenges of neuroscience is to
understand how this continuous stream of information from the
environment is represented and processed by spikes and converted
back into (cid:131)uid behaviors.
One of the most useful tools has been the ability to quantify
information transfer by action potentials through the use of infor-
mation theory [11]. Transfer entropy (TE) [19, 37] is an information-
theoretic measure that is being used extensively for estimating e(cid:130)ec-
tive networks that emerge from task-speci(cid:128)c interactions between
the underlying structural network elements at di(cid:130)erent spatial and
temporal scales [13, 22, 23, 30, 33–36]. Most of the work analyzing
information processing in the brain has been empirical, focusing
on three main techniques: fMRI, in vitro, and in vivo studies. While
fMRI studies have yielded lots of insights into whole-brain orga-
nization, development and pathology, network nodes are de(cid:128)ned
at the macro or meso-scale where each node corresponds to hun-
dreds or thousands of neurons [29, 33]. In vitro studies have helped
understand the micro-level structural organization and (cid:131)ow of in-
formation across small networks of neurons, but it is not possible
to study neural dynamics in the context of behavior [30, 35]. Fi-
nally, in vivo studies make it possible to record micro-scale activity
from behaving animals, but resolving the cortical network that is
involved in a behavior and recording from all neurons involved in
that particular behavior is not feasible yet [36].
Over the past few decades, theoretical neuroscientists have be-
gun to address these issues by studying computational models of
networks of spiking model neurons. However, most of this work
has focused on characterizing the dynamics of the abstract net-
work, without any substantial connections to function [5, 9, 28, 38].
More recently, work has begun to focus on developing functional
spiking neural networks [20]. However, this has been generated
largely for disembodied networks: a network receives a time-series
GECCO '17, July 15-19, 2017, Berlin, Germany
Candadai Vasu et. al.
input, and its task is to generate a speci(cid:128)ed output [8]. (cid:140)ese mod-
els address how speci(cid:128)c pa(cid:138)erns of neural activity are generated
by sensory stimuli or as part of motor actions, but lack the con-
tinuous closed-loop interaction between the brain, body and an
environment. Incorporating these components to address how a
neural circuit works, requires us to develop behaviorally-functional
spiking networks. (cid:140)e view that the body and the environment
play a crucial role in the understanding of behavior is increasingly
accepted [6, 18]. In parallel, there has been some work that has
focused on understanding behavior through the development and
analysis of whole brain-body-environment models [3]. However,
with only a few exceptions [12, 15], the majority of this work has
used non-spiking neural models.
(cid:140)e work presented here takes an entirely di(cid:130)erent approach to
the study of behaviorally-functional spiking neural networks. We
use an evolutionary algorithm to evolve spiking neural controllers
for simulated agents performing a visual categorization task. (cid:140)ere
are numerous bene(cid:128)ts of using this kind of approach to study the
neural basis of behavior. (a) Brain-body-environment models make
it possible to relate neural activity to behavior, because unlike em-
pirical studies, all variables of the system are easily accessible. (cid:140)is
allows us to take seriously the view that cognition is situated, em-
bodied, and dynamical [6, 10, 32]. (b) By designing tasks that are
deliberately minimal and yet of interest to cognitive scientists [2],
we can begin to address questions about information processing
in the brain that are relevant to understanding cognition. (c) (cid:140)e
use of a spiking neuron model, which can replicate the dynamics of
Hodgkin–Huxley–type neurons with the computational e(cid:129)ciency
of integrate-and-(cid:128)re neurons [17], allow us to use the same tools
used by neuroscientists to analyze the model (e.g., Transfer En-
tropy). (d) By evolving agents, instead of hand-designing them, we
are able to make minimal prior assumptions about how various
behaviors must be implemented in the circuit. (cid:140)is maximizes the
potential to reveal counter-intuitive solutions to the production
of behavior. (cid:140)e evolutionary algorithm was used to determine
the values of the electrophysiological parameters that optimize a
behavioral performance measure. Further, the state-of-the-art in
supervised learning of spiking neural networks is not capable of
optimizing the kind of embodied models we are building [40]. (e)
Typically, each successful evolutionary search produces a distinct
set of parameter values, leading to an ensemble of successful models
produced over several runs. (cid:140)e properties of this ensemble can
be analyzed to identify multiple ways of solving the same problem
and to identify recurring principles that underlie the production of
behavior.
(cid:140)is paper has two broad primary aims. First, to develop a
behaviorally-functional spiking neural network for a cognitively-
interesting behavior. Second, to analyze the ensemble of solutions
using transfer entropy and derive robust pa(cid:138)erns that underlie
di(cid:130)erent approaches to performing the behavior. We focus on a
visual categorization behavior used in previous computer simula-
tion studies [3]. Categorical perception involves partitioning the
sensed world through action [14]. (cid:140)at is, the continuous signals
received by the sense organs are sorted into discrete categories,
whose members resemble one another more than they resemble
members of other categories. (cid:140)e behavior involves two tasks:
catching circle-shaped objects and avoiding line-shaped objects. In
speci(cid:128)c, we are interested in the following questions: (1) Does TE
form an explanatory bridge between structure and function? (cid:140)at
is, does the inferred e(cid:130)ective network tell us something that cannot
be observed from the structural network alone about how behavior
is produced? (2) Do di(cid:130)erent aspects of the behavior (e.g., catching
some objects versus avoiding others) have their own specialized
e(cid:130)ective networks? In other words, are the e(cid:130)ective networks for
minor variations of one task (say, circle catching) more similar to
each other than the e(cid:130)ective networks of minor variations of the
other task (say, line avoiding)? (3) Is having task-speci(cid:128)c e(cid:130)ective
networks indicative of the agent's performance? More generally,
can the degree of task-speci(cid:128)city of the e(cid:130)ective networks predict
behavioral performance?
(cid:140)e rest of this paper is organized as follows. In the next section,
we describe the agent, neural network model, evolutionary algo-
rithm, and transfer entropy analysis used here. In the following
section, we discuss the results of the evolutionary simulations, (cid:128)rst
by examining a particular case and then generalizing to an analysis
of the ensemble of evolutionary runs. Finally, we discuss our results
in light of the work on multifunctional networks and on focusing
empirical experimental design, and we outline ongoing and future
work.
2 METHODS
(cid:140)is section outlines the technical speci(cid:128)cations of the agent, en-
vironment, and task; the neural controller; the evolutionary algo-
rithm; and the information-theoretic tools that were used in analy-
sis. (cid:140)e entire simulation was programmed in C++ and analyses
were carried out using MATLAB and Python.
2.1 Agent, Environment, and Task
In previous work [2, 3], model agents were evolved that could
"visually" discriminate between objects of di(cid:130)erent shapes, catching
some while avoiding others. (cid:140)ese experiments were designed to
produce evolved examples of categorical perception [3]. All details
of the agent, environment, and task have been adapted from these
previous studies.
(cid:140)e agent has a circular body with a diameter of 30, and can
move horizontally as objects fall from above (Figure 1A). (cid:140)e agent's
horizontal velocity is proportional to the sum of opposing forces
produced by two motors. (cid:140)e agent also has an "eye" which consists
of 7 vision rays evenly distributed over an angle of π/6. (cid:140)ese rays
extend out from the agent's body with a maximum range of 220.
If an object intersects a ray within this range, an external input
is fed to a corresponding sensory neuron. (cid:140)e value of the input
is inversely proportional to the distance at which the intersection
occurs, normalized from 0 to 10.
(cid:140)ere are two kinds of objects in the world: circular objects and
line objects. Circular objects have a diameter of 30 and line objects
have length 30. (cid:140)ese objects fall from a height of 275 at some
initial horizontal o(cid:130)set with respect to the agent. Objects fall with
a constant vertical velocity of -3 and no horizontal motion.
2.2 Neural Controller
(cid:140)e agent's behavior is controlled by a 3-layer neural network
(Figure 1B). (cid:140)e network architecture consists of seven sensory
Behaviorally-Functional Networks of Spiking Neurons
GECCO '17, July 15-19, 2017, Berlin, Germany
neurons fully connected to N fully interconnected interneurons,
which are in turn fully connected to two motor neurons.
(cid:140)ere are 7 sensory neurons in the top layer which are stimulated
by the agent's vision ray. (cid:140)ey follow the state equation:
i = 1, ..., 7
τs (cid:219)si = −si + Ki(x, y)
(1)
where si is the state of sensory neuron i, τs is the time constant that
is shares across all sensory neurons, Ki(x, y) is the sensory input
th ray due to an object at location (x, y) in agent-centered
from the i
coordinates, and the dot notation over the state variable indicates
the time di(cid:130)erential d
dt
(cid:140)e seven sensory neurons project down to a middle layer of N
fully interconnected Izhikevich spiking neurons with the following
two-dimensional system of ordinary di(cid:130)erential equations [16]:
.
(cid:219)vi = 0.04v
i + 5vi + 140 − ui + Si + Ii
2
i = 1, ..., N
(cid:219)ui = a(bvi − ui)
with the auxiliary a(cid:137)er-spike rese(cid:138)ing
(cid:40)
v ← c
u ← u + d
if v ≥ 30mV,
then
with each interneuron receiving weighted input from each sensory
neuron:
and from other spiking interneurons:
Si =
ji σ(sj + θs)
ws
7
j=1
N
j=1
Ii =
wi
jioi
where vi is representative of the membrane potential of spiking
neuron i, ui represents its membrane recovery variable, ws
is the
ji
strength of the connection from the jth sensory neuron to the ith
spiking interneuron, θs is a bias term shared by all sensory neurons,
σ(x) = 1/(1 + e−x) is the standard logistic activation function,
is the strength of the recurrent connections from the jth to the
wi
ji
ith spiking neuron, and oi is the output of the neuron: 1 if vi ≥
30mV, and 0 otherwise. (cid:140)e sign of all outgoing connections from
an interneuron depend on its excitatory or inhibitory nature, as
identi(cid:128)ed by a binary parameter. Parameters a,b,c and d control the
type of spiking dynamics. For inhibitory neurons a ∈ [0.02, 0.1],b ∈
[0.2, 0.25],c = −65 and d = 2, whereas for excitatory neurons
a = 0.02, b = 0.2, c ∈ [−65,−50] and d ∈ [2, 8].
Finally, the layer of interneurons feeds into the two motor neu-
rons, with the following state equation:
N
hj
j=1
k =0
τm (cid:219)mi = −mi +
¯oj(t) =
1
hj
wm
ji
¯oj
i = 1, 2
oj(t − k)
Figure 1: Basic setup for the categorical perception experi-
ments. [A] Agent, environment, and task. (cid:135)e agent moves
horizontally while an object falls towards it from above. (cid:135)e
object can be one of two shapes: circle or line. (cid:135)e task con-
sists of catching circles and avoiding lines, adapted from [3].
(cid:135)e agent's sensory apparatus consists of an array of seven
distance sensors (dashed lines). [B] Neural architecture. (cid:135)e
distance sensors (black) fully project to a layer of fully in-
terconnected spiking interneurons (red), which in turn fully
project to the two motor neurons (light gray).
where mi represents the motor neurons, wm
is the strength of the
ji
connection from the jth spiking interneuron to the ith motor neuron,
¯oj represents the moving average over a window of length hj for
the output of spiking interneuron j.
Finally, the di(cid:130)erence in output between the motor neurons
results in an instantaneous horizontal velocity that moves the agent
in one direction or the other.
Altogether, a network with N interneurons has a total of P =
2 + 14N + 4 parameters. Unlike previous work [2, 3], bilateral
N
symmetry in network parameters was not enforced. States were
initialized to 0 and circuits were integrated using the forward Euler
method with an integration step size of 0.1.
2.3 Evolutionary Algorithm
A real-valued genetic algorithm was used to evolve the controller
parameters: connection weights, biases, time constants, and in-
trinsic neuron parameters. Agents were encoded as P-dimensional
vectors of real numbers varying from [-1, 1]. Each vector element
linearly mapped to a parameter of the circuit: interneuron and
motor neuron biases ∈ [−4, 4], sensory neuron biases ∈ [−4,−2],
time-constants ranged ∈ [1, 2], and connection weights ∈ [−50, 50].
As mentioned previously, the range of the intrinsic parameters of
the Izhikevich neurons and the polarity of their weights depend on
a binary parameter that decides the inhibitory/excitatory nature of
the neuron. Parents were selected with a rank based mechanism,
with an enforced elitist fraction of 0.04. O(cid:130)spring were generated
from uniform crossover of two parents (probability 0.5). A Gauss-
ian distributed mutation vector was applied to each parent (µ = 0,
2 = 0.5).
σ
Fitness was calculated across 48 trials with objects dropped uni-
formly distributed over the range of horizontal o(cid:130)sets [-50,+50].
j=1 pi/48, where
pi = 1 − di for the objects that need to be caught and pi = di for
the objects that need to be avoided, and di is the horizontal distance
(cid:140)e performance measure to be maximized was:48
(2)
(3)
(4)
(5)
(6)
(7)
BA GECCO '17, July 15-19, 2017, Berlin, Germany
Candadai Vasu et. al.
between the centers of the object and the agent when their vertical
separation goes to zero on the ith trial. (cid:140)e distance di is clipped
to a maximum value of 45 and normalized to run between 0 and 1.
2.4 Transfer Entropy
Transfer entropy (TE) is an information-theoretic measure which
has received recent a(cid:138)ention in neuroscience for its potential to
identify e(cid:130)ective connectivity between neurons [24]. Intuitively,
TE from neuron J to I is a measure of the additional information
provided by the activity in neuron J over and above the information
from I's own history of activity that helps predict the activity of
neuron I. (cid:140)is proportional increase, when high, corresponds to a
causal in(cid:131)uence of J over I. TE is especially useful in taking into
account non-linear interactions between neural units and provides a
directed measure of in(cid:131)uence from one neuron to another. Further,
due to synaptic delays, the causal in(cid:131)uence of one neuron over
another can only be detected if tested at that corresponding delay.
In order to account for this, a modi(cid:128)ed version of TE was proposed
by Ito et. al [35]. We used the MATLAB toolbox provided by the
these authors to estimate TE. (cid:140)is involved utilizing the history of
neuron J over di(cid:130)erent time delays to predict the future activity of
I and then picking the peak-TE over all delays, as follows:
p(it it−1, jt−d)
TEJ→I(d) =
p(it , it−1, jt − d)log2
p(it it−1)
(8)
where it denotes a binary spike/no-spike activity of neuron I at
time t, where jt denotes a binary spike/no-spike activity of neuron
J at time t, d denotes the synaptic time delay, p(x) is the probability
of that set of spiking events occurring at that particular times, and
p(xy) is the conditional probability that a set of spiking events
occur given that certain other events have occurred.
2.5 Cluster Specialization Coe(cid:129)cient
Successful agents were subject to TE analyses on a trial-by-trial
basis. (cid:140)at is, the trial-speci(cid:128)c TE network was estimated for one
trial of falling line or circle. (cid:140)is produces 48 TE networks from the
24 circle and 24 line trials. (cid:140)ese networks are then hierarchically
clustered using a vanilla agglomerative clustering algorithm in
MATLAB. (cid:140)is procedure produces a clustering tree diagram where
the leaves of the tree correspond to one of the 48 networks. (cid:140)e
leaves are successively connected to one another depending on
how close they are in TE network space, until they are all in one
cluster. (cid:140)ese trees are called dendrograms. For each successful
agent we produced a trial-by-trial TE dendrogram.
In order to quantify the task-speci(cid:128)city of the TE networks,
we measured the number of clusters that were unique to each
task. We cut the dendrogram at di(cid:130)erent places to partition the 48
networks into multiple clusters. (cid:140)en the members of the cluster
were sorted based on task. We identi(cid:128)ed the smallest number of
clusters that can be formed, such that all the networks in the cluster
correspond to networks that were inferred from the neural activity
of one and only one of the tasks. (cid:140)ese clusters are called task-
specialized clusters. For example, the dendrogram in Figure 5C
can be dissected at level 1, to partition the data into two task-
specialized clusters, each containing only networks corresponding
to one task. Further dissection of these clusters will yield more
Figure 2: Fitness distributions for the best evolved agents
from 100 evolutionary runs for network sizes 1-6. No net-
work with only one spiking neuron evolved to solve the task.
Networks containing as little as two spiking neurons were
found that could solve the task. Interestingly, larger net-
works worked just as well as networks of size two.
number of task-specialized clusters but the minimum number of
task-specialized that can be formed is 2. CSC is de(cid:128)ned based
on the ratio between the minimum number of task-specialized
clusters to the maximum number of task-specialized clusters that
can be formed. (cid:140)e maximum number of task-specialized clusters
that can be formed is equal to the number of networks i.e. one
network per cluster. (cid:140)erefore, the CSC for an agent is de(cid:128)ned
as CSC = 1 − cMin/cMax, where cMin is the minimum number of
task-specialized clusters and cMax is the maximum number of task-
specialized clusters.
3 RESULTS
3.1 Minimal spiking circuit for object
categorization task
In order to identify the smallest network that could perform the
visual categorization task, we ran 100 evolutionary runs with dif-
ferent spiking network sizes ranging from 1 through 6. In Figure 2
we show the best (cid:128)tness a(cid:137)er 1000 generations for each of the
evolutionary runs, for all the circuit sizes, as a smoothed histogram
distribution. Other than the circuits with only one spiking neuron,
all evolutionary runs produced circuits that could solve the task.
Interestingly, networks with only two spiking neurons evolved just
as reliably as larger networks.
3.2 Analysis of the best agent
We selected the highest (cid:128)tness individual ((cid:128)tness=97.8%) over all
100 runs from the network size N = 3 batch to analyze (cid:128)rst. (cid:140)e
spiking network for this individual was composed of 2 inhibitory
neurons and 1 excitatory neuron (Figure 3). (cid:140)e feed-forward
component shown in Figure 3A captures the sensory neurons to
interneuron weights and the interneuron to motor neuron weights
represented by the thickness of the edges connecting the nodes.
Figure 3B shows the fully-connected recurrent inter-layer weights
where thickness of the edges denote weights and the width of the
nodes denote the self-connection. In both cases, red edges denote
inhibitory weights and blue edges denote excitatory weights.
00.20.40.60.81Fitness051015Frequencyn=2n=3n=4n=5n=6n=1 Behaviorally-Functional Networks of Spiking Neurons
GECCO '17, July 15-19, 2017, Berlin, Germany
(cid:140)is agent perfectly di(cid:130)erentiated the circles and the lines by
"catching" all circles and moving away from all the lines (Figure 4A).
(cid:140)e horizontal o(cid:130)set distance between the agent and falling object
as it approaches the agent is 0 for all circles and is large for all
lines. (cid:140)e (cid:128)tness of the agent is not a perfect 100% simply because
that would entail catching the circles at exactly the center of its
30-unit wide body. (cid:140)e small dispersion at the end of the behavioral
traces in Figure 4A is representative of the agent's small deviation
from catching at its center. Visualizing the corresponding spiking
neuron activity (Figure 4B) shows that the inhibitory neurons (n1
and n2) have a higher (cid:128)ring rate than the excitatory neuron. (cid:140)is
is consistent with relative (cid:128)ring rates in biological neurons [39].
In order to probe the di(cid:130)erence in neural dynamics between
the two tasks, the agent's spiking neuron activity from each of
the 48 trials of falling lines and circles was recorded and analyzed.
We used Transfer Entropy to estimate the task speci(cid:128)c e(cid:130)ective
network for each trial. Since TE quanti(cid:128)es information transfer in
the comparable units of bits, we can directly compare TE networks
from spiking activity recorded while the agent is performing di(cid:130)er-
ent tasks. (cid:140)e e(cid:130)ective network estimated from one line-avoiding
trial (Figure 5A) shows that n3 is not part of the network. (cid:140)is can
be reconciled with its activity shown in magenta in Figure 4B which
shows that n3 is completely dormant during this task. (cid:140)e e(cid:130)ective
network for one circle catching trial (Figure 5B) on the other hand,
shows a fully connected network. Spiking activity during one of
the circle catching trials (orange traces in Figure 4B) shows that
the neural activity in n1 and n2 control the scanning behavior of
the agent, while n3's activity coincides with the agent's (cid:128)ne motor
control at the end of the trial when it needs to center itself with
the circle. (cid:140)is explains the lack of participation by n3 in the line
avoiding task because such (cid:128)ne motor skills are not required.
It is to be noted that depending on the activity in the neural
network, the estimated TE networks signi(cid:128)cantly di(cid:130)er from the
structural network. Another point to note is that the TE networks
also include recurrent self-connections and these are computed
simply based on a neuron's ability to predict its own behavior.
While TE networks for only one circle and one line trial are shown
here, we performed the same analysis for all trials on this best agent.
We found that all the observations made for the individual trials
were consistent across the rest of the trials.
3.3 TE network clusters show task
specialization
(cid:140)e relationship between di(cid:130)erent TE networks across all trials for
the best N = 3 agent was determined by clustering of the 48 TE
networks. (cid:140)is resulted in the networks ge(cid:138)ing organized by task.
(cid:140)is observation was made by (cid:128)rst estimating the task-speci(cid:128)c TE
networks for each of the 48 trials (24 circle TE networks and 24 line
TE networks). (cid:140)ese 48 networks were then clustered using a simple
hierarchical clustering algorithm. (cid:140)e dendrogram of the clustered
TE networks shown in Figure 4C shows an interesting property. All
TE networks of the same task (either circles or lines) fall under one
cluster before being grouped into the cluster of the other task. (cid:140)is
means that the same structural network e(cid:130)ectively presents itself
as very similar networks for variations of one task, which are all
di(cid:130)erent from the very similar e(cid:130)ective networks for the other task.
Figure 3: Structure of the best N = 3 network. [A] Feed-
forward component. Top layer represents the sensory neu-
rons. Middle layer represents spiking interneurons. Bot-
tom layer represents motor neurons. Excitatory connec-
tions shown in blue; inhibitory connections in red. (cid:135)e spik-
ing neurons are also colored according to whether they are
excitatory or inhibitory. [B] Interneuron recurrent compo-
nent. Strength of self-connection is shown by the thickness
of the ring on the node. Relative strengths of connections
are shown by the thickness of the edges in A and B.
In other words, there is high within-task homogeneity and high
across-task heterogeneity in the task-speci(cid:128)c e(cid:130)ective networks.
Naturally, the next step is to look for this phenomenon in other
N = 3 agents.
3.4 High CSC agents have high (cid:128)tness
We developed a metric to quantify and compare task-specialization
in TE network clusters, which henceforth we will refer to as cluster
specialization coe(cid:129)cient, or CSC. N = 3 agents that had high CSC
showed very high behavioral performance. (cid:140)e greater the value
of CSC, the greater the expression of within-task homogeneity
and across-task heterogeneity, with the maximum value being 0.95
corresponding to the 2 specialized clusters that purely have e(cid:130)ective
networks corresponding to one and only one task in each of them.
All agents, from the 100 evolutionary runs that had a CSC of 0.95
were collected and they were all high-performing agents. It is to be
noted that the structural networks of agents that had a high CSC
were highly degenerate in terms of the ratio of inhibitory-excitatory
neurons and yet exhibited this phenomenon.
3.5 High (cid:128)tness with high CSC is independent
of network size
(cid:140)e same analyses that were performed on the best N = 3 agent
and other N = 3 agents, were repeated for smaller (N = 2) and
AB GECCO '17, July 15-19, 2017, Berlin, Germany
Candadai Vasu et. al.
Figure 4: Behavioral dynamics of the best network. [A] Be-
havior. (cid:135)e horizontal o(cid:130)set between the agent and the
falling object is shown along the horizontal access. (cid:135)e ver-
tical axis denotes the time. As the object falls, it can be seen
that the orange traces (representing the circles) end up with
0 o(cid:130)set, meaning the the agent "caught" them and vice versa
for lines. Only 24 of the 48 trials are shown for clarity. Two
traces are highlighted. [B] Neural traces for example runs
highlighted in [A]. (cid:135)e neural activity is similar at the begin-
ning of the object's fall. Note that neuron 3 does not show
any activity for the line avoiding task but does spike during
the circle catching task.
larger networks (N = 4,5 and 6). Irrespective of the network size,
agents that have the maximum possible CSC of 0.95, consistently
showed high behavioral performance (Figure 6A). (cid:140)is establishes
the idea that high within-task homogeneity and high across-task
heterogeneity of task-speci(cid:128)c e(cid:130)ective networks yields high per-
formance in this task. (cid:140)e obvious next question is to ask if the
converse is true. Plo(cid:138)ing the distribution of CSCs for agents that
have ≥90% (cid:128)tness (Figure 6B) revealed that there exists solutions
that have CSCs as low as 0 that still perform well. However, a
majority of agents that perform well have a high CSC.
3.6 CSC positively correlates with behavioral
performance
In order to further ascertain the relationship between CSC and (cid:128)t-
ness we looked at the (cid:128)tness of all agents as a function of their CSC.
As shown in Figure 7 (cid:128)(cid:138)ing a straight line to the points on a CSC
versus Fitness plot, shows a positive correlation between an agent's
CSC and (cid:128)tness for all values of N . Based on this and the previous
result, we conclude that expressing specialized task-speci(cid:128)c e(cid:130)ec-
tive networks is not a necessary condition for high performance but
conversely, it is su(cid:129)cient for high performance. Given the generic
nature of the network and it's evolutionary optimization process,
it can also be said that this is true for all categorization tasks. It
Figure 5: Transfer entropy networks of the best N = 3 agent.
[A] E(cid:130)ective network for one line avoiding trial highlighted
in Figure 4A. Note that neuron 3 is not part of this e(cid:130)ective
network. (cid:135)is is consistent with its lack of activity for this
task, as shown in Figure 4B. [B] E(cid:130)ective network for one
circle catching trial highlighted in Figure 4A. [C] A dendro-
gram of the hierarchical clustering of trial-by-trial TE net-
works. (cid:135)e branches of the tree are color-coded by whether
the TE network was inferred from a circle catching task (or-
ange) or line avoiding task (magenta). It can be seen here
that the tree can be cut at the top level to produce 2 clus-
ters whose members are of only one task. (cid:135)e TE networks
shown in panels [A] and [B] are identi(cid:128)ed in this tree with
color matched asterisks.
is intuitive that networks that can e(cid:130)ectively express themselves
distinctly for each category can perform well.
4 DISCUSSION AND CONCLUSION
In this paper we present results from the evolutionary optimization
of embodied spiking neural networks for cognitively-interesting
behavior. We applied information theoretic tools to extract task
speci(cid:128)c network representations based on spiking dynamics in the
interneuron layer. (cid:140)is information showed the interesting char-
acteristic of being clustered by task. We developed a novel metric,
Cluster Specialization Coe(cid:129)cient (CSC), that quanti(cid:128)es the task-
speci(cid:128)city of TE networks by dissecting the hierarchical clustering
tree of trial-by-trial e(cid:130)ective networks. Further analysis of the
ensemble revealed a trend that shows positive correlation between
CSC and performance. While we also noticed that there are some
cases where agents performed well and had a low CSC, we found
no agents that performed poorly with a high CSC. Further, we only
claim correlation and not causation. However, in combination with
neuroscience literature on the existence of task-speci(cid:128)c e(cid:130)ective
networks in the macro scale [21], we hypothesize that biological
networks self-organize to have high within-task homogeneity and
100 mV10 secsABn1n2n30.0020.0040.0060.0080.010.0120.0140.0160.018BAC** Behaviorally-Functional Networks of Spiking Neurons
GECCO '17, July 15-19, 2017, Berlin, Germany
high across-task heterogeneity for the purposes of e(cid:129)cient catego-
rization of stimuli.
It is widely known that biological neural networks perform mul-
tiple functions using the same underlying structural neural cir-
cuit [4, 7]. Our work provides a framework to study the task-speci(cid:128)c
e(cid:130)ective networks that emerge from these otherwise structurally-
similar networks. While we have analyzed this system for di(cid:130)erent
tasks in a single behavior, it can be easily extended to multiple-
behaviors. (cid:140)e same analysis can be performed to study the task-
relevant variations in the neural dynamics and its robustness across
variations of the task. Another perspective to look at this from,
is that of neural reuse [1]. Presenting new tasks in evolutionary
time, would allow the networks to reuse behaviorally appropriate
functional components developed for previously evolved behaviors
for the new behavior. While it is obvious that the same structural
components are reused, functional reuse can be analyzed by compar-
ing the task-speci(cid:128)c e(cid:130)ective networks. Furthermore, systematic
studies of multiple behaviors can further explain the relationship
between di(cid:130)erent types of behaviors and neural reuse.
Our results have revealed a relationship between di(cid:130)erent catego-
rization behaviors and the allocation of neural resources to perform
them: acquiring category-speci(cid:128)c e(cid:130)ective networks yields high
performance. We intend to further test this by including the CSC
as component in the (cid:128)tness estimation. (cid:140)is will help determine
if encouraging this phenomenon makes it more likely to produce
multifunctional circuits. Based on our results, we hypothesize that
this will in fact be the case.
(cid:140)e model presented here also speaks to the computational abil-
ity of spiking neurons in an embodied context. It has been estab-
lished spiking neurons have greater computational power than
perceptron-like neurons [25]. Furthermore, principles of morpho-
logical computation have shown that using systems that can exploit
the continuous interaction between environment and the agent
e(cid:130)ectively increases the computational abilities of the neural net-
works [31]. In our model, these two concepts are combined because
spiking neurons are now involved in a morphological computation
scenario. Although it is di(cid:129)cult to quantify the computational
power, that only 2 spiking neurons can perform the behavior is
indicative of that.
As a measure of e(cid:130)ective network connectivity, although Trans-
fer Entropy has been most widely used in neuroscience, it has
been used in evolutionary robotics in some instances - synchro-
nization dynamics of neurons and its in(cid:131)uence on evolvability and
behavior has been studied using the same task described in this
paper [27], and TE has also been used in analysis of neural net-
works in supervised and unsupervised learning contexts [26]. One
of the limitations of this methodology is that TE does not scale
elegantly with the size of the networks. While this is not an is-
sue with neuroscientists where TE is used as a post-task analysis
method, an online estimation of TE networks during evolution can
be computationally expensive. Although small spiking networks,
even as small as 2 neurons like we have shown here, have high
computational power [25], this can be a problem in large networks
for complex tasks. (cid:140)erefore, in order to use CSC as a component
of (cid:128)tness to more e(cid:129)ciently develop arti(cid:128)cial systems, a computa-
tionally more e(cid:129)cient metric might be required. (cid:140)is is one of the
directions we plan to explore further. Meanwhile, the limits of TE
Figure 6: Performance and CSC distributions of the high-
performing agents for all N . [A] A histogram of the (cid:128)tness
of all individuals that have the highest CSC of 0.95 showing
that all agents that had a high CSC value performed well.
[B] A histogram of the CSC values for agents that evolved
to have a (cid:128)tness ≥ 90%, showing that while there are agents
that perform well with low CSC values, it is more likely that
an agent with high (cid:128)tness also has a high CSC value.
Figure 7: High task-speci(cid:128)city in the e(cid:130)ective network sug-
gests high performance. Fitness as a function of CSC for
all best-agents across the ensemble, across di(cid:130)erent network
sizes. (cid:135)e high-concentration of points in the right top cor-
ner reveals that a high CSC makes it more likely that behav-
ioral performance is also high. Linear regression lines show
a positive correlation between (cid:128)tness and CSC for all net-
work sizes. Interestingly there is no signi(cid:128)cant di(cid:130)erence
in this trend for di(cid:130)erent network sizes.
00.51Fitness01020Frequencyn=2n=3n=4n=5n=600.51Cluster Specialization Coefficient0204060FrequencyABn=2n=3n=4n=5n=6 GECCO '17, July 15-19, 2017, Berlin, Germany
Candadai Vasu et. al.
Cambridge, Massachuse(cid:138)s.
[21]
[23]
can be pushed to evolving and analyzing increasingly larger net-
works for a number of di(cid:130)erent behaviors and also for performing
multiple behaviors. (cid:140)ese behaviors can be systematically chosen
to be (a) completely independent, (b) overlapping or (c) partially
overlapping. Studying the relationship between CSC and (cid:128)tness in
each of these cases would provide interesting insights into neural
reuse and how neural dynamics shapes behavior. Another limi-
tation of this study is that the TE analysis was performed on a
subset of the evolved parameters: the recurrent connections be-
tween interneuron spiking neurons; sensor-to-interneuron weights
and interneuron-to-motorneuron weights were not included. Con-
straining the optimization space to only the recurrent connections
and se(cid:138)ing other weights constant, would make sure that the TE
analysis is performed on the system that is entirely-responsible
for controlling the di(cid:130)erences in behaviors. Finally, TE averages
neural dynamics over the entire trial. Although this is useful for
comparisons across trials, in order to get an in-depth understanding
of how the network produces behavior, unrolling the information
analysis over time will be required.
Once again, the potential of computational modeling in helping
scientists focus experimental design has been shown here. Speci(cid:128)-
cally, we have shown that evolutionary algorithms can help opti-
mize embodied behaviorally-functional spiking neural networks,
and that analysis of the evolved agents can help generate testable
insights. (cid:140)e role played by the body and the environment in
producing behavior is increasingly acknowledged by neuroscien-
tists [18]. Analysis of computational models of embodied agents can
thus provide an appropriate platform to develop hypotheses that
can then be tested empirically. Looking forward, stimulation and
recording experiments in vitro and calcium imaging experiments
in C. elegans are ideal biological models to test our hypothesis:
high within-task homogeneity and high across-task heterogene-
ity among task speci(cid:128)c clusters of e(cid:130)ective networks yield high
categorization performance.
ACKNOWLEDGMENTS
(cid:140)e authors would like to thank the anonymous referees for their
valuable comments and helpful suggestions.
REFERENCES
[1] Michael L. Anderson. 2010. Neural reuse: A fundamental organizational principle
[3] R. D. Beer. 2003. (cid:140)e dynamics of active categorical perception in an evolved
[2] R. D. Beer. 1996. Toward the evolution of dynamical neural networks for mini-
of the brain. Behavioral and brain sciences 33, 4 (2010), 245–266.
mally cognitive behavior. From animals to animats 4 (1996), 421–429.
model agent. Adaptive Behavior 11, 4 (2003), 209–243.
[4] K. L. Briggman and W. B. Kristan. 2006. Imaging dedicated and multifunctional
neural circuits generating distinct behaviors. Journal of Neuroscience 26, 42
(2006), 10925–10933.
[5] N. Brunel. 2000. Dynamics of networks of randomly connected excitatory and
inhibitory spiking neurons. Journal of Physiology-Paris. 94, 5 (2000), 445–463.
[6] H. J. Chiel and R. D. Beer. 1997. (cid:140)e brain has a body: adaptive behavior
emerges from interactions of nervous system, body and environment. Trends in
neurosciences 20, 12 (1997), 553–557.
[7] M. J. Coleman B. J. Norris E. Marder D. M. Blitz, A. E. Christie and M. P. Nusbaum.
1999. Di(cid:130)erent proctolin neurons elicit distinct motor pa(cid:138)erns from a multifunc-
tional neuronal network. Journal of Neuroscience 19, 13 (1999), 5449–5463.
universal computations with spikes. PLOS Comput. Biol 12, 6 (2016).
[9] S. Ostojic E. S. Scha(cid:130)er and L. F. Abbo(cid:138). 2013. A complex-valued (cid:128)ring-rate
model that approximates the dynamics of spiking networks. PLoS Comput Biol 9,
10 (2013).
[8] H.J. Kappen D. (cid:140)almeier, M. Uhlmann and R. Memmesheimer. 2016. Learning
[20] B. DePasquale L. F. Abbo(cid:138) and R. Memmesheimer. 2016. Building functional
[19] A. Kaiser and T. Schreiber. 2002. Information transfer in continuous processes.
[24] R. Vicente M. Wibral and M. Lindner. 2014. Transfer entropy in neuroscience. In
[25] W. Maass. 1997. Networks of spiking neurons: the third generation of neural
[10] E. (cid:140)ompson F. J. Varela and E. Rosch. 1991. (cid:138)e Embodied Mind. MIT Press,
[11] R. de Ruyter van Steveninck F. Rieke, D. Warland and W. Bialek. 1996. Spikes:
Exploring the Neural Code. MIT Press, Cambridge, Massachuse(cid:138)s.
[12] D. Floreano and C. Ma(cid:138)iussi. 2001. Evolution of spiking neural controllers
for autonomous vision-based robots. International Symposium on Evolutionary
Robotics (2001), 38–61.
[13] B. Gourvitch and J. J. Eggermont. 2007. Evaluating information transfer between
auditory cortical neurons. Journal of Neurophysiology 97 (2007), 2533–2543.
[14] S. Harnad. 1987. Psychophysical and cognitive aspects of categorical perception: A
critical overview. In Categorical perception: (cid:138)e groundwork of cognition. Cam-
bridge University Press.
[15] R. Wood M. (cid:139)inn I. Harvey, E. Di Paolo and E. Tuci. 2005. Evolutionary robotics:
A new scienti(cid:128)c tool for studying cognition. Arti(cid:128)cial Life 11, 1-2 (2005), 79–98.
[16] E. M. Izhikevich. 2003. Simple model of spiking neurons. IEEE Transactions on
neural networks 14, 6 (2003).
[17] E. M. Izhikevich. 2004. Which model to use for cortical spiking neurons?. IEEE
Transactions on neural networks 15, 5 (2004).
[18] E. Izquierdo and R. D. Beer. 2016. (cid:140)e whole worm: brain,body,environment
models of C. elegans. Current Opinion in Neurobiology 40 (2016), 23–30.
Physica D 166 (2002), 43–62.
networks of spiking model neurons. Nature 201, 6 (2016).
J. L. Vincent M. Corbe(cid:138)a D. C. Van Essen M. D. Fox, A. Z. Snyder and M. E. Raichle.
2005. (cid:140)e human brain is intrinsically organized into dynamic, anticorrelated
functional networks. Proceedings of the National Academy of Sciences of the
United States of America 102, 27 (2005), 9673–9678.
[22] P. Massobrio M. Garofalo, T. Nieus and S. Martinoia. 2009. Evaluation of the
performance of information theory-based methods and cross-correlation to
estimate the functional connectivity in cortical networks. PLoS one 4, 8 (2009).
J. M. Beggs. M. Shimono. 2015. Functional clusters, hubs, and communities in
the cortical microconnectome. Cerebral Cortex 25, 10 (2015), 3743–3757.
Directed Information Measures in Neuroscience (2014), 3–36.
network models. Neural Networks 10, 9 (1997), 1659–1671.
[26] Renan C. Moioli and Phil Husbands. 2013. Neuronal assembly dynamics in
supervised and unsupervised learning scenarios. Neural Computation 25, 11
(2013), 2934–2975.
[27] Renan C. Moioli, Patricia A. Vargas, and Phil Husbands. 2012. Synchronisation
e(cid:130)ects on the behavioural performance and information dynamics of a simulated
minimally cognitive robotic agent. Biological Cybernetics (2012), 1–21.
[28] M. A. Whi(cid:138)ington N. Kopell, G. B. Ermentrout and R. D. Traub. 2000. Gamma
rhythms and beta rhythms have di(cid:130)erent synchronization properties. Proceedings
of the National Academy of Sciences 97, 4 (2000), 1867–1872.
[29] R. Goebel N. Kriegeskorte and P. Bande(cid:138)ini. 2006. Information-based functional
brain mapping. Proceedings of the National academy of Sciences of the United
States of America 103, 10 (2006), 3863–3868.
[30] M. Myroshnychenko Fang-Chin Yeh E. Hiolski P. Ho(cid:138)owy N. Timme, S. Ito and
J. M. Beggs. 2014. Multiplex networks of cortical and hippocampal neurons
revealed at di(cid:130)erent timescales. PLoS one 9, 12 (2014).
body, and environment. In Creating Brain-Like Intelligence (2009), 66–83.
biologically inspired robotics. Science 318, 5853 (2007), 1088–1093.
[33] M. Lindner R. Vicente, M. Wibral and G. Pipa. 2011. Transfer entropy(cid:128)(cid:148)a
model-free measure of e(cid:130)ective connectivity for the neurosciences. Journal of
computational neuroscience 30, 1 (2011), 45–67.
[34] M. Rubinov and O. Sporns. 2010. Complex network measures of brain connectiv-
ity: uses and interpretations. Neuroimage 52, 3 (2010).
[35] R. Heiland-A. Lumsdaine A. M. Litke S. Ito, M. E. Hansen and J. M. Beggs. 2011.
Extending transfer entropy improves identi(cid:128)cation of e(cid:130)ective connectivity in a
spiking cortical network model. Plos one 6, 11 (2011).
[36] S. Ito F. Yeh N. Timme M. Myroshnychenko C. C. Lapish P. Ho(cid:138)owy W. C. Smith
S. C. Masmanidis A. M. Litke O. Sporns S. Nigam, M. Shimono and J. M. Beggs.
2016. Rich-club organization in e(cid:130)ective connectivity among cortical neurons.
Journal of Neuroscience 36, 3 (2016), 670–684.
(2000), 461–464.
[37] T. Schreiber. 2000. Measuring information transfer. Physical Review Le(cid:136)ers 85, 2
[38] C. van Vreeswijk and H. Sompolinsky. 1996. Chaos in neuronal networks with
[31] R. Pfeifer and G. Gmez. 2009. Morphological computation(cid:128)(cid:147)connecting brain,
[32] M. Lungarella R. Pfeifer and F. Iida. 2007. Self-organization, embodiment, and
balanced excitatory and inhibitory activity. Science 274, 5293 (1996), 1724.
[39] F. A. Wilson, S. P. O'scalaidhe, and Patricia S. Goldman-Rakic. 1994. Functional
synergism between putative gamma-aminobutyrate-containing neurons and
pyramidal neurons in prefrontal cortex. Proceedings of the National Academy of
Sciences 91, 9 (1994), 4009–4013.
[40] X. Wang X. Lin and Z. Hao. 2016. Supervised learning in multilayer spiking
neural networks with inner products of spike trains. Neurocomputing (2016).
|
1812.06590 | 2 | 1812 | 2019-07-05T20:53:47 | Exploring the Psychological Basis for Transitions in the Archaeological Record | [
"q-bio.NC",
"q-bio.PE"
] | In lieu of an abstract here is the first paragraph: No other species remotely approaches the human capacity for the cultural evolution of novelty that is accumulative, adaptive, and open-ended (i.e., with no a priori limit on the size or scope of possibilities). By culture we mean extrasomatic adaptations--including behavior and technology--that are socially rather than sexually transmitted. This chapter synthesizes research from anthropology, psychology, archaeology, and agent-based modeling into a speculative yet coherent account of two fundamental cognitive transitions underlying human cultural evolution that is consistent with contemporary psychology. While the chapter overlaps with a more technical paper on this topic (Gabora & Smith 2018), it incorporates new research and elaborates a genetic component to our overall argument. The ideas in this chapter grew out of a non-Darwinian framework for cultural evolution, referred to as the Self-other Reorganization (SOR) theory of cultural evolution (Gabora, 2013, in press; Smith, 2013), which was inspired by research on the origin and earliest stage in the evolution of life (Cornish-Bowden & C\'ardenas 2017; Goldenfeld, Biancalani, & Jafarpour, 2017, Vetsigian, Woese, & Goldenfeld 2006; Woese, 2002). SOR bridges psychological research on fundamental aspects of our human nature such as creativity and our proclivity to reflect on ideas from different perspectives, with the literature on evolutionary approaches to cultural evolution that aspire to synthesize the behavioral sciences much as has been done for the biological scientists. The current chapter is complementary to this effort, but less abstract; it attempts to ground the theory of cultural evolution in terms of cognitive transitions as suggested by archaeological evidence. | q-bio.NC | q-bio | Gabora, L., & Smith, C. (2019). Exploring the psychological basis for transitions in the
archaeological record. In T. Henley, E. Kardas, & M. Rossano (Eds.), Handbook of cognitive
archaeology: A psychological framework.
Exploring the Psychological Basis for Transitions in the Archaeological Record
Liane Gabora
Department of Psychology, University of British Columbia
Fipke Centre for Innovative Research, 3247 University Way,
Kelowna, BC Canada, V1V 1V7
[email protected]
https://people.ok.ubc.ca/lgabora/
Department of Anthropology, Portland State University Portland, OR, 97207
Cameron M. Smith
[email protected]
https://www.pdx.edu/anthropology/cameron-smith
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
2
Exploring the Psychological Basis for Transitions in the Archaeological Record
Introduction
No other species remotely approaches the human capacity for the cultural evolution of novelty
that is accumulative, adaptive, and open-ended (i.e., with no a priori limit on the size or scope
of possibilities). By culture we mean extrasomatic adaptations -- including behavior and
technology -- that are socially rather than sexually transmitted. This chapter synthesizes
research from anthropology, psychology, archaeology, and agent-based modeling into a
speculative yet coherent account of two fundamental cognitive transitions underlying human
cultural evolution that is consistent with contemporary psychology. While the chapter
overlaps with a more technical paper on this topic (Gabora & Smith 2018), it incorporates
new research and elaborates a genetic component to our overall argument. The ideas in this
chapter grew out of a non-Darwinian framework for cultural evolution, referred to as the Self-
other Reorganization (SOR) theory of cultural evolution (Gabora, 2013, in press; Smith,
2013), which was inspired by research on the origin and earliest stage in the evolution of life
(Cornish-Bowden & Cárdenas 2017; Goldenfeld, Biancalani, & Jafarpour, 2017, Vetsigian,
Woese, & Goldenfeld 2006; Woese, 2002). SOR bridges psychological research on
fundamental aspects of our human nature such as creativity and our proclivity to reflect on
ideas from different perspectives, with the literature on evolutionary approaches to cultural
evolution that aspire to synthesize the behavioral sciences much as has been done for the
biological scientists. The current chapter is complementary to this effort, but less abstract; it
attempts to ground the theory of cultural evolution in terms of cognitive transitions as
suggested by archaeological evidence.
By the term archaeological evidence we mean the 'material correlates' or
'precipitates' of behavior. While artifacts were once treated as indicators of 'progress', more
contemporary approaches, including the approach taken here, seek to reveal the cognitive
conditions responsible for artifacts (and other material precipitates of behavior; Haidle, 2009,
Wragg-Sykes, 2015). Note that although our theoretical approach is founded on evolutionary
principles, it is not what is often referred to as 'evolutionary psychology' (Cosmides &
Tooby, 1992; Sell et al, 2009) which focuses on biological underpinnings of cultural evolution
as opposed to the impact of culture as an evolutionary process unto itself. Our approach is
more aligned to other evolutionary approaches to the general question of how modern
cognition arose, such as those of Wynn and Coolidge (e.g., Wynn et al, 2017) highlighting
developmental psychology and Bruner (e.g. 2010), highlighting palaeoneurology.
Our use of the term 'transition' in the title is intensional (for as Straus, 2009, observes
the term 'transition' is sometimes used too casually). We begin the chapter with a discussion
of the concept of evolutionary transition, for the 'unpacking' of this term could be of
explanatory value with respect to archaeological change and its cognitive underpinnings.
Evolutionary Transitions
Transitions are common in evolution (Szathmary & Maynard Smith, 1995). Nonlinear
interactions between different information levels (e.g., genotype, phenotype, environment,
and even developmental characteristics) often give rise to emergent outcomes that generate
discontinuities (Galis & Metz, 2007). Research into the mechanisms underlying evolutionary
transitions have made headway into explaining the origins of new varieties of information
organization and unpacking terms such 'adaptation due to natural selection', aiming "to
analyze trends of increasing complexity" (Griesmer, 2000). Szathmary and Maynard-Smith's
account of the eight major transitions in the history of life remains widely accepted today
(Szathmary 2015, Calcott & Sterelny, 2011), with other transitions continue to be identified,
3
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
including the evolution of new sexes (Parker, 2004), and new varieties of ant agriculture
(Schultz & Brady, 2008), animal individuality (Godfrey-Smith 2011), metabolism and cell
structure (DeLong et al., 2010), technology (Geels, 2002) and hominin socialization (Foley &
Gamble 2009).
Research on the dynamics (e.g., rates and types) of evolutionary transitions shows that
despite their diversity they exhibit common features: they are (1) rare, (2) involve new levels
of organization of information, (3) followed by diversification, and (4) incomplete (Wilson
2010). Szathmary and Maynard-Smith include the transition from "primate societies to human
societies " as part of their "Extended Evolutionary Synthesis" (Wilson, 2010), but this
synthesis was formulated just prior to the rise of explicitly evolutionary approaches to modern
cognition.
We suggest that the theory of evolutionary transitions can provide a useful framework
for understanding the cognitive changes culminating in behavioral modernity (BM). In
cognitive evolution, evidence of significant change might be stretched out over time and
space for many reasons, such as lag between initial appearance and demic diffusion,
ambiguities in the archaeological and fossil records. In this chapter we explore two such
transitions. The first is the origin of a richer, post-Pan, post-Australopithecine culture as early
as 2.2 million years ago (Harmand et al., 2015). The second is the explosion of creative
culture in the Middle/Upper Paleolithic.
Origin of Post-Pan, Post-Australopithecine Culture: A First Cognitive Transition
Archaeological and Anthropological Evidence
The minds of Australopithecus and earliest Homo appear to have been anchored to the present
moment of concrete sensory perceptions, i.e., the "here and now". Simple stone (and some
bone and antler) implements indicates that they encoded perceptions of events in memory,
thereby supplying "timely information to the organism's decision-making systems" (Klein et
al., 2002, page 306) -- but had little voluntary access to memories without external cues. The
upshot was minimal innovation and artifact variation.
This is evident in the early archaeological record, beginning with stone tools from
Lomekwi 3 West Turkana, Kenya, 3.3 mya (Harmand et al., 2015), and characterized by
opportunism in highly restricted environments (Braun et al., 2008). Tools were
technologically on par with those of modern chimpanzees (Byrne & Russon, 1998; Blackwell
& d'Errico, 2001; see Read [2008] and Fuentes [2015] for cognitive considerations of
chimpanzee toolmaking). These tools lack evidence of symbolism (d'Errico et al., 2003), and
were transported relatively short distances across landscapes (Potts, 2012). While nut-
cracking and other simple tool use outside Homo may involve sequential chaining of actions,
(and thus sequential chaining of the mental representations underlying these actions), outside
Homo, sequential processing did not occur reliably enough to cross the threshold for abstract
thought (see Gabora & Steel, 2017 for a mathematical model of this). Thus, it appears that
early Homo could not invent or refine complex actions, gestures, or vocalizations, and their
ability to voluntarily shape, modify, or practice skills and actions was at best minimal.
Early Homo evolved into several forms, including H. erectus, dating between 2.8 - 0.3
million years ago (Villmoare et al., 2015), and there was a shift away from biology and
towards culture as the primary means of adaption in this lineage, attended by significant
cultural elaboration. Having expanded out from Africa as early as 2 mya, Homo constructed
tools involving more production steps and more varied raw materials (Haidle, 2009), imposed
symmetry on tool stone (Lepre et al., 2011), used and controlled fire (Goren-Inbar et al.,
2004), crossed stretches of open water up to 20 km (Gibbons 1998), ranged as far north as
latitude 52◦ (Preece et al., 2010), revisited campsites possibly for seasons at a time, built
4
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
shelters (Mania & Mania 2005), transported tool stone over greater distances than their
predecessors (Moutsou, 2014), and ranked moderately high among predators (Plummer,
2004). While the cranial capacity of Homo erectus was approximately 1,000 cc -- about 25%
larger than that of Homo habilis, at least twice as large as that of living great apes, and 75%
that of modern humans (Aiello, 1996) -- brain volume alone cannot explain these
developments. It is widely thought that these signs of a culture richer than that of Pan or
Australopithecus c. 1.7 mya reflect a significant transition in cognitive and/or social
characteristics.
Early Ideas about What Caused this First Transition
We can take as a starting point Donald's (1991) theory of cognitive evolution as it was a
breakthrough that paved the way for much of what followed. Because the cognition of Homo
habilis was primarily restricted to whatever episode one was currently experience, Donald
refers to it as an episodic mode of cognition. He proposed that the enlarged cranial capacity
brought about the onset of what he calls a self-triggered recall and rehearsal loop, which we
abbreviate STR. STR enabled hominins to voluntarily retrieve stored memories independent
of environmental cues (sometimes referred to as 'autocuing') and engage in representational
redescription and the refinement of thoughts and ideas. This was a fundamentally new mode
of cognitive functioning which Donald referred to as the mimetic mind because it could
'mime' or act out past or possible future events, thereby not only temporarily escaping the
present, and communicating that escape to others.
STR also enabled attention to be directed away from the outside world toward ones'
internal representations, paving the way for abstract thought. We use the term abstract
thought to refer to the reprocessing of previous experiences, as in counterfactual thinking,
planning, or creativity, as opposed to direct perception of the 'here and now' (for a review of
abstract thought, see Barsalou, 2005). Note that in much of the cultural evolution literature,
abstract thought and creativity, if mentioned at all, are equated with individual learning,
which is thought to mean 'learning for oneself' (e.g., Henrich & Boyd, 2002; Mesoudi,
Whiten & Laland, 2006; Rogers, 1988). However, they are not the same thing; individual
learning deals with obtaining pre-existing information from the environment through non-
social means (e.g., learning to differentiate different kinds of trees). The information does not
change form just because the individual now knows it. In contrast, abstract thought involves
the reprocessing of internally sourced mental contents, such that they are in flux, and when
this results in the generation of useful or pleasing ideas, behavior, or artifacts that did not
previously exist, it is said to be creative. Indeed, there is increasing recognition of the extent
to which creative outcomes are contingent upon internally driven incremental/iterative honing
or reprocessing (Basadur, 1995; Chan & Schunn, 2015; Feinstein, 2006; Gabora, 2017).
Note that Donald's explanation focuses on neither technical nor social abilities but on
a cognitive ability that facilitated both. STR enabled systematic evaluation and improvement
of thoughts and motor skills such that they could be adapted to new situations, resulting in
voluntary rehearsal and refinement of actions and artifacts. STR also broadened the scope of
social activities to include re-enactive play and pantomime.
Proposed Theory of Cognitive Underpinnings of this First Transition
Leaving aside alternatives to Donald's proposal until the end of this section, we note for now
that although Donald's theory seems reasonable, it does not explain why larger brain size
enabled STR. In what follows, we contextualize Donald's (1991) schema in research on the
nature of associative memory. We will ground the concept of STR in a neural level account of
the mechanisms underlying cognitive flexibility and creativity (Gabora, 2010; Gabora &
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
Ranjan, 2013).
5
We start by summarizing a few well-known features of associative memory. Each
neuron is sensitive to a primitive stimulus attribute, or microfeature, such as lines of a
particular orientation, or sounds of a particular pitch. Items in memory are distributed across
cell assemblies of such neurons; thus each neuron participates in the encoding of many items.
Memory is also content-addressable: there is a systematic relationship between the content of
an item and the neurons that encode it; therefore, items that share microfeatures may be
encoded in overlapping distributions of neurons.
While in and of itself increased brain volume cannot explain the origin of BM, we
suggest that larger brains enabled a transition from coarse-grained to more fine-grained
memory. The smaller the number of neurons a brain has to work with, the fewer attributes of
any given memory item it can encode, and the less able it is to forge associations on the basis
of shared attributes. Conversely, the evolution of finer-grained memory meant that memory
items could be encoded in more detail, i.e., distributed across larger sets of cell assemblies
containing more neurons. Since memory organization was content addressable, this meant
more ways in which distributed representations could meaningfully overlap, and greater
overlap enabled more routes by which one memory could evoke another.
This in turn made possible the onset of STR, and paved the way for the capacity to
engage in recursive recall and streams of abstract thought, and a limited form of insight
(Gabora, 2002, 2010; Gabora & Ranjan, 2013). As a simple example, the reason that the
experience of seeing a leaf floating on a lake could potentially play a role in the invention of a
raft is that both experiences involve overlap in the set of relevant attribute "float on water",
and thus overlap of activated cell assemblies.
Items in memorys could now be reprocessed until they achieved a form that was
acceptably consistent with existing understandings or sufficiently enabled goals and desires to
be achieved (Gabora, 1999. This scenario provides a plausible neural-level account of
Donald's (1991) proposal that abstract thought was a natural consequence of possessing a
self-triggered recall and rehearsal loop, made possible by an increase in brain size.
Comparison with Other Theories
Mithen's (1996) model features the accumulation and overlap of different intelligence
modules. Although in its details his model runs rather counter to much current thinking
including our own, his thoughts on cognitive fluidity and creativity influenced the model
proposed here.
Some theories attribute this transition to social factors. Foley and Gamble (2009)
emphasize enhanced family bonding and the capacity for a more focused style of
concentration, further enhanced by controlled use of fire by at least 400,000 years ago.
Wiessner (2014) suggested that fire not only enabled the preparation of healthier food, but by
providing light after dark, facilitated playful and imaginative social bonding. Others
emphasize an extrication from biologically based to culturally based kinship networks (Leaf
& Read, 2012; Read, Lane & van der Leeuw, 2015). We believe that these social explanations
are correct but that they have their origin in cognitive changes, which altered not only social
interactions but interactions with other aspects of human experience as well.
Our proposal bears some resemblance to Hauser et al.'s (2002) suggestion that the
capacity for recursion is what distinguishes human cognition from that of other species, Penn
et al's (2008) concept of relational reinterpretation, and Read's (2009) claim that recursive
reasoning with relational concepts made possible a conceptually based system of social
relations but may have evolved alongside non-social activities such as toolmaking. However,
our proposal goes further, because it grounds the onset of recursive reasoning in a specific
6
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
cognitive transition. While Read suggests that recursive reasoning was made possible by
larger working memory, we argue that larger working memory in and of itself is not useful; it
must goes hand-in-hand with -- and indeed is a natural byproduct of -- a finer-grained memory.
As an illustrative example, let us suppose that a hominid with a coarse-grained memory
increased its working memory from being able to think only of one thing at a time (e.g., a
leaf) to two (e.g., a leaf and the moon). This would generally be a source of confusion.
However, if it held only one thing in mind at a time but encoded it in richer detail (e.g.,
incorporating attributes of a leaf such as 'thin', 'flat', and so forth), it could forge meaningful
associations with other items based on these attributes (e.g., other thin, flat things).
Our proposal also bears some similarity to Chomsky's (2008) suggestion that this
transition reflects the onset of the capacity for a 'merge' operation. 'Merge' is described as
the forging of associations based on their global similarity, i.e., between items that are highly
similar or that co-occur in space or time. In contrast, for STR the memory must be sufficiently
fine-grained (i.e., items must be encoded in sufficient detail) that the associative process can
operate on the basis of specific attributes to which specific neurons are tuned. STR can forge
associations between items that are related by as few as a single attribute, and do so
recursively such that the output of one such operation is the input for the next, and reliably,
such that encodings are modified in light of each other in the course of streams of thought
(Gabora, 2002, 2013, 2017; detailed examples, such as the invention of a fence made of skis
on the basis of the attributes 'tall', 'skinny' and 'sturdy' [Gabora, 2010], and the generation of
the idea of a beanbag chair on the basis of the single attribute 'conforms to shape' [Gabora,
2018], are provided elsewhere). Thus our proposal (but not 'merge') offers a causal link
between brain size and cognitive ability, i.e., more neurons means they can be tuned to a
broader range of attributes and thereby form more associations on the basis of shared
attributes.
Creative Culture in Middle/Upper Paleolithic: A Second Cognitive Transition
Archaeological and Anthropological Evidence
The African archaeological record indicates another significant cultural transition
approximately 100,000 years ago that shows many material correlates of BM. Though there is
no single definitive indicator of BM (d'Errico et al., 2005; Shea 2011), it is generally thought
to involve (a) a radical proliferation of tool types that better fit tools to specific tasks
(McBrearty & Brooks, 2000), (b) elaborate burial sites indicating ritual (Hovers, Ilani, Bar-
Yosef, & Vandermeersch, 2003) and possibly religion (Rappaport, 1999), (c) artifacts
indicating personal symbolic ornamentation (d'Errico et al., 2009), (d) 'cave art', i.e.,
representational imagery featuring depictions of animals (Pike et al., 2012) and human beings
(Nelson, 2008), (e) complex hearths and highly structured use of living spaces (Otte, 2012),
(f) calorie-gathering intensification that included widespread use of aquatic resources
(Erlandson, 2001), and (g) extensive use of bone and antler tools, sometimes with engraved
designs. BM extended across Africa after 100,000 years ago, and was present in Sub-
Himalayan Asia and Australasia over 50,000 years ago (Mulvaney & Kamminga, 1999) and
Continental Europe thereafter (Mellars, 2006).
It is uncertain whether this archaeological record reflects a genuine transition resulting
in BM because claims to this effect are based on the European Paleolithic record, and largely
exclude the lesser-known African record (Fisher & Ridley, 2013). Artifacts associated with a
rapid transition to BM 40,000-50,000 years ago in Europe are found in the African Middle
Stone Age tens of thousands of years earlier, pushing the cultural transition more closely into
alignment with the transition to anatomical modernity between 200,000 and 100,000 BP.
Nevertheless, it seems clear that BM appeared in Africa between 100,000 to 50,000 years ago,
7
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
and spread to Europe, resulting in displacement of the Neanderthals (Fisher & Ridley, 2013).
Despite an overall lack of increase in cranial capacity, the prefrontal cortex, and more
particularly the orbitofrontal region, increased significantly in size (Dunbar 1993), in what
was likely a time of major neural reorganization (Morgan 2013). Homo sapiens could now
effectively archive information, and adapt it to different needs and circumstances, making
their cultures radically more creative, open-ended, and accumulative than any prior hominin
(Mithen, 1998).
This transition is also commonly associated with the origins of complex language.
Although the ambiguity of the archaeological evidence makes it difficult to know exactly
when language began (Davidson & Noble 1989; Christiansen & Kirby, 2003; Hauser,
Chomsky, & Fitch, 2002), it is widely believed -- based on stone tool symmetry and
complexity of manufacture -- that as long ago as c. 1.7 million years Homo used gestural and
prelinguistic vocalization communications that would have shared some organizational
similarities with modern humans insofar as they differed significantly from other primate
communications. The evolution of grammatically- and syntactically-modern language is
generally placed (depending on whether one is observing it in Africa, sub-Himalayan Asia or
Western Eurasia) after about 100,000 years ago, around the start of the Upper Palaeolithic
(Bickerton, 2014; Dunbar, 1993; Tomasello, 1999).
Proposed Cognitive Mechanism Underlying Second Transition
We propose that the root cause of the cultural explosion of the Middle/Upper Paleolithic was
a fine-tuning of the biochemical mechanisms underlying the capacity to spontaneously shift
between different modes of thought in response to the situation, and that this is accomplished
by varying the specificity of the activated region of memory. The ability to shift between
different modes is referred to as contextual focus (CF) because it requires the capacity to
focus or defocus attention in response to contextual factors (Gabora, 2003), such as a specific
a goal, a particular audience, or aspect of the situation, and do so continuously throughout a
task. Focused attention is conducive to analytical thought (Agnoli, Franchin, Rubaltelli, &
Corazza, 2015; Vartanian, 2009; Zabelina, 2018), wherein activation of memory is
constrained enough to hone in and mentally operate on only the relevant aspects of mental
contents. In contrast, by diffusely activating a wide region of memory, defocused attention is
conducive to associative thought; it enables more obscure (though potentially relevant)
aspects of the situation to be considered. This greatly enhances the potential for insight, i.e.,
the forging of obscure but useful or relevant connections.
Note that associative thought is useful for breaking out of a rut, but would be risky
without the ability to reign it back in; basic survival related tasks may be impeded if
everything is reminding you of everything else. Therefore, it would take considerable time to
fine-tune the mechanisms underlying the capacity to spontaneously shift between these two
processing modes such that one retained the benefits of escaping local minima without
running the risk of being perpetually side-tracked. The time needed to fine-tune this could
potentially be the explanation for the lag between anatomical and BM.
Once the products of one mode of thought could become 'ingredients' for the other,
hominids could reflect on thoughts and ideas not just from different perspectives but at
different levels of granularity, from basic level concepts (e.g., rabbit) up to abstract concepts
(e.g., animal) and down to more detailed levels (e.g., legs), as well as conceive of their
interrelationships. This kind of personal reflection was necessary for, and indeed a precursor
to, the need to come up with names for these things, i.e., the development of complex
languages. Thus, it is proposed that CF paved the way for not just language but the seemingly
diverse collection of cognitive abilities considered diagnostic of BM.
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
8
To see how the onset of CF could result in open-ended cultural complexity, recall that
associative memory has the following properties: distributed representation, coarse coding,
and content addressability. Each thought may activate more or fewer cell assemblies
depending on the nature of the task. Focused attention is conducive to analytic thought
because memory activation is constrained enough to zero in and operate on key defining
properties. Defocused attention, by diffusely activating a broader region of memory, is
conducive to associative thought; obscure (but potentially relevant) aspects of the situation
come into play (Gabora, 2000, 2010). Thus, thinking of, say, the concept SUN in an analytic
mode might bring to mind only the literal sun, in an associative mode of thought it might also
bring to mind other sources of heat, or other stars, or even someone with a sunny disposition.
Once our hominid ancestors could shift between these modes of thought, tasks requiring
either one mode, or the other, or shifting between them in a precisely orchestrated sequence
the fruits of one mode become ingredients for the other. This resulted in the forging of a
richly integrated creative internal network of understandings about the world and one's place
in it, or worldview, which we claim made BM possible. Thus, the notion that diffuse memory
activation is conducive to associative thought while activation of a narrow receptive field is
conducive to analytic thought is not only consistent with the architecture of associative
memory, but suggests an underlying mechanism by which CF enabled the ability to both (1)
stay task-focused, and (2) deviate from the task to make new connections, as needed. In this
view, language did not evolve solely to help people communicate and collaborate (thereby
accelerating the pace of cultural innovation); it also helped people think ideas through for
themselves and manipulate them in a deliberate, controlled manner. Language facilitated the
weaving of experiences into stories, parables, and broader conceptual frameworks, thereby
integrating knowledge and experience (see also, Gabora & Aerts, 2009). Instead of staying
squarely in the narrow regime between order and chaos, thought could now shift between
orderly and chaotic, as appropriate.
Thus, we propose that the emergence of a self-organizing worldview required two
transitions. The onset of STR over 2 mya allowed rehearsal and refinement of skills and made
possible minor modifications of representations. The onset of CF approximately 100,000
years ago made it possible to forge larger bridges through conceptual space that paved the
way for innovations specifically tailored to selective pressures. It enabled a cultural version of
what Gould and Vrba (1982) call exaptation, wherein an existing trait is co-opted for a new
function (Gabora, Scott, & Kauffman, 2013). Exaptation of representations and ideas vastly
enhanced the ability to expand the technological and social spheres of life, as well as to
develop individualized perspectives conducive to fulfilling complementary social roles. This
increase in cognitive variation provided the raw material for enhanced cultural adaptation to
selective pressures.
Comparison with Other Theories
Our proposal is superficially similar to (and predates -- see Gabora 2003) the suggestion that
what distinguishes human cognition from that of other species is our capacity for dual
processing (Evans, 2008; Nosek, 2007). This is the hypothesis that humans engage in (1) a
primitive implicit Type 1 mode involving free association and fast 'gut responses', but (2) an
explicit Type 2 mode involving deliberate analysis. However, while dual processing makes
the split between older, more automatic processes and newer, more deliberate processes, CF
makes the split between an older associative mode based on relationships of correlation and a
newer analytic mode based on relationships of causation. We propose that although earlier
hominids relied on the older association-based system, because their memories were coarser-
grained, there were fewer routes for meaningful associations, so there was less associative
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
processing of previous experiences. Therefore, items encoded in memory tended to remain in
the same form as when they were originally assimilated; rather than engaging in associative or
analytic processing of previously assimilated material, there was greater tendency to leave
mental contents in their original form and instead focus on the present. Thus, while dual
processing theory attributes abstract, hypothetical thinking to the more recent Type 2 mode,
according to our theory abstract thought is possible in either mode but differs in character in
the two modes: logically constructed arguments in the analytic mode versus flights of fancy in
the associative mode. Our theory is rooted in a distinction in the creativity literature between
(1) associative divergent processes said to predominate during idea generation, and (2)
analytic convergent processes said to predominate during the refinement and testing of an idea
(Finke, Ward, & Smith, 1992; see also Sowden, Pringle, & Gabora [2014], for a comparison
of theories of creativity; and Gabora, 2018 for the distinction between associative versus
divergent thought).
Mithen (1996) proposed that BM came about through the integration of previously-
compartmentalized intelligence modules that were specialized for natural history, technology,
socialization and language. This integration, he says, enabled cognitive fluidity: the capacity
to combine concepts and adapt ideas to new contexts, and thereby explore, map, and
transform conceptual spaces across different knowledge systems. A related proposal
emphasizes the benefit of cognitive fluidity for the capacity to make and understand analogies
(Fauconnier & Turner, 2002). Our explanation is consistent with these but goes beyond them
by showing how conceptual fluidity would arise naturally as a function of the capacity to,
when appropriate, shift to a more associative processing mode.
There are many versions of the theory that BM reflects the onset of complex language
9
complete with recursive embedding of syntactic structure (Bickerton & Szathmáry, 2009;
Carstairs-McCarthy, 1999), which enabled symbolic representation and abstract thought
(Bickerton, 2014; Deacon, 1997), narrative myth (Donald, 1991), enhanced communication,
cooperation, and group identity (Voorhees, Read & Gabora, in press). Our proposal is
consistent with the view that complex language lay at the heart of BM, but because STR
followed by CF would have enabled hominids to not just recursively refine and modify
thoughts but consider them from different perspectives at different hierarchical levels, it set
the stage for complex language.
Since, as explained above, we see evidence of recursive reasoning well before BM,
our framework is inconsistent with the hypothesis that onset of recursive thought enabled
mental time travel, distinctly-human cognition, and BM (Corballis, 2011; see also Suddendorf
et al., 2009). Nevertheless, we suggest that the ability to shift between different modes of
thought using CF would have brought on the capacity to make vastly better use of it. The
proposal that BM arose due to onset of the capacity to model the contents of other minds,
sometimes referred to as 'Theory of Mind' (Tomasello, 2014) is somewhat underwritten by
recursion, since the mechanism that allows for recursion is required for modeling the contents
of other minds (though in this case the emphasis is on the social impact of recursion, rather
than the capacity for recursion itself). Our proposal is also consistent with explanations for
BM that emphasize social-ecological factors (Foley & Gamble, 2009; Whiten, 2011), but
places these explanations in a broader framework by suggesting a mechanism that aided not
just social skills but other skills (e.g., technological) as well. While most of these explanations
are correct insofar as they go, we suggest that they do not get to the root of the matter. As Carl
Woese wrote of science at large "...sometimes [there is] no single best representation... only
deeper understanding, more revealing and enveloping representations," (2004, page 173). We
propose that the second cognitive transition necessary for cumulative, adaptive, open-ended
cultural evolution was the onset of CF, because once hominids could adapt their mode of
10
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
thought to their situation by reflecting on mental contents through the lenses of different
perspectives, at different levels of analysis, their initially fragmented mental models of their
world could be integrated into more coherent representations of their world -- i.e., worldviews.
This facilitated not just survival skills, conceptual fluidity, and creative problem-solving, but
also social exchange and the emergence of complex social structures.
In short, the explanation proposed here is the only one we know of that grew out of a
synthesis of archaeological and anthropological data with theories and research from both
psychology and neuroscience, and it appears to underwrite explanations having specifically to
do with dual processing, conceptual fluidity, language, social interaction, or theory of mind.
A Tentative Genetic Basis for Contextual Focus
The hypothesis that BM arose due to the onset of contextual focus (CF, or the capacity to shift
between different modes of thought), leads to the question: what caused CF? In this section
we explore a possible genetic basis for CF. First, we provide historical context by reviewing
both studies that implicated genes such as FOXP2 in the origins of language, and the evidence
that caused this explanation to fall out of favor. Next, we review anthropological and
archaeological evidence that the coming into prominence of creative and cognitive abilities
(including but not limited to those that involve language) coincides with the evolutionary
origins of FOXP2. Then we synthesize these literatures in a new explanation of the role of
FOXP2 in language and cognition.
FOXP2, a transcription gene on chromosome 7 (Reimers-Kipping, Hevers, Paabo, &
Enard 2011), regulates the activity of other genes involved in the development and function of
the brain (Fisher & Ridley 2013; Kovas & Plomin 2006). Among other functions, FOXP2
plays a role in the functioning of the motor cortex, the striatum, and the cerebellum, which
controls fine motor skills (Liegeois, Morgan, Connelly, & Vargha-Khadem 2011). There are
findings of familial resemblance for specific components of linguistic competence and
suggestions of a genetic basis for such competences (e.g., Kovac, Gopnik, & Palmour 2002).
The finding that a mutation in FOXP2 was associated with language comprehension
and production in a family known as the KE Family (Lai, Fisher, Hurst, Vargha-Khadem, &
Monaco 2001) led to the proposal that the gene plays a key role in language acquisition. The
family was diagnosed with Specific Language Impairment (SLI), a severe deficit in language
development that exists despite adequate educational opportunity and normal nonverbal
intelligence (Lai et al. 2001; Morgan 2013). Those afflicted with SLI exhibit deficits in
speech comprehension as well as verbal dyspraxia -- a severe difficulty controlling the
movement and sequencing of orofacial muscles required for the articulation of fluent speech
(Lalmansingh, Karmakar, Jin, & Nagaich, 2012). Thus FOXP2 became prematurely known as
the "language gene" (Bickerton, 2014).
Evidence that FOXP2 is actively transcribed in brain areas where mirror neurons are
present suggested that FOXP2 makes language possible at least partially through its effect on
the capacity to imitate (Corballis, 2004). This hypothesis was strengthened by findings that
mirror neurons play a key role in language development (Arbib, 2011). However, in order to
imitate the language of others there must already be others using language. Therefore, it is
difficult to fully account for the origin of language by positing that FOXP2 affects language
by way of its effect on imitation. Moreover, this cannot explain the existence of defects
associated with FOXP2 that involve neither language nor imitation. Finally, other species
imitate but do not exhibit grammatical, syntactically rich language.
The recognition that FOXP2 was not the "language gene" generated sober discussion
about the simplicity of single gene explanations for complex traits. Indeed, as a transcription
gene, FOXP2 only has an indirect effect on neural structure or behavior (Reimers-Kipping et
11
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
al. 2011). Evidence that FOXP2 affects abilities other than language, such as cognitive
fluidity, and even IQ (Kurt, Fisher, & Ehret 2012; Lai et al. 2001), and plays a broad role in
the modulation of neural plasticity (Fisher & Scharff 2009), suggest that the neurological
basis of FOXP2 deficits lie in the structural and functional abnormalities of cortico-striatal
and cortico-cerebellar circuitries, which are important for learning, memory, and motor
control, not language exclusively. The structure of language is now widely believed to have
come about through ontogenetic human learning and processing mechanisms (Christiansen &
Chater 2008), with diversity of linguistic organization the rule rather than the exception, and
instances of universality reflecting stable engineering solutions satisfying multiple design
constraints rather than natural selection (Evans & Levinson 2009).
Language capacity is now widely thought to be polygenic, i.e., affected by multiple
genes (Chabris et al., 2012; Kovas & Plomin, 2006), transcription genes such as FOXP2 are
often pleiotropic, i.e., affecting multiple traits. It has been firmly established that a small
perturbation (such as a mutation) can percolate through a system resulting in widespread,
large-scale change, a phenomenon known as self-organized criticality (Bak, Tang, &
Wiesenfeld, 1987), and self-organized criticality is particularly widespread in regulatory
genes (Kauffman, 1993). Thus, the possibility that FOXP2 plays a role in cognition that
extends beyond language is worthy of consideration.
While it is clear that FOXP2 is important to language capacity in modern human
populations, a recent re-examination of the gene in a large, global, sample concludes that
there is "no evidence that the original two amino acid substitutions were targeted by a recent
sweep limited to modern humans <200 kya...[and furthermore] recent natural selection in the
ancestral Homo sapiens population cannot be attributed to the FOXP2 locus and thus Homo
sapiens' development of spoken language" (Atkinson et al. 2018, page 9). However, as noted
above, FOXP2 is only one of one of several genes strongly implicated in conditioning
language capacity; Mozzi et al. (2016) indicate that nine other genes are found, in various
states, to compromise skills listed as language, reading, and speech, thus effecting both
cognition (language and reading) and motor control (speech). Furthermore, the evolutionary
history of these genes is under investigation, and promises to enrich our understanding of the
evolution of the cognition responsible for language and, by extension, behavioral modernity.
Given the evidence that FOXP2 plays a role in the evolution of complex cognitive
abilities including language, but this relationship is not solely mediated through its effects on
the capacity for imitation, it seems reasonable to propose that FOXP2 and/or other, associated
genes or transcription factors enabled CF. Interestingly, the appearance of anatomically
modern humans in the fossil record as early as 200,000 years ago coincides with accelerated
FOXP2 evolution (Corballis 2004b; Lai et al. 2001). It has been proposed that the appearance
of anatomically modern humans was due to amino acid substitutions that differentiate the
human FOXP2 gene from that of chimpanzees (Enard et al. 2002; Zhang et al. 2002). Despite
new challenges to the evolutionary history of FOXP2 (see discussion above) the
preponderance of evidence currently suggest that within the last 200,000 years FOXP2
underwent at least two human-specific mutations, at least one of which occurred within the
last 100,000 years (Lai et al. 2001; Morgan 2013). This chronologically aligns modifications
of FOXP2 with the onset of not just anatomical modernity but also BM. Hence we propose
that the Paleolithic transition to BM reflects a mutation to FOXP2 and/or its molecular
associates that facilitated fine-tuning of the capacity to spontaneously shift between
associative and analytic modes (CF) depending on the situation by varying the specificity of
the activated memory region.
We propose that the kickoff point for the explosion of creativity and onset of language
in the Middle/Upper Paleolithic was human-specific amino acid substitutions to FOXP2
12
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
and/or its associates changed modifications of basal ganglia neurons that contribute to
cognitive flexibility. These neurons have longer dendrite length and greater synaptic plasticity
in humans compared to chimpanzees. These changes enhanced the efficiency of neural
cortico-basal ganglia circuits, enabling individuals to spontaneously adjust to what extent the
details of a given item in memory contributed to the flow of thought. For example, if not just
the salient or defining aspects of a particular representation are activated but also peripheral
aspects, then these peripheral features could trigger remote associations, thereby increasing
conceptual fluidity. Our theory is consistent with the proposal that mutations in FOXP2
and/or its associates underwrote the capacity for both the gestures and grammars associated
with BM (see Vicario 2013).
In short, we suggest that while FOXP2 is not the language gene it may have a broad
and identifiable influence on cognition by enabling the capacity for CF. This is not
incompatible with Crow's (2012) proposal that the Protocadherin11XY gene pair played a
key role in establishing cerebral asymmetry and enabling complex language. However, our
proposal is compatible with evidence that FOXP2 and related genes paved the way for not just
language but other features of BM as well. By tuning the mode of thought to match the needs
of the present moment, CF allowed information to be processed at different degrees of
granularity, and from different perspectives. Thus it is by way of CF that hominins became
able to combine actions and words into an infinite variety of cultural outputs, and respond to
changing selective pressures. Individuals could engage in convergent thought for well-defined
tasks, but shift to divergent thought when they were stuck, or when they wanted to express
themselves or explore aesthetic possibilities. This enabled them to connect ideas in new ways,
resulting in advanced tools, elaborate burials, and different forms of creative expression,
including art and jewelry.
This chapter outlined a theory of how the the uniquely human capacity for collectively
generated, open-ended, adaptive cultural evolution could have come about. Although change
occurred in a mosaic fashion in the Homo lineage over more than two million years, two
significant evolutionary transitions stand out. Thus, we propose that the distinctive rich
symbolism and grammatically complex language of the genus Homo reflect two evolutionary
transitions brought about by novel forms of cognitive information processing.
First, the larger brain of H. erectus resulted in finer grained memory with detailed
representations, paving the way for rehearsal of actions, refinement of skills, novel
associations between closely related items in memory. This enabled STR, escape from
episodic proximity, representational redescription, minor improvements in cultural outputs,
and a "cultural ratcheting" that expanded the capacity for open-ended cultural evolution.
The second transition occurred approximately 100,000 BP, a period associated with
the origins of art, science and religion (Mithen, 1998). We suggest that newly-evolved basal
ganglia circuits enabled onset of contextual focus: the ability to shift between convergent and
divergent modes of thought, enabling hominins to process information from different
perspectives and at multiple levels of detail. Hominins could now put their own spin on the
ideas of others, adapting them to individual needs and tastes, leading to cumulative
innovation. Thoughts, impressions, and attitudes could be modified by thinking about them in
the context of each other, and they could be woven into an integrated "worldview" that
defines who we are in relation to the world. This allowed the capacity for self-expression,
creating an environment conducive to the emergence of complex language, including
grammar, recursion, word inflections, and syntactical structure, as well as comprehension.
This theory is consistent with findings that FOXP2 is associated with cognitive
Discussion
13
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
abilities that do not involve language, and with findings that non-language creative abilities
arose at approximately the same time as complex language (Chrusch & Gabora, 2014). It is
also consistent with findings that despite the existence of sophisticated cognitive abilities in
other species such as birds (Emery, 2016), we alone exhibit cumulative cultural evolution.
Cumulative cultural evolution may involve the 'recycling' of cortical maps such that cultural
innovations invade evolutionarily older brain circuits and inherit some of their structural
constraints (Dehaene, 2005; Lieberman, 2016).
Elsewhere we provide support for the proposed two-transition scenario obtained using
an agent-based model of cultural evolution (Gabora & Smith 2018). We note that the origins
of BM are currently being rethought in light of wide dissatisfaction with an archaic 'trait-list'
approach to its understanding (Ames, Riel-Salvatore, & Collins, 2013) and with nonlinear
models of multifaceted cultural evolutionary change (Mesoudi, 2009; McDowell, 2013). The
transitions to possession of the cognitive capacities that we propose made BM possible -- STR
and CF -- exhibit Wilson's (2010) defining characteristics of evolutionary transitions, i.e., they
are rare, incomplete (did not 'throw a switch' resulting in immediate 'turning on' of BM), and
involved new levels of organization. The increased sociality implied by the onset of STR and
CF also meets Wilson's expectation that evolutionary transitions drive "the suppression of
fitness differences within groups, causing between-group selection to become the primary
evolutionary force" (Wilson, 2010:135). It is our hope that the proposed theory of cognitive
evolution reflects an emerging 'Extended Evolutionary Synthesis' (Smith & Ruppell, 2011,
Smith, Gabora, & Gardner-O'Kearny in press; Woese, 2004). We suggest that the origins of
BM be considered an evolutionary transition that culminated in new varieties of information
both within the mind and in artificial memory systems external to it, giving way to new social
arrangements and paving the way for the complex cultural systems in which we are presently
immersed.
This work was supported by a grant (62R06523) to the first author from the Natural Sciences
and Engineering Research Council of Canada.
Acknowledgments
References
Agnoli, S., Franchin, L., Rubaltelli, E., & Corazza, G. E., (2015). An eye-tracking analysis of
irrelevance processing as moderator of openness and creative performance. Creativity
Research Journal, 27, 125-132. doi:10.1080/10400419.2015.1030304
Aiello, L. (1996). Hominin pre-adaptations for language and cognition. In Mellars P., &
Gibson K. (eds): Modeling the early human mind, pp. 89-99. McDonald Institute
Monographs, Cambridge.
Ames C., Riel-Salvatore J., & Collins B. (2013). Why we need an alternative approach to the
study of modern human behaviour. Canadian Journal Archaeology, 37(1), 21-47.
Arbib M. (2011). From mirror neurons to complex imitation in the evolution of language and
tool-use. Annual Review Anthropology, 40, 257-273. doi:10.1146/annurev-anthro-
081309-145722
Atkinson, E. G., Amanda, J. A., Palacios, Bobo D. M., Webb A. E., & Ramachandran S.
Henn, B. M. (2018). No evidence for recent selection at FOXP2 among diverse human
populations. Cell, 174, 1-12. doi:10.1016/j.cell.2018.06.048
Bak P., Tang C., & Weisenfeld K. (1987). Self-organized criticality. Physical Reiew A., 38,
364 DOI:https://doi.org/10.1103/PhysRevA.38.364
Barsalou L.W. (2005). Abstraction as dynamic interpretation in perceptual symbol systems. In
Gershkoff-Stowe L. & Rakison D. (eds): Building object categories (pp. 389-431).
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
14
Earlbaum, Carnegie Symposium Series.
Bickerton D. (2014). More than nature needs: Language, mind and evolution. Harvard
Bickerton D., & Szathmáry E. 2009. Biological foundations and origin of syntax. MIT Press,
University Press, Cambridge.
Cambridge.
Blackwell L. & d'Errico F. 2001. Evidence of termite foraging by Swartkrans early Hominins.
Proc. Natl. Acad. Sci., 98(4), 1358-1363 https://doi.org/10.1073/pnas.98.4.1358
Braun D., Plummer T., Ditchfield P., Ferraro J., Maina D., Bishop L. & Potts R. (2008).
Oldowan behavior and raw material transport: Perspectives from the Kanjera
Formation. Journal of Archaeology Science, 35, 2329-2345.
Bruner E. (2010). The evolution of the parietal cortical areas in the human genus; between
structure and cognition. In Broadfield, D., M. Yuan, K. Schick & N. Toth (eds) The
Human Brain Evolving: Paleoneurological Studies in Honor of Ralph L. Holloway, pp.
83-96. Stone Age Institute Press, Gosport, Indiana.
Byrne R. & Russon A. 1998. Learning by imitation: A hierarchical approach. Behavioral and
Calcott B. & Sterelny K. (2011). The major transitions in evolution revisited. Vienna Series
Brain Sciences 2, 667-721.
Theor. Biol.. MIT Press, Cambridge.
Call J. & Tomasello M. (2008). Does the chimpanzee have a theory of mind? 30 years later.
Trends in Cognitive Sciences, 12(5), 187-192.
Chabris C., Hebert B., Benjamin D., Beauchamp J., Cesarini D., van der Loos M.,
Johannesson M., Magnusson P., Lichtenstein P., Atwood C., Freese J., Hauser T.,
Hauser R., Christakis N., & Laibson D. (2012). Most reported genetic associations with
general intelligence are probably false positives. Psychology Science, 23(11), 1314-
1323.
Christiansen M., & Chater N. (2008). Language as shaped by the brain. Behav. Brain Sci., 31,
489-558. doi:10.1017/S0140525X08004998
Christiansen M., & Kirby S. (2003). Language evolution: Consensus and controversies.
Trends in Cognitive Sciences, 7, 300-307. doi:10.1016/S1364-6613(03)00136-0
Chrusch C. & Gabora L. (2014). A tentative role for FOXP2 in the evolution of dual
processing modes and generative abilities. Proc. 36th Ann. Mtng. Cogn. Sci. Soc.
(pp.499-504). Cognitive Science Society, Austin.
Chomsky, N. (2008). On Phases. In Freidlin, R., C.P. Otero and M.L. Subizarreta (eds)
Foundational Issues in Linguistic Theory: Essays in honor of Jean-Roger Vergnaud,
pp.133-166.
Corballis M. C. (2004). The origins of modernity: Was autonomous speech the critical factor?
Psychological Review, 111(2), 543-552.
Corballis M., (2011). The recursive mind: The origins of human language, thought, and
civilization. Princeton University Press, Princeton.
Cornish-Bowden A, & Cárdenas M. L. (2017). Life before LUCA. Journal Theoretical
Biology, 434, 68 -- 74. doi:10.1016/j.jtbi.2017.05.023
Cosmides L., & Tooby J. (1992). Cognitive adaptations for social exchange. In Barkow, J.H.,
Cosmides L., & Tooby J. (eds): The adapted mind: Evolutionary psychology and the
generation of culture, pp. 163-228. Oxford University Press, USA.
Crow T. (2012). Schizophrenia as variation in the sapiens-specific epigenetic instruction to
the embryo. Clinical Genetics., 81, 319-324. doi:10.1111/j.1399-0004.2012.01830.x
Davidson I. & Noble W. (1989). The archaeology of perception: Traces of depiction and
language. Current Anthropology, 30, 125-155. doi:10.1086/203723
Deacon T. (1997). The symbolic species: The Coevolution of Language and the Brain.
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
15
Norton, New York.
Dehaene S. (2005). Evolution of human cortical circuits for reading and arithmetic: The
"neuronal recycling" hypothesis. In Dehaene S., Duhamel J., Hauser M. & Rozzolatto
G. (eds): From monkey brain to human brain, pp.133-157. MIT Press, Cambridge.
D'Errico F., Henshilwood C., Lawson G., Canhaeren M., Tillier A., Soressi M., . . . Julien M.
(2003). Archaeological evidence for the emergence of language, symbolism and music:
An alternative multidisciplinary perspective. Journal World Prehistory, 17, 1-70.
dio:10.1023/A:1023980201043
D'Errico F., Vanhaeren M. & van Niekerk K. (2005). Nassarius kraussianus shell beads from
Blombos Cave: evidence for symbolic behavior in the Middle Stone Age. Journal of
Human Evolution, 48, 3-24. doi:10.1016/j.jhevol.2004.09.002
D'Errico F., Bartond N., Bouzouggar A., Mienis H., Richter D., . . . Lozouet P. (2009).
Additional evidence on the use of personal ornaments in the middle paleolithic of north
africa. Proceedings of the National Academy Sciences, 106, 16051-16056.
doi:10.1073/pnas.0903532106
DeLong J.P., Okie J., Moses M,. Silby R. & Brown J. (2010). Shifts in metabolic scaling,
production, and efficiency across major evolutionary transitions of life. Proceedings of
the National Academy Sciences, 107, 12941-12945. doi:10.1072/pnas.1007783107
Donald M. (1991). Origins of the modern mind: Three stages in the evolution of culture and
cognition. Harvard University Press, Cambridge.
Dunbar R. (1993). Coevolution of neocortical size, group size, and language in humans.
Behavorial Brain Science, 16(4), 681-735 https://doi.org/10.1017/S0140525X00032325
Enard W., Przeworski M., Fisher S., Lai C., Weibe V. & Kitano T. (2002). Molecular
evolution of FOXP2, a gene involved in speech and language. Nature, 418: 869-872.
Erlandson J. M. (2001). The archaeology of aquatic adaptations: Paradigms for a new
millennium. Journal of Archaeological Research, 9, 287-350.
doi:10.1023/A:1013062712695
Evans J. (2008). Dual-process accounts of reasoning, judgment and social cognition. Annual
Review Psychology, 59, 255-278. doi:10.1146/annurev.psych.59.103006.093629
Fauconnier G. & Turner M. 2002. The way we think: Conceptual blending and the mind's
hidden complexities. Basic Books, New York.
Chan, J., & Schunn, C. (2015). The impact of analogies on creative concept generation:
Lessons from an in vivo study in engineering design. Cognitive Science 39, 126-155.
doi:10.1111/cogs.12127
Feinstein, J.S. 2006. The nature of creative development. Stanford University Press, Stanford.
Finke R., Ward T. & Smith S. 1992. Creative cognition: Theory, research, and applications.
MIT Press, Cambridge.
doi:10.1126/science.1236171
Fisher S. & Ridley M. 2013. Culture, genes, and the human revolution. Science, 340, 929-930.
Fisher S. & Scharff C. (2009). FOXP2 as a molecular window into speech and language.
Trends Genet., 25, 166-177. doi:10.1016/j.tig.2009.03.002
Foley R. & Gamble C. (2009). The ecology of social transitions in human evolution.
Philosophical Transactions of the. Royal Society B: Biological Sciences, 364, 3267-
3279. doi:10.1098/rstb.2009.0136
Fuentes A. (2015). Evolution of human behavior. Oxford University Press, New York.
Finke R., Ward T. & Smith S. (1992). Creative Cognition: Theory, Research, and
Applications. MIT Press, Cambridge.
Gabora, L. (1998). Weaving, bending, patching, mending the fabric of reality: A cognitive
science perspective on worldview inconsistency. Foundations of Science 3(2), 395-428.
16
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
Gabora, L. (2000). Conceptual closure: how memories are woven into an interconnected
worldview. Annals of the New York Academy of Sciences 901, 1, 42-53.
Gabora L. (2002). Cognitive mechanisms underlying the creative process. In Proceedings of
the 4th conference on creativity & cognition, pp. 126-133. Association for Computing
Machinery, New York.
Gabora L. (2003). Contextual focus: A cognitive explanation for the cultural transition of the
Middle/Upper Paleolithic. Proc. Ann. Conf. Cogn. Sci. Soc., pp.432-437. Lawrence
Erlbaum, Mahwah NJ.
Gabora L. (2006). Self-other organization: Why early life did not evolve through natural
selection. Journal of Theoretical Biology, 241, 441 -- 450.
Gabora L. (2010). Revenge of the 'neurds': Characterizing creative thought in terms of the
structure and dynamics of human memory. Creativity Research Journal, 22, 1-13.
doi:10.1080/10400410903579494
Gabora L. (2013). An evolutionary framework for cultural change: Selectionism versus
communal exchange. Physics of Life Reviews, 10(2), 117-145.
Gabora L. (2017). Honing theory: A complex systems framework for creativity. Nonlinear
Dynamics, Psychology, and Life Sciences, 21(1), 35-88.
Gabora L. (2018). The neural basis and evolution of divergent and convergent thought. In
Vartanian O. & Jung R. (eds): The Cambridge handbook of the neuroscience of
creativity. Cambridge University Press, Cambridge.
Gabora, L. (in press.) Creativity: Linchpin in the quest for a viable theory of cultural
evolution. Current Opinion in Behavioral Sciences.
Gabora L. & Aerts D. (2009). A mathematical model of the emergence of an integrated
worldview. Journal of Mathematical Psychology, 53, 434-451.
doi:10.1016/j.jmp.2009.06.004
Gabora L., & Ranjan A. (2013). How insight emerges in distributed, content-addressable
memory. In A. Bristol, O. Vartanian, & J. Kaufman (Eds.) The neuroscience of
creativity, pp. 19-43. Cambridge, MA: MIT Press.
Gabora, L., E.O. Scott and S. Kauffman. (2013). A quantum model of exaptation:
Incorporating potentiality into evolutionary theory. Progress in Biophysics and
Molecular Biology 113, 108-116.
Gabora, L. & Steel, M. (2017). Autocatalytic networks in cognition and the origin of culture.
Journal of Theoretical Biology, 431, 87-95.
Gabora, L. & Smith, C.M. (2018). Two cognitive transitions underlying the capacity for
cultural evolution. Journal of Anthropological Sciences 96, 1-26. Doi:
10.4436/jass.96008
Geels F. (2002). Technological transitions as evolutionary reconfiguration processes: a multi-
level perspective and a case-study. Research Policy, 31, 1257-1274. doi:10.1016/S0048-
7333(02)00062-8
Godfrey-Smith, P. (2011). Darwinian populations and transitions in individuality. In Calcott
B. & Sterelny K. (eds): The major evolutionary transitions revisited. Vienna series in
theoretical biology, pp. 65-81. MIT Press, Cambridge.
Goldenfeld, N., Biancalani, T., Jafarpour, F., (2017). Universal biology and the statistical
mechanics of early life. Philosophical Transactions. Series A, Mathematical, Physical,
and Engineering Sciences, 375(2109), 20160341-20160341. doi:10.1098/rsta.2016.0341
Goren-Inbar N., Alperson N., Kiselv M., Simchoni O., Melamed Y., Ben-Nun, A.. & Werker,
E. (2004). Evidence of Hominin control of fire at Gesher Benot Ya’aqov, Israel.
Science, 304, 725-727. doi:10.1126/science.1095443
Gould, S. & E. Vrba. (1982). Exaptation -- a Missing Term in the Science of Form.
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
17
Palaeobiology 8(1), 4-15.
Haidle M. (2009). How to think a simple spear. In S.A. deBaune, F.L. Coolidge & T. Wynn
(eds): Cognitive archaeology and human evolution, pp.57-73. Cambridge University
Press, Cambridge.
Harmand S., Lewis J., Feibel C., Lepre C., Prat S., Lenoble A., . . . & Roche, H. (2015). 3.3-
million-year-old stone tools from Lomekwi 3, West Turkana, Kenya. Nature, 521, 310-
315. doi:10.1038/nature14464
Hauser, M., Chomsky, N., & Fitch, W. (2002). The faculty of language: What is it, who has it
and how did it evolve? Science, 298, 1569-1579. doi:10.1126/science.298.5598.1569
Henrich J. & Boyd R. (2002). On modeling cognition and culture: Why replicators are not
necessary for cultural evolution. Journal of Cognition and Culture, 2: 87-112.
https://doi.org/10.1163/156853702320281836
Hovers, E., Lani S., Bar-Yosef, O., & Vandermeersch, B. (2003). An early case of color
symbolism: Ochre use by modern humans in Qafzeh Cave. Current Anthropology, 44,
491-522. doi:10.1086/375869
Kauffman S. (1993). Origins of Order. Oxford University Press, Oxford.
Klein, S., Cosmides, L., Tooby, J., & Chance, S. (2002). Decisions and the evolution of
memory: Multiple systems, multiple functions. Psychological. Review, 109, 306-329.
doi:10.1037/0033-295X.109.2.306
Kovas, Y. & Polmin, R. (2006). Generalist genes: implications for the cognitive sciences.
Trends of Cognitive Science, 10, 198-203. doi:10.1016/j.tics.2006.03.001
Kovac, I., Gopnik M. & Palmour R. (2002). Sibling resemblance for specific components of
linguistic competence in families of speech/language impaired children. Journal of
Neurolinguistics, 15, 497-513.
Kurt, S., Fisher, S., & Ehret, G. (2012). FOXP2 mutations impair auditory-motor association
learning. PloS One, 7, 1-5. doi:10.1371/journal.pone.0033130
Lai C., Fisher S., Hurst J., Vargha-Khadem, F., & Monaco, A. (2001). A forkhead-domain
gene is mutated in a severe speech and language disorder. Nature, 413, 519-523.
doi:10.1038/35097076
Lalmansingh, A., Karmakar, S., Jin, Y., & Nagaich, A. K. (2012). Multiple modes of
chromatin remodeling by forkhead box proteins. BBA - Gene Regulatory Mechanisms,
1819(7), 707-715. doi:10.1016/j.bbagrm.2012.02.018
Leaf, M. & Read, D. (2012). Human thought and social organization: Archaeology on a new
plane. Lexington Publishers, Guilford, CT.
Lepre, C. J., Roche, H., Kent, D. V., Harmand, S., Quinn, R. L., Brugal, J., . . . Feibel, C. S.
(2011). An earlier origin for the acheulian. Nature, 477(7362), 82-85. doi:
10.1038/nature10372
Lieberman, P. (2016). The evolution of language and thought. Journal of Anthropological
Sciences, 94, 127-146.
Liegeois, F., Morgan, A.T., A. Connelly, and F. Vargha-Khadem. Endophenotypes of
FOXP2: dysfunction within the human articulatory network. European Journal of
Paediatric Neurology 15(4), 283-288. doi: 10.1016/j.ejpn.2011.04.006
Mania, D. & Mania, U. (2005). The natural and socio-cultural environment of Homo erectus
at Bilzingsleben, Germany. In Gamble C. (ed): The individual Hominid in context:
Archaeological investigations of Lower and Middle Palaeolithic landscapes, locales,
and artifacts (pp. 98-114). Psychology Press, Hove, UK.
McBrearty, S., & Brooks, A. (2000). The revolution that wasn't: A new interpretation of the
origin of modern human behavior. Journal of Human Evolution, 39, 453-563.
doi:10.1006/jhev.2000.0435
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
McDowell, J. J. (2013). A quantitative evolutionary theory of adaptive behavior dynamics.
Psychological Review, 120(4), 731-750. doi:10.1037/a0034244
Mesoudi, A. (2009). How cultural evolutionary theory can inform social psychology and vice
versa.Psychological Review, 116, 929-952. doi:10.1037/a0017062
Mesoudi, A., Whiten, A., & Laland, K. N. (2006). Towards a unified science of cultural
18
evolution. Behavioral and Brain Sciences, 29(4), 329-347.
doi:10.1017/S0140525X06009083
Mithen S. (1996). The Prehistory of the Mind: The Cognitive Origins of Art, Religion and
Science. Thames and Hudson Press, London.
Mellars P. (2006). Going East: New genetic and archaeological perspectives on the modern
human colonization of Eurasia. Science, 313, 796-800. doi:10.1126/science.1128402
Morgan A. 2013. Speech-language pathology insights into genetics and neuroscience: Beyond
surface behaviour. International Journal of Speech-Language Pathology, 15, 245-254.
doi:10.3109/17549507.2013.777786
Moutsou T. 2014. The Obsidian Evidence for the Scale of Social Life During the Palaeolithic.
British Archaeological Reports, Oxford.
Mozzi, A., D. Forni, M. Clerici, U. Pozzoli, S. Mascheretti, F.R. Guerini, S. Riva, N.
Bresolin, R. agliani and M. Sironi.2016. The evolutionary history of genes involved in
spoken and written language: beyond FOXP2.Nature Scientific Reports 6, Article
number: 22157 (2016). DOI https://www.nature.com/articles/srep22157.
Mulvaney, J. & Kamminga, J. (1999). Prehistory of Australia. Smithsonian Institution
Scholarly Press, Washington.
Nosek, B. A. (2007). Implicit-explicit relations. Current Directions in Psychological
Science, 16(2), 65-69. doi:10.1111/j.1467-8721.2007.00477.x
Otte, M. (2012). The management of space during the paleolithic. Quaternary
International, 247, 212-229. doi:10.1016/j.quaint.2010.11.031
Penn, D. C., Holyoak, K. J., & Povinelli, D. J. (2008) Darwin's mistake: Explaining the
discontinuity between human and nonhuman minds. Behavioral and Brain Sciences
31(2), 109-178.
Pike A., Hoffmann D., García-Diez M., Pettitt P., Alcolea J., De Balbin R. & Zilhão J. 2012.
U-series dating of Paleolithic art in 11 caves in Spain. Science, 336, 1409-1413.
A. W. G. Pike, Hoffmann, D. L., García-Diez, M., Pettitt, P. B., Alcolea, J., Balbín, R. D., . . .
Zilhão, J. (2012). U-series dating of paleolithic art in 11 caves in spain. Science, 336,
1409-1413. doi:10.1126/science.1219957
Plummer T. 2004. Flaked stones and old bones: Biological and cultural evolution at the dawn
of technology. Yearbook Physical Anthropol., 47: 118-164
.https://doi.org/10.1002/ajpa.20157
Potts, R. (2012). Environmental and behavioral evidence pertaining to the evolution of early
homo.Current Anthropology, 53, S299-S317. doi:10.1086/667704
Parfitt, S.A., Ashton, N.M. Lewis, S.G., Abel, R.L., Coope, G.R., M.H. Field…&Stringer,
C.B. (2010). Early pleistocene human occupation at the edge of the boreal zone in
northwest europe. Nature, 466, 229-233. doi:10.1038/nature09117
Rappaport R. (1999). Ritual and religion in the making of humanity. Cambridge University
Press, Cambridge.
Read, D. W. (2008). Working memory: A cognitive limit to non-human primate recursive
thinking prior to hominid evolution. Evolutionary Psychology, 6(4), 147470490800600.
doi:10.1177/147470490800600413
Read, D., Lane, D., & van der Leeuw, S. (2009). The innovation innovation. In Lane D.,
Pumain D., van der Leeuw S.E. & West G. (eds). Complexity perspectives in innovation
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
19
and social change (pp. 43-84). Springer, Netherlands.
Read, D., & van der Leeuw, S. (2015). The extension of social relations in time and space
during the Palaeolithic period and beyond. In Wenban-Smith, F. Coward, F., Hosfield
R. & Pope M. (eds): Settlement, society and cognition in human evolution (pp. 31-53).
Cambridge University Press, Cambridge.
Reimers-Kipping, S., Hevers, W., Pääbo, S., & Enard, W. (2011). Humanized Foxp2
specifically affects cortico-basal ganglia circuits. Neuroscience, 175, 75-84.
doi:10.1016/j.neuroscience.2010.11.042
Schultz, T. R., & Brady, S. G. (2008). Major evolutionary transitions in ant
agriculture. Proceedings of the National Academy of Sciences, 105(14), 5435.
doi:10.1073/pnas.0711024105
Sell, A., Cosmides, L., Tooby, J., Sznycer, D., Rueden, C. v., & Gurven, M. (2009;2008;).
Human adaptations for the visual assessment of strength and fighting ability from the
body and face.Proceedings of the Royal Society B: Biological Sciences, 276(1656), 575.
doi:10.1098/rspb.2008.1177
Shea, J. (2011). Homo sapiens is as Homo sapiens was: Behavioral variability versus
"behavioral modernity" in Paleolithic archaeology. Curr. Anthropol., 52, 1-35.
Smith C.M. & Ruppell J. 2011. What anthropologists should know about the new
evolutionary synthesis. Structure Dynamics, 5, 1-13.
https://escholarship.org/uc/item/18b9f0jb
Smith C.M. 2013. Comment on 'An evolutionary framework for cultural change;
Selectionism versus communal exchange'. Physics of Life Reviews 10: 156-157. doi:
10.1016/j.plrev.2013.03.011.
Smith, C.M., L. Gabora and W. Gardner-O'Kearny. In press. The extended evolutionary
synthesis paves the way for an evolutionary model of culture change. Cliodynamics.
Sowden, P. T., Pringle, A., & Gabora, L. (2015). The shifting sands of creative thinking:
Connections to dual-process theory. Thinking & Reasoning, 21, 40-60.
doi:10.1080/13546783.2014.885464
Straus, L.G. (2009). Has the Notion of 'Transitions; in Paleolithic Prehistory Outlived its
Usefulness? The European Record in Wider Context. Pages 3-18 in Camps, M. and
Chauhan, P. (eds). 2009. Sourcebook of Paleolithic Transitions. New York, Springer.
Suddendorf T., Addis D. & Corballis M. 2009. Mental time travel and the shaping of the
human mind. Philos. T. R. Soc. B., 364, 317-24.
Suddendorf, T., Addis, D. R., & Corballis, M. C. (2009). Mental time travel and the shaping
of the human mind. Philosophical Transactions of the Royal Society B: Biological
Sciences, 364, 1317. doi:10.1098/rstb.2008.0301
Szathmary, E., & Smith, J. (1995). The major evolutionary transitions. Nature, 374, 227-232.
Szathmary E. 2015. Toward major evolutionary transitions theory 2.0. Proc. Nat. Acad. Sci.,
Tomasello, M.(1999). The cultural origins of human cognition. Harvard University Press,
Tomasello, M. (2014). A natural history of human thinking. Harvard University Press,
doi:10.1038/374227a0
112, 10104-10111.
Cambridge.
Cambridge.
Vartanian, O. (2009). Variable attention facilitates creative problem solving. Psychology of
Aesthetics, Creativity, and the Arts, 3, 57-59. doi:10.1037/a0014781
Vetsigian, K., Woese, C. & Goldenfeld, N. (2006). Collective evolution and the genetic
code. Proceedings of the National Academies of the United States of America USA, 103,
10696 -- 10701.
PSYCHOLOGICAL BASIS TRANSITIONS ARCHAEOLOGICAL RECORD
Voorhees, B., Read, D., & Gabora, L. (in press). Identity, Kinship, and the evolution of
20
cooperation. Current Anthropology.
Vicario C. 2013. FOXP2 gene and language development: the molecular substrate of the
gestural-origin theory of speech? Front. Behav. Neurosci., 7, 1-3.
Vicario, C. M. (2013). FOXP2 gene and language development: The molecular substrate of
the gestural-origin theory of speech? Frontiers in Behavioral Neuroscience, 7.
doi:10.3389/fnbeh.2013.00099
Villmoare, B., Kimbel, W. H., Seyoum, C., Campisano, C. J., DiMaggio, E. N., Rowan, J., . . .
Reed, K. E. (2015). Early homo at 2.8 ma from ledi-geraru, afar, ethiopia. Science, 347,
1352-1355. doi:10.1126/science.aaa1343
Wiessner, P. W. (2014). Embers of society: Firelight talk among the Ju/'hoansi
bushmen. Proceedings of the National Academy of Sciences, 111, 14027.
doi:10.1073/pnas.1404212111
Whiten A. 2011. The scope of culture in chimpanzees, humans and ancestral apes. Philos. T.
R. Soc. B, 366: 997-1007.
Whiten, A. (2011). The scope of culture in chimpanzees, humans and ancestral apes. The
Royal Society. doi:10.1098/rstb.2010.033
Wilson, D. (2010). Multilevel selection and major transitions. In Pigliucci M. & Miller G.B.
(eds): Evolution: The extended synthesis, pp.81-93. MIT Press, Cambridge.
Woese, C. R. (2002). On the evolution of cells. Proceedings of the National Academy of
Sciences of the United States of America, 99(13), 8742-8747.
doi:10.1073/pnas.132266999
Woese, C. R. (2004). A new biology for a new century. Microbiology and Molecular Biology
Reviews, 68, 173-186. doi:10.1128/MMBR.68.2.173-186.2004
Wragg-Sykes, R.M. (2015). To see a world in a hafted tool: birch pitch composite technology,
cognition and memory in Neanderthals. In Wenban-Smith, F., Pope, M., Coward, F. and
Hosfield, R. (eds) Settlement, society and cognition in human evolution: landscapes in
mind, pp. 117-137. Cambridge, Cambridge University Press. doi:
10.1017/CBO9781139208697.008
Wynn, T., Overmann, K.A., Coolidge, F.L. & Janulis, K. (2017). Bootstraping ordinal
thinking. In Wynn, T. & Coolidge F.L. (eds): Cognitive models in Palaeolithic
archaeology (pp. 197-213). Oxford University Press, Oxford.
Zhang, J., Webb, D. M., & Podlaha, O. (2002). Accelerated protein evolution and origins of
human-specific features: FOXP2 as an example. Genetics, 162, 1825-1835.
|
1709.04300 | 1 | 1709 | 2017-09-12T00:16:30 | Is Smaller Better: A Proposal To Consider Bacteria For Biologically Inspired Modeling | [
"q-bio.NC",
"cs.ET"
] | Bacteria are easily characterizable model organisms with an impressively complicated set of capabilities. Among their capabilities is quorum sensing, a detailed cell-cell signaling system that may have a common origin with eukaryotic cell-cell signaling. Not only are the two phenomena similar, but quorum sensing, as is the case with any bacterial phenomenon when compared to eukaryotes, is also easier to study in depth than eukaryotic cell-cell signaling. This ease of study is a contrast to the only partially understood cellular dynamics of neurons. Here we review the literature on the strikingly neuron-like qualities of bacterial colonies and biofilms, including ion-based and hormonal signaling, and action potential-like behavior. This allows them to feasibly act as an analog for neurons that could produce more detailed and more accurate biologically-based computational models. Using bacteria as the basis for biologically feasible computational models may allow models to better harness the tremendous ability of biological organisms to make decisions and process information. Additionally, principles gleaned from bacterial function have the potential to influence computational efforts divorced from biology, just as neuronal function has in the abstract influenced countless machine learning efforts. | q-bio.NC | q-bio |
Is Smaller Better: A Proposal
To Consider Bacteria For
Biologically Inspired Modeling
Archana Ram*1 and Andrew Lo1,2
1 Department of Electrical Engineering and Computer Science, Massachusetts
Institute of Technology, Cambridge, Massachusetts 02139, 2 Sloan School of
Management, Massachusetts Institute of Technology, Cambridge,
Massachusetts 02139
*Email: [email protected]
Abstract
Bacteria are easily characterizable model organisms with an impressively compli-
cated set of capabilities. Among their capabilities is quorum sensing, a detailed
cell-cell signaling system that may have a common origin with eukaryotic cell-
cell signaling. Not only are the two phenomena similar, but quorum sensing,
as is the case with any bacterial phenomenon when compared to eukaryotes,
is also easier to study in depth than eukaryotic cell-cell signaling. This ease
of study is a contrast to the only partially understood cellular dynamics of
neurons. Here we review the literature on the strikingly neuron-like qualities of
bacterial colonies and biofilms, including ion-based and hormonal signaling, and
action potential-like behavior. This allows them to feasibly act as an analog for
neurons that could produce more detailed and more accurate biologically-based
computational models. Using bacteria as the basis for biologically feasible com-
putational models may allow models to better harness the tremendous ability of
biological organisms to make decisions and process information. Additionally,
principles gleaned from bacterial function have the potential to influence com-
putational efforts divorced from biology, just as neuronal function has in the
abstract influenced countless machine learning efforts.
Introduction
The number of bacteria on Earth is staggering. Conservative estimates claim
that there are nearly half a million bacterial species in just 30 grams of soil.43
The myth persists that bacteria are simple organisms, but this could not be fur-
ther from the truth. The complexity of bacterial function in many ways mirrors
that of eukaryotic cells.19 In this review, we examine the literature demonstrat-
ing that bacterial cells, colonies, and biofilms exhibit notable similarities to
neurons and neuronal networks, including action potential-like behavior, ion-
based signaling, and hormonal signaling.
A natural first question is "why would one even consider using bacteria instead
of neurons?" This can be answered in a few ways. First, bacteria are simple.
Take two well-studied model bacteria, Escherichia coli and Bacillus subtilis.
Their genomes are both slightly over 4Mbp long and code roughly 4000 protein-
coding genes.59 Compare this with the human genome, which is roughly 3 billion
base pairs long and codes for roughly 100,000 proteins. Even C. elegans, one
of the simplest neuroscientific model organisms, has a genome size of 100Mbp
[25x the size of a bacterial genome] and codes for almost 22,000 proteins.61 The
comparative simplicity of bacteria makes them inherently an easier organism to
study. There is, simply put, less they are capable of and, as a result, there is less
to understand about their functionality. Additionally, they are easier to work
with in a laboratory environment. One E. coli cell divides roughly every 20-60
minutes and, as a result, colonies can be grown overnight. This, coupled with
the fact that bacteria are able to become competent and uptake extracellular
1
DNA, allows for easier and quicker genetic experimentation.58 Neurons, on the
other hand, are terminally differentiated cells. As a result, division is much
slower.60
One major benefit of the use of bacteria as a model organism is the fact that
each bacterium is an organism to itself as opposed to a cell constituting part of a
larger organism. Take the phenomenon of bacterial chemotaxis, which can eas-
ily be investigated in a laboratory setting. A chemical attractant can be placed
into either a suspension or a plate of bacteria and each individual organism
will move towards the attractant.18 This sort of well-characterized input-output
behavior cannot be replicated in neuronal cell cultures and, as a result, allows
for more complex behavioral experimentation that may actually help elucidate
comparable functionality in higher-order organisms, as will be discussed in the
next section.
This paper does not necessarily seek to propose the use of bacteria as a model
organism for purely behavioral work but rather to help create more biologically
feasible models. In order to create a biologically plausible model, one must first
possess a thorough understanding of the biological underpinnings of the model,
something that cannot be said to be the case with regards to the mammalian
brain or even lower-order nervous systems owing to their tremendous complex-
ity. Despite the general lack of understanding of neuronal network function,
artificial neural networks use as their basic unit a binary input/output node
that is supposedly an abstraction of a neuron.38 This sort of abstraction is an
oversimplified view of actual neuronal function that makes the network itself
biologically implausible.39 This review begins with an overview of similarities
between mammals and lower-order organisms and proceeds to discuss the ways
in which bacterial communities mirror neuronal circuits and networks. It seeks
to propose the investigation of prokaryotes, e.g. B. subtilis, in order to gain
insight into understanding and better modeling higher-level organisms such as
mammals.
Similarities between bacteria and higher organ-
isms
Continuing with the earlier discussion about bacterial chemotaxis, hunger and
satiety detection is a relatively well-conserved system among many different
species and, as a result, is a good starting point for a discussion about inter-
species similarities.44,45,46,47,48 There exist numerous structural and molecu-
lar similarities between Drosophila melanogaster and mammals, for example,
in this regard, thereby allowing Drosophila to potentially serve as a easier-
to-study proxy for the mammalian brain in this context.49,50,51,52,53,54,55,56,57
These similarities are also seen with respect to eukaryotes such as Drosophila
and lower-level organisms. When hungry, prokaryotes and Drosophila larvae as-
2
cend nutrient gradients in a process known as chemotaxis. Drosophila larvae's
approach is in stages -- first, approach the source then, once near the source, reach
it, overshoot it and then return to the source. Their motion consists of "runs"
and "turns" -- the "runs" predominate their motion and the turns are abrupt --
generally when a decreasing chemical concentration is sensed during forward
motion. The sort of motion is akin to a biased random walk in that the or-
ganism aims to meander towards the center of the nutrition concentration but
may wander slightly along the way.16,17,18 This is similar to the mechanism em-
ployed by E. coli, which also seem to also favor crawls towards higher nutrient
concentrations as opposed to lower ones.18
The similarities between Drosophila and bacteria go further than feeding be-
haviors. Consider quorum sensing, a form of bacterial cell-cell communication
used to sense local bacterial population density. The protein AarA of the Gram-
negative soil bacterium Providencia stuartii is necessary to release the molecular
signals for quorum sensing in that species. This protein, however, appears to
be homologous to the Drosophila protease RHO, which is required to activate
epidermal growth factor receptor ligands in that organism, in addition to being
essential to its proper wing vein development and eye organization. Indeed, the
two are so similar chemically that expressing RHO in P. stuartii AarA acts as a
substitute for AarA expression, as the mutants possess relatively normal quorum
sensing capabilities. Similarly, expressing P. stuartii AarA in Drosophila RHO
mutants allows wing development to proceed normally, again allowing the sub-
stitution of the two homologs, despite their origin in two very different species.26
It appears that many signaling
This homology is not an isolated incident.
mechanisms are shared by prokaryotes and eukaryotes. In fact, the evolution
of cell-cell signaling is hypothesized to have been more reliant on horizontal
gene transfer from bacteria to animals than purely vertical inheritance.32 An
interesting example is glutamate decarboxylase, which catalyzes the amino acid
glutamate to form the neurotransmitter GABA. This enzyme is coded by a gene
acquired by eukaryotes from prokaryotes through horizontal gene transfer.26
Similarities in bacterial and neuron ion-based com-
munication
Bacteria not only have influenced the development of eukaryotic cell-cell signal-
ing, they possess a number of direct similarities to neurons, specifically in their
means of cell-cell communication and the cell membrane.
The neuronal membrane voltage is regulated by the common but important ions
Na+, Cl−, Ca2+ and K+. Briefly, K+ tends to accumulate inside the membrane,
while Na+, Cl− and Ca2+ have higher concentrations outside the membrane.
Notably, K+ is a major influence on membrane voltage.27 There is a growing
3
body of evidence that bacteria also use these ions to regulate voltage across
the bacterial cell membrane. As in neurons, Na+ accumulates on the outside of
the membrane, while K+ accumulates inside. There is even evidence for ionic
Na+/K+ exchange, perhaps mediated by pumps similar to the ones found in
neurons.28 To this end, it should be pointed out that the first structure of the
K+ channel, essential to the function of both bacteria and neurons, was first
determined from a bacterial source due to the ease of bacterial study compared
to neurons.36 The resting membrane potential of E. coli is -75 mV, only about 5
mV lower than that of neurons, suggesting additional electrophysiological par-
allels between the two cell types.29
These ions are used in bacteria not only to regulate membrane voltage, but
also as signals. The PhoP/PhoQ system in Salmonella typhimurium governs
various virulence properties of that organism, and has distinct binding sites for
both Ca2+ and Mg2+. These extracellular ions act as the signals that instigate
the action of the system in a way that appears somewhat analogous to the neu-
ronal calcium channel regulator CaBP1, which also has binding sites for both
Ca2+ and Mg2+.30,31,63
Another example of ion-based signaling is seen in Bacillus subtilis, a Gram-
positive spore-forming bacterium.21 Bacteria tend to produce biofilms when
they are stressed, e.g. when there are limited nutrients in the environment.34
When the species produces biofilms of greater than 1 million cells, the colony
naturally produces electrical oscillations that serve to modulate the biofilm's
voltage as a whole.
Intracellular and extracellular potassium ions produce a
gradient on the given substrate, towards which motile bacteria of various species
are attracted, based on the potassium's capability to alter their resting mem-
brane potential. This potassium-based attraction appears to be coupled to the
biofilm's oscillations, thereby producing a phenomenon reminiscent of graded
action potential-based neuronal signaling in higher organisms [Figure 1].22
Similarities between bacterial quorum sensing and
neuronal communication
Ions, however, are not the only sort of signals present in bacteria. Similar to
neurons, bacteria also use hormonal compounds to communicate. Bacterial quo-
rum sensing is achieved through the use of different peptides and hormones, and
in many ways, it mimics how neurons communicate. There are several types of
quorum sensing systems, classified by signaling molecule. One system involves
hormone-like compounds known as autoinducers (AIs), found in Gram-negative
bacteria. Frequently used and studied among these autoinducer compounds is
the amino acid derivative AHL (N -acyl homoserine lactone). AHLs are mainly
used for intra-species communication between Gram-negative bacteria, useful in
environments where different bacterial species share resources. The most com-
4
Figure 1: A diagram of ion-based communication in biofilms. Image reused with
permission from original publisher.22
5
mon autoinducer used in Gram-negative bacteria, mainly used for communica-
tion within a given colony, is AI-2. This is a relatively universal communication
molecule used in over 40 species of bacteria. A divergent type of quorum sensing
system is found in Gram-positive bacteria. Rather than autoinducer molecules,
it uses modified oligopeptides to signal population density.33
The canonical example of quorum sensing is the system used by Vibrio fischeri,
a bacterium that lives inside the light organ of the squid Euprymna scolopes.
Once these bacteria grow to a high enough cell density, quorum sensing allows
them to induce the expression of genes necessary for bioluminescence.26 At its
most basic, quorum sensing molecules passively diffuse through the bacterial
membrane, accumulating both intra- and extracellularly to a concentration pro-
portional to cell density. Once the signal has reached an appropriate level inside
the cell, the transcription of certain genes will begin. This signal, however, can
also be detected through receptors. Gram-negative/AI2-based systems tend to
use cytoplasmic receptors but, interestingly, gram-positive bacteria like B. sub-
tilis that exhibit oscillatory electrical communication similar to neurons also
possess membrane receptors similar to those in neurons.62 Quorum sensing can
occur in series and in parallel. It can induce the transient expression of genes,
and it can be used by one bacterial colony to "eavesdrop" on other populations.
There even exist hierarchical quorum sensing circuits. Sometimes quorum sens-
ing will produce inhibitory signals within a colony, while in other cases, bacterial
populations will "sabotage" a quorum sensing signal from another colony, and
degrade it in a process known as quorum quenching. B. subtilis, for example,
produces an enzyme called AiiA that is capable of hydrolyzing the AHL of an-
other soil bacterium, thereby inhibiting its external signaling attempts.26,33
There exist important and striking similarities between quorum sensing and
neuronal communication. Both quorum sensing and neuronal circuits operate
in series and in parallel, and both are able to develop into a hierarchical multi-
circuit system.35 Neurons can provide both excitatory and inhibitory signals,
just as quorum sensing does, and neurons can communicate with each other as
well as other cell types, just as quorum sensing can function between bacteria
of the same species or between different species.27 These parallels are unmistak-
able; it is easy to confirm that there are at least abstract similarities between
bacterial colony and neuronal function.
Abstracted bacterial systems are similar to neu-
ronal analogues
Perhaps the most compelling evidence supporting this analogy, especially from
a computational perspective, comes from the bacterial attractant concentration
detection system. When unsaturated, this system can be approximated and
therefore modeled as a linear time-invariant (LTI) system whose impulse re-
6
sponse is depicted below [Figure 2]. This response curve is almost identical to
an action potential, which one might consider to be the "impulse response" of
an individual neuron receiving external stimulation.41
This result demonstrates how modeled bacterial behavior can foreseeably serve
as a realistic proxy to modeled neuronal behavior. Bacteria have many benefits
over neurons in a computational sense -- most importantly, they are functionally
simpler than neurons and easier to observe in a non-computational biological
context. The comparatively easier-to-study nature of a bacterial population,
combined with its complex neuron-like behavior, makes the bacterium an in-
teresting and unconventional candidate for neuroscience research. The relative
simplicity of bacteria allows for ease of biologically faithful modeling, while the
complexity of their cell-cell interactions has the potential to yield insights into
higher-level behaviors from feeding strategies to distributed processing.
Conclusion
The use of bacteria in neuroscience research naturally prompts the question
of why one would use organisms without neurons for neuroscience. The use
of bacteria in cellular neuroscience may be farfetched, but we believe there is
sufficient evidence to merit further study of the similarities between bacteria
and neurons. It is important to note, however, that beyond neuroscience, there
exist numerous benefits to using bacteria instead of neurons as the basis for
biologically inspired computational models.
Consider the artificial neural network. An ANN abstracts neurons to weighted
binary input/output units.38 It is readily apparent that this abstraction is not
only an oversimplification of actual neuronal function, but is also an incorrect
representation of actual neuronal activity. The brains of higher organisms (for
example, mammals) are extraordinarily complicated. It would be a nontrivial
task at best to represent even a sliver of their activity. It is highly unlikely that
ANNs in their current state are in any way biologically faithful models. Rather,
they are a computational abstraction that has become notably divorced from
the reality upon which they are based.39
In contrast, consider the bacterium. While it cannot be abstracted to a weighted
binary input/output unit, its abilities can be far more easily viewed and under-
stood by the biologist than the neuron.
It exhibits clear ion channel-based
communication, while bacterial quorum sensing is notably similar to neuronal
communication.
In addition, when bacteria coalesce into a biofilm, they are
able to produce something akin to the graded action potential activity seen in
C. elegans, the focus of much neuroscientific research over the years.25 Again, it
is important to emphasize that bacterial behavior is not fundamentally dissim-
ilar from that of higher organisms. This allows us to claim that insights gained
from bacterial modeling could apply to similar functions in higher organisms
7
Figure 2: Above, the impulse response for the LTI approximation of a bacte-
rial concentration-detection system, reprinted with permission of the authors.41
Below, a characteristic action potential graph.42
8
such as Drosophila or mammals.
Bacteria are by no means the "perfect" neuroscientific model organism -- they
lack neurons, to begin with -- but they represent a compromise. They exhibit
network-like activity, and as a colony, they are capable of making decisions in
ways similar to primitive lower organisms.40 Their beauty comes in the simplic-
ity and relative ease of study compared to other organisms. For a model to have
a sound biological basis, the biology upon which it is based must be well under-
stood. It is intuitive enough to say that the simpler the organism, the easier it
is to understand it. Following this logic to its conclusion, bacteria -- more specif-
ically, bacterial colonies -- are a possible stepping stone for biologically faithful
neuroscientific modeling. They have the potential to lead the way to modeling
more complex organisms, especially after dedicated behavioral, molecular, and
cellular work makes their neuron-like activity less of a black box and more of a
white box.
The use of modeled bacteria has the potential to go far beyond neuroscience,
however. One should consider the possible benefits to computer science. By
using a modeled bacterium instead of a weighted binary input/output unit, as
does an ANN, one may allow for greater network functionality using fewer nodes
and a smaller amount of training data. While one should not expect any sort
of network based on modeled bacteria to mirror bacterial networks functionally,
it is not inconceivable that the additional complexity could enhance network
functionality in ways that simple weighted binary input/output units might not.
Finally, why should we bother with more feasible biological models at all? Are
the models we have in use today -- Bayesian networks, artificial neural networks,
and so forth -- not enough for our needs? The answer becomes obvious if we
rephrase the question: Is it worth it to allow the insights gained from biological
work to influence computational work? It does not go without saying that the
mammalian brain is an incredibly compact, tremendously powerful organ which
the most powerful models and computers cannot match. It would be foolish not
to work towards a time when the mammalian brain is not only understandable,
but influential in the creation of both software and hardware. Modeling bacte-
rial colonies is a first step in helping ensure this possibility may one day become
a reality.
Acknowledgements
This publication would not have been possible without the dedicated guidance
of Jessie Stickgold-Sarah, Ron Weiss, Patrick Winston, Alan Grossman and
Michaela Ennis. The authors would also like to thank Gurol Suel, Guoping
Feng, Anil Sindhwani, Wei Low, Vineel Adusumilli, Vipul Vachharajani, Robby
Vasen, Ramu Thiagarajan, and Nikhil Kunapuli for their time and help in edit-
ing/conceiving of this publication.
9
References
1 Krashes, M. J., DasGupta, S., Vreede, A., White, B., Armstrong, J. D., &
Waddell, S. (2009). A neural circuit mechanism integrating motivational state
with memory expression in Drosophila. Cell, 139(2), 416-427.
2 Stuber, G. D., & Wise, R. A. (2016). Lateral hypothalamic circuits for feeding
and reward. Nature neuroscience, 19(2), 198-205. Chicago
3 Tessmar-Raible, K., Raible, F., Christodoulou, F., Guy, K., Rembold, M.,
Hausen, H., & Arendt, D. (2007). Conserved sensory-neurosecretory cell types
in annelid and fish forebrain: insights into hypothalamus evolution. Cell, 129(7),
1389-1400.
4 Bouret, S. G., & Simerly, R. B. (2006). Developmental programming of hy-
pothalamic feeding circuits. Clinical genetics, 70(4), 295-301.
5 Jobst, E. E., Enriori, P. J., & Cowley, M. A. (2004). The electrophysiol-
ogy of feeding circuits. Trends in Endocrinology & Metabolism, 15(10), 488-499.
6 Denis, R. G., Joly-Amado, A., Webber, E., Langlet, F., Schaeffer, M., Padilla,
S. L., ... & Martinez, S. (2015). Palatability can drive feeding independent of
AgRP neurons. Cell metabolism, 22(4), 646-657.
7 Dietrich, M. O., Zimmer, M. R., Bober, J., & Horvath, T. L. (2015). Hypotha-
lamic Agrp neurons drive stereotypic behaviors beyond feeding. Cell, 160(6),
1222-1232.
8 Krashes, M. J., Koda, S., Ye, C., Rogan, S. C., Adams, A. C., Cusher, D. S.,
... & Lowell, B. B. (2011). Rapid, reversible activation of AgRP neurons drives
feeding behavior in mice. The Journal of clinical investigation, 121(4), 1424.
9 Bouret, S. G., Draper, S. J., & Simerly, R. B. (2004). Trophic action of lep-
tin on hypothalamic neurons that regulate feeding. Science, 304(5667), 108-110.
10Stuber, G. D., & Wise, R. A. (2016). Lateral hypothalamic circuits for feeding
and reward. Nature neuroscience, 19(2), 198-205. Chicago
11 Alhadeff, A. L., Hayes, M. R., & Grill, H. J. (2014). Leptin receptor sig-
naling in the lateral parabrachial nucleus contributes to the control of food
intake. American Journal of Physiology-Regulatory, Integrative and Compara-
tive Physiology, 307(11), R1338-R1344.
12 Olsen, S. R., & Wilson, R. I. (2008). Cracking neural circuits in a tiny
brain: new approaches for understanding the neural circuitry of Drosophila.
Trends in neurosciences, 31(10), 512-520. Chicago
10
13 De Velasco, B., Erclik, T., Shy, D., Sclafani, J., Lipshitz, H., McInnes, R.,
& Hartenstein, V. (2007). Specification and development of the pars intercere-
bralis and pars lateralis, neuroendocrine command centers in the Drosophila
brain. Developmental biology, 302(1), 309-323.
14 Pool, A. H., Kvello, P., Mann, K., Cheung, S. K., Gordon, M. D., Wang,
L., & Scott, K. (2014). Four GABAergic interneurons impose feeding restraint
in Drosophila. Neuron, 83(1), 164-177.
15 Chiang, A. S., Lin, C. Y., Chuang, C. C., Chang, H. M., Hsieh, C. H.,
Yeh, C. W., ... & Wu, C. C. (2011). Three-dimensional reconstruction of brain-
wide wiring networks in Drosophila at single-cell resolution. Current Biology,
21(1), 1-11.
16 Gomez-Marin, A., Stephens, G. J., & Louis, M. (2011). Active sampling
and decision making in Drosophila chemotaxis. Nature communications, 2, 441.
17 Bargmann, C. I., & Horvitz, H. R. (1991). Chemosensory neurons with
overlapping functions direct chemotaxis to multiple chemicals in C. elegans.
Neuron, 7(5), 729-742.
18 Berg, H. C., & Brown, D. A. (1972). Chemotaxis in Escherichia coli analysed
by three-dimensional tracking. Nature, 239(5374), 500-504.
19 Shapiro, J. A. (1988). Bacteria as multicellular organisms. Scientific Ameri-
can, 258(6), 82-89.
20 Shapiro, J. A. (1998). Thinking about bacterial populations as multicel-
lular organisms. Annual Reviews in Microbiology, 52(1), 81-104.
21 Aguilar, C., Vlamakis, H., Losick, R., & Kolter, R. (2007). Thinking about
Bacillus subtilis as a multicellular organism. Current opinion in microbiology,
10(6), 638-643.
22 Humphries, J., Xiong, L., Liu, J., Prindle, A., Yuan, F., Arjes, H. A., ...
& Suel, G. M. (2017). Species-independent attraction to biofilms through elec-
trical signaling. Cell, 168(1), 200-209.
23 Keller, E. F., & Segel, L. A. (1971). Model for chemotaxis. Journal of
theoretical biology, 30(2), 225-234.
24 Hillen, T., & Painter, K. J. (2009). A user's guide to PDE models for chemo-
taxis. Journal of mathematical biology, 58(1-2), 183.
25 Lockery, S. R., & Goodman, M. B. (2009). The quest for action poten-
11
tials in C. elegans neurons hits a plateau. Nature neuroscience, 12(4), 377-378.
26 Waters, C. M., & Bassler, B. L. (2005). Quorum sensing: cell-to-cell com-
munication in bacteria. Annu. Rev. Cell Dev. Biol., 21, 319-346.
27 Bear, M. F., Connors, B. W., & Paradiso, M. A. (Eds.). (2007). Neuro-
science (Vol. 2). Lippincott Williams & Wilkins.
28 Lanyi, J. K. (1979). The role of Na+ in transport processes of bacterial
membranes. Biochimica et Biophysica Acta (BBA)-Reviews on Biomembranes,
559(4), 377-397.
29 Schuldiner, S., & Kaback, H. R. (1975). Mechanisms of active transport
in isolated bacterial membrane vesicles. 28. Membrane potential and active
transport in membrane vesicles from Escherichia coli. Biochemistry, 14(25),
5451-5461.
30 V A c(cid:13)scovi, E. G., Ayala, Y. M., Di Cera, E., & Groisman, E. A. (1997).
Characterization of the bacterial sensor protein phoq evidence for distinct bind-
ing sites for mg2+ and ca2+. Journal of Biological Chemistry, 272(3), 1440-
1443.
31 V A c(cid:13)scovi, E. G., Soncini, F. C., & Groisman, E. A. (1996). Mg 2+ as
an extracellular signal: environmental regulation of Salmonella virulence. Cell,
84(1), 165-174.
32 Hughes, D. T., & Sperandio, V. (2008).
Inter-kingdom signalling: com-
munication between bacteria and their hosts. Nature reviews. Microbiology,
6(2), 111.
33 Bassler, B. L. (2002). Small talk: cell-to-cell communication in bacteria.
Cell, 109(4), 421-424.
34 Prindle, A., Liu, J., Asally, M., Ly, S., Garcia-Ojalvo, J., & Suel, G. M.
(2015). Ion channels enable electrical communication in bacterial communities.
Nature, 527(7576), 59-63.
35 Krashes, M. J., DasGupta, S., Vreede, A., White, B., Armstrong, J. D.,
& Waddell, S. (2009). A neural circuit mechanism integrating motivational
state with memory expression in Drosophila. Cell, 139(2), 416-427.
36 Yellen, G. (1999). The bacterial K+ channel structure and its implications
for neuronal channels. Current opinion in neurobiology, 9(3), 267-273.
37 Roth, G., & Dicke, U. (2005). Evolution of the brain and intelligence. Trends
in cognitive sciences, 9(5), 250-257.
12
38 Lippmann, R. (1987). An introduction to computing with neural nets. IEEE
Assp magazine, 4(2), 4-22.
39 Staelin, D. H., & Staelin, C. H. (2011). Models for Neural Spike Compu-
tation and Cognition. CreateSpace.[Links].
40 Ben-Jacob, E., Lu, M., Schultz, D., & Onuchic, J. N. (2014). The physics of
bacterial decision making. Frontiers in cellular and infection microbiology, 4.
41 Cobo, L. C., & Akyildiz, I. F. (2010). Bacteria-based communication in
nanonetworks. Nano Communication Networks, 1(4), 244-256.
42 Image sourced from https://upload.wikimedia.org/wikipedia/commons/thumb/e/e0/Action potential schematic.png/984px-
Action potential schematic.png
43 Dykhuizen, D. E. (1998). Santa Rosalia revisited: why are there so many
species of bacteria?. Antonie van Leeuwenhoek, 73(1), 25-33.
44 Krashes, M. J., DasGupta, S., Vreede, A., White, B., Armstrong, J. D.,
& Waddell, S. (2009). A neural circuit mechanism integrating motivational
state with memory expression in Drosophila. Cell, 139(2), 416-427.
45 Stuber, G. D., & Wise, R. A. (2016). Lateral hypothalamic circuits for
feeding and reward. Nature neuroscience, 19(2), 198-205. Chicago
46 Tessmar-Raible, K., Raible, F., Christodoulou, F., Guy, K., Rembold, M.,
Hausen, H., & Arendt, D. (2007). Conserved sensory-neurosecretory cell types
in annelid and fish forebrain: insights into hypothalamus evolution. Cell, 129(7),
1389-1400.
47 Bouret, S. G., & Simerly, R. B. (2006). Developmental programming of
hypothalamic feeding circuits. Clinical genetics, 70(4), 295-301.
48 Jobst, E. E., Enriori, P. J., & Cowley, M. A. (2004). The electrophysiol-
ogy of feeding circuits. Trends in Endocrinology & Metabolism, 15(10), 488-499.
49 Denis, R. G., Joly-Amado, A., Webber, E., Langlet, F., Schaeffer, M., Padilla,
S. L., ... & Martinez, S. (2015). Palatability can drive feeding independent of
AgRP neurons. Cell metabolism, 22(4), 646-657.
50 Dietrich, M. O., Zimmer, M. R., Bober, J., & Horvath, T. L. (2015). Hypotha-
lamic Agrp neurons drive stereotypic behaviors beyond feeding. Cell, 160(6),
1222-1232.
51 Krashes, M. J., Koda, S., Ye, C., Rogan, S. C., Adams, A. C., Cusher,
13
D. S., ... & Lowell, B. B. (2011). Rapid, reversible activation of AgRP neurons
drives feeding behavior in mice. The Journal of clinical investigation, 121(4),
1424.
52 Bouret, S. G., Draper, S. J., & Simerly, R. B. (2004). Trophic action of lep-
tin on hypothalamic neurons that regulate feeding. Science, 304(5667), 108-110.
53Stuber, G. D., & Wise, R. A. (2016). Lateral hypothalamic circuits for feeding
and reward. Nature neuroscience, 19(2), 198-205. Chicago
54 Alhadeff, A. L., Hayes, M. R., & Grill, H. J. (2014). Leptin receptor sig-
naling in the lateral parabrachial nucleus contributes to the control of food
intake. American Journal of Physiology-Regulatory, Integrative and Compara-
tive Physiology, 307(11), R1338-R1344.
55 Olsen, S. R., & Wilson, R. I. (2008). Cracking neural circuits in a tiny
brain: new approaches for understanding the neural circuitry of Drosophila.
Trends in neurosciences, 31(10), 512-520. Chicago
56 De Velasco, B., Erclik, T., Shy, D., Sclafani, J., Lipshitz, H., McInnes, R.,
& Hartenstein, V. (2007). Specification and development of the pars intercere-
bralis and pars lateralis, neuroendocrine command centers in the Drosophila
brain. Developmental biology, 302(1), 309-323.
57 Pool, A. H., Kvello, P., Mann, K., Cheung, S. K., Gordon, M. D., Wang,
L., & Scott, K. (2014). Four GABAergic interneurons impose feeding restraint
in Drosophila. Neuron, 83(1), 164-177.
58 Cooper GM. (2000). The Cell: A Molecular Approach (Vol. 2). Sunder-
land (MA): Sinauer Associates.
59 Harwood, C., & Wipat, A. (2013). Microbial Synthetic Biology (Vol. 40).
Elsevier.
60 Hobert, O. (2011). Regulation of terminal differentiation programs in the
nervous system. Annual review of cell and developmental biology, 27, 681-696.
61 Fraser, A. G., Kamath, R. S., Zipperlen, P., Martinez-Campos, M., Sohrmann,
M., & Ahringer, J. (2000). Functional genomic analysis of C. elegans chromo-
some I by systematic RNA interference. Nature, 408(6810), 325.
62 Ng, W. L., & Bassler, B. L. (2009). Bacterial quorum-sensing network archi-
tectures. Annual review of genetics, 43, 197-222.
63 Wingard, J. N., Chan, J., Bosanac, I., Haeseleer, F., Palczewski, K., Ikura,
M., & Ames, J. B. (2005). Structural analysis of Mg2+ and Ca2+ binding to
14
CaBP1, a neuron-specific regulator of calcium channels. Journal of Biological
Chemistry, 280(45), 37461-37470.
15
|
1902.10825 | 1 | 1902 | 2019-02-27T23:13:06 | Multiscale Fluctuation-based Dispersion Entropy and its Applications to Neurological Diseases | [
"q-bio.NC",
"eess.SP"
] | Fluctuation-based dispersion entropy (FDispEn) is a new approach to estimate the dynamical variability of the fluctuations of signals. It is based on Shannon entropy and fluctuation-based dispersion patterns. To quantify the physiological dynamics over multiple time scales, multiscale FDispEn (MFDE) is developed in this article. MFDE is robust to the presence of baseline wanders, or trends, in the data. We evaluate MFDE, compared with popular multiscale sample entropy (MSE), and the recently introduced multiscale dispersion entropy (MDE), on selected synthetic data and five neurological diseases' datasets: 1) focal and non-focal electroencephalograms (EEGs); 2) walking stride interval signals for young, elderly, and Parkinson's subjects; 3) stride interval fluctuations for Huntington's disease and amyotrophic lateral sclerosis; 4) EEGs for controls and Alzheimer's disease patients; and 5) eye movement data for Parkinson's disease and ataxia. MFDE dealt with the problem of undefined MSE values and, compared with MDE, led to more stable entropy values over the scale factors for pink noise. Overall, MFDE was the fastest and most consistent method for the discrimination of different states of neurological data, especially where the mean value of a time series considerably changes along the signal (e.g., eye movement data). This study shows that MFDE is a relevant new metric to gain further insights into the dynamics of neurological diseases recordings. | q-bio.NC | q-bio | Multiscale Fluctuation-based Dispersion Entropy
and its Applications to Neurological Diseases
Hamed Azami1,∗, Steven E. Arnold2, Saeid Sanei3, Senior Member, IEEE, Zhuoqing Chang4, Guillermo Sapiro5,
Fellow, IEEE, Javier Escudero6, Member, IEEE, and Anoopum S. Gupta7
1
9
1
0
2
b
e
F
7
2
]
.
C
N
o
i
b
-
q
[
1
v
5
2
8
0
1
.
2
0
9
1
:
v
i
X
r
a
Abstract -- Fluctuation-based dispersion entropy (FDispEn) is
a new approach to estimate the dynamical variability of the
fluctuations of signals. It is based on Shannon entropy and
fluctuation-based dispersion patterns. To quantify the physio-
logical dynamics over multiple time scales, multiscale FDispEn
(MFDE) is developed in this article. MFDE is robust to the
presence of baseline wanders, or trends, in the data. We eval-
uate MFDE, compared with popular multiscale sample entropy
(MSE), and the recently introduced multiscale dispersion entropy
(MDE), on selected synthetic data and five neurological diseases'
datasets: 1) focal and non-focal electroencephalograms (EEGs); 2)
walking stride interval signals for young, elderly, and Parkinson's
subjects; 3) stride interval fluctuations for Huntington's disease
and amyotrophic lateral sclerosis; 4) EEGs for controls and
Alzheimer's disease patients; and 5) eye movement data for
Parkinson's disease and ataxia. MFDE dealt with the problem
of undefined MSE values and, compared with MDE,
led to
more stable entropy values over the scale factors for pink noise.
Overall, MFDE was the fastest and most consistent method
for the discrimination of different states of neurological data,
especially where the mean value of a time series considerably
changes along the signal (e.g., eye movement data). This study
shows that MFDE is a relevant new metric to gain further insights
into the dynamics of neurological diseases recordings.
Index Terms -- Complexity, multiscale fluctuation-based disper-
sion entropy, non-linearity, biomedical signals, electroencephalo-
gram, blood pressure.
I. INTRODUCTION
One of the most popular and powerful nonlinear measures
used to evaluate the dynamical characteristics of signals is en-
tropy [1] -- [4]. Shannon entropy (ShEn) and conditional entropy
(ConEn) are two key fundamental concepts in information the-
ory widely used for characterization of physiological signals
[2], [3]. ShEn and ConEn show the amount of information and
rate of information production, respectively, and are related to
the uncertainty or irregularity of data [2] -- [5]. A higher entropy
1Department of Neurology and Massachusetts General Hospital, Harvard
University, Charlestown, MA 02129, USA.
*Corresponding author, email: [email protected].
2S. E. Arnold is with the Department of Neurology and Massachusetts
General Hospital, Harvard Medical School, Charlestown, MA 02129, USA.
3S. Sanei is with the School of Science and Technology, Nottingham Trent
University, Nottingham, UK.
4Z. Chang is with the Department of Electrical and Computer Engineering,
Duke University, Durham, NC 27707, USA.
5G. Sapiro is with the Department of Electrical and Computer Engineering,
Computer Sciences, Biomedical Engineering, and Math, Duke University,
Durham, NC 27707, USA.
6J. Escudero is with the Institute for Digital Communications, School of
Engineering, University of Edinburgh, Edinburgh, UK.
7A. S. Gupta is with the Department of Neurology and Massachusetts
General Hospital, Harvard Medical School, Boston, MA 02114, USA.
value demonstrates higher irregularity, while smaller entropy
values show lower irregularity or uncertainty in a time series
[2], [4], [6].
Existing entropy techniques, such as sample entropy (Sam-
pEn) and permutation entropy (PerEn), are widely used to
quantify the irregularity of signals at one temporal scale
[4], [5]. They assess repetitive patterns and return maximum
values for completely random processes [7] -- [9]. However,
these techniques fail to account for the multiple time scales
inherent in biomedical recordings [8], [10]. To deal with this
limitation, multiscale SampEn (MSE) was proposed [11] and
it has become a prevalent algorithm to quantify the complexity
of univariate time series, especially physiological recordings
[8], [12].
Following [8], [11], the concept of complexity stands for
"meaningful structural richness", which may be in contrast
with uncertainty or irregularity of time series defined by classi-
cal entropy approaches such as SampEn and PerEn [4], [7], [8],
[13]. As mentioned above, these entropy approaches evaluate
repetitive patterns and return maximum values for completely
random processes [7] -- [9]. However, a completely ordered time
series with a low entropy value or a completely disordered
signal with a high entropy value is the least complex [7], [8],
[14]. For instance, white noise is more irregular than pink
noise (1/f noise) even though the latter is more complex since
the pink noise has long-range correlations and its 1/f decay
produces a fractal structure in time [7], [8], [14].
In brief, the concept of complexity builds on three hypothe-
ses: I) the complexity of a physiological time series indicates
its ability to adapt and function in ever-changing environment;
II) a biological time series requires to operate across multiple
temporal scales and so, its complexity is similarly multiscaled
and hierarchical; and III) a wide class of disease states, in
addition to aging, decrease the adaptive capacity of the individ-
ual, thus reducing the information carried by output variables.
Therefore, the MSE focuses on quantifying the information
expressed by the physiologic dynamics over multiple temporal
scales [7], [8].
In spite of its popularity, MSE is undefined or unreliable for
very short signals and computationally complex for real-time
applications as a result of using SampEn [10], [15]. To address
these shortcomings, multiscale PerEn (MPE) was proposed
[15]. Although MPE is able to deal with short signals and
is considerably faster than MSE, it does not fulfill the key
hypotheses of the concept of complexity as described above
[16]. Furthermore, the behaviour of MPE is different from that
of MSE in some cases so, in reality, it is not a replacement. To
overcome the limitations of MPE and MSE at the same time,
we have recently introduced multiscale dispersion entropy
(DispEn - MDE), based on our developed DispEn [4], [17],
to quantify the complexity of signals [18].
Compared with the conventional complexity approaches, 1)
MDE increases the reliability of the results and at the same
time does not lead to undefined values for short signals, 2)
MDE is markedly faster, especially for long signals, and 3)
it yields larger differences between physiological conditions,
such as subjects with epilepsy disorders or Alzheimer's disease
(AD) vs. matched controls [18].
MSE and MDE have been applied in different research
fields,
including biomedical engineering and neuroscience
[12], [19]. MSE was successfully used for the diagnosis of
depression using heart rate variability, speech recordings, and
electroencephalograms (EEGs) [20]. Using MSE, an increased
EEG signal complexity was found in Parkinson's disease
(PD) patients during non-rapid eye movement sleep at high
scale factors [21]. MDE was successfully used for sleep stage
classification using single-channel electrooculography signals
[22]. Miskovic et al. showed that slow sleep EEG data were
characterized by reduced MDE values at
low scales and
increased MDE values at high scale factors [23]. MDE and
MSE were used to discriminate AD patients from age-matched
controls using magnetoencephalogram signals [24]. The differ-
ences between the MDE values for the AD vs. healthy subjects
were more significant than their corresponding MSE-based
values.
In many real-world applications (e.g., in computing the
correlation function and in spectral analysis), the (local or
global) trends from a signal [25], [26] need to be removed. In
such methods, after detrending the local or global trends of
a time series, the fluctuations are evaluated [25], [26]. When
only the fluctuations of data are relevant or local trends of
a time series are irrelevant [25] -- [27], there is no difference
between dispersion patterns {11} , {22} , and {33} or {12}
and {23}. That is, the fluctuations of {11}, {22} , and {33}
or {12} and {23} are equal. Thus, we have very recently
introduced fluctuation-based DispEn (FDispEn) [17]. The po-
tential of FDispEn for characterization of various synthetic
and biomedical data was shown. For example, FDispEn signifi-
cantly discriminated eleven 3-4 years old children from twelve
11-14 years old subjects using their stride interval fluctuations
[17]. However,
this was never extended to multiscale for
covering a wider range of applications.
Therefore, the main contributions of this study are proposing
multiscale FDispEn (MFDE) and evaluating MFDE, MDE,
and MSE on selected synthetic signals and five neurological
datasets: focal and non-focal EEGs, stride interval fluctuations
in PD, young and elderly individuals as well as Huntington's
disease (HD) and amyotrophic lateral sclerosis (ALS), resting-
state EEG activity in AD, and eye movement data in ataxia
vs. PD.
This article is structured as follows. In Section II, the MFDE
algorithm is detailed. The synthetic and real datasets used here
are briefly described in Section III. The results and discussion
are provided in Section IV. After describing future works in
Section V, we conclude the paper in Section VI.
2
II. METHODS
A. Multiscale Fluctuation-based Dispersion Entropy (MFDE)
MFDE is based on the coarse-graining process [8] and
FDispEn [17]. Assume we have a univariate signal of length
L: u = {u1, u2, ..., uL}. In the MFDE algorithm, the original
signal u is first divided into non-overlapping segments of
length τ , named scale factor. Afterwards, the average of each
segment is calculated to derive a coarse-grained time series as
follows [8]:
xj
(τ ) =
1
τ
jτ
Xb=(j−1)τ +1
ub,
1 ≤ j ≤ (cid:22) L
τ (cid:23) = N
(1)
Of note is that other coarse-graining processes can be used
in this step [19], but, for the sake of clarity, we focus on the
original definition in this paper. Finally, the FDispEn of each
coarse-grained signal xj
(τ ) is calculated.
The FDispEn of the univariate signal of length N : x =
{x1, x2, ..., xN} is defined as follows:
Step 1) First, xj(j = 1, 2, ..., N ) are mapped to c classes
with integer indices from 1 to c. To this end, the normal
cumulative distribution function (NCDF) is first utilized to
overcome the problem of assigning the majority of xi to only
few classes, especially when thr maximum or minimum values
are noticeable larger or smaller than the mean/median value
of the signal [4], [17], [18]. For more information about the
reasons behind using NCDF, please see [4], [17].
The NCDF maps x into y = {y1, y2, ..., yN} from 0 to 1 as
follows:
yj =
xj
1
σ√2π
Z
−∞
−(t−µ)2
2σ2
e
dt,
(2)
where σ and µ are the SD and mean of time series x,
respectively. Then, we linearly assign each yi to an integer
from 1 to c. To do so, for each member of the mapped signal,
j denotes the jth
we use zc
member of the classified time series and the rounding operator
involves either increasing or decreasing a number to the next
digit [4], [17], [18].
j = round(c · yj + 0.5), where zc
i
i , zc
i+d, ..., zc
Step 2) Time series zm,c
are defined with respect to em-
bedding dimension m − 1 and time delay d according to
zm,c
i = {zc
i+(m−1)d}, i = 1, 2, ..., N − (m − 1)d
[4], [17]. Each time series zm,c
is mapped to a fluctuation-
based dispersion pattern πv0v1...vm−1 , where zc
i+d =
v1,..., zc
i+(m−1)d = vm−1. The number of possible fluctuation-
based dispersion patterns that can be assigned to each time
series zm,c
i = v0, zc
is equal to (2c − 1)(m−1) [17].
Step 3) For each (2c − 1)m−1 potential dispersion patterns
πv0...vm−1 , relative frequency is obtained as follows:
i
i
p(πv0...vm−1) =
#{i(cid:12)(cid:12)i ≤ N − (m − 1)d, z
N − (m − 1)d
m,c
i
has type πv0...vm−1 }
(3)
,
where # means cardinality. In fact, p(πv0...vm−1) shows the
number of dispersion patterns of πv0...vm−1 that is assigned to
zm,c
, divided by the total number of embedded signals with
i
embedding dimension m.
Step 4) Finally, based on Shannon's definition of entropy,
the FDispEn value is calculated as follows:
F DispEn(x, m, c, d) =
(2c−1)m−1
−
Xπ=1
p(πv0...vm−1) · ln(cid:0)p(πv0...vm−1)(cid:1) ,
(4)
It is worth noting that the mapping based on the NCDF
used in the calculation of FDispEn [4] for the first temporal
scale is maintained across all scales. In fact, in MFDE, µ and
σ of NCDF are respectively set at the average and standard
deviation (SD) of the original signal and they remain constant
for all scale factors. This approach is similar to keeping r
constant (usually 0.15 of the SD of the original signal) in the
MSE-based algorithms [8].
FDispEn deals with the differences between adjacent ele-
ments of dispersion patterns, named fluctuation-based disper-
sion patterns [17]. In this way, we have vectors with length
m − 1, which each of their elements changes from −c + 1
to c − 1. Thus, there are (2c − 1)m−1 potential fluctuation-
based dispersion patterns. For instance, let us have a series
x = {3.6, 4.2, 1.2, 3.1, 4.2, 2.1, 3.3, 4.6, 6.8, 8.4}, shown on
the top left of Fig. 1. We want to calculate the FDispEn
value of x. For simplicity, we set d = 1, m = 2, and
c = 3. The five potential fluctuation-based dispersion patterns
vs. nine potential dispersion patterns are depicted on the
right of Fig. 1. Step 1: xj (j = 1, 2, . . . , 10) are linearly
mapped into three classes with integer indices from 1 to 3,
as can be seen in Fig. 1. Step 2: a window with length
2 (embedding dimension) moves along the signal and the
number of each of the fluctuation-based dispersion patterns
is counted. Step 3: the relative frequency for both DispEn
and FDispEn are shown on the bottom left of Fig. 1. Step
4: using Equation (4), the FDispEn value of x is equal to
−( 4
9 ) + 3
9 ln( 3
9 ln( 4
When all possible fluctuation-based dispersion patterns have
the highest value of FDispEn is
9 )) = 1.0609.
equal probability value,
9 ) + 2
9 ln( 2
obtained, which has a value of ln((2c − 1)m−1). In contrast,
if there is only one p(πv0...vm−1) different from zero, which
demonstrates a completely regular/predictable time series, the
smallest value of FDispEn is obtained [17].
B. Parameters of MFDE
There are four parameters for MFDE, namely the embedding
dimension m, the number of classes c, the time delay d, and
the maximum scale factor τmax.
Based on the existing complexity-based approaches [8],
[10], [11], [15], the time delay was set equal to 1 in this study.
However, if the sampling frequency is noticeably larger than
the highest frequency component of a signal, the first minimum
or zero crossing of the autocorrelation function or mutual
information can be used for the selection of an appropriate
time delay [19], [28].
It is considered that c > 1 to avoid the trivial case of having
only one fluctuation-based dispersion pattern. For MFDE and
MDE, we set c = 6 for all the data in this work, albeit
the range 2 < c < 9 leads to similar findings. For more
information about c, m, and d, please refer to [17], [18].
3
To work with reliable statistics to calculate FDispEn, it is
recommended that the number of potential fluctuation-based
dispersion patterns is smaller than the length of the signal
((2c − 1)m−1 < L) [17]. For MFDE, the coarse-graining
τmaxk,
process causes the length of a signal decreases to j L
it is recommended to have (2c − 1)m−1 < j L
For all the following experiments, we set m = 2 and d = 1
for MFDE, MDE, and MSE. The number of classes is equal
to 6 for both the MDE and MFDE techniques. The threshold r
for MSE, which is used as a benchmark, was chosen as 0.15 of
the SD of a signal [8]. Finally, for consistency, the maximum
τmaxk.
scale factor τmax was set based on cm < j L
complexity techniques used herein [18].
τmaxk for all the
III. EVALUATION SIGNALS
To assess the ability of MFDE, compare it with MSE and
MDE, and to characterize various univariate time series, we
use the following synthetic and neurological datasets.
A. Synthetic Signals
1) The complexity of pink noise (1/f noise) is higher than
white noise, whereas the irregularity or uncertainty of the
former signal is lower than the latter [7], [8], [18]. Thus,
white and pink noise are two suitable data for assessing the
multiscale entropy techniques [7], [8], [14], [16], [29]. For
more information about white vs. pink noise, please refer to
[8], [30].
2) Physiological signals are often corrupted by different
kinds of noise, such as additive white Gaussian noise (WGN)
[31]. A WGN is also considered as a basic statistical model
used in information theory to mimic the effect of random
processes that occur in nature [32]. In order to understand
the relationship between MFDE, MSE, and MDE, and the
level of noise affecting periodic time series, we generated an
amplitude-modulated periodic signal with a WGN with diverse
power. First, we generated a time series as an amplitude-
modulated sum of two cosine waves with frequencies at 0.5
Hz and 1 Hz. The first 20 s of this series (100 s) does not have
any noise. Then, WGN was added to the time series [30].
B. Neurological Datasets
Discrimination of people with neurological diseases from
healthy subjects, or among different neurological diseases,
by analysis of their recorded time series is a long-standing
challenge in the physiological complexity literature [8], [18],
[21], [33] -- [35]. EEGs, walking stride interval time series, and
eye movement are clinical pavements that may be helpful in
diagnosis and tracking of neurological diseases states [6], [18],
[35], [36]. Using these recordings, MFDE, MDE, and MSE
are used to characterize several neurological diseases such as
ALS, AD, PD, cerebellar ataxias, and HD.
1) Dataset of Focal and Non-focal Electroencephalograms
(EEGs): Epilepsy is a common neurological condition. EEG
signals are used to identify areas that generate or propagate by
seizures [35], [37]. Generally, focal EEG signals are recorded
10
Original Signal
s {11, 22, or 33}
4
3
2
1
{12 or 23}
1
2
s
e
s
s
a
C
l
f
o
Embedding Dimension
s
e
s
s
a
C
l
f
o
3
2
1
{21 or 32}
1
2
Embedding Dimension
3
2
2
e
s
s
a
C
l
f
o
r
e
b
m
u
N
s
e
s
s
a
C
l
f
o
r
e
b
m
u
N
1
r
e
b
m
u
N
Embedding Dimension
1
2
{13}
3
1
r
e
b
m
u
N
Embedding Dimension
1
2
e
d
u
t
i
l
p
m
A
5
0
3
2
1
r
e
b
m
u
N
s
s
a
C
l
l
e
u
a
V
y
t
i
l
i
b
a
b
o
r
P
l
e
u
a
V
y
t
i
l
i
b
a
b
o
r
P
0.2
0.1
0
0.5
0
2
2
4
4
6
Sample Point
Classified Signal
Sample Point
6
8
8
10
10
Probability of Each Potential Dispersion Pattern
{11}
{21}
{31}
{12}
{22}
{32}
{13}
{23}
{33}
Dispersion Patterns
Probability of Each Potential Fluctuation-based Dispersion Pattern
s
e
s
s
a
C
l
f
o
r
e
b
m
u
N
3
2
1
{31}
1
2
{13}
{21 or 32}
{12 or 23}
Fluctuation-based Dispersion Patterns
{11, 22, or 33}
{31}
Embedding Dimension
Fig. 1: Illustration of the FDispEn vs. DispEn algorithms using linear mapping of x = {3.6, 4.2, 1.2, 3.1, 4.2, 2.1, 3.3, 4.6, 6.8, 8.4} (top left) with the
time delay 1, number of classes 3, and embedding dimension 2. The nine dispersion patterns {11, 12, 13, 21, 22, 23, 31, 32, 33} and five fluctuation-based
dispersion patterns {11, 12, 13, 21, 31} are shown on the right of Figure. The relative frequency for both DispEn and FDispEn are illustrated on the bottom
left of Figure.
from the epileptic part of the brain, whereas non-focal EEGs
correspond to brain regions unaffected by epilepsy [37]. The
ability of MFDE, MDE, and MSE to discriminate focal from
non-focal signals is evaluated by the use of an EEG dataset
(publicly-available at [38]) [35].
The dataset includes 5 patients and, for each patient, there
are 750 focal and 750 non-focal bivariate time series. The
length of each signal was 20 s with sampling frequency of
512 Hz (10240 samples). Focal and non-focal EEG time series
samples are depicted in Fig. 2. For more information, please,
refer to [35]. All subjects gave written informed consent that
their signals from long-term EEG might be used for research
purposes [35]. Before applying the complexity methods, the
time series were digitally filtered using a Hamming window
FIR band-pass filter of order 200 and cut-off frequencies 0.5
Hz and 40 Hz, a band typically used in the analysis of brain
activity.
2) Dataset of Walking Stride Interval Time Series for Young,
Elderly, and Parkinson's Disease (PD) Subjects: It was shown
that aging leads to less complex recordings of stride [8], [36].
It was also documented that the gait of ALS patients is less
stable and more temporally disorganized in comparison with
that of healthy individuals. Furthermore, advanced ALS, HD,
and PD were associated with certain common, but also distinct,
features of altered stride dynamics [36], [39]. To this end,
we use the walking stride interval fluctuations to distinguish
PD patients from healthy elderly subjects, young from elderly
people, and ALS from HD patients (next dataset).
To compare MFDE, MDE, and MSE, publicly-available
stride interval recordings were used [36], [40]. The signals
were recorded from five young, healthy men (23 - 29 years
old), five healthy old adults (71 - 77 years old), and five elderly
adults (60 - 77 years old) with PD. All the individuals walked
100
0
-100
e
d
u
t
i
l
p
m
A
-200
0
2000
Non-focal Data
4000
6000
Sample Point
Focal Data
8000
10000
100
0
-100
e
d
u
t
i
l
p
m
A
-200
0
2000
4000
6000
Sample Point
8000
10000
Fig. 2: Example of a focal and non-focal EEG time series.
continuously on level ground around an obstacle-free path
for 15 minutes. The stride interval was measured by the use
of ultra-thin, force sensitive resistors placed inside the shoe.
Fig. 3 shows an example of the stride-interval time series for
a young, an elderly, and a PD subject. For more information,
please refer to [40].
3) Dataset of Walking Stride Interval Time Series for
Huntington's Disease (HD) vs. Amyotrophic Lateral Sclerosis
(ALS) Patients: For the HD subjects, there is an increased
randomness in stride interval fluctuations as compared with
the healthy people [36], [39]. On the other hand, gait usually
becomes abnormal during the course of the ALS disease. A
)
c
e
s
(
1.5
Young Subject
50
100
Sample Point
150
Elderly Subject
200
50
100
150
200
Sample Point
Parkinson's Disease (PD) Subject
i
e
m
T
e
d
i
r
t
S
)
c
e
s
(
i
e
m
T
e
d
i
r
t
S
)
c
e
s
(
i
e
m
T
e
d
i
r
t
S
1
0
1.5
1
0
1.5
1
0
5
Huntington's Disease (HD)
Sample Point
5000
Amyotrophic Lateral Sclerosis (ALS)
Sample Point
5000
0
-500
-1000
-1500
-2000
0
0
-500
-1000
-1500
-2000
0
)
c
e
s
(
i
e
m
T
e
d
i
r
t
S
)
c
e
s
(
i
e
m
T
e
d
i
r
t
50
100
150
200
Sample Point
S
Fig. 3: Example of effects of aging and Parkinson's disease on fluctuations
of stride-interval dynamics.
Fig. 4: Example of effects of amyotrophic lateral sclerosis and Huntington's
disease on fluctuations of stride-interval dynamics.
decreased (average) walking velocity was reported in ALS
[41]. It is yet unknown if the loss of motoneurons also changes
the stride-to-stride complexity of gait.
The records, which are available at [42], are from 20 HD and
13 ALS patients. The mean age of the HD and ALS patients
respectively were 47 (range 29-71) and 54.9 years (range 36-
70). Subjects with ALS were able to walk independently for
five minutes and did not use a wheelchair or assistive device
for mobility. The subjects were instructed to walk at their
normal pace along a 77-m-long hallway for 5 minutes. To
measure the gait rhythm and the timing of the gait cycle,
force-sensitive insoles were placed in the patients' shoes. The
sampling frequency of the data was 300 Hz. Fig. 4 shows an
example of the stride-interval time series for a HD and an ALS
subject. Note that all the patients provided informed, written
consent and the study was approved by the Massachusetts
General Hospital (MGH) Institutional Review Board. For more
information about the dataset, please refer to [39].
4) Surface Electroencephalogram (EEG) Dataset of Brain
Activity in Alzheimer's Disease (AD): AD, as a neurodegen-
erative disease, is the most common form of dementia [43],
[44]. AD changes the interaction between neurons in the brain
during its progression. Consequently, it alters brain activity.
Some of these changes may be recorded by the EEG technique
[45] -- [48].
The 16-channel EEG dataset includes 11 AD patients (5
men; 6 women; age: 72.5 ± 8.3 years, all data given as mean
± SD) and 11 age-matched control healthy subjects (7 men; 4
women; age: 72.8 ± 6.1 years) [49]. To screen their cognitive
status, a mini-mental state examination (MMSE) [50] was
done. The MMSE scores for AD patients and healthy subjects
are 13.3 ± 5.6 and 30 ± 0, respectively.
The subjects were recruited from the Alzheimer's Patients'
Relatives Association of Valladolid (AFAVA), Spain. The EEG
time series were recorded with Oxford Instruments Profile
Study Room 2.3.411 EEG equipment at the Hospital Cl´ınico
Alzheimer's Disease (AD)
200
400
600
800
1000
1200
Sample Point
Healthy Control
40
30
20
10
0
e
d
u
t
i
l
p
m
A
-10
0
e
d
u
t
i
l
p
m
A
40
20
0
-20
-40
-60
0
200
400
600
800
1000
1200
Sample Point
Fig. 5: Example of effects of Alzheimer's disease on EEG time series.
Universitario de Valladolid (Spain). The EEGs were recorded
using the international 10-20 system, in an eyes closed and
resting state. All 16 electrodes were referenced to the linked
ear lobes of each individual. The signals were sampled at
256Hz and digitized with a 12-bit analog-to-digital converter.
Informed consent was obtained for all 22 subjects and the
local ethics committee approved the study. Before band-pass
filtering with cut-off frequencies 1 and 40 Hz and a Hamming
window with order 200, the signals were visually examined
by an expert physician to select 5 s epochs (1280 samples)
with minimal artifacts for analysis. On average, 30.0 ± 12.5
epochs (mean±SD) were selected from each electrode and
each subject. An example of an AD EEG signal vs. an age-
matched healthy control's EEG is shown in Fig. 5.
5) Eye Movement Dataset for Parkinsonism and Ataxia Pa-
tients: Neurodegenerative diseases affect oculomotor function
20
10
0
e
d
u
t
i
l
p
m
A
-10
0
5
0
-5
-10
-15
0
e
d
u
t
i
l
p
m
A
Ataxia
1000
2000
Sample Point
Parkinson's Disease (PD)
3000
1000
2000
Sample Point
3000
Fig. 6: Example of eye movements for Parkinson's disease vs. ataxia.
in a variety of ways, which impact vision and also provide
clues into the underlying pathology and diagnosis. Cerebellar
ataxias are an heterogeneous group of inherited and acquired
diseases. As a broad group, ataxias cause profound and charac-
teristic abnormalities in smooth pursuit, saccades, and fixation
[51]. Oculomotor abnormalities in PD are clinically more
subtle, but quantitative testing demonstrates abnormalities in
both saccades and in smooth pursuit [52], [53].
Participants with cerebellar ataxia and parkinsonism were
recruited to participate in eye movement testing in MGH Neu-
rology clinics. Stimuli for the antisaccades task were presented
on an Apple iPad screen, while simultaneously recording each
participant's face from an Apple iPhone camera sampling at
240fps. The video was processed using [54] to extract facial
landmarks, in particular the iris center. 57 participants with
cerebellar ataxia and 20 participants with parkinsonism (18
with Parkinson's disease and 2 with atypical parkinsonism)
were included in this dataset. An example of eye movements
for Parkinson's disease vs. ataxia is depicted in Fig. 6.
IV. RESULTS AND DISCUSSION
A. Synthetic Signals
Fig. 7 demonstrates the results obtained for MFDE, MDE,
and MSE using 40 different white and pink noise signals
with a length of 5,000 sample points. All the results are in
agreement with the fact that pink noise has more complex
structure than white noise, and white noise is more irregular
than pink noise [7], [8], [14]. At short scale factors, the entropy
values of white noise are higher than those of pink noise. At
high scale factors the entropy value for the coarse-grained pink
noise time series stays almost constant, whereas for the coarse-
grained white noise data monotonically decreases. A slightly
decreasing trend in MDE for pink noise is observed, but not so
much in MFDE, showing an advantage of MFDE over MDE.
6
For white noise, when the length of the signal, obtained by
the coarse-graining process, decreases (i.e., the scale factor
increases), the mean value of each segment converges to a
constant value and the SD becomes smaller. Therefore, no new
structures are revealed on higher scales. This demonstrates
white noise signals contain information only at short time
scales [8], [14]. For MSE, MDE and MFDE, we set m = 2
and d = 1, according to Subsection II-B.
The MFDE, MDE, and MSE methods are applied to the
quasi-periodic signals with additive noise using a moving
window of 450 samples (3 s) with 50% overlap. Fig. 8 demon-
strates the MFDE-, MDE- and MSE-based profiles using the
quasi-periodic signal with increasing additive noise power.
As expected, the entropy values for all the three methods
increase along the signal. At high scale factors, the entropy
values decrease due to the filtering nature of the coarse-
graining process [19]. To sum up, the results show that all
the methods lead to the similar findings, although the MDE
and MFDE values are slightly more stable than the MSE ones,
as demonstrated by the smoother nature of variations for MDE
and MFDE, compared with MSE. Therefore, when a high level
of noise is present, MDE and MFDE result in more stable
profiles than MSE.
To evaluate the computation time of MFDE (with m=2 and
3 for completeness), MDE (m=2 and 3), and MSE (m=2
and 3), we use white noise signals with different lengths,
changing from 100 to 100,000 sample points. The results are
shown in Table I. The simulations were carried out using a
PC with Intel (R) Xeon (R) CPU, E5420, 2.5 GHz and 8-
GB RAM by MATLAB R2015a. For 100 and 300 sample
points, MSE (m = 2 and 3) results in undefined values at
least at several scale factors. This does not happen for MDE
and MFDE, demonstrating the advantage of these methods
over MSE for short time series. There is no major difference
between the computation time for the MSE with m=2 and
3. The results show that for the different number of sample
points, MFDE and MDE are considerably faster than MSE.
This computational advantage of MFDE and MDE increases
markedly with the data length. It is consistent with the fact
that the computational cost of SampEn, FDispEn, and DispEn
are O(N 2), O(N ), and O(N ), respectively [4], [16], [17].
Note that
the MSE and MDE codes used in this paper
are publicly-available at http://dx.doi.org/10.7488/ds/1477 and
http://dx.doi.org/10.7488/ds/1982, respectively.
B. Neurological Datasets
In the physiological complexity literature, it is hypothesized
that healthy conditions correspond to more complex states
due to their ability to adapt to adverse conditions, exhibiting
long range correlations, and rich variability at multiple scales,
while aged and diseased individuals demonstrate complexity
loss. That is, they lose the capability to adapt to such adverse
conditions [8]. Therefore, we employ MFDE, compared with
MDE and MSE, to characterize different pathological states
using several neurological datasets. Note that we use these
standard datasets only to evaluate the complexity methods,
not to compete with other signal processing approaches.
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2.5
2
1.5
1
0
MDE
White Noise
Pink Noise
MFDE
White Noise
Pink Noise
e
r
u
s
a
e
M
y
p
o
r
t
n
E
3.5
3
2.5
2
1.5
10
20
Scale Factor
0
10
20
Scale Factor
3
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2.5
2
1.5
1
0.5
0
7
MSE
White Noise
Pink Noise
10
20
Scale Factor
Fig. 7: Mean value and SD of the MFDE, MDE, and MSE results for 40 different pink and white noise time series. The MSE values are undefined at several
high scale factors.
MFDE
MDE
60
40
20
w
o
d
n
W
i
l
a
r
o
p
m
e
T
0
0
1.5
1
0.5
60
40
20
w
o
d
n
W
i
l
a
r
o
p
m
e
T
0
0
5
Scale Factor
10
5
Scale Factor
10
2
1.5
1
0.5
60
40
20
w
o
d
n
W
i
l
a
r
o
p
m
e
T
0
0
MSE
1.5
undefined
1
values
0.5
5
Scale Factor
10
Fig. 8: Mean value and SD of the MFDE, MDE, and MSE results for the quasi-periodic time series with increasing additive noise power using a window
moving along the signal (temporal window). The MSE values at several temporal scale factors are undefined.
TABLE I: The computational time of MFDE, MDE, and MSE.
Number of samples →
MFDE(m = 2)
MFDE (m = 3)
MDE (m = 2)
MDE (m = 3)
MSE (m = 2)
MSE (m = 3)
100
0.0028 s
0.0049 s
0.0028 s
0.0053 s
300
0.0038 s
0.0061 s
0.0041 s
0.0070 s
undefined at all scales undefined at several scales
undefined at all scales
undefined at all scales
3,000
10,000
1,000
100,000
30,000
0.4157 s
0.0169 s 0.0463 s 0.1290 s
0.4945 s
0.0211 s 0.0541 s 0.1501 s
0.4189 s
0.0176 s 0.0478 s 0.1336 s
0.0224 s 0.0598 s 0.1673 s
0.5446 s
0.0743 s 0.7031 s 6.0879 s 72.1888 s
undefined at several scales 0.0681 s 0.6546 s 5.6362 s 62.3229 s
0.0073 s
0.0097 s
0.0078 s
0.0111 s
0.0113 s
1) Dataset of Focal and Non-focal Electroencephalograms
(EEGs): The ability of the MFDE, MDE, and MSE techniques
to distinguish the focal from non-focal signals is evaluated
here. The results, depicted in Fig. 9, show that the non-focal
signals are more complex than the focal ones. This fact is in
agreement with previous studies [35], [55]. Note that because
the entropy-based methods are used for stationary signals [2],
[17], we separated each signal into segments of length 2 s
(1024 sample points) and applied the algorithms to each of
them. The results demonstrate that all the techniques lead to
the similar findings, albeit MDE and MFDE are significantly
faster than MSE ones, as illustrated in Subsection III-A. It
should be mentioned that the average entropy values over 2
channels for these bivariate EEG signals are reported for these
univariate complexity techniques.
The non-parametric Mann-Whitney U-test was employed to
evaluate the differences between results for focal vs. non-focal
signals at each scale factor. In this study, the scale factors with
p-values between 0.01 and 0.05, and smaller than 0.01 are
respectively shown with + and *. The p-values demonstrate
that MFDE is the only complexity method with significant
differences at all scale factors, showing its advantage over
MSE and MDE.
2) Dataset of Walking Stride Interval Time Series for Young,
Elderly, and Parkinson's Disease (PD) Subjects: As shown
in Fig. 10, for most scale factors the average MFDE, MDE,
and MSE values are smaller in elderly subjects compared
with young subjects. This is consistent with those obtained
by transfer entropy [56] and the fact that recordings from
healthy young subjects correspond to more complex states
due to their ability to adapt to adverse conditions, whereas
older individuals' signals demonstrate complexity loss [7], [8],
[57]. The results also show that the PD patients' stride interval
recordings are less complex than those for the elderly subjects,
which is in agreement with the fact that some diseases lead
to lower complexity values [8], [11]. Since the length of each
stride interval signal was between 200 to 700 samples, we did
not separate the signals into smaller epochs.
The non-parametric Mann-Whitney U-test was employed to
evaluate the differences between results for young vs. elderly
individuals and elderly vs. PD patients at each scale factor.
The p-values demonstrate that
the best algorithm for the
discrimination of PD from elderly subjects and elderly from
young persons is MDE.
3) Dataset of Walking Stride Interval Time Series for
Huntington's Disease (HD) vs. Amyotrophic Lateral Sclerosis
(ALS) Patients: Due to their long length, the signals were
separated into epochs with length 3 s. The MFDE- and MSE-
8
MFDE
MDE
MSE
2
1.5
1
e
r
u
s
a
e
M
y
p
o
r
t
n
E
3.2
3
2.8
2.6
2.4
e
r
u
s
a
e
M
y
p
o
r
t
n
E
Focal Data
Non-focal Data
Focal Data
Non-focal Data
0
10
20
Scale Factor
0
10
20
Scale Factor
2
1.5
1
e
r
u
s
a
e
M
y
p
o
r
t
n
E
0.5
0
Focal Data
Non-focal Data
10
20
Scale Factor
Fig. 9: Mean value and SD of the results obtained by the MFDE, MDE, and MSE computed from the focal and non-focal EEGs. The scale factors with
p-values between 0.01 and 0.05, and smaller than 0.01 are respectively shown with + and *. The MSE values are undefined at high scale factors. The MSE
values are undefined at high scale factors.
4
3.5
3
2.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2
0
4
3.5
3
2.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2
0
MFDE
+
+
Elderly Subjects
Young Subjects
5
Scale Factor
MFDE
10
Elderly Subjects
Old Parkinson's Patients
5
Scale Factor
10
e
r
u
s
a
e
M
y
p
o
r
t
n
E
3.5
3
2.5
0
3.5
3
2.5
2
e
r
u
s
a
e
M
y
p
o
r
t
n
E
1.5
0
MDE
+ * + * +
Elderly Subjects
Young Subjects
5
Scale Factor
10
MDE
+
+
+
*+*
Elderly Subjects
Old Parkinson's Patients
5
Scale Factor
10
3
2.5
2
1.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
1
0
3
2.5
2
1.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
1
0
MSE
++
*
Elderly Subjects
Young Subjects
5
Scale Factor
10
MSE
+
Elderly Subjects
Old Parkinson's Patients
5
Scale Factor
10
Fig. 10: Mean value and SD of the results obtained by the MFDE, MDE, and MSE techniques computed from the young, elderly, and old Parkinson's subjects'
stride interval recordings. The scale factors with p-values between 0.01 and 0.05, and smaller than 0.01 are respectively shown with + and *. The MSE values
are undefined at high scale factors.
based results, depicted in Fig. 11, show that the stride interval
fluctuations for HD are more complex than those for the ALS
patients walking without any wheelchair or assistive device for
mobility. This is in agreement with [36], [39]. The p-values
show that both MFDE and MSE, unlike MDE, significantly
discriminated the ALS from HD patients.
4) Surface Electroencephalogram (EEG) Dataset
in
Alzheimer's Disease (AD): As the length of each EEG is 5 s,
we do not separate the signals into smaller epochs. MFDE,
MDE, and MSE were used to characterize the time series
recorded from 11 AD patients vs. 11 age-matched healthy
controls. The results are depicted in Fig. 12. The average of
MFDE, MDE, and MSE values for AD patients was smaller
than those for healthy controls at short-time scale factors,
while the AD subjects' EEGs had larger entropy values at
long-time scale factors. Herein, short-time (or low) scale
factors mean the temporal scales that are smaller than or
equal to the scale of crossing point of the curves for AD
patients vs. controls. Long-time (or high) scale factors denote
the temporal scales that are larger than the scale of crossing
point of the curves for AD patients vs. controls. For example,
short-time and long-time scale factors are 1-12 and 13-30,
respectively, for MFE in Fig. 12. All the results are consistent
with [18], [33], [34], [58], [59]. Nevertheless, for MSE, unlike
MDE and MFDE, values at high scale factors are undefined,
showing an advantage of MFDE and MDE over MSE. Another
advantage of MFDE and MDE over MSE is that these methods
lead to larger differences at a number of temporal scale factors.
Of note is that the average of the entropy values for all
the channels is reported for the univariate multiscale entropy
methods herein.
5) Eye Movement Dataset
for Parkinsonism vs. Ataxia
Patients: To deal with the stationarity of signals, we separated
each signal into epochs with length 1 s. The mean and SD
of MFDE, MDE, and MSE values for parkinsonism vs. ataxia
patients are depicted in Fig. 13. The results show that the
mean values for all the complexity methods computed from the
parkinsonism subjects are higher than those recorded from the
9
1.5
+++
MFDE
e
r
u
s
a
e
M
y
p
o
r
t
n
E
1
0.5
0
0
ALS
Huntington
10
20
Scale Factor
2.5
2
1.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
1
0
MDE
ALS
Huntington
10
20
Scale Factor
0.8
0.6
0.4
0.2
e
r
u
s
a
e
M
y
p
o
r
t
n
E
0
0
MSE
+
+++
+++ ++ +
+++
ALS
Huntington
10
20
Scale Factor
Fig. 11: Mean value and SD of results obtained by the MFDE, MDE, and MSE techniques computed from the HD and ALS subjects' stride interval recordings.
The scale factors with p-values between 0.01 and 0.05, and smaller than 0.01 are respectively shown with + and *.
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2
1.5
1
0
MFDE
MDE
MSE
3.5
3
2.5
e
r
u
s
a
e
M
y
p
o
r
t
n
E
AD Patients
Control Subjects
AD Patients
Control Subjects
10
20
Scale Factor
30
0
10
20
Scale Factor
30
e
r
u
s
a
e
M
y
p
o
r
t
n
E
2
1.5
1
0
AD Patients
Control Subjects
10
20
Scale Factor
30
Fig. 12: Mean value and SD of results of the MFDE, MDE, and MSE for 11 AD subjects vs. 11 age-matched controls. The scale factors with p-values
between 0.01 and 0.05, and smaller than 0.01 are respectively shown with + and *. The MSE values are undefined at high scale factors.
ataxia patients. This is consistent with the fact that oculomotor
impairment is dramatic and a core clinical feature of cerebellar
ataxia, whereas eye movement abnormalities in Parkinson's
disease are relatively mild.
Like the other results, the MSE values are undefined at high
scale factors. The Mann-Whitney U-test p-values show that
only MFDE was significantly different in parkinsonism and
ataxia patients across the range of scale factors. This shows
that where the mean value of a time series noticeably changes
along the signal, MFDE may be better than MSE and MDE
in detecting different states of physiological data.
On the whole, the results support that, in general, MDE
and MFDE perform better than MSE. MSE achieves better p-
values for the discrimination of ALS vs. HD subjects (Fig. 11),
but in the other datasets, it fails because it cannot be computed.
We also showed that MSE is considerably slower than MDE
and MFDE in Table I. Thus, we recommend MFDE and
MDE over MSE for the analysis of physiological recordings.
Between MDE and MFDE, based on the p-values, MDE was
better than MFDE only for the dataset of walking stride
interval signals for young, elderly, and PD subjects (Fig. 10).
However, MFDE outperformed MDE for the characterization
of three neurological datasets: 1) focal vs. non-focal EEGs
(Fig. 9); 2) stride interval fluctuations for Huntington's disease
vs. amyotrophic lateral sclerosis (Fig. 11); and 3) eye move-
ment data for parkinsonism vs. ataxia (Fig. 13). In addition,
MFDE results for pink noise were more stable than those
for MDE (Fig. 7). Furthermore, MFDE was slightly faster
than MDE (Table I). In sum, the results indicate that MFDE
was the fastest and most consistent technique to distinguish
various dynamics of the synthetic and real data, especially
when dealing with the presence of baseline wanders, or trends,
in signals.
V. FUTURE WORK
In spite of the promising findings based on MFDE and
MDE, these novel signal processing approaches should be
employed on various physiological datasets with a higher
number of subjects in order to evaluate their ability for
detection of dynamical variability of different kinds of timer
series.
The physiological nature of the findings for AD vs. con-
trols needs to be further investigated to understand why AD
patients' EEGs are less complex at low scale factors while the
controls' recording are less complex at high temporal scales.
With regard to eye movement, the higher complexity signal
in PD compared with ataxia can be coarsely explained by the
fact that eye movements are more impaired in ataxia. However,
in future work we hope to better understand more precisely
how and why abnormalities seen in ataxia result in a lower
complexity signal.
In this article, the most commonly used coarse-graining
process was used [8], [10], [15], [18]. The alternative coarse-
graining processes based on empirical mode decomposition
and finite impulse response (FIR) filters [19] can be employed
instead of the classical implementation of coarse-graining pro-
cess used herein. Refined composite MFDE based on refined
composite MDE [18] can be proposed for very short univariate
signals. The multivariate extension of MFDE dealing with both
the time and spatial domain at the same time can also be
developed.
MFDE
MDE
+
+
+
+
+
*
+
3.5
l
e
u
a
V
y
p
o
r
t
n
E
2
1.8
1.6
1.4
1.2
1
0.8
l
e
u
a
V
y
p
o
r
t
n
E
3
2.5
2
0
Ataxia Patients
Parkinson's Disease Patients
0
2
4
Scale Factor
6
10
MSE
+
+
Ataxia Patients
Parkinson's Disease Patients
2.4
2.2
2
1.8
1.6
1.4
1.2
1
l
e
u
a
V
y
p
o
r
t
n
E
Ataxia Patients
Parkinson's Disease Patients
2
4
Scale Factor
6
0
2
4
Scale Factor
6
Fig. 13: Mean value and SD of results obtained by the MFDE, MDE, and MSE techniques computed from the ataxia' and parkinsonism subjects' eye
movement recordings. The scale factors with p-values between 0.01 and 0.05, and smaller than 0.01 are respectively shown with + and *. The MSE values
are undefined at high scale factors.
VI. CONCLUSIONS
In this paper, we introduced MFDE to quantify the complex-
ity of time series based on their fluctuation-based dispersion
patterns. The results on synthetic data showed that MFDE,
MDE, and MSE lead to similar findings although MSE val-
ues were undefined at high scales. This fact, together with
their much faster computation time, makes us recommend
MFDE and MDE over MSE for the analysis of biomedical
signals. Based on the Mann-Whitney U-test p-values, MDE
outperformed MFDE only for the dataset of walking stride
interval signals for young, elderly, and PD subjects. Both
the MDE and MFDE methods significantly discriminated the
AD patients from healthy controls. However, MFDE was
better than MDE for the characterization of three neurolog-
ical datasets: 1) focal vs. non-focal EEGs; 2) stride interval
fluctuations for Huntington's disease vs. amyotrophic lateral
sclerosis; and 3) eye movement data for Parkinson's disease
vs. ataxia, potentially because MFDE is robust to changes in
the mean value of a time series, as seen in the eye movement
dataset. Additionally, MFDE, compared with MDE, led to
more stable entropy values over the scale factors for pink
noise. These observations suggest that MFDE may be better
than MSE and MDE in detecting different states of synthetic
and physiological recordings. We expect MFDE, in addition to
MDE, to be widely used for the characterization of different
physiologic data in various neurological diseases.
ACKNOWLEDGMENT
We would like to thank Dr. Pedro Espino (Hospital Clinico
San Carlos, Madrid, Spain) for his help in the recording
and selection of EEG epochs. We would like to thank Dr.
Jeremy Schmahmann, Dr. Albert Hung, and Dr. Christopher
Stephen for their help in recruiting research participants, and
Mary Donovan for her help in collecting eye movement
data from participants at MGH. We also thank Dr. Daniel
Ab´asolo (University of Surrey, Guildford, UK) for making the
Alzheimer's disease dataset available. G. Sapiro and Z. Chang
are partially supported by NSF, NIH, and DoD.
REFERENCES
[2] J. S. Richman and J. R. Moorman, "Physiological time-series analysis
using approximate entropy and sample entropy," American Journal
of Physiology-Heart and Circulatory Physiology, vol. 278, no. 6,
pp. H2039 -- H2049, 2000.
[3] C. Bandt and B. Pompe, "Permutation entropy: a natural complexity
measure for time series," Physical review letters, vol. 88, no. 17,
p. 174102, 2002.
[4] M. Rostaghi and H. Azami, "Dispersion entropy: A measure for time
series analysis," IEEE Signal Processing Letters, vol. 23, no. 5, pp. 610 --
614, 2016.
[5] C. E. Shannon, "A mathematical
theory of communication," ACM
SIGMOBILE Mobile Computing and Communications Review, vol. 5,
no. 1, pp. 3 -- 55, 2001.
[6] S. Sanei and J. A. Chambers, EEG signal processing. John Wiley &
Sons, 2007.
[7] H. C. Fogedby, "On the phase space approach to complexity," Journal
of statistical physics, vol. 69, no. 1-2, pp. 411 -- 425, 1992.
[8] M. Costa, A. L. Goldberger, and C.-K. Peng, "Multiscale entropy
analysis of biological signals," Physical Review E, vol. 71, no. 2,
p. 021906, 2005.
[9] Y.-C. Zhang, "Complexity and 1/f noise. a phase space approach,"
Journal de Physique I, vol. 1, no. 7, pp. 971 -- 977, 1991.
[10] H. Azami, A. Fern´andez, and J. Escudero, "Refined multiscale fuzzy
entropy based on standard deviation for biomedical signal analysis,"
Medical & Biological Engineering & Computing, vol. 55, no. 11,
pp. 2037 -- 2052, 2017.
[11] M. Costa, A. L. Goldberger, and C.-K. Peng, "Multiscale entropy
analysis of complex physiologic time series," Physical Review Letters,
vol. 89, no. 6, p. 068102, 2002.
[12] A. Humeau-Heurtier, "The multiscale entropy algorithm and its variants:
A review," Entropy, vol. 17, no. 5, pp. 3110 -- 3123, 2015.
[13] Y. Bar-Yam, Dynamics of complex systems, vol. 213. Addison-Wesley
Reading, MA, 1997.
[14] L. E. V. Silva, B. C. T. Cabella, U. P. da Costa Neves, and L. O. M.
Junior, "Multiscale entropy-based methods for heart rate variability com-
plexity analysis," Physica A: Statistical Mechanics and its Applications,
vol. 422, pp. 143 -- 152, 2015.
[15] F. C. Morabito, D. Labate, F. La Foresta, A. Bramanti, G. Morabito,
and I. Palamara, "Multivariate multi-scale permutation entropy for
complexity analysis of Alzheimer's disease EEG," Entropy, vol. 14,
no. 7, pp. 1186 -- 1202, 2012.
[16] S.-D. Wu, C.-W. Wu, and A. Humeau-Heurtier, "Refined scale-
dependent permutation entropy to analyze systems complexity," Physica
A: Statistical Mechanics and its Applications, vol. 450, pp. 454 -- 461,
2016.
[17] H. Azami and J. Escudero, "Amplitude-and fluctuation-based dispersion
entropy," Entropy, vol. 20, no. 3, p. 210, 2018.
[18] H. Azami, M. Rostaghi, D. Ab´asolo, and J. Escudero, "Refined com-
posite multiscale dispersion entropy and its application to biomedical
signals," IEEE Transactions on Biomedical Engineering, vol. 64, no. 12,
pp. 2872 -- 2879, 2017.
[1] S. M. Pincus, "Approximate entropy as a measure of system complex-
ity.," Proceedings of the National Academy of Sciences, vol. 88, no. 6,
pp. 2297 -- 2301, 1991.
[19] H. Azami and J. Escudero, "Coarse-graining approaches in univariate
multiscale sample and dispersion entropy," Entropy, vol. 20, no. 2,
p. 138, 2018.
11
[43] Alzheimer's Association, "2017 Alzheimer's disease facts and figures,"
Alzheimer's & Dementia, vol. 13, no. 4, pp. 325 -- 373, 2017.
[44] R. S. Wilson, E. Segawa, P. A. Boyle, S. E. Anagnos, L. P. Hizel, and
D. A. Bennett, "The natural history of cognitive decline in Alzheimer's
disease," Psychology and Aging, vol. 27, no. 4, p. 1008, 2012.
[45] J. Dauwels, F. Vialatte, and A. Cichocki, "Diagnosis of Alzheimer's
disease from EEG signals: where are we standing?," Current Alzheimer
Research, vol. 7, no. 6, pp. 487 -- 505, 2010.
[46] S. Bhat, U. R. Acharya, N. Dadmehr, and H. Adeli, "Clinical neuro-
physiological and automated EEG-based diagnosis of the Alzheimer's
disease," European Neurology, vol. 74, no. 3-4, pp. 202 -- 210, 2015.
[47] D. Ab´asolo, R. Hornero, P. Espino, D. Alvarez, and J. Poza, "Entropy
analysis of the EEG background activity in Alzheimer's disease pa-
tients," Physiological measurement, vol. 27, no. 3, p. 241, 2006.
[48] D. Ab´asolo, J. Escudero, R. Hornero, C. G ´omez, and P. Espino,
"Approximate entropy and auto mutual
information analysis of the
electroencephalogram in Alzheimer's disease patients," Medical & bio-
logical engineering & computing, vol. 46, no. 10, pp. 1019 -- 1028, 2008.
[49] J. Escudero, D. Ab´asolo, R. Hornero, P. Espino, and M. L ´opez, "Analysis
of electroencephalograms in Alzheimer's disease patients with multi-
scale entropy," Physiological measurement, vol. 27, no. 11, p. 1091,
2006.
[50] T. N. Tombaugh and N. J. McIntyre, "The mini-mental state exami-
nation: a comprehensive review," Journal of the American Geriatrics
Society, vol. 40, no. 9, pp. 922 -- 935, 1992.
[51] N. Buttner, D. Geschwind, J. C. Jen, S. Perlman, S. M. Pulst, and R. W.
Baloh, "Oculomotor phenotypes in autosomal dominant ataxias," Arch
Neurol, vol. 55, pp. 1353 -- 7, Oct 1998.
[52] M. Vidailhet, S. Rivaud, N. Gouider-Khouja, B. Pillon, A. M. Bonnet,
B. Gaymard, Y. Agid, and C. Pierrot-Deseilligny, "Eye movements in
parkinsonian syndromes," Ann Neurol, vol. 35, pp. 420 -- 6, Apr 1994.
[53] K. G. Rottach, D. E. Riley, A. O. DiScenna, A. Z. Zivotofsky, and R. J.
Leigh, "Dynamic properties of horizontal and vertical eye movements in
parkinsonian syndromes," Ann Neurol, vol. 39, pp. 368 -- 77, Mar 1996.
[54] https://www.ri.cmu.edu/publications/intraface/
[55] R. Sharma, R. B. Pachori, and U. R. Acharya, "Application of entropy
measures on intrinsic mode functions for the automated identification of
focal electroencephalogram signals," Entropy, vol. 17, no. 2, pp. 669 --
691, 2015.
[56] S. Nemati, B. A. Edwards, J. Lee, B. Pittman-Polletta, J. P. Butler,
and A. Malhotra, "Respiration and heart rate complexity: effects of
age and gender assessed by band-limited transfer entropy," Respiratory
physiology & neurobiology, vol. 189, no. 1, pp. 27 -- 33, 2013.
[57] M. U. Ahmed and D. P. Mandic, "Multivariate multiscale entropy: A
tool for complexity analysis of multichannel data," Physical Review E,
vol. 84, no. 6, p. 061918, 2011.
[58] J. Escudero, E. Acar, A. Fern´andez, and R. Bro, "Multiscale entropy
analysis of resting-state magnetoencephalogram with tensor factorisa-
tions in Alzheimer's disease," Brain research bulletin, vol. 119, pp. 136 --
144, 2015.
[59] H. Azami, D. Ab´asolo, S. Simons, and J. Escudero, "Univariate and
multivariate generalized multiscale entropy to characterise EEG signals
in Alzheimer's disease," Entropy, vol. 19, no. 1, p. 31, 2017.
[20] A. L. Goldberger and M. D. Costa, "Complexity based methods and
systems for detecting depression," Aug. 7 2014. US Patent App.
14/233,999.
[21] C.-C. Chung, J.-H. Kang, R.-Y. Yuan, D. Wu, C.-C. Chen, N.-F. Chi,
P.-C. Chen, and C.-J. Hu, "Multiscale entropy analysis of electroen-
cephalography during sleep in patients with parkinson disease," Clinical
EEG and neuroscience, vol. 44, no. 3, pp. 221 -- 226, 2013.
[22] M. M. Rahman, M. I. H. Bhuiyan, and A. R. Hassan, "Sleep stage
classification using single-channel eog," Computers in biology and
medicine, vol. 102, pp. 211 -- 220, 2018.
[23] V. Miskovic, K. J. MacDonald, L. J. Rhodes, and K. A. Cote, "Changes
in eeg multiscale entropy and power-law frequency scaling during the
human sleep cycle," Human brain mapping, vol. 40, no. 2, pp. 538 -- 551,
2019.
[24] H. Azami, E. Kinney-lang, A. Ebied, A. Fern´andez, and J. Escudero,
"Multiscale dispersion entropy for the regional analysis of resting-state
magnetoencephalogram complexity in alzheimer's disease," in Engi-
neering in Medicine and Biology Society (EMBC), 2017 39th Annual
International Conference of the IEEE, pp. 3182 -- 3185, IEEE, 2017.
[25] K. Hu, P. C. Ivanov, Z. Chen, P. Carpena, and H. E. Stanley, "Effect of
trends on detrended fluctuation analysis," Physical Review E, vol. 64,
no. 1, p. 011114, 2001.
[26] Z. Wu, N. E. Huang, S. R. Long, and C.-K. Peng, "On the trend,
detrending, and variability of nonlinear and nonstationary time series,"
Proceedings of the National Academy of Sciences, vol. 104, no. 38,
pp. 14889 -- 14894, 2007.
[27] C.-K. Peng, S. Havlin, H. E. Stanley, and A. L. Goldberger, "Quantifi-
cation of scaling exponents and crossover phenomena in nonstationary
heartbeat time series," Chaos: An Interdisciplinary Journal of Nonlinear
Science, vol. 5, no. 1, pp. 82 -- 87, 1995.
[28] F. Kaffashi, R. Foglyano, C. G. Wilson, and K. A. Loparo, "The effect
of time delay on approximate & sample entropy calculations," Physica
D: Nonlinear Phenomena, vol. 237, no. 23, pp. 3069 -- 3074, 2008.
[29] A. Humeau-Heurtier, C.-W. Wu, S. De Wu, M. Guillaume, and P. Abra-
ham, "Refined multiscale hilbert-huang spectral entropy and its applica-
tion to central and peripheral cardiovascular data," IEEE Transactions
on Biomedical Engineering, pp. 1 -- 11, 2016.
[30] H. Azami and J. Escudero, "Improved multiscale permutation en-
tropy for biomedical signal analysis: Interpretation and application to
electroencephalogram recordings," Biomedical Signal Processing and
Control, vol. 23, pp. 28 -- 41, 2016.
[31] J. Lam, "Preserving useful info while reducing noise of physiological
signals by using wavelet analysis," pp. 1 -- 20, 2011.
[32] C. Houdr´e, D. M. Mason, P. Reynaud-Bouret, and J. Rosinski, High
Dimensional Probability VII. Springer, 2016.
[33] D. Labate, F. La Foresta, G. Morabito, I. Palamara, and F. C. Morabito,
"Entropic measures of EEG complexity in Alzheimer's disease through a
multivariate multiscale approach," Sensors Journal, IEEE, vol. 13, no. 9,
pp. 3284 -- 3292, 2013.
[34] A. C. Yang, S.-J. Wang, K.-L. Lai, C.-F. Tsai, C.-H. Yang, J.-P. Hwang,
M.-T. Lo, N. E. Huang, C.-K. Peng, and J.-L. Fuh, "Cognitive and
neuropsychiatric correlates of EEG dynamic complexity in patients
with Alzheimer's disease," Progress in Neuro-Psychopharmacology and
Biological Psychiatry, vol. 47, pp. 52 -- 61, 2013.
[35] R. G. Andrzejak, K. Schindler, and C. Rummel, "Nonrandomness,
nonlinear dependence, and nonstationarity of electroencephalographic
recordings from epilepsy patients," Physical Review E, vol. 86, no. 4,
p. 046206, 2012.
[36] J. M. Hausdorff, P. L. Purdon, C. Peng, Z. Ladin, J. Y. Wei, and
A. L. Goldberger, "Fractal dynamics of human gait: stability of long-
range correlations in stride interval fluctuations," Journal of Applied
Physiology, vol. 80, no. 5, pp. 1448 -- 1457, 1996.
[37] U. R. Acharya, Y. Hagiwara, S. N. Deshpande, S. Suren, J. E. W. Koh,
S. L. Oh, N. Arunkumar, E. J. Ciaccio, and C. M. Lim, "Characterization
of focal eeg signals: a review," Future Generation Computer Systems,
2018.
[38] http://ntsa.upf.edu/
[39] J. M. Hausdorff, A. Lertratanakul, M. E. Cudkowicz, A. L. Peterson,
D. Kaliton, and A. L. Goldberger, "Dynamic markers of altered gait
rhythm in amyotrophic lateral sclerosis," Journal of applied physiology,
vol. 88, no. 6, pp. 2045 -- 2053, 2000.
[40] https://www.physionet.org/physiobank/database/gaitdb
[41] B. Goldfarb and S. Simon, "Gait patterns in patients with amyotrophic
lateral sclerosis.," Archives of physical medicine and rehabilitation,
vol. 65, no. 2, pp. 61 -- 65, 1984.
[42] https://physionet.org/physiobank/database/gaitndd/
|
1606.03592 | 2 | 1606 | 2016-08-18T09:52:31 | Balanced activation in a simple embodied neural simulation | [
"q-bio.NC"
] | In recent years, there have been many computational simulations of spontaneous neural dynamics. Here, we explore a model of spontaneous neural dynamics and allow it to control a virtual agent moving in a simple environment. This setup generates interesting brain-environment feedback interactions that rapidly destabilize neural and behavioral dynamics and suggest the need for homeostatic mechanisms. We investigate roles for both local homeostatic plasticity (local inhibition adjusting over time to balance excitatory input) as well as macroscopic task negative activity (that compensates for task positive, sensory input) in regulating both neural activity and resulting behavior (trajectories through the environment). Our results suggest complementary functional roles for both local homeostatic plasticity and balanced activity across brain regions in maintaining neural and behavioral dynamics. These findings suggest important functional roles for homeostatic systems in maintaining neural and behavioral dynamics and suggest a novel functional role for frequently reported macroscopic task-negative patterns of activity (e.g., the default mode network). | q-bio.NC | q-bio | Balanced activation in a simple embodied neural
simulation
Authors: Peter J. Hellyer 1,2,3, Claudia Clopath 1, Angie A. Kehagia2, Federico
E. Turkheimer 2, Robert Leech3
1. Department of Bioengineering, Imperial College London, Exhibition
Road, London, SW7 8TZ, UK
2. Centre for Neuroimaging Sciences, Institute of Psychiatry, Psychology
and Neuroscience, King's College London. De Crespigny Park,
London, SE5 8AF, UK
3. Computational, Cognitive and Clinical Neuroimaging Laboratory
(C3NL), Imperial College London, Hammersmith Hospital, Du Cane
Road, London, W12 0NN, UK
Corresponding author: Robert Leech ([email protected]), C3NL, Division of
Brain Sciences, Imperial College London, Hammersmith Hospital, Du Cane Road,
London, W12 0NN, UK
Abstract
In recent years, there have been many computational simulations of
spontaneous neural dynamics. Here, we explore a model of spontaneous
neural dynamics and allow it to control a virtual agent moving in a simple
environment. This setup generates interesting brain-environment feedback
interactions that rapidly destabilize neural and behavioral dynamics and
suggest the need for homeostatic mechanisms. We investigate roles for both
local homeostatic plasticity (local inhibition adjusting over time to balance
excitatory input) as well as macroscopic "task negative" activity (that
compensates for "task positive", sensory input) in regulating both neural
activity and resulting behavior (trajectories through the environment). Our
results suggest complementary functional roles for both local homeostatic
plasticity and balanced activity across brain regions in maintaining neural and
behavioral dynamics. These findings suggest important functional roles for
homeostatic systems in maintaining neural and behavioral dynamics and
suggest a novel functional role for frequently reported macroscopic "task-
negative" patterns of activity (e.g., the default mode network).
Introduction
In recent years, there has been increasing evidence that homeostatic systems
play an important role in regulating neural activity. At the microscopic level,
experimental and theoretical work suggests that the balance of local excitation
and inhibition (E/I) has important computational properties [1-2]. Further, such
E/I balance can be maintained with relatively simple local homeostatic
inhibitory plasticity (e.g., [3-6]). At the macroscopic level, there is evidence
from functional MRI that there is some level of balance in activity over regions.
Networks of brain regions showing increased activity matched by other
networks showing reduced activity (e.g., [7-9]). In our previous work, we have
suggested that macroscopic brain networks may act to counterbalance task
activation in other brain regions [10-12]. This may constitute (at the
macroscopic scale) an analogue of inhibitory mechanisms and computational
motifs seen at smaller scales [13].
Computational simulations are useful for understanding the functional roles of
these homeostatic systems; however, computational models typically simulate
the brain at rest or under constrained task settings (e.g., [14-15]). In the
present work, we explore the regulatory role of homeostatic mechanisms by
embodying a well-known neural simulation to control a simulated agent
moving through a simple environment. This setup generates interesting brain-
environment interactions that require homeostatic mechanisms to maintain
rich neural and behavioral dynamics.
The computational simulation is based on the Greenberg-Hastings model [16],
incorporating information about human structural connectivity [17] that has
been previously shown to approximate empirical functional connectivity
patterns [18]. In the model, a node was set to the ON state, with either a small
random probability or if incoming activity was greater than a local threshold
value (analogous to the amount of local inhibition).
In order to explore the interaction between brain and environment, we
embodied the computational model in a simple environment. We began by
defining an 'agent' that can move within a 2-dimensional plane, bounded by
surrounding walls (see Figure 1). Within this framework, we defined a group of
task-positive nodes (TP), which activated in response to simulated "sensory"
input. Two pairs of bilateral nodes reacted to "visual" input from the
environment to the simulated brain, and a pair reacted to "somatosensory"
input: if input was detected these nodes were set to ON. A further pair of
nodes simulated "motor" output from the simulated brain to the agent in the
environment; their activity determined the movements of the agent.
This simple setup allows us to observe interesting interactions between the
simulated neural network and the environment. Specifically, the following
environment/agent closed-loop interaction occurs: (i) different parts of the
environment evoke different amounts of visual and sensory stimulation, which
subsequently (ii) alter regional/microscopic? neural activity, (iii) leading to
altered neural dynamics macroscopically, across the entire model; (iv) these
altered dynamics in turn change the motor output from the model, v) changing
the agent's trajectory in the environment which in turn alters subsequent
sensory input. This presents a challenge for network models, especially
models focusing on spontaneous, rich dynamics, that often require careful
parameterization to remain in a specific dynamic regime, and so typically are
investigated in static situations (i.e., where the input to the model is stationary,
such as Guassian noise). In such models, changes to the model input
typically lead to destabilization of the dynamics (i.e., a shift to either random,
saturated, or absent patterns of activity).
Balanced activity has been shown to facilitate a rich, spontaneous dynamical
regime that is robust to different parameter values, [14,19]; this has not been
explored in non-stationary scenarios, such as occur in this type of embodied
models where brain-model interactions can occur. Therefore, we incorporated
two mechanisms in the model to balance activity: first, a simplified version of
the local (i.e., within-node) homeostatic plasticity rule presented in [3] and
employed in a similar macroscopic neural model in [20]. This mechanism
adjusts the local threshold (inhibition) at each node, balancing against
incoming excitatory activity from other nodes, and so driving time-averaged
local activity to approximate a pre-specified, small target activity rate. In
addition to the local mechanism, we also employed a macroscopic balancing
mechanism across brain regions, such that the activity of the task positive
nodes (i.e., the six "sensory" nodes) was balanced over time by bilateral task
negative nodes (TN); i.e., task negative nodes were ON but switched to off
when a given sensory node was turned ON. The choice of the TN nodes was
loosely based on the default mode network pattern of task-evoked relative
deactivations from fMRI/PET [21], which we have previously suggested, may
constitute a macroscopic balancing system [10].
Here, we explore the complementary roles that these two balancing systems
may play in maintaining flexible dynamics in the embodied situation with
sensory input from and motor output to the environment. In particular, we
assess the agent's neural dynamics and trajectory through the environment,
and demonstrate that these balancing mechanisms allow the agent to escape
constrained environment-brain feedback loops, and more completely traverse
the environment.
Figure 1: Illustration of the neural model (left) and two perspectives of the
model-controlled agent (grey figure, with its colored trail over time) placed in
the environment (right). "Visual" or "somatosensory" sensory input to the
agent, depends on proximity to the wall around the edge of the environment.
The agent moves based activity in "motor" nodes of the model. The model
(either compiled or as a unity project) can be downloaded from the GitHub
repository https://github.com/c3nl-neuraldynamics/Avatar/releases).
Results
target activity, !=0.1 (Figure 2A). Moreover,
We start by considering simulated neural and movement dynamics, without
the explicit task negative mechanism. Consistent with previous results [20],
we observe that over time the homeostatic model adapts the threshold
weights such that time-averaged excitatory activation approximates the pre-
specified
the model
demonstrates high levels of persistent variability around this average values
(Figure 2A and 2C), even though there is relatively little intrinsic noise in the
system. The model also displays non-zero but relatively weak positive
correlations between nodes' activity timecourses (Figure 2D).
Figure 2: How the simulation with local inhibitory plasticity changes over time
(timecourses are of average values from 1000 epochs). A: mean activation
approaches the pre-specified target value !=0.1. B: this is accomplished by
reductions and then gradual increases in local thresholds (consistent with
[20]). C: We observe increased variability of activity (standard deviation of
activity over time, averaged across nodes). Finally, D: we also observe an
00.050.10.150.26.577.588.59Time02000040000600008000010000012345678Time02000040000600008000010000000.050.10.150.20.250.3ABCDAvg activity Avg threshold SD activityAvg correlation increase in connectivity across the network (measured as mean correlation
between nodes). (Results are presented for a single 100,000 long training run,
although qualitatively similar results were found for different initial conditions
and random seeds).
We also observed the simulated trajectory: over time the model moves into a
regime with generally higher levels of movement (i.e., left/right rotation and/or
forward motion) (Figure 3A), although there is considerable variability (i.e., the
mean level of movement and activity varies considerably over time). Several
example trajectories (each 1000 epochs long) are presented in (Figure 3B).
Figure 3: A "Motion" output (either nodes linked to "forward" output to the
agent or nodes linked to "turn" over time. This is calculated from the activity
of the "motor" nodes). Note that this figure presents the average output
generated by the model rather than the motion of the actual agent, which can
be impeded by obstacles (i.e., walls) in the environment, in this case, the
simulated brain could be sending a move forward command, but this cannot
be achieved because of the wall. B Example trajectories of the agent over
012345678910Average forward movement00.020.04Average amount of turning00.20.4ABTime x 104 nine randomly chosen time windows of 1000 epochs. Light colors are earlier
in the time window, warm colors are later in the time window.
However, we observe that the dynamic regimes of the simulated neural
activity and motion are not stationary; the model can be characterized as
alternating between periods of high and low activity, with correspondingly high
and low amounts of movement. Similarly, it never settles down into a stable
regime, with a single distribution of movements/activity (see Figure 4 for an
illustrative shorter 1000 epoch time period). This alternating pattern reflects a
feedback loop arising from how the agent interacts with the environment. The
level of sensory input: i) alters the level of simulated activity, which ii) alters
the level of motor output, iii) that manifests in agent movement, which in turn,
iv) may alter subsequent sensory input.
Figure 4: An illustration of the brain/environment interaction. A: At short time
periods, there are brief, but sustained periods of elevated activity substantially
above the target rate. B: The homeostatic rule is continuously adapting to the
level of activity, tightly matched sustained changes in the local threshold. C:
Model alternating between low and high average threshold levels (or
inhibition). D: The sustained changes are being driven by the sensory input
the model receives, which closely match overall activity; E: This demonstrates
how sensory input and activity are related to how close the agent is to a wall
(where sensory
"visual" and
"somatosensory" sensors. F: Strong relationship between proximity to the wall
and level of activity. G/H the threshold change (G) and the sensory input (H)
(left axes) relate to average movement (right axis), over time.
is highest, given
input
triggering of
video:
A simple illustration of this feedback is presented in Figure 4 (and
supplementary
https://github.com/c3nl-
neuraldynamics/Avatar/blob/master/Figures/FigureS1.gif; there are high levels
of simulated visual and somatosensory input at the edges of the environment
(that is touching and/or looking at the wall) and low or no sensory input in the
center. Therefore, high sensory input triggers higher average activity which
leads to elevated motor output, which on average moves the agent forward
into the wall and in turn keeps the input activity high. This feedback cycle
means the agent remains trying to run into the wall, trapped next to the wall
and with sustained elevated activity patterns. In contrast, away from the wall,
where there is little or no sensory feedback, the levels of activity remain low,
and motor output and movement are either low or not present, meaning that
the agent remains in a very low simulated neural activity and movement
regime. The presence of homeostatic plasticity mechanism that tunes the
threshold at each node to balance excitatory input from connected nodes
ensures that the agent does not stay trapped in either state for long. As the
threshold for activity (varying depending on the local homeostatic mechanism)
at individual nodes increases (in the high activity state) or decreases (in the
low activity state), the average activity level adapts to the target level. This
results in the agent escaping the 'trap', with resulting activity levels closer to
the target level ! and, consequently, more stable simulated neural and
movement dynamics.
The model without local homeostasis is typically unable to cope with the
sensory/motor feedback system. Local thresholds can be chosen to allow
interesting dynamics (i.e., variable movements/neural activity); however,
these must be chosen to either allow rich dynamics in the presence of
sensory input (i.e., with higher local inhibition) or dynamics in the absence of
sensory input (i.e., with lower local inhibition). Therefore, over time the agent
will tend to either: a) remain approximately stationary in a low-sensory area
with local thresholds too great to allow much exploration (i.e., near stationary,
see Figure 5C); or b) become trapped in a high-sensory area running into the
wall (e.g., a corner) (see Figure 5A and 5B).
Figure 5: Example trajectories (from initialization) of the agent without local
homeostatic plasticity, and threshold weights set uniformly at 5, 7 or 9 and run
for 1000 epochs. With lower weights, the agent receives high levels of
excitatory activity across the brain, and walks into the wall and is trapped.
With higher weights, activity within the network is very low and driven mainly
by
through
connections; as a result the agent moves very little over the course of the
1000 epochs. Cool colors are earlier in the time window, warm colors are later
in the time window.
than activity propagating
rather
random
local excitation
While the model with local homeostasis is able to deal with this sensory-motor
interaction, the addition of an explicit task-negative system, alongside the
homeostatic learning rule, facilitates stable simulated neural dynamics and
behavioral trajectories through the environment. This occurs because the task
negative system balances changes in external input to the model so that the
number of activated units (sensory nodes or task negative nodes) remains
constant irrespective of interactions with the environment.
To demonstrate the complementary role of the two homeostatic systems, we
compared the model with just the local homeostatic mechanism with the
model with the local homeostatic mechanism and the task-negative system on
a range of measures assessing the model's dynamics (simulated neural
activity, simulated movement, and threshold changes: all differences reported
were significant at t>3). We see that for both types of model, there are similar
levels of mean simulated activity across the network (excluding sensory or
task-negative nodes), approximately equal to the target rate (Figure 6A).
However, we see that activity and threshold weight changes are less variable
for the TN (Figure 6C) model. This is the case when considering both the
standard deviation and the coefficient of variation (s.d./mean) of activity. Also,
there is a strong negative relationship between distance from the wall and the
amount of activity in the local homeostasis model (Spearman Rho=-0.45),
whereas this relationship becomes much smaller in the task-negative model
(Spearman Rho=-0.25),
between
brain/environment is less influential.
suggesting
loop
the
feedback
Figure 6: Comparing simulations with and without TN (blue is without, red is
with TN). We see that while the level of activation is similar in the two types of
model (A, mean activity for non-TN or task positive nodes), there are higher
average thresholds (B, again for non-TN or TP nodes), possibly because
there are never completely quiescent periods;
is
importantly
there
Avg activity 00.10.2AAvg threshold 6810BSD activity0510CAvg correlation 00.20.4DTime020000400006000080000100000Trajectory entropy00.20.4ETime020000400006000080000100000Fractal dimension11.21.4F substantially reduced variability in activity over time for the TN model (C); and
lower average correlations between nodes (D). All results presented here and
in the text comparing simulations with and without TN were highly statistically
significant (t>3), with data sampled from each of the 1000 epoch blocks. Also,
(E) image entropy from plotted trajectories and (F) fractal dimension from
plotted trajectories calculated over 1000 epoch blocks.
When considering motor output (i.e., "motor signals" sent to control movement
of the agent, e.g., turn left, go forward), we observe that there is significantly
higher entropy for the plotted trajectories of the TN model. Similarly, when
observing the movement of the agent we see that there is more movement in
general (Figure 7A,B), and that the path of the agent has a higher fractal
dimension and higher entropy (taking the 2d entropy of the image of the path
over 1000 epochs) for the TN model (and has visited significantly more of all
possible locations, see Figure 7C), suggesting it enjoys a more complex
pattern covering more of the environment and is less affected by the feedback
system.
Figure 7: A and B (left and center) present forward motion (i.e., average
distance moved per epoch), and absolute average turning over time for both
the TN (red) and non TN (blue) simulations. There is generally more
movement for the TN model. C, (right), location of the agent averaged across
the last 50,000 epochs contrasting the TN model with the non-TN model
(smoothed with a Gaussian kernel, sigma=0.5). We see that in general the TN
model visits more of the environment than the non-TN (i.e., pixels with warm
colors, indicate TN > non-TN; cold colors are the reverse).
Time020000400006000080000100000Avg forward movement0.0050.010.0150.020.0250.030.0350.040.0450.05Time020000400006000080000100000Avg rotations00.050.10.150.20.250.3Locations (TN - No TN)ABC The location of the sensory input systems. motor output systems task-
negative nodes were chosen relatively arbitrarily. This is because the coarse
resolution of the parcellation means that assigning sensory or motor labels to
nodes is inherently very approximate. As such, we do not wish to draw
parallels with specific brain regions or networks (e.g., from the functional
imaging literature). However, it is interesting to look at the relationships
between task positive and task negative nodes and how this affects the neural
simulations. Therefore, we repeated the simulations, randomly varying the
location of the task negative nodes. We ran 65 simulations that were the
identical to the simulations detailed previously except: 1) they were shorter
(5000 epochs, for practical reasons); 2) the TN were randomly chosen from
any of the nodes that had not been defined as "sensory" or "motor". We
observed that the stability of the model is dependent on the choice of the
individual TN nodes. Specifically, we see that the solution is more stable when
TP are linked by more walks to TN nodes (calculated using the Brain
Connectivity Toolbox [22]). This was assessed by counting the number of
walks (of <6 steps) between TP and TN nodes. This resulted in distance
measures between somatosensory TP-TN nodes, and between visual TP-TN
nodes. These distances were entered into general linear models to predict
measures of neural dynamics. This analysis revealed that the standard
deviation and coefficients of variation for threshold weights and activation as
well as the correlation between distance to wall and activation were all
significantly related to distance between TP-TN (p<0.01; F(3,63)>4.5).
Therefore, as TP-TN nodes became more closely connected, the model finds
a more stable solution, and neural dynamics are less affected by brain-
environment feedback.
Discussion
This model is unequivocally not intended to be a detailed model of all aspects
of real embodied cognition or of actual neural and sensorimotor systems;
instead, in both regards, it is highly simplified. We acknowledge that there
have been many arbitrary design choices, and do not intend this to be a
definite presentation of how to model brain/environment/behavior interactions.
Such interactions systems are likely to be far more complex, possibly non-
stationary, and will depend on the complexity both of the neural system, but
also the complexity of the motor and sensory systems. Instead, the example
we present here is a useful toy example; the simplification allows us to
consider the interactions between macroscopic brain networks, neural
dynamics and the environment to better understand possible functional roles
of homeostatic systems in the brain.
With the above as a strong caveat, however, our findings highlight the
challenges that feedback between environment and brain presents to neural
models. Further, our results suggest that modeling spontaneous dynamics at
rest (e.g., [23-25]) or with a simple task such as encoding a sensory stimulus
(e.g., [14-15]) is different to modeling sensori-motor interactions with an
environment; further, the existence of a closed-loop feedback made the roles
of homeostatic mechanisms more important and obvious. In our case, we
observed that without the local homeostatic plasticity, the agent in the
environment would become trapped in either a stationary state (with high
levels of local inhibition) or would be in a permanent state of motion (with too
little local inhibition). Instead, we observe that plasticity is a constant feature
of the system. Initially, there are large changes in local thresholds across time
points, as the model approximately balances average incoming excitation at
each node. As time progresses, however, the weight changes become
smaller, but never drop completely to zero.
We also observed that local homeostatic plasticity could be complemented by
adding a macroscopic, task-negative system to compensate for sensory-
induced activity, across the whole simulated brain. The simple system we
implemented, modeled on patterns of task negative deactivation from the
fMRI/PET literature (e.g., [10]) counteracted the destabilizing effects from the
changing amount of sensory input that the model receives in different
locations in the environment. Without the task negative system, the overall
level of activity within the model is more dependent on the level of sensory
input (i.e., "touching" or "seeing" the wall). This makes the task of the local
homeostatic plasticity mechanism harder, since exogenous input to the
system varies considerably. Instead, the task negative system simply
balances the level of exogenous activity to a constant amount, such that task
negative input decreases as sensory input increases and visa versa. This
means that the environment/brain feedback loop does not change the overall
level of incoming activity to the model, therefore facilitating the local
homeostatic plasticity to find a more stable solution, i.e., one that requires the
smallest weight changes to approximate the target activation rate. Further,
what we observe are different balancing systems operating at different spatial
and temporal scales and with different specific mechanisms. This is consistent
with the proposed description of normalization found in many neural systems
[26], which provides a canonical computation across scales and
implementations, and results in improved neural coding efficiency and
sensitivity.
frequently observed and
From a traditional cognitive neuroscience perspective, this way of thinking
about task negative systems may sit somewhat uncomfortably. What we have
been describing as task negative may provide a partial functional explanation
for the default mode network. The default mode network is a well-
characterized,
relatively poorly understood
macroscopic brain network located in areas of the brain not associated with
sensorimotor activity; the default mode network has been observed across
ontogeny [27], phylogeny [28], and found across different cognitive and
sensorimotor tasks [29] and implicated through abnormal function in many
disorders [30]. According to our findings, the default mode network can be
thought of as acting as a counterweight, or as an endogenous generator of
neural activity that allows the neural system to remain relatively stable in an
inherently unstable world. One analogy could be with the vascular system of
warm-blooded animals, which attempts
to maintain a constant body
temperature, irrespective of the temperature outside, in order to maintain a
stable environment for chemical reactions to take place, ultimately allowing
more flexible behavior. (We note that the proposed balancing functional role
for task negative brain networks does not preclude more traditional cognitive
roles ascribed to them, such as internal mentation. We hypothesise that task
negative systems could have initially evolved to perform some basic neural
function, such as balancing incoming sensory activity, and eventually been
exaptively repurposed over evolution to perform more specific cognitive
functions that occur when external input is not present).
Following this explanation of the task negative systems in general and the
default mode network more specifically, we see that task-negative systems
may not strictly be "necessary" for accomplishing any task. As such, lesioning
task negative regions is unlikely to disturb any associated function entirely,
and as such task negative systems may appear to be epiphenomenal.
However, just as a sailing boat does not require a keel to move (the keel
counterbalances the forces on the sail, facilitating stability and allowing a
wider range of movement and greater speed), the brain may have a greater
range of neural state and potentially be more controllable, when it is properly
counterbalanced. It might only be over longer time periods when initially
adapting to a novel environment or across development that damage to task
negative systems becomes particularly disabling, failing to facilitate other
adaptive systems as efficiently.
We find that the model finds a more stable solution if TN nodes are strongly
linked to task positive nodes. This is consistent with the presence of multiple
TN systems in the brain ([9,31]) rather than a single TN. This would be
consistent with the brain being configured to involve active counterbalancing
systems, such that in the optimal case each task positive configuration would
have a matched task negative one, to balance it. However, we acknowledge
that there are likely to be trade-offs between having a perfectly balanced
system and having a functional one. In our model, the nodes do not carry out
any actual computations and are assigned unitary roles (in terms of
sensorimotor function), which is unlikely to be true; both of these, and other
(e.g., onto- or phylogenetic) considerations could affect the type of TN that
evolution has arrived at.
Finally, in order to achieve a relatively stable solution with rich spontaneous
dynamics and interactions with the environment, the system may have to
encode (in the local inhibitory weights) information about the world, and the
agent's movement in it. Given the relative simplicity of the environment in the
current simulation, the presence of local thresholds is adequate to facilitate a
relatively stable solution. However, as the environment (and sensory input
systems) becomes more complex, it will be necessary to use more
sophisticated models with more flexibility. If the repertoire of brain states is to
be more fully explored in the face of this increasing complexity, then it will be
necessary to capture more information about the environment/sensory
systems. This leaves open questions about the roles of other types of learning
(e.g., longer-distance excitatory and reinforcement learning) and their roles in
supporting the system staying in a rich dynamical regime in a complex
environment with complex sensorimotor systems and with more cognitive
control mechanism.
Methods
Empirical Structural Connectivity
Simulated activity patterns were generated from a computational model
constrained by empirical measures of white-matter structural connectivity
between 66 cortical regions of the human brain, defined by diffusion tensor
imaging (DTI) [17]. This structural network has been used in a range of
previous computational models to demonstrate emergent properties of resting
state functional connectivity [14,20,24,32]. A full methodology, describing the
generation of this matrix ! is available in [17]. In brief: measures of length
and strength of stream-line based connectivity were estimated using
Deterministic tractography of DSI datasets (TR=4.2s, TE=89s, 129 gradient
directions max b-value 9000s/mm2) of the brain in 5 healthy control subjects.
A high-dimensional ROI based connectivity approach was projected though
the
(FreeSurfer
http://surfer.nmr.mgh.harvard.edu/), such that !!,! is the number of streamlines
connecting nodes ! and !.
the Desikan-Killianey
regions
atlas
66
of
Computational Model
Neural Dynamics
To simulate brain activity, we defined a simple model based on the
Greenberg-Hastings model, which has been shown in previous work to
approximate patterns of empirical functional connectivity [18]. At each time
point, t, each node, i, in the model can be in one of three states, Si,t :
excitatory (E), quiescent (Q), or refractory (R). Nodes changed state
according the following simple probabilities: !!!→! =1; !!!→! =1;
!!!→! =10!!. Importantly, nodes would also change from Q->E if the
!!. The strength of the activation threshold, Ti , could
summed input from n connected nodes, j, was greater than a local threshold
!!,! !!,!!!>
value:
!!!!
be tuned to separately at each node (see below). Si,t was binarized so that E
was coded as 1, R or Q as 0.
Homeostatic plasticity
For most of the simulations, we used a local homeostatic plasticity
mechanism as follows: we allowed the activation threshold to vary by a small
amount based on the activity in each node at the previous time-step,
according to the following rule similar (but simplified) to that introduced in [3]
and used in [20] and with a similar (but simpler) effect of balancing incoming
excitation from connected nodes:
!"!=!(S!,!−!)
where ! is a target activation and ! is a learning rate. Thus in the case that
the activity of ! is 1, and !<1 there is an increase in the threshold whereas,
will approximate !.
otherwise the threshold decreases. Therefore, the time-averaged activity of Si,
Environmental Embedding
The motor activity (movement) of the agent was defined by two commands;
Turn (h) in radians per update step and Move (v) which moved the agent
forward v world units. The activity within these two parameters at each time-
step was determined by the simulated neural activity at four nodes (two rotate
and two advance nodes) of the computational model. We chose these nodes
to be bilaterally symmetrical such that they approximately correspond to motor
related regions in the brain (n.b., we make no claims that this anatomical
correspondence is correct or that the results are dependent on this). If a rotate
node was active, the agent would attempt to turn ≈30° in that direction. If both
nodes were active, then the effect would cancel out this out. If a single
forward node was active, the agent would move forward 1/10 of a unit the
arbitrary world space, if both forward nodes were active, the unit would move
forward 1 unit of world space. In addition, we added some temporal
smoothing across time for activity within the move such that the move
command described was 7/8 of the activity of the relevant assigned node, and
1/8 of the activity of the previous time step. (The amount of this smoothing
and the values of how nodes translated into movement were chosen semi-
arbitrarily, to produce agent motion that appeared superficially plausible, i.e.,
neither very fast or slow).
Sensory information ('visual' perception) from the environment was integrated
into the computational model through the use of two horizontal ray-traces
emanating from each "eye" of the agent and offset by ±10° from the vertical. A
distance threshold was defined, such that if an object (i.e., the wall, in this
simple environment) was less <2 units of world space then a specific node
("near visual") of the model was set to the E state, if an object was detected
between 2 and 10 world units away then the "far visual" node was set to
excitatory. In addition, to "visual" input, we also defined a rudimentary
"somatosensory" input, whereby if model had collided with any other object in
the environment then specific "somatosensory" nodes for collisions on either
of the Left or Right side of the agent within the computational model were set
to the E state.
Task-Negative nodes
In order to explore the effect of balance between task positive (TP) and task
negative (TN) networks, we defined for some simulations, a collection of TN
nodes that were anti-correlated with the TP nodes described above. These
TN nodes were defined as two (bilateral) pairs of task negative nodes
approximately corresponding to regions that consistently show relative
deactivation across many empirical fMRI tasks were chosen (although, this
was still a relatively arbitrary decision and we do not wish to make any claims
based on anatomical precision). These nodes were set to the E state if TN
nodes (i.e., the "visual", or "somatosensory" nodes were in the Q or R states,
and Q, when the TN nodes were activated, such that TP and TN nodes were
anti-correlated.
Further, given that TN activity is task specific [10,33,34] we defined two TN
nodes (one on the left and its homologous region on the right) that were set to
be excitatory when "visual" activity was not excitatory; and, a separate pair of
nodes (again bilateral, homologous) were set
to be activated when
"somatosensory" activity was silent. For most simulations, the location of the
task negative nodes were kept constant. However, in an additional set of
simulations, the task negative nodes were randomly re-positioned by picking
random bilateral homologous pairs of nodes from the network. Unless stated
otherwise, the results presented are from a single model run for 100,000
epochs. However, we repeated the model two further times with different
random seeds (so different patterns of excitatory noise, resulting simulated
dynamics and movements), replicating the results presented below.
Acknowledgements and Supplementary Materials
The code/implementation (in Unity) and compiled versions of the model are
available to at https://github.com/c3nl-neuraldynamics/Avatar/releases.
Special thanks go to Eva Papaeliopoulos for getting us started in making
computer games.
References
1 van Vreeswijk, C., & Sompolinsky, H. (1996). Chaos in neuronal networks with
balanced excitatory and inhibitory activity. Science, 274(5293), 1724-1726.
2 Doiron, B., Litwin-Kumar, A., Rosenbaum, R., Ocker, G. K., & Josić, K. (2016). The
mechanics of state-dependent neural correlations. Nature Neuroscience, 19(3),
383-393.
3 Vogels, T. P., Sprekeler, H., Zenke, F., Clopath, C., & Gerstner, W. (2011). Inhibitory
plasticity balances excitation and inhibition in sensory pathways and memory
networks. Science, 334(6062), 1569-1573.
4 Woodin, M. A., Ganguly, K., & Poo, M. M. (2003). Coincident pre-and postsynaptic
activity modifies GABAergic synapses by postsynaptic changes in Cl−
transporter activity. Neuron, 39(5), 807-820.
5 Haas, J. S., Nowotny, T., & Abarbanel, H. D. (2006). Spike-timing-dependent plasticity
of inhibitory synapses in the entorhinal cortex. Journal of Neurophysiology,
96(6), 3305-3313.
6 D'amour, J. A., & Froemke, R. C. (2015). Inhibitory and excitatory spike-timing-
dependent plasticity in the auditory cortex. Neuron, 86(2), 514-528.
7 Fox, M.D., Synder, A.Z., Vincent, J., Corbetta, M., Van Essen, D.C., & Raichle, M.E
(2005) The human brain is intrinsically organized into dynamic, anticorrelated
functional networks. Proceedings of the Academy of Sciences, 102(207), 9673-
9678.
8 Geranmayeh, F., Wise, R. J., Mehta, A., & Leech, R. (2014). Overlapping networks
engaged during spoken language production and its cognitive control. The
Journal of Neuroscience, 34(26), 8728-8740.
9 Xu, J. (2015). Implications of cortical balanced excitation and inhibition, functional
heterogeneity, and sparseness of neuronal activity in fMRI. Neuroscience &
Biobehavioral Reviews, 57, 264-270.
10 Leech, R., Scott, G., Carhart-Harris, R., Turkheimer, F., Taylor-Robinson, S. D., &
Sharp, D. J. (2014). Spatial dependencies between large-scale brain networks.
PloS one, 9(6), e98500.
11 Braga, R. M., & Leech, R. (2015). Echoes of the Brain Local-Scale Representation of
Whole-Brain Functional Networks within Transmodal Cortex. The
Neuroscientist, 21(5), 540-551.
12 Scott, G., Hellyer, P. J., Hampshire, A., & Leech, R. (2015). Exploring spatiotemporal
network transitions in task functional MRI. Human brain mapping, 36(4), 1348-
1364.
13 Turkheimer, F. E., Leech, R., Expert, P., Lord, L. D., & Vernon, A. C. (2015). The
brain's code and its canonical computational motifs. From sensory cortex to
the default mode network: A multi-scale model of brain function in health and
disease. Neuroscience & Biobehavioral Reviews, 55, 211-222.
14 Hellyer, P. J., Shanahan, M., Scott, G., Wise, R. J., Sharp, D. J., & Leech, R. (2014).
The control of global brain dynamics: opposing actions of frontoparietal
control and default mode networks on attention. J Neurosci.
15 Ponce-Alvarez, A., He, B. J., Hagmann, P., & Deco, G. (2015). Task-Driven Activity
Reduces the Cortical Activity Space of the Brain: Experiment and Whole-Brain
Modeling. PLoS Comput Biol, 11(8), e1004445.
16 Greenberg, J. M., & Hastings, S. P. (1978). Spatial patterns for discrete models of
diffusion in excitable media. SIAM Journal on Applied Mathematics, 34(3), 515-
523.
17 Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Wedeen, V. J., &
Sporns, O. (2008). Mapping the structural core of human cerebral cortex. PLoS
Biol, 6(7), e159.
18 Haimovici, A., Tagliazucchi, E., Balenzuela, P., & Chialvo, D. R. (2013). Brain
organization into resting state networks emerges at criticality on a model of
the human connectome. Physical review letters, 110(17), 178101.
19 Deco, G., Ponce-Alvarez, A., Hagmann, P., Romani, G. L., Mantini, D., & Corbetta, M.
(2014). How local excitation–inhibition ratio impacts the whole brain dynamics.
The Journal of Neuroscience, 34(23), 7886-7898.
20 Hellyer, P. J., Jachs, B., Clopath, C., & Leech, R. (2016). Local inhibitory plasticity
tunes macroscopic brain dynamics and allows the emergence of functional
brain networks. NeuroImage, 124, 85-95.
21 Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., &
Shulman, G. L. (2001). A default mode of brain function. Proceedings of the
National Academy of Sciences, 98(2), 676-682.
22 Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity:
uses and interpretations. Neuroimage, 52(3), 1059-1069.
23 Honey, C. J., Kötter, R., Breakspear, M., & Sporns, O. (2007). Network structure of
cerebral cortex shapes functional connectivity on multiple time scales.
Proceedings of the National Academy of Sciences, 104(24), 10240-10245.
24 Messé, A., Rudrauf, D., Benali, H., & Marrelec, G. (2014). Relating structure and
function in the human brain: relative contributions of anatomy, stationary
dynamics, and non-stationarities. PLoS Comput Biol, 10(3), e1003530.
25 Váša, F., Shanahan, M., Hellyer, P. J., Scott, G., Cabral, J., & Leech, R. (2015). Effects
of lesions on synchrony and metastability in cortical networks. Neuroimage,
118, 456-467.
26 Carandini, M., & Heeger, D. J. (2012). Normalization as a canonical neural
computation. Nature Reviews Neuroscience, 13(1), 51-62.
27 Doria, V., Beckmann, C. F., Arichi, T., Merchant, N., Groppo, M., Turkheimer, F. E.,
..., & Edwards, D.. (2010). Emergence of resting state networks in the preterm
human brain. Proceedings of the National Academy of Sciences, 107(46),
20015-20020.
28 Mantini, D., Gerits, A., Nelissen, K., Durand, J. B., Joly, O., Simone, L., ... &
Vanduffel, W. (2011). Default mode of brain function in monkeys. The Journal
of Neuroscience, 31(36), 12954-12962.
29 Gusnard, D. A., & Raichle, M. E. (2001). Searching for a baseline: functional imaging
and the resting human brain. Nature Reviews Neuroscience, 2(10), 685-694.
30 Zhang, D., & Raichle, M. E. (2010). Disease and the brain's dark energy. Nature
32 Cabral, J., Hugues, E., Sporns, O., & Deco, G. (2011). Role of local network
oscillations in resting-state functional connectivity. Neuroimage, 57(1), 130-
139.
31 Leech R, Braga R, Sharp DJ, 2012, Echoes of the Brain within the Posterior Cingulate
Reviews Neurology, 6(1), 15-28.
Cortex, J Neurosci.
33 Shmuel, A., Augath, M., Oeltermann, A., & Logothetis, N. K. (2006). Negative
functional MRI response correlates with decreases in neuronal activity in
monkey visual area V1. Nature neuroscience, 9(4), 569-577.
34 Seghlier, M.L., & Price, C.J., (2012). Functional heterogeneity within the default
mode network during semantic processing and speech production. Frontiers in
psychology, 3, p281.
|
1311.3586 | 2 | 1311 | 2013-12-03T03:20:41 | Spike timing prediction with active dendrites | [
"q-bio.NC"
] | A complete single-neuron model must correctly reproduce the firing of spikes and bursts. We present a study of a simplified model of deep pyramidal cells of the cortex with active dendrites. We hypothesized that we can model the soma and its apical tuft with only two compartments, without significant loss in the accuracy of spike-timing predictions. The model is based on experimentally measurable impulse-response functions, which transfer the effect of current injected in one compartment to current reaching the other. Each compartment was modeled with a pair of non-linear differential equations and a small number of parameters that approximate the Hodgkin-and-Huxley equations. The predictive power of this model was tested on electrophysiological experiments where noisy current was injected in both the soma and the apical dendrite simultaneously. We conclude that a simple two-compartment model can predict spike times of pyramidal cells stimulated in the soma and dendrites simultaneously. Our results support that regenerating activity in the dendritic tuft is required to properly account for the dynamics of layer 5 pyramidal cells under in-vivo-like conditions. | q-bio.NC | q-bio |
Spike-timing prediction with active dendrites
Richard Naud1, Brice Bathellier2 and Wulfram Gerstner3
1 Department of Physics, University of Ottawa,
150 Louis Pasteur, ON, K1N 6N5, Canada.
2 Unit of Neuroscience Information and Complexity (UNIC) CNRS UPR-3239,
1 av. de la Terasse, Gif-sur-Yvette, 91198, France.
3 School of Computer and Communication Sciences and School of Life Sciences,
Ecole Polytechnique Federale de Lausanne,
Building AAB Lausane-EPFL, 1015, Switzerland.
(Dated: August 2, 2021)
A complete single-neuron model must correctly reproduce the firing of spikes and bursts. We
present a study of a simplified model of deep pyramidal cells of the cortex with active dendrites. We
hypothesized that we can model the soma and its apical tuft with only two compartments, without
significant loss in the accuracy of spike-timing predictions. The model is based on experimentally
measurable impulse-response functions, which transfer the effect of current injected in one com-
partment to current reaching the other. Each compartment was modeled with a pair of non-linear
differential equations and a small number of parameters that approximate the Hodgkin-and-Huxley
equations. The predictive power of this model was tested on electrophysiological experiments where
noisy current was injected in both the soma and the apical dendrite simultaneously. We conclude
that a simple two-compartment model can predict spike times of pyramidal cells stimulated in the
soma and dendrites simultaneously. Our results support that regenerating activity in the dendritic
tuft is required to properly account for the dynamics of layer 5 pyramidal cells under in-vivo-like
conditions.
I.
INTRODUCTION
Partially neglected for a long time, dendrites have been
recently shown to treat synaptic input in a surprising va-
riety of modes[1]. One particularly striking example is
found in pyramidal cells of deep cortical layers. In these
cells, a coincidence between a back-propagating action
potential and dendritic input can trigger voltage-sensitive
ion channels situated on the apical dendrite more than
300 µm from the soma [2, 3]. The somatic membrane
potential increases only after the activation of dendritic
ion channels. This often resulting in a burst of action
potentials. Bursts in these cells can therefore signal a
coincidence of input from the soma (down) with inputs
in the apical dendrites (top). Such top-down coincidence
detection is one computation that is attributed to den-
dritic processes. Other allegedly dendritic computations
include subtraction [4], direction selectivity [5], temporal
sequence discrimination [6], binocular disparity [7], gain
modulation [8] and self-organization of neuron networks
[9].
Models of large pyramidal neurons that are active at
the tuft of their apical dendrites were first described by
Traub et al. (1991) [10] for the hippocampus. This model
of the large CA3 pyramidal neurons included voltage-
dependent conductances on the dendrites. It is a model
based on the Hodgkin-Huxley description of ion channels.
Cable properties of dendrites are taken into account by
segmenting the dendrite into smaller compartments. The
resulting set of equations is solved numerically. A sim-
plified version of this model was advanced by Pinsky and
Rinzel (1994) [11]. They have reduced the model to a
dendritic compartment and a somatic compartment con-
nected by an effective conductance. The model has a
restricted set of five ion channels and accounts for burst-
ing of CA3 pyramidal cells.
Models specific to deep cortical cells have been de-
scribed by extending the approach of Traub et al. (1991).
Schaefer et al. (2003) [12] used morphological reconstruc-
tion to define compartments. This model could repro-
duce the top-down coincidence detection.
Using a simplified approach similar to Pinsky and
Rinzel (1994) [11], Larkum et al. (2004) [8] have mod-
elled dendrite-based gain modulation. The parameters in
the model could be tuned to quantitatively reproduce the
firing rate response of layer 5 pyramidal cells stimulated
at the soma and the dendrites simultaneously. Larkum
et al. (2004) concluded that a two-compartment model
was sufficient to explain the time-averaged firing rate.
A more stringent requirement for neuron model vali-
dation, however, is to predict spike times [13 -- 18]. Given
the low spike-time reliability of pyramidal neurons, spike
time prediction is compared to the intrinsic reliability
[15]. This approach can be seen as predicting the instan-
taneous firing rate [19]. Generalized integrate-and-fire
models can predict instantaneous firing rate of layer 5
pyramidal neurons with substantial precision[16, 18, 20]
in the absence of dendritic stimulation. The question
remains whether a neuron model can predict the spike
times of layer 5 pyramidal neurons when both the den-
drites and the soma are stimulated simultaneously.
We present a study of a simplified model of layer 5
pyramidal cells of the cortex with active dendrites. Fol-
lowing Larkum et al. (2004) [8], we hypothesized that
we can model the soma and its apical tuft with two
compartments, without significant loss in the accuracy
of spike-timing predictions. We introduce experimen-
tally measurable impulse-response functions [21], which
transfer the effect of current injected in one compartment
to current reaching the other. The impulse-response
functions replace the instantaneous connection used in
previous two-compartment models [8, 11] and acts as
a third, passive, compartment. Each compartment was
modeled with a pair of non-linear differential equations
with a small number of parameters that approximate the
Hodgkin-and-Huxley equations. The predictive power of
this model was tested on electrophysiological experiments
where noisy current was injected in both the soma and
the apical dendrite simultaneously [8].
II. METHODS
Methods are separated in four parts. First we present
the model, second the experimental protocol, then fitting
methods and finally the analysis methods.
FIG. 1.
Schematic representation of the two-compartment
model. A Somatic and dendritic compartment communicate
through passive and active propagation. Passive communi-
cation filters through a convolution (denoted by an asterisk)
the current injected in the other compartment. Active com-
munication in the soma introduces a perturbation propor-
tional to the dendritic current ICa. Active communication
to the dendrites introduces a stereotypical back-propagating
action potential current (BAPC). The somatic compartment
has spike-triggered adaptation and a moving threshold. The
dendritic compartment has an activation current and recovery
current. B Associated experimental protocol with current in-
jection both in soma and apical dendrite of layer 5 pyramidal
cells of the rat somato-sensory cortex. Variables are defined
in the main text.
2
A. Description of the Model
Fig. 1 shows a schematic representation of the two-
In details, the model follows the
compartment model.
system of differential equations:
Cs
dVs
dt
Cd
dVd
dt
τm
dm
dt
τx
dx
dt
dVT
dt
τT
(cid:88)
(cid:88)
{ti}
{ti}
= −gs(Vs − Es) + αm + Is
IA(t − ti) + ds ∗ Id
+
= −gd(Vd − Ed) + g1m + g2x + Id
IBAP (t − ti) + sd ∗ Is
+
=
(cid:16)− Vd−Em
1
Dm
1 + exp
= m − x
= −(VT − ET ) + DT
(cid:17) − m
(cid:88)
{ti}
δ(t − ti)
(1)
(2)
(3)
(4)
(5)
where Is is the current injected in the soma, Id the cur-
rent injected in the dendrites, Vs is the somatic voltage,
Vd is the dendritic voltage, m is the level of activation
of a putative calcium current (ICa = g1m), x is the level
of activation of a putative calcium-activated potassium
current (IK(Ca) = g2x), VT is the dynamic threshold for
firing somatic spikes, IA is a spike-triggered current medi-
ating adaptation, IBAP is the the current associated with
the back-propagating action potential, sd is the filter re-
lating the current injected in the soma to the current
arriving in the dendrite and ds is the filter relating the
current injected in the dendrite to the current arriving in
the soma. The spikes are emitted if Vs(t) > VT (t) which
results in t(last) = t while Vs → Er and t → t + τR. The
parameters are listed in Table I.
As a control, we also consider an entirely passive model
of dendritic integration. In this model, the current in-
jected in the dendrite is filtered passively to reach the
soma. The generalized passive model has and instanta-
neous firing rate:
κs ∗ Is + κds ∗ Id +
(cid:88)
{ti}
(6)
ηA(t − ti)
λ(t) = λ0 exp
where λ0 is a constant related to the reversal potential,
κs somatic membrane filter, κds is the filter relating the
current injected in the dendrite to the voltage change in
the soma, and ηA is the effective spike-triggered adapta-
tion.
B. Experimental Protocol
Parasagittal brain slices of the somato-sensory cortex
(300-350 m thick ) were prepared from 28-35 day-old Wis-
VdgdCdIK[Ca]CadIVsgsCsIAsIvoltage resetpassiveback-prop.passiveprop.Iactive prop.BAPCαCaISomatic CompartmentDendritic CompartmentA Value Units
Variable
nS
22
gs
Somatic leak conductance
pF
379
Cs
Somatic capacitance
mV
-73
Es
Somatic reversal potential
mV
-53
ET
Threshold baseline
mV
2.0
DT
Spike-triggered jump in threshold
ms
27
τT
Time-constant of dynamic threshold
pA
567
Maximum 'Ca' current
g1
n.u.
337
Maximum effect of 'Ca' current in soma α
nS
22
gd
Dendritic leak conductance
pF
86
Dendritic capacitance
Cd
mV
-53
Ed
Dendritic reversal potential
ms
6.7
τm
Time-constant for variable m
ms
τx
Time-constant for variable x
49.9
ms
Dm 5.5
Sensitivity of 'Ca' Current
Maximum 'K(Ca)' Current
pA
-207
g2
Half-activtion potential of 'Ca' current Em -0.6 mV
TABLE I. List of parameters and their fitted value for the
two-compartment model.
tar rats. Slices were cut in ice-cold extracellular solution
(ACSF), incubated at 34oC for 20 min and stored at room
temperature. During experiments, slices were superfused
with in ACSF at 34oC. The ACSF contained (in mM) 125
NaCl, 25 NaHCO3, 25 Glucose, 3 KCl, 1.25 NaH2PO4, 2
CaCl2 , 1 MgCl2 , pH 7.4, and was continuously bubbled
with 5 % CO2 / 95 % O2. The intracellular solution con-
tained (in mM) 115 K+-gluconate, 20 KCl, 2 Mg-ATP, 2
Na2-ATP, 10 Na2-phosphocreatine, 0.3 GTP, 10 HEPES,
0.1, 0.01 Alexa 594 and biocytin (0.2%), pH 7.2.
Recording electrodes were pulled from thick-walled
(0.25 mm) borosilicate glass capillaries and used with-
out further modification (pipette tip resistance 5-10 MΩ
for soma and 20-30MΩ for dendrites). Whole-cell voltage
recordings were performed at the soma of a layer V pyra-
midal cell . After opening of the cellular membrane a flu-
orescent dye, Alexa 594 could diffuse in the entire neuron
allowing to perform patch clamp recordings on the api-
cal dendrite 600-700 µm from the soma. Both recordings
were obtained using Axoclamp Dagan BVC-700A ampli-
fiers (Dagan Corporation). Data was acquired with an
ITC-16 board (Instrutech) at 10 kHz driven by routines
written in the Igor software (Wavemetrics).
The injection waveform consisted of 6 blocks of 12 sec-
onds. Each block is made of three parts: 1) one second
of low-variance colored noise injected only in the soma,
2) one second of low-variance colored noise injected only
in the dendritic injection site, 3) ten seconds of high-
variance colored noise whose injection site depends on
the block: In the first block, the 10-second stimulus is
injected only in the dendritic site, the second block de-
livers the 10-second stimulus in the soma only, and the
four remaining blocks deliver simultaneous injections in
the soma and the dendrites. The colored noise was sim-
3
ulated with MATLAB as an Ornstein-Uhlenbeck process
with a correlation time of 3 ms. The six blocks make a
72 seconds stimulus that was injected repeatedly with-
out redrawing the colored noise (frozen-noise). Twenty
repetitions of the 72-second stimulus were carried out,
separated by periods of 2-120 seconds. Out of the twenty
repetitions, a set of seven successive repetitions were se-
lected on the basis of high intrinsic reliability.
C. Fitting Methods
κs(t) = (cid:80)
Each kernel (κs, κds, ηA, ds,sd, IA, IBAP ) is ex-
pressed as a linear combination of nonlinear basis (i.e.
i aifi(t)). The rectangular function was cho-
sen as the nonlinear basis. The parameters weighting
the contributions of the different rectangular functions
are then linear in the derivative of the membrane po-
tential for the two-compartment model and generalized
linear for the passive model.
For the two-compartment model, we use a combina-
tion of regression methods and exhaustive search to max-
imize the mean square-error of the voltage derivative.
The regression methods are similar to those previously
used for estimating parameters with intracellular record-
ings. These methods are described in more details in
[15, 22 -- 24]. The fit of the somatic compartment essen-
tially follows Jolivet et al. (2006) [15] but using multi-
linear regression to fit the linear parameters. The fit of
the dendritic compartment needs to iterate through the
restricted set of nonlinear parameters (τm, Dm, Em, τx).
All fits are performed only on the part of the data re-
stricted for training the model.
1: Fit of the dendritic compartment, knowing the in-
jected currents and the somatic spiking history:
1a: Compute the first-order estimate of dVd/dt;
1b: Find the best estimates of the dendritic pa-
rameters linear in dVd/dt given a set of non-
linear parameters (τm, Dm, Em, τx). The best
estimates are chosen through multi-linear re-
gression to minimize the mean square error of
dVd/dt.
1c: Compute iteratively step 1b on a grid of the
nonlinear parameters (τm, Dm, Em, τx) and
find the nonlinear parameters that yield the
minimum mean square error of dVd/dt.
2: Fit of the somatic compartment using the fitted den-
dritic compartment.
2a: Compute the first-order estimate of dVs/dt.
2b: Find the best estimates of the somatic param-
eters linear in dVs/dt given a set of nonlinear
parameters (DT , τT , ET ). The best estimates
are chosen through linear regression to mini-
mize the mean square error of dVd/dt.
4
FIG. 2. The two-compartment model fits qualitatively and quantitatively the electrophysiological recordings. A, B Overlay
of the model (red) and experimental (black) somatic voltage trace. The dashed box indicates an area stretched out for higher
precision. C, D The overlay of model (red) and experimental (blue) dendritic voltage is shown for the stretched sections in
A and B. Left (A,C) and right (B,D) columns show two different injection regimes contrasting by the amount of dendritic
activity which is high for A, C and medium for B, D. E Residuals from the linear regression are shown for the somatic (black)
and dendritic (blue) compartment. F For each repetition the Γ Coincidence factor is plotted against the intrinsic reliability of
the cell. Grey points show the performance of the model on the test set and black points show the performance of the model
on the training set. G Comparison of the inter-spike interval histogram for the model (red) and the experiment (black). H
Comparison of the generalized passive (Pas), and the full two-compartment model (Full) with the intrinsic reliability (R) of
the neuron in terms of the Γ coincidence factor. The averaged Γ factor is shown for the training set (black) and test set (Gray)
2c: Compute iteratively step 2b on a grid of the
nonlinear parameters and simulate the model
with each set of nonlinear parameters in order
to compute the coincidence rate Γ (see Sect.
II D).
2d: Take the parameters that yield the maximum
Γ coincidence factor.
For the generalized linear model, we use maximum
likelihood methods [14, 25]. Expressing the kernels as
a linear combination of rectangular bases we recover the
generalized linear model. Here the link-function is ex-
ponential so that the likelihood is convex. We therefore
performed a gradient ascent of the likelihood to arrive at
the optimal parameters.
D. Analysis Methods
When one focuses on spike timing, one may want to
apply methods that compare spike trains in terms of a
spike-train metric [26] or the coincidence rate [27]. Both
measures can be used to compare a recorded spike train
with a model spike train. A model which achieve an
optimal match in terms of spike-train metrics will auto-
matically account for global features of the spike train
such as the interspike interval distribution.
Here we used the averaged coincidence rate Γ [27].
It can be seen as a similarity measure between pairs of
spike trains, averaged on all possible pairs. To compute
the pairwise coincidence rate, one first finds the num-
ber of spikes from the model that fall within an inter-
val of ∆ =4 ms after or before a spike from the real
neuron. This is called the number of coincident events
Nnm between neuron repetition n and model repetition
m. The coincidence rate is the ratio of the number of
coincident events over the averaged number of events
40 ms40 ms300 ms300 ms70 mV70 mV70 mV70 mV 0.5(Nn+Nm), where Nn is the number of spikes in the
neuron spike train and Nm is the number of spikes in the
model spike train. This ratio is then scaled by the num-
ber of chance coincidences NPoisson = 2∆NmNn/T . This
formula comes from the number of expected coincidences
assuming a Poisson model at a fixed rate Nm/T where
T is the time length of each individual spike trains. The
scaled coincidence rate is
Γnm =
Nnm − NPoisson
0.5(1 − NPoisson/Nn)(Nn + Nm)
.
(7)
The pairwise coincidence rate Γnm is then averaged
across all possible pairings of spike trains (trials) gen-
erated from the model with those from the neuron and
gives the averaged coincidence rate Γ. Averaging across
all possible pairings of spike trains from the neuron with
a distinct repetition of the same stimulus given to the
same neuron gives the intrinsic reliability R.
III. RESULTS
Dual patch-clamp recordings were performed in L5
Pyramidal cells of Wistar rats (see Experimental Meth-
ods). A simplified two-compartment model (see Model
Description) was fitted on the first 36 seconds of stimu-
lation for all repetitions. The rest of the data (36 sec) was
reserved to evaluate the model's predictive power. The
predictive power of the two-compartment model with ac-
tive dendrites was then compared to a model without
activity in the dendrites (see Sect. II A), the generalized
linear passive model.
Figure 2 summarizes the predictive power of the
two-compartment model. The somatic and dendritic
voltage traces are well captured (Fig. 2 A-D). The main
cause for erroneous prediction of the somatic voltage
trace is extra or missed spikes (Fig. 2 A and B lower
panels). The dendritic voltage trace of the model follows
the recorded trace both in a low dendritic-input regime
(Fig. 2 C) and in a high dendritic-input regime with
dendritic 'spikes' (Fig.
2 D). The greater spread of
voltage-prediction-error (Fig. 2) is mainly explained by
the larger range of voltages in the dendrites (somatic
voltage prediction is
strictly subthreshold whereas
dendritic voltage prediction ranges from -70 mV to
+40 mV). The interspike interval distribution is well
predicted by the model (Fig. 2 G).
The generalized passive model does not predict as
many spike times ( Fig. 2 H). The intrinsic variabil-
ity in the test set was 68% and the two-compartment
model predicted 50%. The prediction falls to 36 % in the
absence of a dendritic non-linearity (Fig. 2 H).
The fitted kernels show that spike triggered adapta-
tion is a monotonically decaying current that starts very
strongly and decays slowly for at least 500 ms (Fig. 3
A). The back-propagating action potential is mediated
by a strong pulse of current lasting 2-3 ms (Fig. 3 B).
5
FIG. 3. Fitted kernels of the two-compartment model. A
The kernel IA(t) for spike-triggered adaptation is negative
and increases monotonically between 6 and 600 ms. B The
back-propagating current IBAP(t)reaching the dendrites is a
short (2ms) and strong (900 pA) pulse. C The convolution
kernel ds(t) linking the current injected in the dendrite to
the current reaching the soma. D The convolution kernel
sd(t) linking the current injected in the soma to the current
reaching the dendrite.
The coupling ds from dendrite to soma has a maximal
response after 2-3 ms and then decays so as to be slightly
negative after 35 ms (Fig. 3 C). The coupling sd from
soma to dendrite follows qualitatively ds with smaller
amplitudes and slightly larger delays for the maximum
and minimum peaks (Fig. 3 D), consistent with the
larger membrane time-constant in the soma than in the
dendrites.
The two-compartment model can reproduce qualita-
tive features associated with the dendritic non-linearity
in the apical tuft of L5 pyramidal neurons. We study two
of these features: the critical frequency [2] and the gain
modulation [8]. The first relates to the critical somatic
firing frequency above which a non-linear response is seen
in the soma, reflecting calcium channel activation in the
dendrites. To simulate the original experiment, we force
5 spikes in the soma at different frequencies and plot the
integral of the dendritic voltage. The critical frequency
for initating a non-linear increase in summed dendritic
voltage is 138 Hz (Fig. 4 A). Perez-Garci et al. (2006)
[28] reported a critical frequency of 105 Hz while Larkum
et al. (1999) [2] reported 85 Hz. This appears to vary
across different cells and pharmacological conditions.
The model also appears to perform gain modulation
as in [8] (Fig. 4 B). The relation between somatic firing
rate and mean somatic current depends on the dendritic
excitability. The onset (or shift) but also the gain (or
slope) of the somatic frequency versus somatic current
curve depend on the mean dendritic current. The gain
modulation is attributed to a greater presence of bursts
6
(Fig. 4 B) caused by dendritic calcium-current activa-
tion at higher dendritic input. The link between burst
and dendritic activity is reflected in the burst- and spike-
triggered average injected current (Fig. 4 C-D) similar
to Ref. [8]. The burst-triggered current is greater for the
dendritic injection, whereas the spike-triggered current is
larger for somatic injection. The greater correlation, rel-
ative to somatic current, of the dendritic current with the
observation of bursts indicate that the two-compartment
model performs a type of top-down coincidence detection
with bursts.
IV. CONCLUSION
Using a two-compartment model interconnected with
temporal filters, we were able to predict a substan-
tial fraction of spike times. The predicted spike trains
achieved an averaged coincidence rate of 50%. The scaled
coincidence rate obtained by dividing by the intrinsic re-
liability [16, 29] was 72%, which is comparable to the
state-of-the performance for purely somatic current in-
jection which reaches up to 76%[20]. Comparing with a
passive model for dendritic current integration, we found
that the predictive power decreased to a scaled coinci-
dence rate of 53%. Therefore we conclude that regener-
ating activity in the dendritic tuft is required to properly
account for the dynamics of layer 5 pyramidal cells under
in-vivo-like conditions.
ACKNOWLEDGMENTS
The authors would like to thank Matthew Larkum for
helpful suggestions.
FIG. 4. The model reproduces the qualitative features of
active dendrites reported in [2] and [8]. A Dendritic non-
linearity is triggered by somatic spiking above a critical fre-
quency. Somatic spike-trains of 5 spikes are forced in the
soma of the mathematical model at different firing frequen-
cies. The normalized integral of the dendritic voltage is shown
as a function of the somatic spiking frequency. B Dendritic in-
jection modulates the slope of the somatic spiking-frequency
vs. current curve. The slope of the frequency vs mean so-
matic current as measured between 5 and 50 Hz is plotted as
a function of the mean dendritic current. Both somatic and
dendritic currents injected are Ornstein-Uhlenbeck processes
with a correlation time of 3 ms and a standard deviation of 300
pA. C Spike-triggered average of the current injected in the
soma (black) and in the dendrites (blue). D Burst-triggered
average of the current injected in the soma (black) and in
the dendrites (blue). The fact that the blue curve is higher
than the black curve, and that this relation is inverted in C,
indicate that the two-compartment model performs a type of
top-down coincidence detection with bursts.
[1] G. Stuart, N. Spruston, and M. Hausser, Dendrites, 2nd
[11] P. Pinsky and J. Rinzel, Journal of Computational Neu-
ed. (Oxford University Press, Oxford, 2007).
roscience (1994).
[2] M. Larkum, J. Zhu, and B. Sakmann, Nature 398, 338
[12] A. Schaefer, M. Larkum, B. Sakmann, and A. Roth,
(1999).
Journal of Neurophysiology (2003).
[3] M. Larkum, J. Zhu,
and B. Sakmann, J. Physiology
[13] J. Keat, P. Reinagel, R. C. Reid, and M. Meister, Neuron
(London) 447-466 (2001).
30, 803 (2001).
[4] F. Gabbiani, H. G. Krapp, C. Koch, and G. Laurent,
Nature 420, 320 (2002).
[5] W. R. Taylor, S. He, W. R. Levick, and D. I. Vaney,
Science 289, 2347 (2000).
[6] T. Branco, B. Clark, and M. Hausser, Science 12 (2010).
[7] K. Archie and B. Mel, Nature Neuroscience (2000).
[8] M. E. Larkum, W. Senn, and H.-R. Luscher, Cerebral
Cortex 14, 1059 (2004).
[9] R. Legenstein and W. Maass, The Journal of Neuro-
science 31, 10787 (2011).
[14] J. Pillow, L. Paninski, V. Uzzell, E. Simoncelli,
and
E. Chichilnisky, Journal of Neuroscience 25, 11003
(2005).
[15] R. Jolivet, A. Rauch, H. Luscher, and W. Gerstner, Jour-
nal of Computational Neuroscience 21, 35 (2006).
[16] R. Jolivet, R. Kobayashi, A. Rauch, R. Naud, S. Shi-
nomoto, W. Gerstner, Journal of Neuroscience Methods
169, 417 (2008).
[17] R. Jolivet, F. Schurmann, T. Berger, R. Naud, W. Gerst-
ner, and A. Roth, Biological Cybernetics 99, 417 (2008).
[10] R. D. Traub, R. K. S. Wong, R. Miles, and H. Michelson,
[18] W. Gerstner and R. Naud, Science 326, 379 (2009).
J. Neurophysiol. 66, 635 (1991).
Mean Dendritic Current (nA) 7
[19] R. Naud, F. Gerhard, S. Mensi, and W. Gerstner, Neural
[25] L. Paninski, Network: Computation in Neural Systems
Computation 23, 3016 (2011).
15, 243 (2004).
[20] R. Naud, T. Berger, B. Bathellier, M. Carandini, and
W. Gerstner, in Front. Neur. Conference Abstract: Neu-
roinformatics 2009 1 -- 8 (2009).
[26] J. D. Victor and K. Purpura, Journal of Neurophysiology
76, 1310 (1996).
[27] W. Kistler, W. Gerstner, and J. Hemmen, Neural Com-
[21] I. Segev, W. Rall, and J. Rinzel, The theoretical founda-
putation 9, 1015 (1997).
tion of dendritic function (MIT Press, 1995).
[28] E. P´erez-Garci, M. Gassmann, B. Bettler,
and
[22] L. Paninski, J. Pillow, and E. Simoncelli, Neurocomput-
M. Larkum, Neuron 50, 603 (2006).
ing 65-66, 379 (2005).
[23] S. Mensi, R. Naud, M. Avermann, C. C. H. Petersen,
and W. Gerstner, Journal of Neurophysiology 107, 1756
(2012).
[24] C. Pozzorini, R. Naud, S. Mensi, and W. Gerstner, Na-
ture Neuroscience 16, 942 (2013).
[29] R. Naud and W. Gerstner, "Spike timing: Mechanisms
and function," (CRC Press, 2012) Chap. Can We Predict
Every Spike.
|
1809.09321 | 1 | 1809 | 2018-09-25T05:15:40 | Automated, predictive, and interpretable inference of C. elegans escape dynamics | [
"q-bio.NC",
"q-bio.QM",
"stat.ML"
] | The roundworm C. elegans exhibits robust escape behavior in response to rapidly rising temperature. The behavior lasts for a few seconds, shows history dependence, involves both sensory and motor systems, and is too complicated to model mechanistically using currently available knowledge. Instead we model the process phenomenologically, and we use the Sir Isaac dynamical inference platform to infer the model in a fully automated fashion directly from experimental data. The inferred model requires incorporation of an unobserved dynamical variable, and is biologically interpretable. The model makes accurate predictions about the dynamics of the worm behavior, and it can be used to characterize the functional logic of the dynamical system underlying the escape response. This work illustrates the power of modern artificial intelligence to aid in discovery of accurate and interpretable models of complex natural systems. | q-bio.NC | q-bio | Automated, predictive, and interpretable inference of C. elegans
escape dynamics
Bryan C. Daniels
ASU -- SFI Center for Biosocial Complex Systems,
Arizona State University, Tempe, AZ 85281, USA
William S. Ryu
Department of Physics and The Donnelly Centre,
University of Toronto, Toronto, ON M5S 1A7, Canada
Ilya Nemenman
Department of Physics, Department of Biology,
and Initiative in Theory and Modeling of Living Systems,
Emory University, Atlanta, GA 30322, USA
Abstract
The roundworm C. elegans exhibits robust escape behavior in response to rapidly rising tem-
perature. The behavior lasts for a few seconds, shows history dependence, involves both sensory
and motor systems, and is too complicated to model mechanistically using currently available
knowledge. Instead we model the process phenomenologically, and we use the Sir Isaac dynamical
inference platform to infer the model in a fully automated fashion directly from experimental data.
The inferred model requires incorporation of an unobserved dynamical variable, and is biologically
interpretable. The model makes accurate predictions about the dynamics of the worm behavior,
and it can be used to characterize the functional logic of the dynamical system underlying the es-
cape response. This work illustrates the power of modern artificial intelligence to aid in discovery
of accurate and interpretable models of complex natural systems.
8
1
0
2
p
e
S
5
2
]
.
C
N
o
i
b
-
q
[
1
v
1
2
3
9
0
.
9
0
8
1
:
v
i
X
r
a
1
The quantitative biology revolution of the recent decades has resulted in an unprecedented
ability to measure dynamics of complex biological systems in response to perturbations
with the accuracy previously reserved for inanimate, physical systems. For example, the
entire escape behavior of a roundworm Caenorhabditis elegans in response to a noxious
temperature stimulus can be measured for many seconds in hundreds of worms [1, 2]. At
the same time, theoretical understanding of such living dynamical systems has lagged behind,
largely because, in the absence of symmetries, averaging, and small parameters to guide our
intuition, building mathematical models of such complex biological processes has remained
a very delicate art. Recent years have seen emergence of automated modeling approaches,
which use modern machine learning methods to automatically infer the dynamical laws
underlying a studied experimental system and predict its future dynamics [3 -- 14]. However,
arguably, these methods have not yet been applied to any real experimental data with
dynamics of a priori unknown structure to produce interpretable dynamical representations
of the system. Thus their ability to build not just statistical but physical models of data
[15], which are interpretable by a human, answer interesting scientific questions, and guide
future discovery, remains unclear.
Here we apply the Sir Isaac platform for automated inference of dynamical equations
underlying time series data to infer a model of the C. elegans escape response, averaged over
a population of worms. We show that Sir Isaac is able to fit not only the observed data, but
also to make predictions about the worm dynamics that extend beyond the data used for
training. The inferred optimal model is fully interpretable, with the identified interactions
and the inferred latent dynamical variable being biologically meaningful. And by analysing
the dynamical structure of the model -- number of dynamic variables, number of attractors
(distinct behaviors), etc. -- we can generalize these results across many biophysical systems.
RESULTS
Automated Dynamical Inference
Sir Isaac [7, 8] is one of the new generation of machine learning algorithms able to infer
a dynamical model of time series data, with the model expressed in terms of a system of
differential equations. Compared to other approaches, Sir Isaac is able to infer dynamics
2
(at least for synthetic test systems) that are (i) relatively low-dimensional, (ii) have un-
observed (hidden or latent) variables, (iii) have arbitrary nonlinearities, (iv) rely only on
noisy measurements of the system's state variables, and not of the rate of change of these
variables, and (v) are expressed in terms of an interpretable system of coupled differential
equations. Briefly, the algorithm sets up a complete and nested hierarchy of nonlinear dy-
namical models. Nestedness means that each next model in the hierarchy is more complex
(in the sense of having a larger explanatory power) [16 -- 18] than the previous one, and in-
cludes it as a special case. Completeness means that any sufficiently general dynamics can
be approximated arbitrarily well by some model within the hierarchy. Two such hierarchies
have been developed, one based on S-systems [19] and the other on sigmoidal networks [20].
Both progressively add hidden dynamical variables to the model, and then couple them to
the previously introduced variables using nonlinear interactions of specific forms. Sir Isaac
then uses a semi-analytical formulation of Bayesian model selection [8, 18, 21, 22] to choose
the model in the hierarchy that best balances the quality of fit versus overfitting and is,
therefore, expected to produce the best generalization. The sigmoidal network hierarchy
is especially well-suited to modeling biological systems, where rates of change of variables
usually show saturation, and it will be the sole focus of our study.
Experimental model system
Nociception evokes a rapid escape behavior designed to protect the animal from potential
harm [23, 24]. C. elegans, a small nematode with a simple nervous system, is a classic model
organism used in the study of nociception. A variety of studies have used C. elegans to
elucidate genes and neurons mediating nociception to a variety of aversive stimuli including
high osmolarity and mechanical, chemical, and thermal stimuli [25 -- 28]. However, a complete
dynamic understanding of the escape response at the neuronal, let alone the molecular, level
is not fully known. Recent studies have quantified the behavioral escape response of the
worm when thermally stimulated with laser heating [1, 2], and these data will be the focus
of our study. The response is dynamic: when the stimulus is applied to the animal's head, it
quickly withdraws, briefly accelerating backwards, and eventually returns to forward motion,
usually in a different direction. Various features of this response change with the level of
laser heating, such as the length of time moving in the reverse direction and the maximum
3
FIG. 1. The escape response behavior is fitted and predicted well by the inferred
model. (A) Colored lines and shaded bands represent the empirical mean and the standard devi-
ation of the mean, respectively, of the escape response velocity for five groups of worms stimulated
with laser currents in different ranges (38, 42, 39, 41, and 41 subjects in each group). Dashed
lines and bands of the corresponding color show means and standard deviations of the mean (see
Materials and methods) of fits to these empirical data by the chosen model. Only the velocity in
the range of time [−1, 2.25] s relative to the time of the onset of the laser stimulus was used for
fitting. While most worms are still moving backward during this range of time, the inferred model
predicts, without any additional free parameters, the time at which a worm's speed again becomes
positive both as a function of (B) applied laser current and (C) peak observed worm reverse speed.
These predictions (red curves) agree well with the binned averages (orange points with error bars
representing the standard error of the mean if the bin has more than five worms) of individual
worms' behavior (blue dots).
speed attained.
Fits and Predictions
We use the worm center-of-mass speed, v, as the variable whose dynamics needs to be
explained in response to the laser heating pulse. We define v > 0 as the worm crawling
forward and v < 0 as the worm retreating backwards. The input to the model is the
underlying temperature, h(t), which can be approximated as h(t) = Ih0(t), where I is the
experimentally controlled laser current, and h0 is the temperature template, described in
Materials and Methods. Based on trajectories of 201 worms in response to laser currents
4
BCAFigure1.pdf ranging between 9.6 and 177.4 mA, we let Sir Isaac determine the most likely dynamical
system explaining this data within the sigmoidal networks model class [8] (see Materials
and Methods for a detailed description of the modeling and inference). The inferred model
has a latent (unobserved) dynamical variable, hereafter referred to as x2, in addition to the
speed. v and x2 are coupled by nonlinear interactions. However, some of these nonlinear
interactions may be insignificant, and may be present simply because the nested hierarchy
introduces them before some other interaction terms that are necessary to explain the data.
Thus we reduce the model by setting parameters that are small to zero one by one and in
various combinations, refitting such reduced models, and using Bayesian Model Selection to
choose between the reduced model and the original Sir Isaac inferred model. The resulting
model is
= − v
+ V1h(t) +
τ1
= −x2 + V2 h(t),
dv
dt
dx2
dt
W11
1 + ev+θ1
+
W12
1 + ex2
,
x2(t = 0) = 0.
(1)
(2)
Here V1, V2, W11, W12 are constants inferred from data, while the model uses the default value
of 1.0 s for the characteristic time scale of the dynamics of x2 (see Materials and Methods
for the values of the parameters). Interestingly, the inferred model reveals that the latent
dynamical variable x2 is a linear low-pass filtered (integrated) version of the heat signal.
The fits produced by this model are compared to data in Fig. 1(A), showing an excellent
agreement (see Materials and Methods for quantification of the quality of fits). Surprisingly,
the quality of the fit for this automatically generated model is better than that of a manually
curated model [2]: only about 10% of explainable variance in the data remains unexplained
by the model for times between 100 ms and 2 s after the stimulus, compared to about 20%
for the manual model, cf. Fig. 5(A).
However, the quality of the fit is not surprising in itself since the Sir Isaac model hierarchy
can fit any dynamics using sufficient data. A utility of a mathematical model is in its ability
to make predictions about data that were not used in fitting. Thus we use the inferred
model to predict when the worm will return to forward motion, which usually happens well
after the temporal range used for fitting. Figure 1(B,C) compares these predictions with
experiments, showing very good predictions. Such ability to extrapolate beyond the training
range is usually an indication that the model captures the underlying physics, and is not
purely statistical [15], giving us confidence in using the model for inferences about the worm.
5
FIG. 2. Phase space structure of inferred model. With one hidden variable x2, the model
dynamics can be visualized in the two-dimensional phase space. As the instantaneous heat input h
returns to zero after a brief pulse (red curve in insets), a single fixed point in the two-dimensional
(v, x2) dynamics moves from negative speed (escape) to positive (forward motion). As the velocity
of the worm trails the fixed point, this produces first a fast escape and then a slow return to forward
motion (speed trajectory in orange in inset and in the phase portrait plots). Blue arrows indicate
flow lines, circles indicate stable fixed points, and green lines indicate nullclines (dark green where
dv/dt = 0 and light green where dx2/dt = 0).
Model analysis
The algorithm has chosen to include a single latent dynamical variable, which is a linear
leaky integrator of the experienced temperature. Having access to both the instantaneous
stimulus and its integral over the immediate past allows the worm to estimate the rate of
change in the stimulus. This agrees with the observation [1] that both the current tempera-
ture, as well as the rate of its increase, are noxious for the worm. From this, one could have
guessed, perhaps, that at least one latent variable (temperature derivative, or temperature
at some previous time) is required to properly model the escape response. However, the fact
that Sir Isaac inferred this from time series data alone and was able to model the data with
exactly one hidden variable is surprising.
Figure 2 shows the phase portraits of the inferred dynamical model, Eqs. (1, 2), as well
as the dynamics of the speed and the heat stimulus h. Crucially, we see that there is only
one fixed point in the phase space at any instant of time, and the position of this fixed point
6
-60-40-20020020406080Speedv(px/s)x2050100Heath01234-400Time(s)v(px/s)-60-40-20020020406080Speedv(px/s)x2050100Heath01234-400Time(s)v(px/s)-60-40-20020020406080Speedv(px/s)x2050100Heath01234-400Time(s)v(px/s)t = 0.25 st = 4 st = 1 s is affected by the current laser stimulus value. This suggests that, at least at the level of the
population-averaged response, the behavior does not involve switching among alternative
behaviors defined dynamically as multiple fixed points or limit cycles (e. g., forward and
backward motion) with the switching probability influenced by the stimulus [29], but rather
the stimulus controls the direction and the speed of the single dominant crawling state.
The network diagram of the model in Eqs. (1, 2) is shown in Fig. 3, where we omit
the linear degradation terms for v and x2. With the maximum likelihood parameter val-
ues, Tbl. III, the model can be interpreted as follows. The hidden variable x2 is a linear
leaky integrator of the heat signal, storing the average recent value of the stimulus over
about 1 s. While we do not know which exact neuron can be identified with x2, the ther-
mosensory neurons AFD, FLP located in head of the worm are strong candidates [30, 31].
The thermosensory neurons AFD respond to changes in temperature and are the primary
sensors responsible for thermotaxis [30]. The sensory neuron FLP also is thermosensitive
and has a role in the thermal sensory escape response [31]. When x2 is near zero, the W12
term is large, and, together with the linear relaxation of the speed, −v/τ1, it establishes a
constant positive forward motion. We identify this term with the forward drive command
interneurons AVB, PVC [32]. After the temperature increases, x2 grows. It rapidly increases
the denominator in the W12 term and hence shuts down the forward drive. This is again
consistent with the literature indicating that the worms pause with even reasonably small
temperature perturbations [1]. An additional effect of the stimulus is to directly inject a
negative drive −V1h into the dynamics of the velocity. When the stimulus is large, the V1
term is sufficiently negative to result in the velocity overshooting the pause into the negative,
escape range. We identify the V1 term with the reverse command interneurons AVA, AVD,
and AVE, activated by the thermosensory neurons AFD, FLP [32]. When v > −θ1 ≈ −42
px/s, the W11 term is suppressed. However, during fast escapes this suppression is lifted,
activating the positive drive W11, which leads to faster recovery of the forward velocity. We
identify the W11 term with internal recovery dynamics of the reverse command interneurons
whose molecular mechanisms of activity are only partially understood [33]. The velocity
does not just relax to zero over some characteristic time, but crosses back into the forward
crawl once x2 has decreased sufficiently to reactivate the W12 term. Overall, the biological
interpretability of the model is striking. And where there is no direct match between known
worm biology and the model, the model strongly suggests that we should be looking for
7
FIG. 3. Network diagram of the inferred model. Variables, interactions, parameter values,
and biological mechanisms are described in the main text.
specific predicted features, such as the neural and molecular mechanisms for both sensing
the heat stimulus and its recent average.
Another notable feature of the network diagram is that it is similar to other well-known
sensory networks, namely chemotaxis (and the related thermotaxis) in E. coli [34, 35] and
chemotaxis in D. discoideum [36, 37]. In all these cases, the current value of the stimulus
is sensed in parallel to the stimulus integrated over the recent time. They are later brought
together in a negative feedback loop (E. coli) or an incoherent feedforward loop (D. dis-
coideum), resulting in various adaptive behaviors. In contrast, the worm's behavior is more
complicated: the ambient temperature participates in a negative feedback loop on the speed
through W11, which results in adaptation. At the same time, the integrated temperature
(through W12) and the current temperature add coherently to cause the escape in response
to both the temperature and its rate of change. This illustrates the difference between sens-
ing, when an organism needs to respond to stimulus changes only, and escape, where more
complicated dynamics is needed.
DISCUSSION
In this article, we have used modern machine learning to learn the dynamics underlying
the temperature escape behavior in C. elegans. The resulting automatically inferred model
is more accurate than the model curated by hand.
It uses the dynamics within what is
normally considered as discrete behavioral states to make extremely precise verifiable (and
verified) predictions about the behavior of the worm beyond the range of time used for
training. The model is fully interpretable, with many of its features having direct biological,
8
Laser, h(t)Speed, vx2Forward Drive, W12Forward Recovery, W11Backward Drive, V1Integration, V2 mechanistic equivalents. Where such biological equivalents are unknown, the model makes
strong predictions of what they should be, and suggests what future experiments need to
search for.
One can question if describing the C. elegans nociceptive behavior, which typically is
viewed as stochastic [1, 2] and switching between discrete states, with the deterministic
dynamics approach of Sir Isaac is appropriate. The quality of the fits and predictions is
an indication that it is. This is likely because (i) the escape is, to a large extent, deter-
ministic, becoming more and more so as the stimulus intensity increases [1]; (ii) on the
scale of individual worms, the discretized behavior states (forward, backward, pause, . . . )
have their own internal dynamics, with different time-dependent velocities, which Sir Isaac
models well, and the boundaries between the states are not highly pronounced; and (iii) our
equations model a population of worms, and so even if individual worms were dominated by
stochasticity, Sir Isaac would do just fine modeling the dynamics of the mean behavior.
Crucially, the model discovers that the behavior, at least of an average worm, is not a
simple one-to-one mapping of the input signal: the instantaneous stimulus and its temporally
integrated history (one latent variable) are both important for driving the behavior. The
behavior is driven by one fixed point in the velocity-memory phase space, and the worm
changes its speed while chasing this fixed point, which in turn changes in response to the
stimulus. This is in contrast to other possibilities, such as the worm being able to exhibit
both the forward and the backward motion at any stimulus value, and the stimulus and its
history merely affect the probability of engaging in either of these two behaviors.
This is the first successful application of automatic phenomenological inference to model-
ing dynamics of complex biological systems, without using the helping constraints imposed
by (partial) knowledge of the underlying biology. The emphasis on interpretable, physical
models allows extrapolation well beyond the data used for training, which is hard for purely
statistical methods. The automation allows for a comprehensive search through the model
space, so that the automatically inferred model is better than the human-assembled one,
especially when faced with only partial knowledge of the complex system. Even for the best
studied biological systems, we do not have the necessary set of measurements to model them
from the ground up. Our work illustrates the power of phenomenological modeling approach,
which allows for top down modeling, adding interpretable constraints to our understanding
of the system.
9
MATERIALS AND METHODS
Data collection
A detailed description of the experimental methods has been previously published [2]. In
summary, we raised wild-type, N2 C. elegans using standard methods, incubated at 20◦C
with food. Individual worms were washed to remove traces of food and placed on the surface
of an agar plate for 30 minutes at 20◦C to acclimatize. Worms were then transferred to an
agar assay plate seeded with bacteria (food) and left to acclimatize for 30 more minutes.
The worms were then stimulated with an infra-red laser focused to a diffraction limited
spot directed at the "nose" of the worm. The intensity of the stimulus was randomized by
selecting a laser current between 0 and 200 mA, with a duration of 0.1s. The temperature
increase caused by laser heating was nearly instantaneous and reached up to a maximum of
2◦C for our current range. Each worm was stimulated only once and then discarded. Video
of each worm's escape response was recorded at 60 Hz and processed offline using custom
programs in LabVIEW and MATLAB.
Input data
Data used for dynamical inference are as described in Ref. [2]. Speed data for 201 worms
were extracted from video frames at 60 Hz and smoothed using a Gaussian kernel of width
500 ms. We use data between 0.5 s before and 2.25 s after the start of the laser stimulus.
Aligning the data by laser start time, the stimulus happens at the same time in each trial.
Naively used, this can produce models that simply encode a short delay followed by an
escape, without requiring the stimulus input. To ensure that instead the stimulus causes
the response in the model, for each trial we add a random delay between 0 and 1 s to the
time data.
Additionally, we are not interested in capturing any dynamics in the pre-laser free crawling
state. If we only use the small amount of pre-laser data measured in this experiment, the
inference procedure is free to include models with complicated transient behavior before the
stimulus. For this reason, we include a copy of pre-laser data at a fictitious "equilibration
time" long before the stimulus time (10 s), artificially forcing the model to develop a pre-
stimulus steady state of forward motion. Finally, we weight the pre-laser data such that
10
it appears with equal frequency as post-laser data in fitting, in order that the inference
algorithm is not biased toward capturing post-laser behavior more accurately than the pre-
laser one.
Estimating explainable variance
The observed variance in worm speed can be partitioned into that caused by the input
(changes in laser current) and that caused by other factors (individual variability, experi-
mental noise, etc.). As in Ref. [2], we focus on our model's ability to capture the former,
"explainable" variance. We treat the latter variance as "unexplainable" by our model, and
it is this variance that we use to define uncertainties on datapoints for use in the inference
procedure. We estimate these variances by splitting the data into trials with similar laser
current Iµ (5 bins), producing variances σ2
µ(t) that depend on the laser current bin and on
the time relative to laser start. For simplicity, we use a constant uncertainty for all data-
points that is an average over laser current bins and times relative to the stimulus onset,
which is equal to σ = 14.2 px/s. We use this uncertainty when calculating the χ2 that
defines the model's goodness-of-fit:
χ2(Θ) =
ND(cid:88)
i=1
(cid:18)xmodel
i
(cid:19)2
(Θ) − xdata
σ
i
,
(3)
where the index i enumerates data points used for the evaluation, ND is the number of
points, and Θ is the vector of all parameters.
Sir Isaac inference algorithm
We use the Sir Isaac dynamical inference algorithm [8] to find a set of ODEs that best
describes the data without overfitting. Based on previous studies of simulated biological
systems, we use the continuous-time sigmoidal network model class, which produces a set
of J ODEs of the form
= −xi/τi + Vi h(t) +
dxi
dt
J(cid:88)
j=1
Wij ξ(xj + θj),
(4)
where ξ(y) = 1/(1 + exp(y)) and h(t) is the sensory input defined in Eq. (7). The algorithm
infers both the number of parameters (controlled by the total number of dynamical variables
11
FIG. 4. Monitoring goodness of fit in the process of model inference. Sir Isaac adds
data gradually to aid in parameter fitting. As data is added, the selected model includes more
detail (number of model parameters in black) until it saturates to the 10 parameter model we use.
The goodness of fit (χ2 per degree of freedom, in yellow) is measured using all data, including time
points not used in model fitting.
J) and the parameter values themselves (the timescales τi, interaction strengths Vi and Wij,
and biases θj). The first dynamical variable x1 is taken to be the signed speed of the worm's
center of mass, in pixels per second, with negative values corresponding to backward motion.
Further dynamical variables xi with i > 1 correspond to latent (unmeasured) dynamical
variables.
The fitting procedure starts by using one datapoint (one random time) from each of a few
trials, then gradually adds trials and eventually multiple datapoints per trial, refitting model
parameters at each step. The resulting model fits are scored based on their performance
in predicting the entire time series (see Fig. 4 for the fit quality). When the performance
and model complexity of the winning model saturate, we use the resulting model as our
description of the system.
In this way, parameters are fit using only a small subset of
the available data -- we find that using ∼ 2 randomly chosen timepoints per trial (out of
the total 165) is sufficient, cf. Fig. 4. This approach significantly reduces computational
effort (which scales linearly in the number of datapoints used) and minimizes the effects of
correlations between datapoints that are close to one another in time. Finally, it prevents
the optimization from getting stuck at local minima and saddles, which change as new
data points are added, or data is randomized. These reasons are similar to the reasons
behind stochastic gradient descent approaches [38]. The developed software is available
from https://github.com/EmoryUniversityTheoreticalBiophysics/SirIsaac.
12
0.00.51.01.52.0Number of datapoints per trial1.01.52.02/ dof2/ dof024681012Number of model parametersNum. parameters TABLE I. Parameters used for input data, dynamical model inference, and the dynamics of the
sensory input. The inference parameters quoted are used in the Sir Isaac algorithm: "complexity
stepsize" sets which models in the model hierarchy are tested, with 1 indicating that no models
are skipped as parameters are added to make models more complex; here "ensemble" refers to the
parameter ensemble used to avoid local minima during parameter fitting; "avegtol" and "maxiter"
are parameters controlling the local minimization phase of parameter fitting; "stopfittingN" sets
the number of higher-complexity models needed to be shown to have poorer performance before
selecting a given model and finishing the optimization; "connection order" and "type order" set the
order of adding parameters within the model hierarchy [7]; and "prior σ" is the standard deviation
of the Gaussian priors on all parameters.
Input data parameters
equilibration time 10 s
Inference parameters
complexity stepsize 1
tested ensemble members 10
maxiter
avegtol
stopfittingN
100
10−2
5
type order
'last'
prior σ
Sensory input parameters
τon
0.1 s
τdecay
Inferred models
ensemble temperature
100
ensemble generation steps 1000
connection order
'node'
100
0.25 s
The inference procedure produces the following differential equations:
= − v
+ V1 h(t) +
τ1
= −x2 + V2 h(t) +
dv
dt
dx2
dt
W11
1 + exp(speed + θ1)
W21
1 + exp(speed + θ1)
+
+
W12
1 + exp(x2)
W22
1 + exp(x2)
,
.
(5)
(6)
The model includes one latent dynamical variable x2. The maximum likelihood fit param-
eters are shown in Table II. (Note that the selected model did not include variables τ2 and
θ2, so they are set to their default values: τ2 = 1 and θ2 = 0.)
We notice that some of the inferred parameters are close to zero, and so we check whether
13
TABLE II. Maximum-likelihood parameters for the model inferred by Sir Isaac. We do not calcu-
late parameter co-variances since posteriors are sloppy [39] and non-Gaussian around the maximum
likelihood. Instead we estimate the standard deviation of the fits themselves, cf. Fig. 1(A). Param-
eter values are reported to an accuracy of about one digit beyond the least statistically significant
one.
τ1
V1
1.33 s
−2.30 px/(s2 mA)
81.0 px/s2
W11
W21 −0.1 s−1
v(t = 0) 12.6 px/s
θ1
V2
W12
W22
42.1 px/s
1.7 1/(s mA)
44.1 px/s2
5.5 s−1
x2(t = 0) 1.6
TABLE III. Maximum-likelihood parameters for the reduced model.
τ1
V1
W11
θ1 42.1 px/s
1.33 s
−2.30 px/(s2 mA) V2 1.5 1/(s mA)
76. px/s2
W12 19.7 px/s2
v(t = 0) 12.1 px/s
the model can be simplified by setting each of them to zero one at a time and in combi-
nations and then measuring the approximate Bayesian model selection posterior likelihood
score [8] for the original Sir Isaac inferred model and for each of the reduced models. The
original model has the log-likelihood of −197.2, and the best model, Eqs. (1, 2), has the
highest Bayesian likelihood of −192.9. This becomes our chosen model, maximum likelihood
parameters for which can be found in Table III.
Bayesian model of uncertainty
To quantify uncertainty in model structure and parameters, we take a Bayesian approach
and sample from the posterior distribution over parameters. Assuming independent Gaus-
sian fluctuations in the data used in fits, the posterior is simply exp (−χ2(Θ)), with χ2(Θ)
from Eq. (3). We use a standard Metropolis Monte Carlo sampler, as implemented in Slop-
pyCell [40], to take 100 samples (each separated by 1000 Monte Carlo steps) used to quantify
uncertainty in the fit, cf. Fig 1(A).
14
FIG. 5. Comparison of the automated model to the manually curated model of Ref. [2].
(A) The fraction of the overall explainable velocity variance not explained by our model; the
unexplained variance is roughly half of that of Ref. [2].
(B, C) The escape response has been
modeled previously as stereotypical, with an overall scaling that is laser current dependent (for
laser currents > 25 mA). Perfect stimulus-dependent rescaling of the escape would correspond to
a correlation of 1 between velocity traces for different stimuli. (B) shows these correlations, over
time, between mean response speeds in different laser current bins. (C) shows the same correlations
in the inferred model. To the extent that correlations in both panels are nearly the same, our model
recovers the approximate stereotypical nature of the response [2].
Model of sensory input
Each experimental trial begins with a forward moving worm, which is stimulated with a
laser pulse of duration τon = 0.1 s starting at time t = 0. The worm's nose experiences a
quick local temperature increase h(t), which we model as a linear increase during the pulse,
with slope proportional to the laser current I. The time scale of the temperature decay of
the heated area to the ambient temperature due to heat diffusion is 0.15 s [41]. However,
the stimulation area is broad (220µm, FWHM), and as the worm retreats, its head with the
sensory neurons first moves deeper into the heated area, before the temperature eventually
decreases. Thus the dynamics of the sensory input is complex and multiscale. However, since
each individual behaves differently, and we do do not measure individual head temperatures,
we model the average sensory stimulus past the heating period as a single exponential decay
15
compare_to_manual_0034_reduced.pdfBCABCAcompare_to_manual_0034_reducedThree.pdf with the longer time scale of τdecay = 0.25 s:
h(t)/α =
0
It/τon
I exp [−(t − τon)/τdecay]
t ≤ 0,
0 ≤ t ≤ τon,
t ≥ τon.
(7)
Here α has units temperature per unit of laser current. For convenience and without loss of
generality, we set α = 1 and absorb its definition into the Vi parameters that multiply h in
Eqs. (5, 6) giving Vi units of the time derivative of xi per unit current.
Comparison to the model of Ref. [2]
A quantitative model of C. elegans nociceptive response was constructed in Ref. [2].
It involved partitioning worms into actively escaping and (nearly) pausing after a laser
stimulation, estimating the probability of pausing as a function of the applied laser current,
and finally noticing that the mean response of the active worms is nearly stereotypical, with
the response amplitude depending nonlinearly on the stimulus intensity. The same stimulus
causes varied response trajectories in individual worms, and this variability is unexplainable
in any model that only relates the stimulus to the population averaged response velocity. Of
the variability in the response that is explainable by the stimulus, the model of Ref. [2] could
not account for ∼ 20% in the range from a few hundred ms to about 2 s post-stimulation.
Figure 5(A) shows that Sir Isaac our selected model captures more of the data variance,
leaving only 10% of the explainable variance unexplained over much of the same time
range. Figure 5(B,C) illustrates that Sir Isaac has recovered the approximate stereotypy in
the response to the same accuracy as it is present in the data.
ACKNOWLEDGMENTS
This work was partially supported by NIH Grants EB022872 and NS084844, James S.
McDonnell Foundation Grant 220020321, NSERC Discovery Grant, and NSF Grant IOS-
1456912. We are grateful to the NVIDIA corporation for donated Tesla K40 GPUs. We
are grateful to the Aspen Center for Physics, supported by NSF grant PHY-1607611, for
16
supporting this research within "Physics of Behavior" programs.
[1] A. Mohammadi, J. Byrne Rodgers, I. Kotera, and W. Ryu, BMC Neurosci 14, 66 (2013).
[2] K. Leung, A. Mohammadi, W. Ryu, and I. Nemenman, PLoS Comput Biol 12, e1005262
(2016).
[3] M. Schmidt and H. Lipson, Science 324, 81 (2009).
[4] M. Schmidt, R. Vallabhajosyula, J. Jenkins, J. Hood, A. Soni, J. Wikswo, and H. Lipson,
Physical Biology 8, 055011 (2011).
[5] D. Sussillo and L. Abbott, Neuron 63, 544 (2009).
[6] G. Neuert, B. Munsky, R. Tan, L. Teytelman, M. Khammash, and A. van Oudenaarden,
Science (New York, NY) 339, 584 (2013).
[7] B. Daniels and I. Nemenman, Plos One 10, e0119821 (2015).
[8] B. Daniels and I. Nemenman, Nature Commun 6, 8133 (2015).
[9] S. Brunton, J. Proctor, and J. Kutz, Proc Natl Acad Sci (USA) 113, 3932 (2016).
[10] N. Mangan, S. Brunton, J. Proctor, and J. Kutz, IEEE Transactions on Molecular, Biological
and Multi-Scale Communications 2, 52 (2016).
[11] Z. Lu, J. Pathak, B. Hunt, M. Girvan, R. Brockett, and E. Ott, Chaos: An Interdisciplinary
Journal of Nonlinear Science 27, 041102 (2017).
[12] J. Pathak, B. Hunt, M. Girvan, Z. Lu, and E. Ott, Phys Rev Lett 120, 024102 (2018).
[13] C. Pandarinath, D. O'Shea, J. Collins, R. Jozefowicz, S. Stavisky, J. Kao, E. Trautmann,
M. Kaufman, S. Ryu, L. Hochberg, J. Henderson, K. Shenoy, L. Abbott, and D. Sussillo,
bioRxiv , 152884 (2017).
[14] A. Henry, M. Hemery, and P. Fran¸cois, PLoS Comput Biol 14, e1006244 (2018).
[15] P. Nelson, Physical Models of Living Systems (W. H. Freeman and Co., New York, NY, 2015).
[16] V. Vapnik, Statistical learning theory (Wiley, 1998).
[17] J. Rissanen, Stochastic Complexity in Statistical Inquiry Theory (World Scientific, 1989).
[18] W. Bialek, I. Nemenman, and N. Tishby, Neural Computation 13, 2409 (2001).
[19] M. Savageau and E. Voit, Mathematical biosci 87, 83 (1987).
[20] R. Beer and B. Daniels, "Saturation probabilities of continuous-time sigmoidal networks,"
(2010), arXiv: 1010.1714.
17
[21] D. J. MacKay, Neural Comput 4, 415 (1992).
[22] V. Balasubramanian, Neural Comput 9, 349 (1997).
[23] R. Eaton, Neural Mechanisms of Startle Behavior (Springer US, 2013).
[24] J. Pirri and M. Alkema, Curr Opin Neurobiol 22, 187 (2012).
[25] C. Bargmann, J. Thomas, and H. Horvitz, Cold Spring Harbor Symposia on Quantitative
Biology 55, 529 (1990).
[26] M. Hilliard, A. Apicella, R. Kerr, H. Suzuki, P. Bazzicalupo, and W. Schafer, The EMBO
Journal 24, 63 (2005).
[27] J. Kaplan and H. Horvitz, Proc Natl Acad Sci USA 90, 2227 (1993).
[28] N. Wittenburg and R. Baumeister, Proc Natl Acad Sci USA 96, 10477 (1999).
[29] G. J. Stephens, B. Johnson-Kerner, W. Bialek, and W. S. Ryu, PLoS computational biology
4, e1000028 (2008).
[30] M. B. Goodman and P. Sengupta, Pflugers Archiv : European journal of physiology 470, 839
(2018).
[31] S. Liu, E. Schulze, and R. Baumeister, PloS one 7, e32360 (2012).
[32] J. G. White, E. Southgate, J. N. Thomson, and S. Brenner, Philosophical transactions of the
Royal Society of London. Series B, Biological sciences 314, 1 (1986).
[33] S. Gao, L. Xie, T. Kawano, M. D. Po, S. Guan, M. Zhen, J. K. Pirri, and M. J. Alkema,
Nature communications 6, 6323 (2015).
[34] V. Sourjik and N. S. Wingreen, Current opinion in cell biology 24, 262 (2012).
[35] A. Paulick, V. Jakovljevic, S. Zhang, M. Erickstad, A. Groisman, Y. Meir, W. S. Ryu, N. S.
Wingreen, and V. Sourjik, eLife 6 (2017), 10.7554/eLife.26607.
[36] A. Jilkine and L. Edelstein-Keshet, PLoS Comput Biol 7, e1001121 (2011).
[37] A. Levchenko and P. Iglesias, Biophys J 82, 50 (2002).
[38] P. Mehta, M. Bukov, C. Wang, A. Day, C. Richardson, C. Fisher, and D. Schwab, arXiv.org
(2018), 1803.08823v1.
[39] M. Transtrum, B. Machta, K. Brown, B. Daniels, C. Myers, and J. Sethna, J Chem Phys
143, 010901 (2015).
[40] C. Myers, R. Gutenkunst, and J. Sethna, Computing in Sci Engin 9, 34 (2007).
[41] A. Mohammadi, Quantitative Behavioral Analysis of Thermal Nociception in Caenorhabditis
Elegans: Investigation of Neural Substrates Spatially Mediating the Noxious Response, and the
18
Effects of Pharmacological Perturbations, Ph.D. thesis, University of Toronto (2013).
19
|
1205.0321 | 2 | 1205 | 2012-05-09T12:29:23 | The span of correlations in dolphin whistle sequences | [
"q-bio.NC",
"physics.data-an"
] | Long-range correlations are found in symbolic sequences from human language, music and DNA. Determining the span of correlations in dolphin whistle sequences is crucial for shedding light on their communicative complexity. Dolphin whistles share various statistical properties with human words, i.e. Zipf's law for word frequencies (namely that the probability of the $i$th most frequent word of a text is about $i^{-\alpha}$) and a parallel of the tendency of more frequent words to have more meanings. The finding of Zipf's law for word frequencies in dolphin whistles has been the topic of an intense debate on its implications. One of the major arguments against the relevance of Zipf's law in dolphin whistles is that is not possible to distinguish the outcome of a die rolling experiment from that of a linguistic or communicative source producing Zipf's law for word frequencies. Here we show that statistically significant whistle-whistle correlations extend back to the 2nd previous whistle in the sequence using a global randomization test and to the 4th previous whistle using a local randomization test. None of these correlations are expected by a die rolling experiment and other simple explanation of Zipf's law for word frequencies such as Simon's model that produce sequences of unpredictable elements. | q-bio.NC | q-bio |
The span of correlations in dolphin whistle
sequences
Ramon Ferrer-i-Cancho1,∗ and Brenda McCowan2
1 Complexity & Quantitative Linguistics Lab
Departament de Llenguatges i Sistemes Inform`atics,
TALP Research Center, Universitat Polit`ecnica de Catalunya,
Campus Nord, Edifici Omega Jordi Girona Salgado 1-3.
08034 Barcelona, Catalonia (Spain)
2 Population Health & Reproduction,
School of Veterinary Medicine,
University of California,
One Shields Ave 1024 Haring Hall
Davis, CA 95616, United States
E-mail: [email protected] and [email protected]
Abstract. Long-range correlations are found in symbolic sequences from human
language, music and DNA. Determining the span of correlations in dolphin whistle
sequences is crucial for shedding light on their communicative complexity. Dolphin
whistles share various statistical properties with human words, i.e. Zipf's law for word
frequencies (namely that the probability of the ith most frequent word of a text is about
i−α) and a parallel of the tendency of more frequent words to have more meanings. The
finding of Zipf's law for word frequencies in dolphin whistles has been the topic of an
intense debate on its implications. One of the major arguments against the relevance of
Zipf's law in dolphin whistles is that is not possible to distinguish the outcome of a die
rolling experiment from that of a linguistic or communicative source producing Zipf's
law for word frequencies. Here we show that statistically significant whistle-whistle
correlations extend back to the 2nd previous whistle in the sequence using a global
randomization test and to the 4th previous whistle using a local randomization test.
None of these correlations are expected by a die rolling experiment and other simple
explanation of Zipf's law for word frequencies such as Simon's model that produce
sequences of unpredictable elements.
Keywords: Zipf's law, die rolling, random typing, dolphin whistles.
PACS numbers: 89.70.-a Information and communication theory
89.75.Da Systems obeying scaling laws
05.45.Tp Time series analysis
The span of correlations in dolphin whistle sequences
2
1. Introduction
Long-range correlations have been reported in different kinds of symbolic sequences:
human language [1, 2, 3, 4], DNA [5] and music [6, 4]. A few studies have studied
the span of correlations in sequences of behavior produced by other species [7, 8, 9].
Rather long-correlations (extending back at least to the 7th previous element) have
reported in sequences of dolphin surface behavioral patterns [7]. A preliminary analysis
of constraints in sequences of dolphin whistles was performed in Ref.
[8] but strong
conclusions were not reached due to the small size of the dataset. Determining the
actual span of correlations in dolphin sequences is crucial for shedding light on the
communicative complexity of dolphins whistles.
Various similarities between human words and dolphin whistles have been reported.
A parallel of Zipf's law of meaning distribution, the tendency or more frequent words
to have more meanings [10], has been found in dolphins whistles [11]. Zipf's law for
word frequencies, namely, the probability of the ith most frequent word of a text,
is ∼ i−α, where α is the exponent of the law [10], has also been found in dolphin
whistles [8]. However, the finding remains controversial because it has been argued that
simply rolling a die could explain the presence of the law in dolphin whistles [12]. The
experiment consists of generating a sequence of faces by rolling a die. One of the faces
of the die plays the role of a word delimiter. For instance, a die of six faces could
produce 1, 5, 2, 6, 3, 2, 4, 3, 5, 6... Treating 6 as a pseudo-word delimiter, the previous
sequence of faces becomes the sequence of pseudo-words 152, 32435, .... The experiment
is an abstraction of the popular monkey typing experiment, that consists of typing at
random on a keyboard (the face that plays the role of the word delimiter is the space
and the other faces are letters) [13]. Die rolling and monkey typing have been argued
to explain or mimic Zipf's law for word frequencies in human words [13] and dolphin
whistles [12]. Another die rolling experiment, where the probability of the ith face is
the probability of the ith most frequent word, has been proposed [14].
The hypothesis of die rolling as an explanation for Zipf's law for word frequencies
in human language in other species can be tested in at least two different ways:
(i) By comparing the actual distribution of 'word' frequencies with the one that is
actually produced by the die rolling process [15]. Concerning the die rolling
experiment of Ref.
[12], the parameters of the model that provide a satisfactory
fit to actual word frequencies according to a statistically rigorous test, are indeed
unknown [15], in spite of the many previous claims about the good fit of the model
using qualitative arguments [13, 12]. Concerning the die rolling experiment or Ref.
[14], the model is able to trivially provide a perfect fit to any theoretical or empirical
discrete distribution, being Zipf's law for word frequencies a particular case.
(ii) By comparing the statistical properties of the sequence of words produced with those
of the sequence that is produced by the die rolling process. The rolling experiments of
Refs. [12, 14] produce a sequence of independent 'words' in the sense that elements
that have already been produced carry no information about the next element. In
The span of correlations in dolphin whistle sequences
3
contrast human language, shows long range correlations in texts using words (e.g.
[2, 3]) or letters (e.g., [1, 16]) as units of the sequence.
Here we will follow the second track for dolphin whistles. The aim of the present article
is to determine the span of correlations in dolphin whistle sequences and evaluate the
suitability of die rolling [12, 14]. The challenge of the analysis is facing the statistical
problems arising from the rather small size of the dataset of Ref.
[8]. The danger of
undersampling in the context of dolphins whistles has already been discusssed [8]. The
next section presents the information theoretic measure approach that will be used to
study correlations in dolphin whistle sequences in that dataset.
2. Mutual information
We consider pairs of whistles in a sequence and three related random variables: X for
the whistle that is the 1st member of the pair, Y for the whistle that is the second
member of the pair and D for their distance. We adopt the convention that consecutive
elements are at distance 1, elements separated by one element are at distance 2, and
so on...[17]. Given a collection of sequences of whistles, p(X = x, Y = yD = d) is
defined as the probability that whistle x is followed by whistle y knowing that they are
at distance d from each other and dmax is defined as the maximum distance considered in
the analysis (thus 1 ≤ d ≤ dmax). Given a certain distance d, the marginal conditional
probabilities are defined as
p(X = xD = d) = Xy
p(Y = yD = d) = Xx
p(X = x, Y = yD = d)
p(X = x, Y = yD = d).
I(X; Y D = d), the conditional mutual information between X and Y given a concrete
distance d, is defined as [18]
I(X; Y D = d) = Xx,y
where
p(X = x, Y = yD = d) log q(X = x, Y = yD = d),
(1)
q(X = x, Y = yD = d) =
p(X = x, Y = yD = d)
p(X = xD = d)p(Y = yD = d)
.
I(X; Y D = d) has been used to study long-range correlations in DNA, texts and music
[5, 4].
f (X = x, Y = yD = d) is defined as the number of times that x has been followed
by y at distance d. The marginal conditional frequencies are defined as
f (X = xD = d) = Xy
f (Y = yD = d) = Xx
f (X = x, Y = yD = d)
f (X = x, Y = yD = d),
(2)
(3)
and the total number of pairs at distance d is defined as
F (d) = Xx
f (X = xD = d) = Xy
f (Y = yD = d).
The span of correlations in dolphin whistle sequences
4
In a finite collection of sequences, p(X = x, Y = yD = d), p(X = xD = d) and
p(Y = yD = d) are relative frequencies, i.e.
p(X = x, Y = yD = d) =
p(X = xD = d)
p(Y = yD = d)
=
=
f (X = x, Y = yD = d)
F (d)
f (X = xD = d)
F (d)
f (Y = yD = d)
F (d)
.
(4)
(5)
(6)
In a real collection of sequences, p(X = x, Y = yD = d), p(X = xD = d) and
p(Y = yD = d) are estimated from these relative frequencies. Applying Eqs. 4, 5 and
6 to Eq. 1, yields the sample mutual information between X and Y given a concrete
distance d,
Is(X; Y D = d) = log F (d) +
σ1 − σ2 − σ3
F (d)
,
(7)
where
σ1 = Xx,y
σ2 = Xx
σ3 = Xy
f (X = x, Y = yD = d) log f (X = x, Y = yD = d)
f (X = xD = d) log f (X = xD = d)
f (Y = yD = d) log f (Y = yD = d).
Next two useful properties of Is(X; Y D = d) are presented:
(i) If marginal conditional frequencies at distance d are boolean, i.e. f (X = xD =
d) = f (Y = yD = d) ∈ {0, 1} for any x and y then Is(X; Y D = d) = log F (d).
To see it, notice that f (X = xD = d) = f (Y = yD = d) = 1 implies, thanks to
Eqs. 2 and 3, that the joint frequencies are all boolean, i.e. f (X = x, Y = yD =
d) ∈ {0, 1} for any x and y. According to Eq. 7, the fact that both marginal and
joint conditional frequencies do not exceed one yields σ1 = σ2 = σ3 = 0 and thus
Is(X; Y D = d) = logF (d) as we wanted to prove.
(ii) If all whistles at cooccurring at distance d are identical (only one whistle type has
non-zero frequency), then σ1 = σ2 = σ3 = F (d) log F (d) and then, according to Eq.
7, Is(X; Y D = d) = 0.
3. Methods
We reused the collection of whistle sequences employed to study Zipf's law in dolphin
whistles [8]. A summary of the elementary statistical properties of the sequences is
provided in Table 1. Sequences of length smaller than 2 where filtered out.
3.1. Distances where correlations are significant
For each dolphin in the dataset, his/her collection of sequences was analyzed to extract
a list of distances in the interval [2, d2
max] at which Is(X; Y D = d) is statistically
The span of correlations in dolphin whistle sequences
5
significant at a significance level of a = 0.05. For each dolphin and each distance, we
used a Monte Carlo procedure for estimating a p-value indicating the probability that
the value of Is(X; Y D = d) from a randomized version of the data Is(X; Y D = d) is
at least as large as that of the original collection of sequences:
(i) R = 107 randomized versions of the original data were generated.
(ii) R≥, the number of times the value of Is(X; Y D = d) is at least as large as that of
max] and the p-value was estimated
the original data was calculated for each d ∈ [1, d2
as R≥/R.
(iii) For each dolphin, all distances d such that p-value = R≥/R ≤ a were added to the
list of distances.
3.2. Upper bounds for the span of correlations
We say that Is(X; Y D = d) is constant if Is(X; Y D = d) is the same for any
randomization of data upon which Is(X; Y D = d) is computed. We consider two
situations in which Is(X; Y D = d) cannot be significantly high: (a) Is(X; Y D = d) is
constant in the sense above or (b) Is(X; Y ; D = d) = 0 as 0 is actually the minimum of
mutual information [18]. Accordingly, we consider three different ways of defining dmax:
• d0
• d1
• d2
max, defined through the number of pairs of elements at distance d in a sequence
of length l, i.e. π(d, l) = l − d. The maximum distance d that can be considered
in a collection of sequences where the longest sequence has length lmax is obtained
from π(dmax, lmax) ≥ 1, which gives d0
max, defined as the smallest distance at which Is(X; Y D = d) is constant in the
sense above for any d ∈ (d1
max, defined as the smallest distance beyond which Is(X; Y D = d) cannot
be significantly high for d ∈ (d2
max because
Is(X; Y D = d) cannot be significantly high if Is(X; Y D = d) is constant.
max]. Notice that d2
max = lmax − 1.
max ≤ d1
max, d0
max].
max, d0
The interest of these definitions of dmax is two-fold. First, bounding a priori the span of
correlations between whistles. Second, reducing the computational cost of evaluating the
significance of Is(X; Y D = d) at distances where the result of the test is straightforward.
Distances greater than d2
max can be discarded.
We did not check if Is(X; Y D = d) = 0 using Eq. 7 as this is problematic due to
finite numerical precision for real numbers. Eq. 1 indicates that Is(X; Y D = d) = 0 if
and only if q(X = x; Y = yD = d) = 1 for any x and y such that p(X = x; Y =
yD = d) > 0. Applying Eqs.
it is easy to see that the condition
q(X = x; Y = yD = d) = 1 is equivalent to a more numerically convenient condition,
i.e. F (d)f (X = x, Y = yD = d) = f (X = xD = d)f (Y = yD = d).
4, 5 and 6,
3.3. Kinds of randomization
Two kind of randomizations of the data upon Is(X; Y D = d) is calculated were
considered: global and local randomization. A precise calculation of d1
max and d2
max
The span of correlations in dolphin whistle sequences
6
Table 1. Summary of the elementary statistical properties of the collections of
sequences of each dolphin. For each dolphin, the following information is shown:
dolphin's name, T , the total number of whistle types, V , the number of different
whistle types, S (the number of sequences), hli = T /S (the mean sequence length in
whistle types), lmax (the maximum length). Dolphins are sorted decreasingly by T .
Name
Liberty
Norman
Panama
Chelsea
Sadie
Stormy
Delphi
Neptune
Circe
Desmond
Sam
Tasha
Bayou
Terry
Schooner
ECB
Gordo
T
334
235
148
110
110
94
65
53
47
40
39
35
27
21
19
15
6
V
31
39
42
12
13
12
18
9
4
3
4
9
10
7
4
5
6
S
hli
lmax
102
74
43
34
33
27
24
9
15
7
12
12
8
5
6
6
2
3.27
3.18
3.44
3.24
3.33
3.48
2.71
5.89
3.13
5.71
3.25
2.92
3.38
4.20
3.17
2.50
3.00
11
21
10
8
8
9
7
13
7
10
13
7
8
7
6
4
4
depends on the kind of randomization.
3.3.1. Global randomization Here the data that is randomized is the whole collection
of sequences. The randomization procedure consisted of copying all whistles in a vector
of length
S
T =
li,
Xi=1
where li is the length of the ith sequence and S is the number of sequences of the
collection. Then, every randomized version of the collection is obtained by generating
a uniformly distributed random permutation of the vector [19] and cutting that vector
in pieces of lengths l1, ..., li, ...lS (cut always in this order) to produce the randomized
collection of sequences. Notice that the randomization procedure preserves the sequence
lengths and the frequencies of each whistle in the original collection of sequences.
Concerning the computation of d1
max, notice that constant Is(X; Y D = d) with regard
to any global randomization can occur in three circumstances:
(i) If all the whistle types are hapax legomena,
they occur only once in the
whole collection as all the marginal conditional frequencies are boolean and thus
I(X; Y D = d) = log F (d), as it has been shown above, for any d ∈ [1, d0
i.e.
max].
The span of correlations in dolphin whistle sequences
7
(ii) If the repertoire size is one (there is only one whistle type of non-zero frequency)
then all the whistles cooccurring at a certain distance are identical and thus
I(X; Y D = d) = 0, as it has been shown above, for any d ∈ [1, d0
max].
(iii) When F (d) = 1, all the marginal conditional frequencies are boolean in this case
and then I(X; Y D = d) = log F (d) = log 1 = 0.
These conditions lead to the following specific procedure for calculating d1
condition (i) (this is the case of the dolphin 'Gordo') or (ii) implies d1
these two conditions is met, d1
warrant F (d) > 1, d1
length and d1
max = d0
sequences of maximum length, then F (d1
F (d) ≥ 2 if and only if Smax ≥ 2.
max. Satisfying
max = 0. If none of
max is calculated by means of condition (iii). In order to
max must be d0
max if there is more than one sequence of maximum
max − 1 otherwise. To see it, notice that if the collection has Smax
max) = Smaxπ(lmax − 1, l) = Smax and thus
3.3.2. Local randomization Here the data that is randomized are the pairs of whistles
occurring at a certain distance d. The randomization procedure consisted of copying all
the pairs of whistles occurring at distance d in a vector of length 2F (d). Then, every
randomized version of the collection was obtained by generating a uniformly distributed
random permutation of the vector [19] and then taking the ith and the (i + 1)th whistles
of the vector, with 1 ≤ i ≤ 2F (d), to form the jth pair of whistles of the randomized
pairs of whistles, with 1 ≤ j ≤ F (d) and j = ⌈i/2⌉. Notice that the randomization
procedure preserves the frequencies of each whistle in the original pairs of whistles at
distance d.
Concerning the computation of d1
max for local normalization notice that constant
Is(X; Y D = d) with regard to any local randomization for distance d, adds two new
relevant conditions with regard to global randomization:
(iv) When all whistle types are locally hapax legomena,
all the whistle types
forming the pairs at distance d occur only once in the ensemble of pairs at
distance d. In that case, all marginal conditional frequencies are boolean and thus
Is(X; Y ; D = d) = F (d) as it has been shown above.
i.e.
(v) When the local repertoire has size one, i.e. all the pairs at distance d are made of
a single whistle type. In that case, all Is(X; Y ; D = d) = 0 as it has been shown
above.
The procedure for calculating d1
compute d1
max
while d1
max:
(a) all whistle types are locally hapax legomena or (b) the local repertoire size is one.
max with the same procedure used for global normalization. Decrease d1
max > 0 and at least one of the two following conditions is met at distance d1
max under local normalization is the following. First,
4. Results
Table 2 shows the values of d where Is(X; Y D = d) is significantly high in the dolphins
d is
from Ref.
[8] at a significance level of 0.05 for global and local randomization. na
The span of correlations in dolphin whistle sequences
8
Table 2. Analysis of the span of correlation between whistles at a certain distance d
using global and local randomization. For the collection of whistles sequences of each
dolphin, the following information is shown: the dolphin's name, d0,1,2
max (the maximum
distance according to different definitions) and the list of distances d ∈ [1, d2
max] such
that Is(X; Y D = d) is significantly high at a significance level of 0.05. Dolphins are
sorted decreasingly by T .
Global randomization
Local randomization
Name
d0
max
d1
max
d2
max Distances
d1
max
d2
max Distances
Liberty
Norman
Panama
Chelsea
Sadie
Stormy
Delphi
Neptune
Circe
Desmond
Sam
Tasha
Bayou
Terry
Schooner
ECB
Gordo
10
20
9
7
7
8
6
12
6
9
12
6
7
6
5
3
3
9
19
8
6
6
7
5
11
5
8
11
5
6
5
4
2
0
7
15
7
3
6
6
5
7
1
3
9
5
6
3
2
2
0
1, 2
5
1
1, 2
1
2
1
7
1
1
1, 4, 6, 7
1, 2, 3, 4
1, 2, 3
1, 2
1, 2
1, 2, 3
1
1, 2, 3, 4
1, 2
1
1
8
19
8
4
6
7
4
11
1
8
10
5
5
4
4
2
0
7
15
7
3
6
6
4
7
1
3
9
5
5
3
2
2
0
defined as the number of dolphins for which Is(X; Y D = d) is significantly high at a
significance level a. Table 2 yields n0.05
7 = 1 for global
randomization and n0.05
7 = 1 for local
randomization.
1 = 11, n0.05
2 = 3, n0.05
1 = 7, n0.05
4 = 3, n0.05
6 = n0.05
2 = 7, n0.05
3 = 4, n0.05
5 = n0.05
d
d
Next we will perform a meta-analysis to determine when n0.05
is significantly high
because n0.05
could be the outcome of false positives (type I statistical errors) from
the test that determines if Is(X; Y D = d) is significantly high for a certain dolphin.
The p-value of a continuous statistic is known to be uniformly distributed under the
null hypothesis [20]. In our case, Is(X; Y D = d) is approximately continuous and the
quality of the approximation is expected to increase with the size of the permutation
space of the collection of sequences. This implies that na
d, the number of dolphins for
which Is(X; Y D = d) is significantly high at a significance level a follows approximately
a binomial distribution with parameters Nd and a, where Nd is the total number of
dolphins for whom Is(X; Y D = d) can be significantly high according to the criteria
above, and a = 0.05 is the significance level. Notice that the dolphin 'Gordo' is a
special case because he does no have distances at which mutual information can be
significantly high (he has d1
max = 0 because all the whistles in his collection
are hapax legomena; notice T = V in Table 1). Thus, Nd is not simply the number of
max = d2
The span of correlations in dolphin whistle sequences
9
d
cannot be significantly high (as n0.05
Table 3. Summary of the test of the significance of n0.05
, the number of dolphins for
which Is(X; Y D = d) is significantly high at a significance level of 0.05. Distances
where n0.05
d = 0 is minimum) are excluded from
the analyses and filled with −. Nd is the number of dolphins for whom Is(X; Y D = d)
can be significantly high. The p-values indicate the probability of reaching a value of
n0.05
at least as high as the actual one simply by chance. The p-values were estimated
by means of two different procedures: a binomial approximation and a Monte Carlo
method. p-values were rounded to leave only one significant digit.
d
d
Global randomization
Local randomization
n0.05
d
Nd
p-value
n0.05
d
Nd
p-value
d
1
2
3
4
5
6
7
7
3
0
0
1
0
1
Binomial Numerical
Binomial Numerical
16
15
−
−
10
−
5
6 × 10−6 < 0.001
0.04
−
−
0.4
−
0.2
0.02
−
−
0.3
−
0.2
11
7
4
3
0
1
1
16
15
13
10
−
7
5
2 × 10−11 < 0.001
3 × 10−6
< 0.001
0.001
0.003
0.009
0.01
−
−
0.2
0.3
0.2
0.2
dolphins in the dataset.
A binomial test can be used to asses if na
d is significantly large. The p-value of this
test is
p − value =
Nd
Xx=na
d
x !ax(1 − a)Nd−x.
Nd
(8)
Concerning global randomization, a binomial test indicates that the number of dolphins
showing a significant correlation is significantly high for d = 1 and d = 2 but not for
d > 2 (Table 3). Concerning local randomization, a binomial test indicates that the
number of dolphins showing a significant correlation is significantly high for 1 ≤ d ≤ 4
but not for d > 4 (Table 3).
The meta-analysis based upon a binomial test assumes that 0.05 is the probability
of rejecting the null hypothesis by chance but indeed the accuracy of the assumption
depends on the properties of the collection of whistle types. For instance, we have seen
that the mutual information cannot be significantly high for the dolphin 'Gordo' and
thus the probability of rejecting the null hypothesis by chance is thus 0 for him. A Monte
Carlo meta-analysis allows one to improve the approximate calculation of the p-value
offered by Eq. 8 to some degree of numerical precision. Consider that a randomized
ensemble of collections of sequences consists of a randomized version of the collection
of each of the dolphins (excluding 'Gordo'). For each distance d, the p-value of na
d is
estimated by the proportion of times in Rm randomized ensembles of collections that
the value of na
d of the original (non-randomized)
ensemble. Rm = 1000 and R = 1000 were used. Concerning global randomization,
d is equal or greater than the value of na
The span of correlations in dolphin whistle sequences
10
d
Table 4. Summary of the support for significance of each distance (based upon Table
3). n0.05
is defined as the number of dolphins for which Is(X; Y D = d) is significantly
high at a significance level of 0.05. For each kind of randomization (global or local)
and meta-analysis (binomial approximation or Monte Carlo method), the distances at
which n0.05
is significantly high are indicated.
d
Distances
Global randomization Local randomization
d > 0
n0.05
Meta-analysis (binomial approx.)
Meta-analysis (Monte Carlo method)
1, 2, 5, 7
1, 2
1, 2
1, 2, 3, 4, 6, 7
1, 2, 3, 4
1, 2, 3, 4
d
one obtains that n0.05
is significantly high for d = 1 and d = 2 but not for d > 2
(Table 3). Concerning local randomization, one obtains that n0.05
is significantly high
for 1 ≤ d ≤ 4 but not for d > 4 (Table 3). To sum up, the conclusions of the two
meta-analyses coincide from a qualitative point of view (Table 4). The binomial test
provides a good enough approximation for both global and local randomization.
d
From a global randomization perspective, correlations are short range: the presence
of significant correlations at low distances (d = 2 and specially d = 1) is unquestionable
but significant correlations at higher distances could be false positives (Table 4). From a
local randomization perspective, correlations extending back to the 4th back whistle type
are unquestionable but farther significant correlations could be false positives (Table 4).
5. Discussion
We have demonstrated that, for the majority of individuals, a dolphin whistle carries (on
average) a significant amount of information about the next whistles of the sequence
(Tables 2 and 3). Global randomization indicates that a whistle carries information
about at least one of the next two whistles of the sequence whereas local randomization
indicates that a whistle carries information about at least one of the next four whistles of
the sequence (Table 4). This is a property that is inconsistent with die-rolling, where a
pseudo-word carries no information at all about the next pseudo-words of the sequence.
The fact that this is also a feature also shared by Simon's model for word frequencies
[21], questions the validity of a popular argument against the utility of Zipf's law for
frequencies, namely that the law is of practically no help in assessing the complexity of
a communication system because there are many ways of reproducing it [22, 23, 14, 12].
Indeed, when the statistical properties of the sequence are taken into account, the
number of candidate explanations drops down. The multiplicity of explanations for
Zipf's law in a species such as dolphins depends on how many statistical features,
besides Zipf's law for word frequencies, are used to break the tie between candidates.
The big question that future research on dolphins whistles must address is: what is the
communicative complexity of a system whose units (e.g., whistles types) are distributed
The span of correlations in dolphin whistle sequences
11
following Zipf's law for word frequencies [8], show a parallel of Zipf's law of meaning
distribution [11] and form sequences with correlations that defy a simple explanation
such as die rolling or Simon's model? We hope that our research stimulates further
data collection to determine if the rather short range correlation discovered here are an
intrinsic property of dolphin whistle communication or a consequence of the small size
of our dataset.
Acknowledgments
We are grateful to L. Doyle for helpful discussions and to M. Piattelli-Palmarini for
making us aware of Ref.
[14]. This work was supported by the grant Iniciaci´o i
reincorporaci´o a la recerca from the Universitat Polit`ecnica de Catalunya and the grants
BASMATI (TIN2011-27479-C04-03) and OpenMT-2 (TIN2009-14675-C03) from the
Spanish Ministry of Science and Innovation.
References
[1] F. Moscoso del Prado Mart´ın. The universal 'shape' of human languages: spectral analysis beyond
speech. PLoS ONE, page in press, 2011.
[2] M. A. Montemurro and D. H. Zanette. PLoS ONE, 6:e19875, 2011.
[3] M. Montemurro and P. A. Pury. Long-range fractal correlations in literary corpora. Fractals,
10:451 -- 461, 2002.
[4] W. Ebeling, T. Poschel, and K.-F. Albrecht. Entropy, transinformation and word distribution of
information-carrying sequences. International Journal of Bifurcation and Chaos, 5:51 -- 61, 1995.
[5] W. Li and K. Kaneko. Long-range correlation and partial 1/f α spectrum in a noncoding DNA
sequence. Europhysics Letters, 17(7):655 -- 660, January 1992.
[6] Z.-Y. Su and T. Wu. Music walk, fractal geometry in music. Physica A: Statistical Mechanics
and its Applications, 380(0):418 -- 428, 2007.
[7] R. Ferrer i Cancho and D. Lusseau. Long-term correlations in the surface behavior of dolphins.
Europhysics Letters, 74(6):1095 -- 1101, 2006.
[8] B. McCowan, S. F. Hanser, and L. R. Doyle. Quantitative tools for comparing animal
communication systems: information theory applied to bottlenose dolphin whistle repertoires.
Anim. Behav., 57:409 -- 419, 1999.
[9] S. A. Altmann. Sociobiology of rhesus monkeys. II: Stochastics of social communication. J. Theor.
Biol., 8:490 -- 522, 1965.
[10] G. K. Zipf. Human behaviour and the principle of least effort. Addison-Wesley, Cambridge (MA),
USA, 1949.
[11] R. Ferrer-i-Cancho and B. McCowan. A law of word meaning in dolphin whistle types. Entropy,
11(4):688 -- 701, 2009.
[12] R. Suzuki, P. L. Tyack, and J. Buck. The use of Zipf's law in animal communication analysis.
Anim. Behav., 69:9 -- 17, 2005.
[13] G. A. Miller and N. Chomsky. Finitary models of language users. In R. D. Luce, R. Bush, and
E. Galanter, editors, Handbook of Mathematical Psychology, volume 2, pages 419 -- 491. Wiley,
New York, 1963.
[14] P. Niyogi and R. C. Berwick. A note on Zipf's law, natural languages, and noncoding DNA regions.
A.I. Memo No. 1530 / C.B.C.L. Paper No. 118, 1995.
[15] R. Ferrer-i-Cancho and B. Elvevag. Random texts do not exhibit the real Zipf's-law-like rank
distribution. PLoS ONE, 5(4):e9411, 2009.
The span of correlations in dolphin whistle sequences
12
[16] W. Ebeling and T. Poschel. Entropy and long-range correlations in literary English. Europhysics
Letters, 26(4):241 -- 246, 1994.
[17] R. Ferrer i Cancho. Euclidean distance between syntactically linked words. Physical Review E,
70:056135, 2004.
[18] T. M. Cover and J. A. Thomas. Elements of information theory. Wiley, New York, 2006. 2nd
edition.
[19] R. Durstenfeld. Algorithm 235: Random permutation. Communications of the ACM, 7(7):420,
1964.
[20] J. A. Rice. Mathematical statistics and data analysis. Duxbury, Belmont, CA, 2007. 3rd edition.
[21] H. A. Simon. On a class of skew distribution functions. Biometrika, 42:425 -- 440, 1955.
[22] G. A. Miller. Some effects of intermittent silence. Am. J. Psychol., 70:311 -- 314, 1957.
[23] A. Rapoport. Zipf's law re-visited. Quantitative Linguistics, 16:1 -- 28, 1982.
|
1311.3952 | 1 | 1311 | 2013-11-15T19:07:41 | A Method for Neuronal Source Identification | [
"q-bio.NC"
] | Multi-sensor microelectrodes for extracellular action potential recording have significantly improved the quality of in vivo recorded neuronal signals. These microelectrodes have also been instrumental in the localization of neuronal signal sources. However, existing neuron localization methods have been mostly utilized in vivo, where the true neuron location remains unknown. Therefore, these methods could not be experimentally validated. This article presents experimental validation of a method capable of estimating both the location and intensity of an electrical signal source. A four-sensor microelectrode (tetrode) immersed in a saline solution was used to record stimulus patterns at multiple intensity levels generated by a stimulating electrode. The location of the tetrode was varied with respect to the stimulator. The location and intensity of the stimulator were estimated using the Multiple Signal Classification (MUSIC) algorithm, and the results were quantified by comparison to the true values. The localization results, with an accuracy and precision of ~ 10 microns, and ~ 11 microns respectively, imply that MUSIC can resolve individual neuronal sources. Similarly, source intensity estimations indicate that this approach can track changes in signal amplitude over time. Together, these results suggest that MUSIC can be used to characterize neuronal signal sources in vivo. | q-bio.NC | q-bio |
A Method for Neuronal Source Identification
Chang Won Lee1, Agnieszka A. Szymanska2, Shun Chi Wu3, A. Lee Swindlehurst4, and
Zoran Nenadic2,4
1Samsung Dallas Technology R&D Lab, Richardson, TX, USA
2Department of Biomedical Engineering, University of California, Irvine, CA, USA
3Covidien, Costa Mesa, CA, USA
4Department of Electrical Engineering and Computer Science, University of California,
Irvine, CA, USA
Abstract
Multi-sensor microelectrodes for extracellular action potential recording have significantly improved
the quality of in vivo recorded neuronal signals. These microelectrodes have also been instrumental in the
localization of neuronal signal sources. However, existing neuron localization methods have been mostly
utilized in vivo, where the true neuron location remains unknown. Therefore, these methods could not be
experimentally validated. This article presents experimental validation of a method capable of estimating
both the location and intensity of an electrical signal source. A four-sensor microelectrode (tetrode)
immersed in a saline solution was used to record stimulus patterns at multiple intensity levels generated
by a stimulating electrode. The location of the tetrode was varied with respect to the stimulator. The
location and intensity of the stimulator were estimated using the Multiple Signal Classification (MUSIC)
algorithm, and the results were quantified by comparison to the true values. The localization results, with
an accuracy and precision of ∼10 µm, and ∼11 µm respectively, imply that MUSIC can resolve individual
neuronal sources. Similarly, source intensity estimations indicate that this approach can track changes
in signal amplitude over time. Together, these results suggest that MUSIC can be used to characterize
neuronal signal sources in vivo.
1
Introduction
Multi-sensor microelectrodes, such as tetrodes, are increasingly being used for the extracellular recording of
action potentials (APs). In addition to improving AP detection and classification by increasing the signal-
to-noise ratio (SNR) [1], multi-sensor probes can be used to localize and characterize single neurons [2, 3, 4].
This function has, however, been largely unexamined and remains underutilized in experimental neuroscience.
As more studies are performed using these probes, the need to localize and characterize neurons based on
recorded APs is becoming increasingly important in deducing neuronal function.
The prospect of localizing neural signal sources has specific benefits for both acute and chronic extracel-
lular recoding experiments. Electrode positioning and guidance during acute recording is a very involved,
time consuming process. Furthermore, electrode micromanipulators have only one degree of freedom, and
data is often lost as neurons migrate away from the probe track.
If these migration trends could be es-
timated over a short period of time, experimentalists could determine if a neuron is a good candidate for
further analysis or likely to travel away from the probe tack. Additionally, autonomous algorithms for single
neuron isolation and tracking [5, 6, 7] could be greatly improved with neuron location data.
In chronic
recording, localization can be used to track the migration trends, as well as firing patterns of neurons and
neuronal populations. This could have significant implications for the study of neural plasticity [8], cell
assembly functional organization [9], scar tissue formation [10], as well as neural network connectivity and
communication [11].
The intensity of the extracellular potential field created when a neuron fires an AP is another important
neuronal characteristic that is often overlooked due to inadequate data processing techniques. As large
1
neurons emit a stronger signal, they are often over-represented in population studies [12]. Currently it is not
possible to distinguish a small nearby neuron from a large distant one. The ability to determine a neuron's
transmembrane current intensity as well as location may lead to better understanding of size-related neural
functional properties.
This article presents a statistical signal processing technique suitable for localization and current intensity
estimation of single neuronal sources, based on multi-sensor measurements of their extracellular APs. To
prove the validity of our technique beyond in silico models, a simple experiment was performed testing the
algorithm's ability to localize and estimate the strength of an artificially generated monopole-like source.
Stimulating signals were generated at various strengths and recorded with a commercial tetrode. The success
of our technique makes it a good candidate for further in vitro testing and in vivo use.
2 Preliminary Work
Source localization is an important problem in experimental neuroscience, and several methods have been
developed to localize source neurons during extracellular recording [4, 11, 13, 14, 15]. However, most of these
methods [11, 13, 15] are not capable of localizing sources in three-dimensional (3D) space. Furthermore, as
these were in vivo studies, none of these methods were fully validated. Our preliminary work has shown
that a closed-form solution to the neuronal source localization problem can be found given a 3D sensor
arrangement, and assuming a monopole forward model [16]. This method, while relatively tractable, is
sensitive to noise, and therefore, a solution may not always exist. Additionally, the geometry of the problem
always leads to two solutions. Most often, the spurious solution lies inside the sensor array and can be
easily identified and discarded. However, sometimes the two solutions cannot be disambiguated. Due to
these limitations another method relying on statistical signal processing was considered and tested in silico
against the closed-form solution. This statistical signal processing method relies on the Multiple Signal
Classification (MUSIC) algorithm [17], which has been previously used in electroencephalogram [18, 19]
and magnetoencephalogram [20] source localization. MUSIC therefore presents a promising technique for
neuronal source localization.
To underscore the differences between the two methods, we compared closed-form and MUSIC-derived
solutions, based on a computational model of a neuron [21]. Using this model, three differently shaped
APs of a bursting neuron, labelled A, B and C in Fig. 1a, were generated at four locations arranged to
mimic a commercially available tetrode (Thomas Recording, Geissen, Germany). These signals were used
for localization. The closed-form as well as MUSIC-derived localization results are shown in Fig. 1b. The
closed-form approach uses a single amplitude measurement, marked in green on the three APs in Fig. 1a,
to estimate source locations, plotted as green squares in Fig. 1b. Contrarily, the MUSIC algorithm takes
as input a time series describing the shape of the AP, as outlined in red in Fig. 1a. The corresponding
calculated source locations are plotted as red circles in Fig. 1b. The MUSIC algorithm, as seen in Fig. 1,
better estimated the location of the simulated neuron, and was therefore chosen for further validation in
vitro.
3 Materials and Methods
As an action potential propagates, positive sodium ions enter the neuron through voltage gated ion channels.
This process starts near the soma, and incrementally travels down the axon and up the dendritic tree,
causing the neuron to polarize. The neuron then re-establishes equilibrium by actively pumping out the
sodium ions. In this way, the neural membrane can be described as a distributed, dynamic network of ion
current sources and sinks. However, this multiple source and sink model of the neural membrane proves to be
overly complex for localization purposes, prompting the need for a simpler approximation. In the most basic
case, we can treat the neuron as a point source and the surrounding medium as an isotropic, homogeneous
volume conductor. Although simplistic, the monopole model has been used in application to neural source
localization [13, 11, 3, 14]. The potential at a specific sensor Si is then
ψi =
I(t)
4πσdi(r)
2
(1)
(a)
(b)
Figure 1: Comparison of closed-form and MUSIC localization solutions. (a) Three differently shaped APs,
labelled A, B and C, were generated for all four sensors of a simulated tetrode, and used to localize the
source. When calculating source locations, the closed-form approach uses only one data point from the
AP peak, marked in green, while MUSIC takes as input a time series describing the AP, marked in red.
(b) Localization results for each AP, labelled A, B, and C, plotted with the model neuron. The simulated
tetrode and neuron soma are potted respectively as a silver cone with sensors S1−4, and a purple sphere.
Closed-form and MUSIC derived localization solutions are shown as green and red points respectively.
where I(t) is the time varying source current, σ is the conductivity of the medium per unit length, and
di is the distance between the sensor Si and the source located at r = [x, y, z]T. We will use this as the
basis of our forward model. However, given that a homogeneous volume conductor is difficult to generate
experimentally, the behavior of the model when this assumption is violated will also be addressed.
3.1 Multiple Signal Classification Algorithm
If signals are generated by a single source, the MUSIC algorithm models measurements from a c-sensor array,
ψ(t) ∈ Rc×1, as an output of the static linear system
(2)
where t is the time instant, m ∈ Rc×1 is the lead field vector (LFV) [18, 19] representing the system's
response to a unitary signal input, s(t) ∈ R is the signal amplitude, and w(t) ∈ Rc×1 is zero-mean noise. In
the case of a single monopole-like source with current I(t), the LFV becomes
ψ(t) = ms(t) + w(t)
(cid:20) 1
m(r) =
1
(cid:21)T
4πσ
d1(r)
d2(r)
1
···
1
dc(r)
(3)
and s(t) = I(t). Eq. (2) then takes the form of the initial forward model with added zero mean noise. The
components of ψ(t) for any sensor i ∈ [1, 2,··· , c] are then
The MUSIC algorithm proceeds by finding the source location r(cid:63) for which the LFV is most orthogonal to
the noise subspace [17]. More formally, the optimal source location r(cid:63) is found by
ψi =
I(t)
4πσdi(r)
+ wi(t)
(4)
r(cid:63) = arg min
r
mT(r)EN ET
N m(r)
mT(r)m(r)
3
(5)
where EN ∈ Rc×(c−1) is the noise subspace. This subspace can be obtained by the following singular value
decomposition,
(6)
where Ψ := [ψ(1) ψ(2) ··· ψ(T )] ∈ Rc×T and T is the number of samples in the time series data. If T ≥ c
under the single-source assumption, the noise subspace can be defined as EN := [u(2) u(3) ··· u(c)], where
u represents the columns of U corresponding to the c − 1 smallest singular values of Ψ. In other words, we
assume that the first singular value of Ψ makes up the signal subspace, and the remaining values make up
the noise subspace EN . Note that σ cancels out in (5) so localization is independent of medium conductivity.
Ψ = U SV T
3.2 Inhomogeneity Correction Factor
In general, the conducting medium may not be homogeneous and isotropic, which implies that σ in Eq. (3)
may not be constant. Allowing each source-sensor path to have its own conductivity value redefines the LFV
as
¯m(r) =
(7)
where ki, i ∈ [1, 2,··· , c], is a constant making each conductivity a multiple of some baseline value σ. To
determine ki, referred to as the inhomogeneity correction factor (ICF), we set the conductivity of the sensor
closest to the source, as this baseline, σi(cid:63) = σ and therefore ki(cid:63) = 1. After combining Eqs. (2) and (7), and
taking the expectation over the zero mean noise distribution, we obtain
k1d1(r)
k2d2(r)
kcdc(r)
1
1
···
1
(cid:20)
1
4πσ
(cid:21)T
E{ψi(t)} =
I(t)
4πkiσdi
i = 1, 2, ··· , c
,
which can be used to solve for ki [22]
ki =
di(cid:63) E{ψi(cid:63) (t)}
diE{ψi(t)
,
i = 1, 2, ··· , c
(i (cid:54)= i(cid:63))
The forward model (2) can then be rewritten as
¯ψ(t) = m(r)s(t) + ¯w(t)
(8)
(9)
(10)
where the components of ¯ψ and ¯w are ¯ψi(t) = kiψ(t), and ¯wi(t) = kiw(t) respectively. The MUSIC algorithm
[Eqs. (2-6)] can then be executed with ¯ψ instead of ψ.
3.3 Accuracy and Precision
The accuracy of an estimate is measured by the difference between the estimated value of the parameter and
its true value . In the case of 3D source localization, this difference, also referred to as an error or bias, is
defined as
(11)
where (cid:107).(cid:107) represents the Euclidean norm, and re, rt ∈ R3×1 are the locations of the estimated and true
source, respectively.
ε := (cid:107)re − rt(cid:107)
The precision of an estimate is measured by its spread (standard deviation) around its mean value . To
generalize this notion to multivariate parameters, we define a standard radius as
(cid:113)
δ :=
δ2
x + δ2
y + δ2
z
(12)
where δx, δy and δz are the standard deviations of the x, y and z components of re, respectively.
4
3.4 Experimental Setup
To test the performance of our method on experimental data, a metal microelectrode (Alpha Omega Co.
USA, Alpharetta, GA) was connected to an AC voltage source (MP150, Biopac, Goleta, CA), with a circular
reference electrode made out of 0.20 mm2 wire. The microelectrode, which served as a stimulator, was placed
in the center of a Petri dish filled with 0.9% (by volume) saline solution (Fig. 2). The reference electrode
was placed against the Petri dish wall along its circumference. Due to radial symmetry, this electrode
arrangement produces a monopole-like electric potential field. The voltage source was then programmed
to produce a 7-Hz sine wave at amplitudes of 0.5, 0.7, 1.0, 1.2, and 1.5 V. This stimulation frequency was
chosen because it is not a subharmonic of power line noise (60 Hz), thus facilitating unambiguous signal
detection. The amplitude of the stimulus was varied to additionally test the algorithm's ability to estimate
the intensity of the source signal. Field potentials were then measured by bringing a tetrode (Thomas
Recording, Geissen, Germany) in close proximity to the stimulator. The tetrode was positioned under
microscope guidance (Olympus IX51, Olympus America, Center Valley, PA) using a combination of course
and fine movements performed by a micromanipulator (Narishige International USA, East Meadow, NY)
and a motorized microdrive (MiniMatrix, Thomas Recording, Geissen, Germany) respectively. The position
was deemed satisfactory once the tetrode tip was in the same microscope focal plane as the stimulator tip.
For each stimulation amplitude, 100 cycles of the sine wave were generated, and then collected, by the
tetrode. The setup was connected to an optically isolated data acquisition system (RX7, Tucker-Davis
Technologies, Alachua, FL). Signals were sampled at 25 KHz, digitized at 16-bit resolution, low-pass filtered
with a cut-off frequency of 3000 Hz, and notch filtered to eliminate 60 Hz noise. After recording, the collected
100 cycles were aligned and averaged for each stimulating amplitude. All experiments were performed at
room temperature (20◦C).
(Left) A stimulator (metal microelectrode) connected to a
Figure 2: Schematic of experimental setup.
voltage source with a circular reference. The stimulator is placed in a Petri dish with 0.9% saline solution.
A tetrode is placed in close proximity to the stimulator. (Right) A top view of the setup with the assumed
electric potential field.
3.4.1 Virtual Sensor Arrays
Measurements were taken at five different tetrode positions, as shown in Fig. 3. The tetrode position was
changed using a coarse movement micromanipulator, while keeping the tetrode tip and the stimulator in
the same focal plane. This approach was necessary in order to precisely estimate the position of the sensors
with respect to the stimulator. Note that based on microscope images, only the location of the tip sensor
5
can be determined unambiguously, while the other three sensors are in general not visible. The MUSIC
algorithm can be used to determine the relative locations of the remaining three sensors, based on knowledge
of the tip sensor location and the manufacturer's specifications (Sec. 5.3). However, this would confound
the validation process. To make sure all sensor locations are precisely identified, a virtual sensor array
was constructed. The five tetrode-position setup allows for multiple sensor arrangements to be considered.
Combining measurements from four out of five tip sensors defines a tetrode-like arrangement and will be
referred to as a virtual tetrode (VT). Specifically, the arrays consisting of tip sensors 1, 2, 4, and 5, and
2, 3, 4, and 5, will be referred to as VT1 and VT2, respectively (Fig. 3). Similarly, to ascertain whether
performance is improved by the addition of sensors, data from all five tip sensors was combined into a virtual
pentode (VP) array. The distances between the stimulator tip and the tetrode tips at positions 1 - 5 were
163.6, 94.6, 95.5, 38.3, and 81.2 µm, respectively. This experimental setup ensures that all "ground truth"
parameters are precisely estimated, including the relative locations of the stimulator and the sensors in each
virtual array, allowing the MUSIC algorithm's performance to be properly assessed. Note that since only
one tetrode was used for recording, multi-sensor VT and VP data could not be acquired simultaneously.
4 Results
The MUSIC algorithm was used to estimate source locations given collected signals. The localization so-
lutions were then used to estimate source current amplitudes. MUSIC source localization results showed
the algorithm's ability to accurately and precisely predict the true source location. However, analysis also
indicated that the common assumption of a homogeneous medium did not hold, even in the simple saline
medium used here. Current amplitude calculations proved that our method was also capable of accurately
tracking changes in source intensity. These results imply our algorithm's ability to resolve small nearby
neurons form large distant ones, which may help uncover size-specific neuronal functions.
4.1 Source Localization
4.1.1 Virtual Tetrode
The performance of the algorithm was first tested based on data from two virtual tetrodes, VT1 and VT2.
These two configurations shared data from tip sensors 2, 4, and 5. However, due to its use of sensor 3 as
opposed to sensor 1, VT2 was relatively closer to the stimulator than VT1 (Fig. 3).
Aligned and averaged data was used as input to the MUSIC algorithm for each stimulation amplitude,
yielding estimated source locations. The stimulator tip S was defined as the origin. Table 1 shows these
parameters, as well as the average location of the estimated sources for both VT1 and VT2. For VT1, the
estimated source error ranged between ∼19 µm and ∼25 µm, with the average estimated source location
(blue circle in Fig. 3) being 21.52 µm from the source. The error did not appear to correlate with the
stimulation amplitude (ρ = −0.16, p-value = 0.81). For VT2, the estimated source error ranged between
∼15 µm and ∼44 µm, with the average estimated source location (yellow circle in Fig. 3) being 18.40 µm
from the source. In this case, the error was inversely correlated with the stimulation amplitude (ρ = −0.97,
p-value = 0.0063). Another difference between the two architectures was a lower precision for VT2 (δ = 24.86
µm) than VT1 (δ = 4.10 µm). While the performance of the two virtual tetrodes was similar, as can be seen
by the overlapping estimated source locations in Fig. 3, VT1 was slightly less accurate, but substantially
more precise, than VT2. However, in both cases, the estimates suffered from a relatively large bias.
The presence of the bias indicates that the forward model defined by Eqs. (2) and (3) is inconsistent
with the measurements. The cause of this discrepancy is elaborated on in Section 5.4.
In the simplest
scenario, this discrepancy can be mitigated by relaxing the constant conductivity constraint, as outlined
in Section 3.2. According to this procedure, the ICF of the sensor closest to the source, tip sensor 4, was
chosen as k4 = 1. Note that this choice is arbitrary, as conductivities are expressed relative to one another.
While the ICF is defined as a time-dependent quantity [Eq. (9)], its values are remarkably constant except
at instances when E{ψi(t)} is near zero or passes through zero [22]. These outliers can be removed by
taking the median value of ki over time. Following this procedure, the remaining ICFs were calculated as:
k1 = 0.92, k2 = 1.05, k3 = 0.78, k5 = 0.75. These values can be interpreted as relative conductivities; the
conductivity of tip sensors 2 was higher than that of tip sensor 4, while those of tip sensors 1, 3 and 5 were
6
Figure 3: Superposition of five microscope images showing five tetrode positions (1 - 5) with tip positions
marked by green circles. The stimulating electrode is on the left with the source, S, defined at the tip and
marked by a pink circle. The tetrode and stimulator were placed in the same focal plane for each depicted
position. MUSIC estimated source locations, averaged over 5 stimulus amplitudes, are shown as blue (VT1)
and yellow (VT2) circles. Blue and yellow squares show the same parameters estimated after accounting for
medium inhomogeneity.
Table 1: Estimated source locations based on data acquired for VT1 and VT2 at different stimulation
amplitudes. Estimated source location error, ε, is defined in Eq. (11), with the true source located at the
origin, rt := [0, 0, 0]T. The last two columns show the average estimated source location and its standard
deviation, respectively, across all stimulation amplitudes. The standard radius, δ, was calculated using
Eq. (12).
Stimulus amp. (V)
0.5
0.7
1.0
1.2
1.5
Ave.
Std. dev.
VT1
VT2
x (µm)
y (µm)
z (µm)
-20.04
0.85
0.00
-22.10
0.82
0.00
-25.11
0.10
0.00
-22.45
-1.03
0.00
-17.68
-0.56
-6.48
-21.48
0.04
-1.29
2.78
0.83
2.90
ε (µm)
20.06
22.12
25.11
22.47
18.84
21.52
δ = 4.10
x (µm)
y (µm)
z (µm)
-17.21
17.85
-36.38
-21.12
6.45
22.86
-24.57
-1.14
0.00
-16.74
-10.91
0.00
-11.33
-9.21
0.00
-18.19
0.61
-2.70
4.99
11.87
21.27
ε (µm)
44.03
31.78
24.59
19.98
14.59
18.40
δ = 24.86
7
lower. The collected signals were then corrected by their corresponding ICFs (Sec. 3.2), and MUSIC source
localization was repeated using the corrected signals.
The performance of the MUSIC algorithm improved significantly after scaling signals to correct for
medium inhomogeneity (Table 2). For VT1, the range of estimated source errors was reduced from ∼[19, 25]
µm to ∼[1, 25] µm. In addition, a right-tailed t-test indicates that the mean error before signal correction
was significantly higher than the mean error after signal correction (p-value = 0.0163). As in the before-
correction case, there was no significant correlation between the error and the stimulation amplitude. After
ICF correction, the average estimated source location (blue square in Fig. 3) for VT1 was only 9.23 µm
away from the source (down from 21.51 µm), with a standard radius of 10.70 µm (up from 4.10 µm). The
improvement for VT2 was even more significant, as the error range was reduced from ∼[15, 44] µm to ∼[1, 4]
µm. A right-tailed t-test confirmed that the reduction in error upon ICF signal correction was significant
(p-value = 0.00058). In addition, no significant correlation was found between the error and the stimulation
amplitude. The average estimated source location (yellow square in Fig. 3) was only 1.30 µm away from the
source (down from 18.40 µm), with a standard radius of 1.87 µm (down from 24.86 µm).
Table 2: Estimated source locations based on VT1 and VT2 data after correcting for inhomogeneity. The
table setup is identical to that of Table 1.
Stimulus amp. (V)
VT1
VT2
0.5
0.15
2.14
0.00
0.7
3.24
10.94
22.57
1.0
-1.01
0.25
0.00
1.2
0.21
1.08
6.31
1.5 Ave.
Std. dev.
-0.96
3.78
13.50
0.33
3.64
8.48
1.73
4.29
9.65
x (µm)
y (µm)
z (µm)
ε (µm)
2.15
25.29
1.04
6.41
14.05
9.23
δ = 10.70
x (µm)
y (µm)
z (µm)
0.53
0.85
0.00
ε (µm)
1.91
3.28
0.82
0.00
4.14
-0.78
0.10
0.00
0.82
-1.03
0.00
0.47
-0.56
0.00
0.79
0.84
0.54
0.86
0.97
0.00
1.30
1.48
1.13
0.00
δ = 1.87
In summary, correcting for inhomogeneity significantly improved the performance of the MUSIC al-
gorithm, with both average estimated source locations being less than 10 µm away from the true source
location. The solutions were also very precise (δ < 11 µm). As a typical neuron soma is 10-50 µm in
diameter [12], these results are promising for neural source localization applications. Specifically the method
may be capable of resolving the locations of neighboring neurons.
4.1.2 Virtual Pentode
To investigate whether the addition of more sensors significantly impacts the localization results, a virtual
array of all five sensors, the VP, was also adopted for analysis (Fig. 4). Table 3 depicts all localization results
for the VP. Given raw signals, estimated source location errors ranged from ∼14 µm to ∼25 µm, and were
very weakly, inversely correlated with the stimulation amplitude (ρ = −0.58, p-value = 0.30). The average
estimated source location was 19.38 µm from the source (blue circle, B in Fig. 4) with a standard radius of
7.37 µm (VP Before in Table 3).
This result has a relatively large bias, which can be attributed to medium inhomogeneity. As in Sec. 4.1.1,
the signals were therefore corrected using derived ICF values. The resulting range of estimated source errors
was reduced from ∼[14, 25] µm to only ∼[1, 4] µm. A right-tailed t-test indicated that the mean error
after signal correction was significantly reduced from the mean error before correction (p-value = 0.000004).
Likewise, after ICF correction the errors' inverse correlation with the stimulation amplitude became slightly
more pronounced (ρ = −0.68, p-value = 0.21), but still relatively weak. The average estimated source
location, given corrected signals, was only 1.30 µm from the true source (red circle, A in Fig. 4), with a
standard radius of 1.88 µm (VP After in Table 3), as compared with an error and standard radius of 19.38
µm and 7.37 µm, respectively, before correction.
8
Figure 4: Superposition of five microscope images showing five tetrode positions (1 - 5) with tip positions
marked by green circles. On the left is the stimulating electrode with the source, S, defined at the tip, and
marked by a pink circle. The tetrode and stimulator were placed in the same focal plane for each depicted
position. The blue and red circles represent VP MUSIC estimated source locations, averaged over 5 stimulus
amplitudes, before (B) and after (A) ICF signal correction. The source location become more accurate after
ICF correction.
Table 3: Estimated source locations based on data acquired for the VP at different stimulation amplitudes.
The top row depicts results given raw data (VP Before), while the bottom row depicts results after signal
correction for medium inhomogeneity (VP After).
Stimulus amp. (V)
0.5
0.7
1.0
1.2
1.5
Ave.
Std. dev.
VP Before
VP After
x (µm)
y (µm)
z (µm)
-20.07
0.82
0.00
-22.03
0.73
0.00
-24.62
-0.98
0.00
-16.93
-10.21
0.00
-11.50
-8.70
0.00
-19.03
-3.67
0.00
5.06
5.36
0.00
ε (µm)
20.09
22.04
24.64
19.77
14.42
19.38
δ = 7.37
x (µm)
y (µm)
z (µm)
ε (µm)
0.48
1.85
0.00
1.91
3.35
2.46
0.00
4.16
-0.80
0.12
0.00
0.81
0.80
0.17
0.00
0.82
0.43
0.30
0.00
0.53
0.85
0.98
0.00
1.30
1.52
1.10
0.00
δ = 1.88
9
Overall, it appears that using an additional sensor does not necessarily imply an improvement in local-
ization results. The VP performed better than VT1, but comparably to VT2.
4.2 Source Current Amplitude Estimation
I(t) = E{ψi(t)} × 4πkiσdi,
Source current amplitude was estimated given ICF corrected signals and the corresponding VP localization
results. Rearranging Eq. (8) to solve for I(t), given the five sensors, yields
i = 1, 2, ··· , 5
(13)
where di represents the distance between the estimated source location and sensor i, and σ = 1.5 × 10−6
S/µm, based on the ionic concentration of the medium. Although I(t) can be calculated separately for each
sensor i, these value should be consistent as there is only one true source. Likewise, although a time varying
quantity, the current was only calculated given the peak value of E{ψi(t)}. The resulting I∗ then reflects
the hypothesized peak current value, and was calculated for each stimulation amplitude. The average I∗
over the five tetrode tip positions is shown as a function of stimulation amplitude in Fig. 5a. The estimated
current was quadratically related to the stimulating amplitude, y = 0.53x2 +0.48x−0.25 (χ2 = 0.00001).The
same relationship, up to three significant figures, was also observed when I∗ was calculated using either VT1
or VT2 localization results. As shown in Fig. 5a, I∗ increased with stimulating amplitude, while the source
location remained constant. Likewise, the precision of estimated currents, ranging from ±0.03 µA to ±0.29
µA (error bars in Fig. 5a) indicated that changes in source current can be detected. The algorithm's ability
to estimate both source current and location, suggests it may be able to resolve changes in extracellular
signal amplitude due to neuronal migration from changes in extracellular signal amplitude due to internal
firing variations.
(a)
(b)
Figure 5: (a) Estimated current amplitude (µA) as a function of stimulation amplitude (V). Each data
point is an average of estimated currents across five tetrode tip positions. The data was fit to a polynomial:
y = 0.53x2 + 0.48x − 0.25 (χ2 = 0.00001). (b) Recorded signal amplitude (µV) as a function of stimulation
amplitude (V). Note that the signal amplitudes for each sensor also follow a quadratic trend, and has a
negative y-intercept. This agrees with the estimated current results in (a), and implies that the phenomenon
is not caused by our method but rather by a signal transformation occurring between the stimulator, medium,
and sensors.
Due to nonlinear interactions between the stimulating electrode and the medium, we cannot compare the
estimated current at each stimulation amplitude to the true current being generated in the medium. However,
10
a rough calculation using Ohm's Law, V = IR, indicates that given our stimulating amplitudes, and the
corresponding estimated currents, the resistance of our setup is on the order of 1 MΩ. This value agrees with
the manufacturer specified impedance of our stimulating electrode, and is therefore a promising indicator
that the estimated current amplitudes are accurate. It is expected that given no stimulation the resulting
current should be zero, however the model in Fig. 5a indicates a negative y-intercept. When working with a
dielectric solution such as saline, the response of the system to low stimulation is highly nonlinear [23], and
the model can be expected to reflect this behavior given more data approaching zero. Additionally, note that
a similar relationship was also observed when comparing the collected signal amplitude to the stimulation
amplitude (Fig. 5b). This indicates that the observed behavior is not a consequence of our method, but
rather a ramification of the signal transformation occurring at the stimulator, medium, and sensor interfaces.
In general, our method accurately captures the behavior of the system, which is nonlinear especially at small
stimulating amplitudes.
5 Discussion
5.1 MUSIC Source Identification
According to the presented results, the MUSIC algorithm can accurately localize a source given a set of
signals from four or more sensors. The estimated source locations proved to be accurate, less than 10 µm
from the true source. As the diameter of a typical neuron soma is 10 - 50 µm [12], these results imply
that the MUSIC algorithm has the ability to successfully localize neurons. The precision of the solutions,
with standard radii ranging from 1.87 µm to 10.70 µm, suggests the MUSIC algorithm may also be able
to resolve the location of several closely spaced neurons. Furthermore, the algorithm's ability to accurately
localize a source given five different stimulation amplitudes, indicates localization consistency even if the
underlying signal changes. The proportional relationship between estimated source current amplitude and
stimulation amplitude also showed that this method can accurately track changes in source amplitude. Note
that the algorithm only relies on the relative signal amplitudes across sensors. Its performance will therefore
not be affected given signals on the nA range typically observed in vivo [24], as opposed to the µA range
presented here. These results imply that the MUSIC algorithm can disambiguate a stationary neuron with
varying firing amplitude from a migrating neuron with a steady firing pattern. Additionally, MUSIC may be
able to resolve large distant neurons from small ones nearby. Our algorithm may therefore have significant
implications for both acute and chronic electrophysiological studies.
5.2 Sensor Array Design Implications
Virtual sensor array experiments indicated that increasing the number of sensors does not necessarily result
in improved localization. The VP performed better than VT1, but comparably to VT2. As data from three
sensors was shared between the VTs, the discrepancy between results from VT1 (average error = 9.23 µm)
and VT2 (average error = 1.30 µm) can be attributed to their use of the sensors 1 and 3, respectively (Fig. 3).
The fact that VT2 performed significantly better than VT1, indicates that data from sensor 1 is inferior
relative to the other four sensors. This could be due to experimental error, the sensor and stimulator's
failure to be in the same focal plane, or the sensor's relatively large distance from the source. These results
are not surprising given that the minimum number of sensors necessary to perform source localization is
four, therefore we can expect the algorithm to be sensitive to outlying data in a tetrode arrangement. The
VP, however, poses an overdetermined localization problem by using all five sensors. The resulting location
estimate, with and average error of 1.30 µm, was relatively unaffected by the inclusion of sensor 1. This
implies that although a superior localization result is not guaranteed with a higher number of sensors, arrays
with a higher number of sensors are more robust against outlying data. With this in mind, microelectrodes
with more than four sensors may be preferable for use in localization studies.
5.3 Calibrating Sensor Locations
Although for the purposes of validation, exact locations for the three circumferential sensors of a tetrode were
not determined, the MUSIC algorithm can be used to establish these locations, as mentioned in Section 3.4.
11
A setup similar to that presented in this paper, with a visible stimulator and tetrode tip, both of whose
locations can be determined empirically, can be used to calibrate the remaining sensor locations. With
the manufacturer's specifications the relative locations of the three circumferential sensors can be easily
determined, the only unknown left being the tetrode's rotation about its own axis. MUSIC localization
can then be performed for incrementally increasing tetrode rotation angles. This will result in a family of
solution arranged in an oval, one segment of which will overlap with the true source location. The rotation
angle responsible for the estimated source location closest to, or overlapping with, the true source location
can then be accepted as the tetrode rotation angle. As the tetrode is not expected to change orientation
while mounted in a holder or stereotactic device, this arrangement can be adopted as accurate throughout
recording. We expect this kind of calibration to be performed prior to in vivo and in vitro recording, making
the constraint of precisely knowing all sensor locations not an issue for neurophysiological experiments.
5.4 ICF Signal Correction
Our results show that adjusting for medium inhomogeneity significantly improved localization results. The
inhomogeneity in our system is most likely coming from the presence of both the stimulating electrode and the
recording tetrode whose dimensions are not negligible. This causes the spherical field symmetry characteristic
of a monopole in a homogeneous isotropic medium to break. Furthermore the platinum-tungsten tetrode
sensors generate concentration overpotential when immersed in an electrolyte, which changes the distribution
of ions at the electrode-electrolyte interface. These sensors therefore behave as capacitors [25], further
violating the purely resistive monopole model presented in Eq. (1). As the assumption of a homogeneous
medium is commonly made in neurophysiological studies [26], the fact that this assumption does not hold,
even in a simple saline medium, indicates that it may also be a factor during in vivo and in vitro experiments.
Therefore, it may be worthwhile to explore adjusting existing techniques to take medium inhomogeneity into
account.
Although we believe medium inhomogeneity to be the main cause of the observed bias, ICF signal scaling
could have also accounted for medium anisotropy, or varying sensor impedances. The virtual sensor array
presented here used the tip sensor of a single electrode, making varying impedance an unlikely reason for
the observed bias. However, medium anisotropy cannot be ruled out by this experiment.
Lastly, ICF adjustment was applied for purposes of validation only. As ICFs are derived using the known
location of the source [Eq. (9)], this adjustment cannot be performed on blind in vitro or in vivo collected
data. We hypothesize that ICF signal correction can be made superfluous by prewhitening the signal prior
to analysis. However, even if prewhitening does not mitigate an observed bias, this is not strictly an issue for
neurophysiological experiments. As long as individual neuronal sources can be disambiguated, their location
bias relative to the recording electrode is irrelevant.
6 Conclusion
The results presented here aptly demonstrate the MUSIC algorithm's ability to both localize and track the
intensity of single sources. To the best of our knowledge this is the first experimental validation of electrical
source localization and intensity characterization using multi-sensor arrays. Our localization accuracy, < 10
µm from the source, is promising for future neuron localization studies in vitro and in vivo. Likewise, the
precision of estimated source locations, < 11 µm, indicates that MUSIC can be used to resolve adjacent
neuronal sources. Furthermore, estimated source intensity results imply that MUSIC can also be used to
track changes in source signal amplitude. These outcomes present our method as a good candidate for source
identification in both acute and chronic electrophysiological experiments, allowing much needed insight to
be gained on neural migration patterns and size-related neuronal functionality.
Acknowledgment
This study was partially supported by the National Science Foundation under Grant 1056105.
12
References
[1] C. M. Gray, P. E. Maldonado, M. Wilson, and B. McNaughton, "Tetrodes markedly improve the relia-
bility and yield of multiple single-unit isolation from multi-unit recordings in cat striate cortex," Journal
of Neuroscience Methods, vol. 63, pp. 43 -- 54, 1995.
[2] M. S. Jog, C. I. Connolly, Y. Kubota, D. R. Iyengar, L. Garrido, and R. Harlan, "Tetrode technology: ad-
vances in implantable hardware, neuroimaging, and data analysis techniques," Journal of Neuroscience
Methods, vol. 117, pp. 141 -- 152, 2002.
[3] D. Aur, C. I. Connolly, and M. S. Jog, "Computing spike directivity with tetrodes," Journal of Neuro-
science Methods, vol. 149, pp. 57 -- 63, 2005.
[4] F. Mechler, J. D. Victor, I. Ohiorhenuan, A. M. Schmid, and Q. Hu, "Three-dimensional localization of
neurons in cortical tetrode recordings," Journal of Neurophysiology, vol. 106, pp. 828 -- 848, 2011.
[5] Z. Nenadic and J. W. Burdick, "A control algorithm for autonomous optimization of extracellular
recordings," IEEE Transactions on Bio-medical Engineering, vol. 53, pp. 941 -- 955, 2006.
[6] J. G. Cham, E. A. Branchaud, Z. Nenadic, B. Greger, R. A. Andersen, and J. W. Burdick, "Semi-
chronic motorized microdrive and control algorithm for autonomously isolating and maintaining optimal
extracellular action potentials," Journal of Neurophysiology, vol. 93, pp. 570 -- 579, 2005.
[7] J. W. Burdick, J. G. Cham, Z. Nenadic, E. A. Branchaud, M. T. Wolf, and R. A. Andersen, "Prosthetic
device and methods and systems related thereto," 2012. US Patent 8,095,210.
[8] S. Kuboshima-Amemori and T. Sawaguchi, "Plasticity of the primate prefrontal cortex," The Neuro-
scientist, vol. 13, pp. 229 -- 240, 2007.
[9] K. D. Harris, J. Csicsvari, H. Hirase, and G. Dragoi, "Organization of cell assemblies in the hippocam-
pus," Letters to Nature, vol. 424, pp. 552 -- 556, 2003.
[10] V. S. Polikov, P. A. Tresco, and W. M. Reichert, "Response of brain tissue to chronically implanted
neural electrodes," Journal of Neuroscience Methods, vol. 148, pp. 1 -- 18, 2005.
[11] P. Barth´o, H. Hirase, L. Monconduit, M. Zugaro, K. D. Harris, and G. Buzs´aki, "Characterization of
neocortical principal cells and interneurons by network interactions and extracellular features," Journal
of Neurophysiology, vol. 92, pp. 600 -- 608, 2004.
[12] D. R. Humphrey, Extracellular, single-unit recording methods. Department of Physiology, Emory Uni-
versity School of Medicine, 1979.
[13] J. Csicsvari, D. A. Henze, B. Jamieson, K. D. Harris, A. Sirota, P. Barth´o, K. D. Wise, and G. Buzs´aki,
"Massively parallel recording of unit and local field potentials with silicon-based electrodes," Journal of
Neurophysiology, vol. 90, pp. 1314 -- 1323, 2003.
[14] M. I. Chelaru and M. S. Jog, "Spike source localization with tetrodes," Journal of Neuroscience Methods,
vol. 142, pp. 305 -- 315, 2005.
[15] Z. Somogyv´ari, L. Zal´anyi, I. Ulbert, and P. Erdi, "Model-based source localization of extracellular
action potentials," Journal of Neuroscience Methods, vol. 147, pp. 126 -- 137, 2005.
[16] C. W. Lee, H. Dang, and Z. Nenadic, "An efficient algorithm for current source localization with
tetrodes," in Proceedings of the 29th Annual International Conference of the IEEE EMBS, pp. 1282 --
1285, 2007.
[17] R. Schmidt, "Multiple emitter location and signal parameter estimation," IEEE Transactions on An-
tennas and Propagation, vol. 34, pp. 276 -- 280, 1986.
13
[18] S. C. Wu, A. L. Swindlehurst, P. T. Wang, and Z. Nenadic, "Efficient dipole parameter estimation
in EEG systems with near-ML performance," IEEE Transactions on Bio-medical Engineering, vol. 59,
pp. 1339 -- 1348, 2012.
[19] S. C. Wu, A. L. Swindlehurst, P. T. Wang, and Z. Nenadic, "Projection versus prewhitening for EEG
interference suppression," IEEE Transactions on Bio-medical Engineering, vol. 59, pp. 1329 -- 1338, 2012.
[20] J. C. Mosher and R. M. Leahy, "Recursive MUSIC: A framework for EEG and MEG source localization,"
IEEE Transactions on Bio-medical Engineering, vol. 45, pp. 1342 -- 1354, 1998.
[21] Z. F. Mainen and T. J. Sejnowski, "Influence of dendritic structure on firing pattern in model neocortical
neurons," Nature, vol. 382, pp. 363 -- 366, 1996.
[22] C. W. Lee, C. E. King, S. C. Wu, A. L. Swindlehurst, and Z. Nenadic, "Signal source localization with
tetrodes: experimental verification," in Proceedings of the 33rd Annual International Conference of the
IEEE EMBS, pp. 67 -- 70, 2011.
[23] B. Zhang, Y. Liu, S. Agarwal, M. Yeh, and H. E. Katz, "Structure, sodium ion role, and practical issues
for β-alumina as a high-k solution-processed gate layer for transparent and low-voltage electronics,"
Applied Materials and Interfaces, vol. 3, pp. 4254 -- 4261, 2011.
[24] J. J. B. Jack, D. Noble, and R. W. Tsien, Electric current flow in excitable cells. USA: Oxford University
Press, 1983.
[25] M. Neuman, Biopotential Electrodes, ch. Instrumentation, Application and Design, pp. 187 -- 190. John
Wiley and Sons, 1997.
[26] G. R. Holt and C. Koch, "Electrical interactions via the extracellular potential near cell bodies," Journal
of Computational Neuroscience, vol. 6, pp. 169 -- 184, 1999.
14
|
1304.0485 | 2 | 1304 | 2013-04-14T18:38:27 | Multi-scale community organization of the human structural connectome and its relationship with resting-state functional connectivity | [
"q-bio.NC"
] | The human connectome has been widely studied over the past decade. A principal finding is that it can be decomposed into communities of densely interconnected brain regions. This result, however, may be limited methodologically. Past studies have often used a flawed modularity measure in order to infer the connectome's community structure. Also, these studies relied on the intuition that community structure is best defined in terms of a network's static topology as opposed to a more dynamical definition. In this report we used the partition stability framework, which defines communities in terms of a Markov process (random walk), to infer the connectome's multi-scale community structure. Comparing the community structure to observed resting-state functional connectivity revealed communities across a broad range of dynamical scales that were closely related to functional connectivity. This result suggests a mapping between communities in structural networks, models of communication processes, and brain function. It further suggests that communication in the brain is not limited to a single characteristic scale, leading us to posit a heuristic for scale-selective communication in the cerebral cortex. | q-bio.NC | q-bio | Title: Multi-scale community organization of the human structural connectome and its
relationship with resting-state functional connectivity
Authors: Richard F. Betzel1,2, Alessandra Griffa3,4, Andrea Avena-Koenigsberger1,2, Joaquín
Goñi1, Jean-Phillippe Thiran3,4, Patric Hagmann3,4, Olaf Sporns1,2.
Affiliations:
1Department of Psychological and Brain Sciences – Indiana University, Bloomington, USA
2Program in Cognitive Science – Indiana University, Bloomington, USA
3Department of Radiology, University Hospital Center and University of Lausanne (CHUV) –
Lausanne, Switzerland
4Signal Processing Laboratory (LTS5), Ecole Polytechnique Fédérale de Lausanne (EPFL) –
Lausanne, Switzerland
Abstract – 176 words
Body – 6,135 words
Figures – 8
Tables – None
Supplement – community_assignments.txt, names.txt, roi.txt
1
Abstract
The human connectome has been widely studied over the past decade. A principal finding is that
it can be decomposed into communities of densely interconnected brain regions. This result,
however, may be limited methodologically. Past studies have often used a flawed modularity
measure in order to infer the connectome’s community structure. Also, these studies relied on the
intuition that community structure is best defined in terms of a network’s static topology as
opposed to a more dynamical definition. In this report we used the partition stability framework,
which defines communities in terms of a Markov process (random walk), to infer the
connectome’s multi-scale community structure. Comparing the community structure to observed
resting-state functional connectivity revealed communities across a broad range of dynamical
scales that were closely related to functional connectivity. This result suggests a mapping
between communities in structural networks, models of communication processes, and brain
function. It further suggests that communication in the brain is not limited to a single
characteristic scale, leading us to posit a heuristic for scale-selective communication in the
cerebral cortex.
Key words: connectome, community structure, dynamics, Markov process, resting-state
2
Introduction
Many complex networks exhibit community structure, defined for example by clustered edge
distributions such that vertices (nodes) in the same community preferentially link to one another
(Guimera & Amaral, 2005; Girvan & Newman, 2004; Newman, 2006). Examples of community
structure can be found in society as groups of friends, workplaces, cities, and states (Moody &
White, 2003; Freeman, 2004); in protein interaction networks as groups of co-functioning
proteins (Jonsson et al., 2006); and in the World Wide Web (WWW) as webpages sharing many
hyperlinks (Albert et al., 1999; Flake et al., 2002).
Detecting community structure is an important endeavor in network science. Though there exist
many methods for doing so, none has emerged as clearly preeminent (for a review, see
Fortunato, 2010). One of the most widely used methods is to identify the vertex partition that
maximizes the modularity quality function, defined as the difference between observed and
expected intra-community edge density (Newman & Girvan, 2004). Despite its widespread
usage, modularity suffers from several drawbacks: maximization is NP-hard and identifying the
optimal partition is often practically impossible (Fortunato, 2010); solutions are sometimes
degenerate, with multiple partitions corresponding to the maximum modularity (Good et al.,
2010); and a resolution limit renders modularity “blind” to communities below some
characteristic scale (Fortunato & Barthelemy, 2007).
The first two problems are ubiquitous to most community detection algorithms and are generally
unavoidable. The resolution limit is more specific to quality functions like modularity, but has
been remedied in several instances by the inclusion of a tunable resolution parameter, which
controls the scale at which communities are detected (Reichardt & Bornholdt, 2006; Arenas et
al., 2008; Ronhovde & Nussinov, 2009). One such example is the partition stability framework,
which defines communities in terms of a Markov process based on a random walk model
(Lambiotte et al, 2008; Delvenne et al., 2010; Lambiotte, 2010). As this process evolves, a
random walker makes progressively longer walks and explores more distant parts of the network.
Intuitively, communities can be thought of as groups of vertices that effectively “trap” the flow
of random walkers over a particular dynamical scale. In this way, the stability framework has
been dubbed a “zooming lens,” whereby focusing in on shorter or longer dynamical timescales
reveals communities of correspondingly smaller or larger diameter (Schaub et al., 2012).
The human connectome, i.e. the full set of neural elements and connections of the human brain,
can be modeled as a complex network (Sporns et al., 2005; Bullmore & Sporns, 2009). The
topological properties of this network have been studied for nearly a decade, revealing key
features including small-world architecture (Gong et al., 2009), hub regions and cores (Hagmann
et al., 2008), rich club organization (van den Heuvel & Sporns, 2011), modular architecture
(Chen et al., 2008; Meunier et al., 2010; Wu et al., 2011), and economical wiring (Bassett et al.,
2010; Bullmore & Sporns, 2012). While these results characterize and contextualize the
connectome among all complex networks, the role of these topological features in shaping
communication processes and dynamic couplings among brain regions remains an area of active
research (Honey et al., 2009; van den Heuvel et al., 2012, Haimovici et al., 2013).
3
Empirical and computational studies suggest that the human connectome underpins complex
neural dynamics and facilitates the communication and integration of information between brain
regions. In functional magnetic resonance imaging (fMRI), such “functional connectivity” is
reflected by the magnitude of statistical dependence, commonly measured as linear correlation,
between blood oxygen-level dependent (BOLD) signals recorded from different brain regions. A
growing body of literature describes patterns of resting-state functional connectivity (rsFC), i.e.
spontaneous, endogenous fluctuations of the BOLD signal in the absence of any explicit
cognitive task. This work has revealed consistent connectivity patterns, dubbed resting-state
networks (RSN), which resemble networks of regions (e.g. somato-motor, visual, default mode,
etc.) that are coherently engaged in various cognitive and behavioral domains (Greicius et al.,
2003; Damoiseaux et al., 2006; Fox & Raichle, 2007, Smith et al., 2009).
This article has two principal aims. The first aim consists of detecting and characterizing the
multi-scale community structure of the human connectome using the partition stability
framework. The second aim is to assess the relationship of such structure to empirically observed
rsFC. Where the first aim attempts to answer the question ‘what communities of brain regions
are poised for diffusion-like communication’, the second aim attempts to answer the question
‘what evidence is there that the previously-identified communities match observed patterns of
functional coupling in the brain and at which dynamical scales is the correspondence between
community structure and functional coupling most salient?’
Methods
Forty (40) healthy human volunteers (24 males and 16 females, 25.3±4.9 years old) underwent
an MRI session on a 3-Tesla scanner (Trio, Siemens Medical, Germany) with a 32-channel head-
coil. The session consisted of (i) a magnetization-prepared rapid acquisition gradient echo
(MPRAGE) sequence sensitive to white/grey matter contrast (1 mm in-plane resolution, 1.2 mm
slice thickness.), (ii) a diffusion spectrum imaging (DSI) sequence (128 diffusion weighted
volumes + 1 reference volume, maximum -value 8000 s/mm2, mm voxel
size), and (iii) a gradient echo EPI sequence sensitive to BOLD contrast (3.3 mm in-plane
resolution and slice thickness with a 0.3 mm gap, TR 1920 ms). During the fMRI acquisition,
subjects were lying in the scanner with eyes open, resting but awake and cognitively alert, thus
recording-resting-state fMRI (rs-fMRI).
DSI, rs-fMRI and MPRAGE data were processed using the Connectome Mapping Toolkit
(Daducci et al., 2012). Each participant’s gray and white matter compartments were segmented
from the MPRAGE volume. In the present study the entire cortical volume was subdivided into
1000 equally-sized regions of interest (Cammoun et al., 2012). Whole brain streamline
tractography was performed on reconstructed DSI data (Wedeen et al., 2008), and participant-
wise, right hemisphere connectivity matrices were estimated by selecting the streamlines
connecting each pair of 501 cortical regions in the right hemisphere. Connectivity strength was
quantified as fiber density, as described in (Hagmann et al., 2008). Each structural connectivity
(SC) matrix can be interpreted as the adjacency matrix of a graph , with vertices
corresponding to cortical regions of interest, and weighted, undirected edges
representing anatomical connections. This network is often referred to as a
4
structural connectome. It is convenient at this point to define a connectome’s vertex strength and
total weight as and
, respectively.
Functional data were pre-processed according to state of the art pipelines (Murphy et al., 2009;
Power et al., 2012), which included motion correction, white matter, cerebrospinal fluid (CSF),
global and movement signals regression, linear detrending, scrubbing and low-pass filtering.
Average signals were then computed for each cortical region. The functional connectivity
between each pair of regions was estimated as the Pearson correlation between the corresponding
average signals. A single, right hemisphere rsFC matrix representative of the whole group of
participants was computed by averaging the 40 individual correlation matrices. The group-mean
structural connectome (weighted and binary) and functional connectivity matrix for the right
hemisphere of cerebral cortex are shown in Figure 1.
One of the main aims of this article was to detect the connectome’s multi-scale community
structure using the partition stability framework. This framework defines communities in terms
of a Markov process – in this article, a continuous-time random walk. Such a process can also be
used to model the diffusion of information or energy over a network, and is defined by the
dynamical system:
∑
where is the probability of finding a random walker on vertex and is the
normalized graph Laplacian matrix. If the network is undirected and connected, this system
.
evolves to the equilibrium state
The stability framework detects community structure at different timescales of the random walk
by identifying the vertex partition that maximizes the quality function “stability”. Let
be a partition of into communities, such that and . The
stability of at a given Markov time is defined as:
]
( ) ∑ ∑ [( )
where the summation extends over all communities and the edges that fall within each
, is the probability that a random walker
community. The first term in the summation, ( )
is
starting in community will be in that community at Markov time . The second term
the probability that two independent random walkers will be in at equilibrium. The difference
in these terms represents the density of random walkers in a community in excess of what is
expected at equilibrium. Key to this framework, the stability measurement depends not only on
the partition , but also on Markov time . In general, different partitions will maximize stability
at different times in the random walk. Varying across timescales and maximizing stability
recovers the community structure at each scale.
The process of actually optimizing ( ) can be accomplished in a number of ways. One
appealing option, and the one used in this article, makes use of the relationship between partition
stability and a more widely used modularity measure. Lambiotte et al. (2008, 2011)
demonstrated that stability could be recast in terms of the modularity of a weighted, symmetric
5
]
( )
“flow graph.” A flow graph is a transformation of in which the dynamics of a Markov
process are embedded into the edges of a new graph. In the case of a continuous-time random
( ) ( ) whose elements are
walk, the flow graph is represented by the full matrix
proportional to the probabilistic flow of random walkers between vertices at time . Examples of
flow graphs for one of the participants are shown in Figure 2. The relationship between stability
and modularity is such that instead of directly maximizing ( ), one can equivalently
maximize the modularity of a flow graph evaluated at :
∑ ∑ [ ( )
This result confers the practical advantage that any heuristic previously used to maximize
modularity can now be repurposed in order to maximize stability.
To apply this set of principles to the connectomes obtained from 40 healthy individuals , a range
of Markov times had to be selected over which to maximize stability. After experimentation, it
was determined that at Markov times below , every participant’s community structure was
characterized by a partition into communities, i.e. every vertex was assigned to its own
community, and at times greater than every participant exhibited a two-community
division. Using these times as lower and upper boundaries, we selected 185 logarithmically-
spaced time points over the interval [ ] at which to maximize partition stability. This
( ) for each participant, evaluating that flow graph at
process entailed defining a flow graph
each of the pre-selected Markov times, and subsequently maximizing each flow graph’s
modularity (or equivalently, stability) by applying the Louvain algorithm 750 times (Blondel et
al., 2008). Then, at a given time point, a participant’s community structure was represented by
the partition corresponding to the maximum modularity (stability).
Rather than analyze the multi-scale community structure of each participant individually,
communities were aggregated across all participants to shift focus onto the group-average
community structure. This operation was summarized by the weighted, symmetric and time-
dependent agreement matrix ( ), whose elements indicated the percentage of all participants
in which, at time , assigned vertices and to the same community. Thus, the elements of
( ) were interpreted as the probability that two vertices belonged to the same community.
Values ranged from ‘0’ in the case of two vertices that never appeared in the same community, to
‘1’ for two vertices that always appeared in the same community across all participants.
Conceptually, ( ) reflected the extent to which participants’ community structures at each
Markov time coincided with one another.
The agreement matrix ( ) proffered a probabilistic description of the connectome’s
community structure at a range of Markov times. It was of practical interest to obtain at each
time point a single partition corresponding to the connectome’s consensus communities, i.e. the
communities that were common to the majority of participants. To obtain such community
structure, an iterative thresholding/clustering algorithm was applied to ( ) (Lancichinetti &
Fortunato, 2012). This procedure consisted of two steps: (i) All edges in ( ) below a threshold
were regarded as “noise” and were set to zero. Imposing such a threshold resulted in the
( ) whose remaining edges linked only those vertices assigned to the same community
matrix
6
( ) was then clustered 100 times using the Louvain
greater than half of the time; (ii)
algorithm. If the resulting partitions were identical, then the algorithm had reached consensus
and it terminated. Otherwise, a meta-agreement matrix was built from the new partitions and the
algorithm returned to (i). At every time step, then, this procedure found a consensus partition
( ) that captured only the features of community structure common to the majority of
participants and ignored the features unique to individuals. The ordered set of consensus
partitions was summarized as ( ) ( ) . It should be noted that, in
general, the process of obtaining a consensus partition resulted in a loss of information, removing
the features of community structure that were expressed infrequently in participants. While such
information might reveal meaningful differences between subjects, it was disregarded in all
further analyses, which focus on characterizing features of community structure common among
the cohort of participants.
A major aim of this article was to gain insight into how the connectome’s multi-scale community
structure related to brain function. To map this relationship, we compared the agreement matrix
( ) and the sequence of consensus partitions to the empirically-measured rsFC matrix .
First, as a measure of correspondence between rsFC and community structure, we computed the
Pearson correlation of the elements in the empirically-measured rsFC matrix with ( ) at
each Markov time. Larger correlation values implied that the propensity for two vertices to share
a community assignment was a good predictor of whether those same vertices were functionally
coupled.
As a second measure of correspondence, each consensus community’s “goodness” was
determined by imposing it upon the rsFC matrix . Conceptually, a “good” community was one
whose internal density of functional connections was greater than expected by chance. This
intuition of a community’s “goodness” was in line with the way communities are defined by the
modularity function. Therefore, a consensus community’s “goodness” with respect to rsFC was
estimated by computing its modularity score. Before doing so, we defined two additional
and
comprised only of ’s positive and negative edge weights, respectively.
matrices
were
and
, respectively. For a
The strength and total weight of
given consensus community ( ), its positive and negative modularity scores were
computed as:
∑ [
]
Large communities, because they consisted of many vertices, also tended to have large
modularity scores. To remove this bias, the procedure described above was repeated 5000 times
but with vertices randomly assigned to communities , each time resulting in a measurement
]) and variance ( [
. This score enabled estimates of the expected value ( [
])
of each community’s modularity score to be obtained. From these estimates, each community’s
score was standardized and expressed as a z-score:
[
(
]
[
])
7
was interpreted as an indicator of how well each community ( )
The sum
mapped onto a functional community. A large, positive indicated that a community’s
modularity was much greater than would be expected given its size.
In addition to identifying communities that were more modular than by chance, the scores
were useful for answering several important questions about how stability-derived communities
related to rsFC: (i) On average which cortical regions contributed the most standardized
modularity; (ii) which pairs of vertices and (iii) which pairs of cortical regions, when assigned to
the same community, portended a large standardized modularity score for that community.
From the set of standardized modularity scores, it was straightforward to compute the total
contribution of each cortical region by summing the scores of each vertex over all
communities in which participated and then aggregating these scores by cortical region.
Another important question was how the co-assignment of groups of vertices or cortical regions
to a given community influenced that community’s modularity. For example, does assigning
vertices and to the same community portend a higher or lower modularity for that
community? To identify such groups, a matrix was built and subsequently clustered. Initially,
the weights of were set to zero. The weights were updated by considering each community
and strengthening the connections among all of the vertices assigned to by an amount .
Thus, the weights of were equal to the sum of standardized modularity over all communities
in which vertices and both appeared. An agglomerative, hierarchical clustering algorithm
was then used to extract groups of vertices that collectively influenced a community’s
modularity. This algorithm treated each row in as a feature vector of the matching vertex.
Starting with every vertex in its own cluster, clusters were merged over a series of steps until
only two clusters remained. At each step, the relationship between every pair of clusters was
defined by the average Euclidean distance between the feature vectors of their respective
elements. The heuristic for merging clusters was to identify the two clusters whose distance was
smallest and to combine their elements, forming a larger joint cluster at the next step. The result
of this procedure was a hierarchical tree of related vertices. The tree can be thresholded,
revealing a finer or coarser clustering depending on the level of the threshold. At any level,
however, these clusters were interpreted as groups of vertices that collectively participated in
communities with large standardized modularity scores.
To identify pairs of cortical regions whose co-assignment contributed to a community’s having
small or large modularity, was down-sampled by aggregating its rows and columns according
to cortical region. The elements of the down-sampled matrix contained the sum of modularity
accounted for by communities in which each pair of cortical regions’ constituent vertices
appeared in together.
Results
The previous section described a procedure for identifying the connectome’s multiscale
community structure using the partition stability framework. The association between this
8
structure and observed functional connectivity was assessed by correlating it with rsFC and,
separately, by measuring how well consensus communities modularized the rsFC matrix.
Maximizing partition stability at 185 time points logarithmically–spaced over the range
[ ] generated a sequence of partitions for each participant. Each sequence began at
the shortest dynamical scale with the partition in which every vertex comprised its own
community. A division of the vertex set into two communities characterized the final partition in
each sequence, corresponding to the largest dynamical scale. From these partition sequences, a
number of statistics were computed at each time point, specifically, the mean and standard
deviation partition stability, number of communities, community size, and number of singleton
communities. Partition stability and the number of communities declined monotonically while
the size of communities increased. The number of singleton communities declined initially,
before the maximum number of non-singleton communities reached a maximum value at
Markov time (Figure 3).
An iterative thresholding/reclustering algorithm was used to generate from the participants’
partitions a sequence of consensus partitions, which represented the “backbone” community
structure. Examples of consensus partitions at selected Markov times are shown in Figure 4A-F.
Each partition represents a division of the cerebral cortex into communities of brain regions
poised for information exchange at the corresponding dynamical scale via diffusion dynamics.
(For a list of all vertices and their community assignments at each Markov time, see the
Supplementary Information).
To compare community structure to brain function, the Pearson correlation of the time-
dependent agreement matrix ( ) with the empirically-measured rsFC matrix was
computed, peaking at a value of at time (Figure 5). This level of
correlation is comparable to earlier studies reporting correspondence between the structural
connectome and rsFC, as well as correlations between functional connectivity generated in
computational models and empirical rsFC (Honey et al., 2009). An interesting observation was
that the peak correlation occurred at a Markov time where there were many communities made
up of many vertices ( communities with an average vertices per
community) whose stability ( ) had not yet begun to decay. Earlier Markov times had
slightly greater stability values but were characterized by having far more singleton
communities, while partitions at later Markov times had no singleton communities but were
marked by extremely low stability values. The consensus partition at this time consisted of
fourteen communities, whose topographical arrangement and composition are shown in Figure 6.
As a second means of relating community structure to observed functional connectivity, we
measured the extent to which consensus communities were also good functional communities.
This process consisted of estimating every consensus community’s standardized modularity
score. Over a range of Markov times from to , a number of communities
had much greater-than-expected modularity (Figure 7A). Mapping these scores onto brain
anatomy and summing across Markov times, it was observed that every cortical region
contributed positive modularity, though some contributed disproportionately more (Figure 7C).
The regions with the greatest modularity (both in terms of peak value and total contribution)
were found to be the precentral and postcentral cortex, the lateral occipital cortex, the superior
9
parietal cortex, and the superior frontal cortex. Other regions also had large values, including the
rostral middle-frontal cortex, the inferior parietal cortex, as well as the superior temporal,
lingual, and fusiform cortex.
The matrix was constructed to measure the pairs of vertices that collectively participated in
communities with greater-than-expected modularity (Figure 7B). This matrix was mapped to the
space of cortical regions by aggregating the values of vertex pairs that linked cortical regions.
This process revealed a number of areas of cortex that, together, participated in communities
with large modularity. (Figure 7D). The largest contributing pairs were precentral/postcentral,
lateral occipital/fusiform, supra-marginal/postcentral, and
rostral-middle-frontal/precentral
cortices.
At the vertex level, was clustered to reveal groups of vertices that collectively participated in
communities with greater-than-expected modularity (Figure 8A-F). Clustering produced a
hierarchical tree, which was cut at different levels to reveal greater or fewer clusters. At a coarse
scale, was decomposable into three spatially contiguous modules (Figure 8A): The first
module (orange) spanned most of the cortical midline and included superior frontal and
paracentral cortices as well as the precuneus; the second module (yellow) was composed of both
precentral and postcentral cortices as well as the rostral middle-frontal cortex; the third module
(blue) included superior temporal, lateral occipital, lingual, and inferior parietal cortices along
with the fusiform area.
Cutting the hierarchical tree at a lower level divided these macro-scale modules into smaller sub-
modules. The first module survived largely unchanged for the range of clusters shown in Figure
8, but did fragment slightly as parts of the superior frontal and precentral cortices were split off.
The second module was eventually halved into two sub-modules: The first of which was
composed primarily of sub-regions of the frontal and precentral cortices; the second module
consisted mostly of precentral, postcentral, and supramarginal cortex. The third module
underwent a series of splits into four smaller sub-modules (Figure 8D-F): the first sub-module
(green) contained primarily lingual gyrus, cuneus, and pericalcarine regions; the second (yellow)
was dominated by lateral occipital, fusiform, and superior parietal cortex; the third (cyan)
consisted of large portions of temporal cortex, as well as some lateral portions of the
orbitofrontal cortex and the insula; and the fourth (blue) was made up of mostly inferior parietal
cortex, but also the banks of the superior temporal sulcus along with other areas of the temporal
cortex. In principle, could be decomposed further until the number of modules was equal to
the number of vertices in the network.
Discussion
In this paper we used the partition stability framework to infer multi-scale community structure
in the human cerebral cortex. This procedure generated a series of communities over a range of
dynamical time scales, beginning with a large number of small communities (fine-scale) and
ending with a small number of large communities (coarse-scale). We compared communities to
brain function using two approaches: (i) We identified a scale at which the Pearson correlation of
an agreement matrix with empirically measured rsFC became maximal, with a peak correlation
10
of approximately ; (ii) We evaluated the modularity of each consensus community
when it was imposed on rsFC, and identified a number of communities that overlapped with
rsFC modules. These results suggest that a community’s position in both space and time, i.e.
where it is located physically/topologically as well as the range of Markov times over which the
community appears, provide complementary information for assessing its importance and
relevance to network function.
Before discussing these results, it is worth explicitly stating our views of stability maximization
and its relationship to the human connectome and communication processes in the brain.
Communities derived from maximizing partition stability can be interpreted in multiple ways.
The first interpretation regards such communities as being dynamically important – communities
reflect the structural properties of a network that bias the trajectory of a random walker exploring
the network. If the random walk is taken as a model of message passing or communication
among vertices, then stability-derived communities take on even more significance. In such a
case, communities represent groups of nodes
that more readily exchange messages
(communicate) with one another than with the rest of the network. On the other hand is the view
that stability maximization, despite its dynamical underpinnings, is simply a useful methodology
for identifying community structure across multiple scales; no special functional significance is
attributed to communities detected this way. In this discussion, we adopt a view more closely
aligned with the first interpretation. Our argument for doing so stems from our conceptualization
of the connectome as a communication network – brain regions at different scales signal to one
another, passing information from region to region over their anatomical connections.
Furthermore, we assert that it is the connection topology that prescribes a region’s preference to
dynamically link to another region, or for a group of regions to mutually couple. Hence, from our
point of view, identifying communities of vertices that are likely to communicate with one
another is of great practical interest and can illuminate true functional dependencies.
Our first major finding was that the correlation of the time-dependent agreement matrix, which
reflected the community ties between vertices at a given time point, with rsFC reached a clear
peak value. This result owed its significance to our interpretation of community structure as a
marker of vertices’ propensity to dynamically couple with one another. Given this point of view,
we can think of the peak correlation as the time point at which the empirical functional couplings
became most closely aligned with the communication preference of the structural network as a
whole. This result is significant, as the functional coupling of different brain regions has been
interpreted as integration of information (van den Heuvel & Hulshoff Pol, 2010). This result is
also surprising, as it suggests that an overwhelming proportion of vertices are “tuned” or
“poised” for communication at a specific scale.
In terms of the random walk dynamics, this scale is of particular relevance. The time at which
the peak correlation occurred coincided with a regime of Markov time characterized by partitions
associated with large stability values but also many non-singleton communities. At much earlier
Markov times, the average partition stability was slightly greater, but the number of singleton
communities vastly outnumbered the non-singleton communities, implying that very few vertices
communicate with one another. At later Markov times, stability tended toward zero, so that even
though the communities at those times were much fewer in number and made up of more
vertices, they were also more ill-defined than those at earlier times. This suggests that the peak
11
correlation occurs within a fairly narrow regime of Markov time in which the number and size of
communities strikes a balance with stability, giving rise to a diverse, well-defined repertoire of
communities.
The first result, however, only reveals part of the story. Linear correlation is a global measure of
correspondence between two variables. When we look at the relationship between community
structure and rsFC at a local scale, for instance at the level of individual communities, a different
picture emerges. Specifically, we found substantial variability in the timing of and extent to
which individual communities contributed to the modularization of the rsFC matrix. Certain
communities, even when the partition as a whole was only weakly related to rsFC, had
exceptionally large standardized modularity. At the time of the peak correlation, a large number
of communities collectively had substantial modularity, which likely contributed to the peak in
the correlation coefficient. This result, however, suggests that the relationship between rsFC and
community structure is not restricted to a single dynamical scale.
Together these two results paint an interesting picture of a possible communication strategy in
the brain. The peak correlation represents a “sweet spot” of communication, a timescale where a
large number of vertex pairs are simultaneously poised to take advantage of a diffusion process
to communicate information between one another. For any given community, however, the scale
at which that community becomes most important likely deviates, even if only slightly, from the
time of the peak correlation. This observation prompts the interesting hypothesis that brain
dynamics may use a tuning mechanism that grants priority to certain scales of communication
while suppressing communication at others. As an example, suppose that in order to execute
some cognitive function two regions of the brain need to relay information back and forth. It is
not unreasonable to suppose that these regions appear in the same community at some scales but
not at others. To facilitate communication between these regions, the tuning mechanism turns
from whatever dynamical scale it started in, to the one at which these regions become members
of the same community. By moving the regions closer together in this community space,
cognitive function is ultimately supported. This sort of selectivity of scale has been observed
empirically in the frequency-specific coupling of neuronal oscillators in cortical networks
(Buzsáki & Draguhn, 2004). In such networks, oscillators can become coupled over a range of
high (short-cycle) to low (long-cycle) frequencies. It was observed that coupling in the high-
frequency range tended to recruit oscillators from small, local neuronal pools. In contrast, low-
frequency oscillations recruited from and bound together neurons from broader pools and
supported the emergence of large-scale, spatially distributed cortical networks. Interestingly,
because the time series obtained from any given oscillator can simultaneously have power at
both high and low frequency, both short- and long-range coupling is possible. On the surface,
these results mirror our own: communities at small Markov times (analogous to high-frequency
oscillations) tend to have small diameter and correspond to local, spatially contiguous groups of
vertices; communities at long Markov times (analogous to low-frequency oscillations) are more
spatially distributed and are comprised of non-adjacent vertices whose relationships are more
often indirect.
This tuning hypothesis also has interesting implications related to the specificity of brain regions
for cognitive processes. When the brain tunes in to a particular dynamical scale, the communities
at that scale tend to trap the flow of information. Thus, each community can be regarded as a
12
complex of vertices which readily exchange information with one another but only sparingly
with vertices belonging to other communities. At a given scale, then, it may not be meaningful to
refer to an individual vertex’s role, but instead refer to the function of the community to which
the vertex belongs. This suggests that there may be cases where communication between only
small numbers of brain regions is necessary in order to perform some cognitive function, but that
the dynamical scale at which these regions communicate necessitates engaging the function of a
much larger community. Tuning in to such a dynamical scale would have the desired resul t of
heightened communication between the required regions, but also gives rise to incidental
communication between brain regions that are unnecessary for performing the cognitive function
in question.
A third interesting finding concerns the relationship of the observed community structure to
empirically observed resting state networks (RSN). The communities we observed are generally
spatially contiguous, due to the fact that the majority of structural connections identified so far in
the human connectome are short-distance and connect nearby regions to one another. Therefore
it is unsurprising that we did not observe distributed and spatially non-contiguous RSNs, e.g. the
right parietal-frontal and default mode network. However, we do observe highly modular
communities in occipital cortex, corresponding quite closely to the primary and extra-striate
visual RSNs (lingual gyrus, cuneus, lateral occipital, fusiform). We also observe large
modularity in communities involving areas that are typically associated with the somatomotor
network (pre- and postcentral cortex). Interestingly, both visual and somatomotor RSNs are
generally classified as unimodal networks and thought to comprise tightly coupled and
functionally related regions that jointly participate in sensorimotor processes (Bassett et al.,
2012; Sepulcre et al., 2012). While the community structures that we observed do not directly
correspond to the boundaries of these RSNs, their approximate resemblance is suggestive of a
relationship between the dynamic process used to identify communities and their strong
functional couplings observed in the course of endogenously driven neural activity. Future work
is needed to further clarify the nature of this relationship.
There are a number of limitations of this work that should also be discussed. We exploited a
random-walk dynamics to identify vertex communities. Random walks as a simple class of linear
dynamics are limited in the types of behaviors they can exhibit. Other classes of dynamical
systems, especially non-linear systems, can exhibit far more complex behaviors, including
deterministic chaos and parameter sensitivity (bifurcations), among others. We defend our
decision to use random walks on the grounds that the partition stability measure depends on this
choice of dynamics to analytically define community structure. Furthermore, the well-
documented behavior of a random walk drastically simplifies the analysis and interpretation of
results.
Even when we restrict ourselves to the random walk class of dynamics, we still have room to
complicate our model further with the addition of parameters that bias the random walk, altering
the transition preference or the rate at which random walkers leave vertices. These parameters
are defined on a vertex-wise basis and afford us means of artificially introducing heterogeneity
into the random walk dynamics (Lambiotte et al., 2011). In this article, these parameters were set
to fairly conservative values, i.e. a random walker’s transition was made without bias and the
rate at which transitions took place was equal for all vertices. Note that this parameterization is
13
the most naïve we could have selected – choosing to bias the random walk or to imbue certain
vertices with faster or slower transition rates would both require justification.
Lastly, our use of the partition stability framework differed in some aspects from the way in
which it has most frequently been applied. Typically, an important step in the framework is to
use several robustness tests to verify the measured community structure (Lambiotte, 2010;
Delmotte et al., 2011). These tests usually involve measuring the sensitivity of community
structure to small changes in edge weight or variations in time – partitions that fail these tests are
regarded as being irrelevant. In using the partition stability framework, we chose not to engage in
robustness testing. We justify this choice by noting that each participant’s connectome can be
regarded as a sample drawn from a population of closely related connectome networks.
Therefore, obtaining the consensus community structure is an implicit test of robustness. At no
point, however, did we test the robustness of community structure as time changed. We argue
that establishing the robustness of a whole partition is of little practical interest and that a more
revealing test is to identify individual communities that persist over time. We note that our
analysis revealed a number of communities that appeared at early Markov times and that
remained disjoint for long ranges of time.
Conclusion
This article offers novel insight into the relationship between models of dynamical processes
unfolding on the structural connectome, community structure, and empirically measured
dynamic couplings. By maximizing partition stability, communities at multiple dynamical scales
were identified. These communities were then compared to observed patterns of neural activity
through a simple correlation measure and also by assessing each community’s contribution to the
modularization of rsFC. It was observed that a number of anatomical regions contributed
disproportionately to this modularization, suggesting that compared to other regions, they might
be more likely to engage in a diffusion-like communication process over the timescales when
these contributions occurred. Future work is needed to further illuminate the role of different
classes of dynamic processes in generating patterns of rsFC in complex brain networks.
Acknowledgements
Supported in part by: The National Science Foundation/IGERT Training Program in the
Dynamics of Brain-Body-Environment Systems at Indiana University (RB); Swiss National
Science Foundation, SNF grant 320030-130090 (AG); Spanish Government grant, contract
number E-28-2012-0504681 (JG); Leenards Foundation, Switzerland (PH); JS McDonnell
Foundation (OS).
14
References
Albert, R., Jeong, H., & Barabási, A. L. (1999). Internet: Diameter of the world-wide
web. Nature, 401(6749), 130-131.
Arenas, A., Fernandez, A., & Gomez, S. (2008). Analysis of the structure of complex networks
at different resolution levels. New Journal of Physics, 10(5), 053039.
Bassett, D. S., Greenfield, D. L., Meyer-Lindenberg, A., Weinberger, D. R., Moore, S. W., &
Bullmore, E. T. (2010). Efficient physical embedding of topologically complex information
processing networks in brains and computer circuits. PLoS Computational Biology, 6(4),
e1000748.
Bassett, D. S., Wymbs, N. F., Rombach, M. P., Porter, M. A., Mucha, P. J., &
Grafton, S. T. (2012). Core-Periphery Organisation of Human Brain
Dynamics.arXiv preprint arXiv:1210.3555.
(2004). Neuronal oscillations
Blondel, V. D., Guillaume, J. L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of
Statistical Mechanics: Theory and
of
large networks. Journal
communities
in
Experiment, 2008(10), P10008.
Bullmore, E., & Sporns, O. (2009). Complex brain networks: graph theoretical analysis of
structural and functional systems. Nature Reviews Neuroscience, 10(3), 186-198.
Bullmore, E., & Sporns, O. (2012). The economy of brain network organization. Nature Reviews
Neuroscience, 13(5), 336-349.
Buzsáki, G., & Draguhn, A.
Science, 304(5679), 1926-1929.
Cammoun, L., Gigandet, X., Meskaldji, D., Thiran, J. P., Sporns, O., Do, K. Q., Maeder P.,
Meuli R., & Hagmann, P. (2012). Mapping the human connectome at multiple scales with
diffusion spectrum MRI. Journal of Neuroscience Methods, 203(2), 386-397.
Chen, Z. J., He, Y., Rosa-Neto, P., Germann, J., & Evans, A. C. (2008). Revealing modular
architecture of human brain structural networks by using cortical thickness from MRI. Cerebral
Cortex, 18(10), 2374-2381.
Daducci A., Gerhard S., Griffa A., Lemkaddem A., Cammoun L., Gigandet X., Meuli R.,
Hagmann P., & Thiran J. P. (2012). The Connectome Mapper: an open-source processing
pipeline to map connectomes with MRI. PloS One, 7(12), e48121
Damoiseaux, J. S., Rombouts, S. A. R. B., Barkhof, F., Scheltens, P., Stam, C. J., Smith, S. M.,
& Beckmann, C. F.
(2006). Consistent
resting-state networks
across healthy
subjects. Proceedings of the National Academy of Sciences, 103(37), 13848-13853.
in cortical networks.
15
Delmotte, A., Tate, E. W., Yaliraki, S. N., & Barahona, M. (2011). Protein multi -scale
organization through graph partitioning and robustness analysis: application to the myosin–
myosin light chain interaction. Physical Biology, 8(5), 055010.
Delvenne, J. C., Yaliraki, S. N., & Barahona, M. (2010). Stability of graph communities across
time scales. Proceedings of the National Academy of Sciences, 107(29), 12755-12760.
Flake, G. W., Lawrence, S., Giles, C. L., & Coetzee, F. M. (2002). Self-organization and
identification of web communities. Computer, 35(3), 66-70.
Fortunato, S., & Barthelemy, M. (2007). Resolution limit in community detection. Proceedings
of the National Academy of Sciences, 104(1), 36-41.
Fortunato, S. (2010). Community detection in graphs. Physics Reports, 486(3), 75-174.
Fox, M. D., & Raichle, M. E. (2007). Spontaneous fluctuations in brain activity observed with
functional magnetic resonance imaging. Nature Reviews Neuroscience, 8(9), 700-711.
Freeman, L. C. (2004). The development of social network analysis. Vancouver: Empirical Press.
Girvan, M., & Newman, M. E. (2002). Community structure in social and biological
networks. Proceedings of the National Academy of Sciences, 99(12), 7821-7826.
Good, B. H., de Montjoye, Y. A., & Clauset, A. (2010). Performance of modularity
maximization in practical contexts. Physical Review E, 81(4), 046106
Gong, G., He, Y., Concha, L., Lebel, C., Gross, D. W., Evans, A. C., & Beaulieu, C . (2009).
Mapping anatomical connectivity patterns of human cerebral cortex using in vivo diffusion
tensor imaging tractography. Cerebral Cortex, 19(3), 524-536.
Greicius, M. D., Krasnow, B., Reiss, A. L., & Menon, V. (2003). Functional connectivity in the
resting brain: a network analysis of the default mode hypothesis. Proceedings of the National
Academy of Sciences, 100(1), 253-258.
Guimera, R., & Amaral, L. A. N. (2005). Cartography of complex networks: modules and
universal roles. Journal of Statistical Mechanics: Theory and Experiment, 2005(02), P02001.
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Wedeen, V. J., & Sporns, O.
(2008). Mapping the structural core of human cerebral cortex. PLoS Biology, 6(7), e159.
Haimovici, A., Tagliazucchi, E., Balenzuela, P., & Chialvo, D. R. (2012). Brain organization into
the connectome at criticality. arXiv preprint
resting state networks emerges
from
arXiv:1209.5353.
16
Honey, C. J., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J. P., Meuli, R., & Hagmann, P.
(2009).
structural
from
connectivity
functional
resting-state
human
Predicting
connectivity. Proceedings of the National Academy of Sciences,106(6), 2035-2040.
Jonsson, P. F., Cavanna, T., Zicha, D., & Bates, P. A. (2006). Cluster analys is of networks
generated through homology: automatic identification of important protein communities
involved in cancer metastasis. BMC Bioinformatics, 7(1), 2.
Lambiotte, R., Delvenne, J. C., & Barahona, M. (2008). Laplacian dynamics and multiscale
modular structure in networks. arXiv preprint arXiv:0812.1770.
Lambiotte, R. (2010, May). Multi-scale modularity in complex networks. In Modeling and
Optimization in Mobile, Ad Hoc and Wireless Networks (WiOpt), 2010 Proceedings of the 8th
International Symposium on (pp. 546-553). IEEE.
Lambiotte, R., Sinatra, R., Delvenne, J. C., Evans, T. S., Barahona, M., & Latora, V. (2011).
Flow graphs: Interweaving dynamics and structure. Physical Review E, 84(1), 017102.
Lancichinetti, A., & Fortunato, S. (2012). Consensus clustering in complex networks. Scientific
Reports, 2.
Meunier, D., Lambiotte, R., & Bullmore, E. T. (2010). Modular and hierarchically modular
organization of brain networks. Frontiers in Neuroscience, 4.
Moody, J., & White, D. R. (2003). Structural cohesion and embeddedness: A hierarchical
concept of social groups. American Sociological Review, 103-127.
Murphy, K., Birn, R. M., Handwerker, D. A., Jones, T. B., & Bandettini, P. A. (2009). The
impact of global signal regression on resting state correlations: are anti-correlated networks
introduced?. Neuroimage, 44(3), 893.
Newman, M. E., & Girvan, M. (2004). Finding and evaluating community structure in
networks. Physical Review E, 69(2), 026113.
Power, J. D., Barnes, K. A., Snyder, A. Z., Schlaggar, B. L., & Petersen, S. E. (2012). Spurious
but systematic correlations in functional connectivity MRI networks arise from subject
motion. Neuroimage, 59(3), 2142-2154.
Reichardt, J., & Bornholdt, S. (2006). Statistical mechanics of community detection. Physical
Review E, 74(1), 016110.
Ronhovde, P., & Nussinov, Z. (2009). Multiresolution community detection for megascale
networks by information-based replica correlations. Physical Review E, 80(1), 016109.
17
Schaub, M. T., Delvenne, J. C., Yaliraki, S. N., & Barahona, M. (2012). Markov dynamics as a
zooming lens for multiscale community detection: non clique-like communities and the field-of-
view limit. PloS One, 7(2), e32210.
Sepulcre, J., Sabuncu, M. R., Yeo, T. B., Liu, H., & Johnson, K. A. (2012). Stepwise
connectivity of the modal cortex reveals the multimodal organization of the human brain. The
Journal of Neuroscience, 32(31), 10649-10661.
Smith, S. M., Fox, P. T., Miller, K. L., Glahn, D. C., Fox, P. M., Mackay, C. E., Filippini, N.,
Watkins, K. E., Toro, R., Laird, A. R., & Beckmann, C. F. (2009). Correspondence of the brain's
functional architecture during activation and rest. Proceedings of the National Academy of
Sciences, 106(31), 13040-13045.
Sporns, O., Tononi, G., & Kötter, R. (2005). The human connectome: a structural description of
the human brain. PLoS Computational Biology, 1(4), e42.
van den Heuvel, M. P., & Hulshoff Pol, H. E. (2010). Exploring the brain network: a review on
resting-state fMRI functional connectivity. European Neuropsychopharmacology, 20(8), 519-
534.
van den Heuvel, M. P., & Sporns, O. (2011). Rich-club organization of the human
connectome. The Journal of Neuroscience, 31(44), 15775-15786.
van den Heuvel, M. P., Kahn, R. S., Goñi, J., & Sporns, O. (2012). High-cost, high-capacity
backbone for global brain communication. Proceedings of
the National Academy of
Sciences, 109(28), 11372-11377.
Wedeen, V. J., Wang, R. P., Schmahmann, J. D., Benner, T., Tseng, W. Y. I., Dai, G., Pandya,
D. N., Hagmann, P., D’Arceuil, H., & De Crespigny, A. J. (2008). Diffusion spectrum magnetic
resonance imaging (DSI) tractography of crossing fibers. Neuroimage, 41(4), 1267-1277.
Wu, K., Taki, Y., Sato, K., Sassa, Y., Inoue, K., Goto, R., Okada, K., Kawashima, R., He, Y.,
Evans, A.C., & Fukuda, H. (2011). The overlapping community structure of structural brain
network in young healthy individuals. PloS One, 6(5), e19608.
18
Figure 1. Structural and functional matrices averaged over 40 participants; A) SC weighted:
entries indicate the average density of white matter fiber tracts connecting brain regions; B) SC
binary: entries indicate the presence {absence} of connecting fibers present in any of the 40
participants; C) rsFC: entries are averaged correlations of the processed BOLD time series
during resting state.
19
Figure 2. Examples of flow graphs obtained at different Markov times in a single participant; A)
Original SC matrix; B-F) flow graphs at times , respectively,
depicting the probabilistic flow of random walkers between pairs of vertices.
20
Figure 3. Statistics from community structure obtained from stability optimization procedure (all
curves depict mean plus/minus two standard deviations); partition stability (blue); number of
communities (green); number of communities with more than one vertex (red); average
community size (magenta). Community size, number of singleton communities, and number of
communities are normalized by size of the network (501 cortical regions in the right hemisphere)
so that they scale between 0 and 1.
21
Figure 4. Consensus partitions obtained at different Markov times; A) with twelve
communities; B) with ten communities; C) with eight communities; D)
with six communities; E) with four communities; F) with two
communities. Vertex size is proportional to strength.
22
Figure 5. Summary of correlation between agreement matrix ( ) and resting-state functional
connectivity (rsFC); A) Pearson’s correlation value (black) at each time point. Superimposed on
this plot are the mean stability (blue) and mean number of community (green) curves. The time
at which the max correlation value occurs falls within a narrow regime where the number of
communities and stability are both substantial; B) ( ) at peak correlation value (top) and rsFC
(bottom); C) Scatter plot of ( ) and rsFC at time of peak correlation (linear fit - red line). The
color of each “o” denotes the number of participants in which the corresponding structural
connection was also present.
23
Figure 6. Summary of 14 consensus communities at the time of peak correlation. In each plot,
vertices are colored gray if they do not belong to the corresponding community. The color of any
non-gray vertex indicates the cortical region that that vertex belongs to. Once again, vertex size
is proportional to strength.
24
Figure 7. Summary of modularity scores of communities at different times; A) sum of positive
and negative normalized scores by vertex over time; B) weighted matrix where each community
co-assignment was weighted by its score; C) anatomical region-by-region average and peak
standardized modularity score; D) weighted agreement matrix aggregated across by anatomical
region.
25
Figure 8. Summary of hierarchical tree cut at different levels; A-F) Cuts reveal spatially
contiguous maps of three to eight clusters.
26
|
1611.00509 | 2 | 1611 | 2017-08-15T07:57:26 | Multiscale modeling via split-step methods in neural firing | [
"q-bio.NC"
] | Neuronal models based on the Hodgkin-Huxley equation form a fundamental framework in the field of computational neuroscience. While the neuronal state is often modeled deterministically, experimental recordings show stochastic fluctuations, presumably driven by molecular noise from the underlying microphysical conditions. In turn, the firing of individual neurons gives rise to an electric field n extracellular space, also thought to affect the firing pattern of nearby neurons.
We develop a multiscale model which combines a stochastic ion channel gating process taking place on the neuronal membrane, together with the propagation of an action potential along the neuronal structure. We also devise a numerical method relying on a split-step strategy which effectively couples these two processes and we experimentally test the feasibility of this approach. We finally also explain how the approach can be extended with Maxwell's equations to allow the potential to be propagated in extracellular space. | q-bio.NC | q-bio |
Multiscale modeling via split-step methods in
Pavol Bauer ∗
neural firing
Stefan Engblom ∗†
Aleksandar Senek ∗
August 16, 2017
Sanja Mikulovic ‡
Abstract
Neuronal models based on the Hodgkin-Huxley equation form a funda-
mental framework in the field of computational neuroscience. While the
neuronal state is often modeled deterministically, experimental recordings
show stochastic fluctuations, presumably driven by molecular noise from
the underlying microphysical conditions. In turn, the firing of individual
neurons gives rise to an electric field in extracellular space, also thought
to affect the firing pattern of nearby neurons.
We develop a multiscale model which combines a stochastic ion channel
gating process taking place on the neuronal membrane, together with
the propagation of an action potential along the neuronal structure. We
also devise a numerical method relying on a split-step strategy which
effectively couples these two processes and we experimentally test the
feasibility of this approach. We finally also explain how the approach
can be extended with Maxwell's equations to allow the potential to be
propagated in extracellular space.
Keywords: Computational neuroscience, Hodgkin-Huxley equation, Stochas-
tic simulation, Multiphysics coupling, Lie-Trotter splitting.
AMS subject classification: 65C20, 92C20 (primary); 65C40, 65L99
(secondary).
1
Introduction
Neurons are responsible for encoding information in the central nervous system.
Lower level functions are many times gathered at specific parts of the brain,
∗Pavol Bauer, Stefan Engblom, and Aleksandar Senek are with the Division of Sci-
entific Computing, Department of Information Technology, Uppsala University, Box 337,
[email protected], Aleksan-
SE-751 05 Uppsala, Sweden, e-mail:
[email protected].
†Corresponding author, address as above. Office phone +46 18 471 27 54, fax +46 18 51
‡Sanja Mikulovic, is with the Department of Neuroscience, Biomedical Centrum, Uppsala
[email protected],
19 25.
University, Box 593, SE-751 24 Uppsala, Sweden, e-mail: [email protected].
1
processing information received from neurons throughout the body, often by
means of signaling nerves, which are bundles of axons that reach out from
the neurons. Higher level functions depend on remarkably larger and complex
networks of neurons with various types of feedback loops.
The chemical connections between neurons are handled via synapses, where
neurotransmitters are extruded into the extracellular space by the presynaptic
neuron. The neurotransmitters form a part of a chemical process which may
initiate a potential wave into the postsynaptic neuron - this wave is known
as the action potential. During the action potential, the membrane potential
quickly rises and falls, and the resulting signal propagates along the cell [29]. The
process that underlies this propagation is the regulation of ion concentrations in
both the intracellular cytoplasm and the extracellular space caused by integral
membrane proteins called ion channels.
There are many variants of ion channel proteins, whose functions are only
recently better understood through studies using experimental techniques such
as X-ray crystallography [30]. To our current understanding, ion channels reside
at one of many conformal states, which are either closed (non-conducting) or
open (conducting).
If the conformal state is open, ions are allowed to pass
through pathways called pores from the extra- to the intracellular space, or in
the opposite direction. If the conformal state is closed, ions are blocked from
entering the channel [21]. Ion channels become activated either in response to
a chemical ligand binding, or in response to voltage changes on the membrane
[21, 29], so-called voltage-gated ion channels. In this work, we focus on voltage-
gated ion channels, which are important for the initiation and the propagation
of action potentials along the neuronal fiber.
Mathematical models of neurons were initially formed by measuring elec-
trically induced responses of a neuron using techniques such as the voltage- or
current clamp. From those experiments, transition properties and parameters
of single channel gating could be identified [21].
Historically, models of firing neurons were formulated as ordinary differential
equations (ODEs). However, in some specific cases experimental as well as
theoretical findings suggest that the gating of ion channels is more accurately
described by a stochastic process. The variance of the gating process, also known
as the channel noise, is thought to be important for information processing in the
dendrites and explain different phenomena regarding action potential initiation
and propagation [5, 15, 32, 35]. For example, it has been shown that intrinsic
noise is essential for the existence of subthreshold oscillations in stellate cells
[10], that it contributes to irregular firing of cortical interneurons [33], and that
it can explain firing correlation in auditory nerves [27].
Another important aspect of neuronal modeling is the investigation of ac-
tion potentials propagating outward the neuron into the extracellular space.
Simulations of such extracellular fields is one of the important methods used
in computational neuroscience [12]. Common usage includes the validation of
experimental methods such as EEG and extracellular spike recordings, or in
the modeling of physiological phenomena which can not be easily investigated
empirically [11].
2
In this paper we present a novel three-stage multiscale modeling framework
consisting of the following components: (a) on the microscale, the gating pro-
cess of the ion channels is governed by a continuous-time discrete-state Markov
chain, (b) on the intermediate scale, the current-balance and the cable equa-
tion which are responsible for the action potential initiation and propagation
is integrated in time as an ODE, (c) on the macroscale, the propagation of the
trans-membrane current into an electrical field in extracellular space is achieved
using partial differential equations (PDEs).
In §2 the three modeling layers are explained in some detail. The numerical
method via split-step modeling is summarized in §3, where we also explain how
the spatial coupling is achieved efficiently. We illustrate our method by some
relevant examples in §4 and offer a concluding discussion in §5.
2 Neuronal modeling at different scales
We now describe the modeling framework at the individual scales, including
the associated modeling assumptions. The microscale physics in the form of
continuous-time Markov chains (CTMC) for the ion channels and the associated
ion currents are discussed in §2.1, the ODE-model for the action potential along
the neuronal geometry in §2.2, and the PDE-model of the extracellular electric
field via Maxwell's equations in §2.3. For convenience, a schematic explanation
of the modeling framework is found in Figure 2.1.
Figure 2.1: Schematic overview of the proposed multiscale modeling framework.
The CTMC solution on the microscale depends on the membrane voltage that
is computed on the mesoscale. This coupling is bidirectional; the mesoscale
solution depends on the microscopic stochastic ion current. The macroscopic
solution considered here depends solely on the mesoscopic membrane current.
3
Ion channel gating (cid:127)CTMC€Membrane dynamics (cid:127)ODE€Local !eld potential (cid:127)PDE€Conductivity GTMembrane current ImMicroscaleMesoscaleMacroscale Membrane potential Vm m0h0
β
h
α
h
m0h1
3αm
βm
3αm
βm
m1h0
β
h
α
h
m1h1
2αm
2βm
2αm
2βm
m2h0
β
h
α
h
m2h1
αm
3βm
αm
3βm
m3h0
β
h
α
h
m3h1
Figure 2.2: Kinetic scheme for the m3h1 sodium channel gating proposed by
Hodgkin and Huxley [22]. Only the gating state m3h1 represents an open ion
channel, for all other states the ion channel is closed.
2.1 Microscale: ion channel gating
The gating of voltage-dependent ion-channels is modeled by a Markov process.
Hodgkin and Huxley [22] first proposed that the gating is governed by a set of
gating variables. We assume that each gating variable takes values in a discrete
state space and that a single combination of gating variables corresponds to an
open and conducting ion channel [21].
The gating states and the transitions between them can be written in the
form of a kinetic scheme. An example for the voltage-gated m3h1 sodium chan-
nel is depicted in Figure 2.2. The notation for the channel states indicates that
there are two involved gating variables m and h, where variable m takes one of
four states and, independently, variable h takes one of two states. In total we
thus arrive at 8 states with a total of 10 reversible transitions, see Figure 2.2.
The total number of channels in the open state m3h1 implies a certain con-
ductivity as will be detailed below. The transition rates between the states
depend on the membrane voltage and on the state itself. When these transi-
tions take place in a microscopic environment where molecular noise is present,
a continuous-time Markov chain (CTMC) is the most suitable model.
S1
S5
S2
S6
S3
S7
S4
S8
Figure 2.3: An equivalent scheme to Figure 2.2, where each combination of
gating variables corresponds to a discrete state of a Markov chain.
We rewrite such schemes in a general, yet more manageable notation as
follows. Assign the different combinations of gating variables an individual
state Si, i = 1, 2, . . . , Mstates. For the sodium scheme in Figure 2.2 there are 8
states and only the state S8 corresponds to an open and conducting ion channel.
The resulting equivalent scheme is shown in Figure 2.3 and an exemplary set of
4
transitions and rates is
S3 −→ S4, rate: αm(Vm),
S3 −→ S2, rate: 2βm(Vm),
S3 −→ S7, rate: αh(Vm),
(2.1)
(2.2)
(2.3)
where Vm is the membrane potential at the location of the ion channel. For this
particular example, the transition rates used by Hodgkin and Huxley are [22]
0.1(Vm + 40)
1 − e−(Vm+40)/10
αm(Vm) =
βm(Vm) = 4e−(Vm+65)/18,
αh(Vm) = 0.07e−(Vm+65)/20,
,
βh(Vm) =
1
1 + e−(Vm+35)/10
.
(2.4)
(2.5)
(2.6)
(2.7)
The rates (2.4)–(2.7) were originally obtained by empirical fitting to exper-
imental data and are formulated relative to the resting potential of the partic-
ular neuron model. We consider a small neural compartment a, obtained from
a discretization of the neuronal fiber into small enough segments such that the
potential is approximately constant within each compartment. Let there be a
total of N a ion channels of the considered type (e.g. the m3h1 sodium channel)
on the compartmental surface. At any instant in time t, let si be the number
of channels in state Si, i = 1, 2, . . . , Mstates. The scheme in Figure 2.3 now
directly translates into Markovian transition rules for the states s ∈ Z8
+, i.e. the
ion-channel counts, thus defining a CTMC (t, s) ∈ R+ × Z8
+.
To model the electrophysiological properties of the ion channel, we define
the single open channel conductance γ, as well as the neuronal membrane area
Aa such that the area density of ion channels is a = N a/Aa. If Oa is the count
of the open ion channels, e.g., for the above example this is Oa = Sa
8 , we arrive
at the total conductance
Ga =
Oa
N a aAaγ.
(2.8)
Next, the trans-membrane current produced by the ion channel is given by
Ohm's law,
I a = Ga(V a
m − E),
(2.9)
where E is the reverse potential of the ion channel, that is, the membrane
potential under which the trans-membrane current I a is zero, and where V a
m is
the membrane potential in compartment a.
In the macroscopic setting proposed by Hodgkin and Huxley [22], the tran-
sitions between the gating states happened according to first order reaction
kinetics. The fraction of closed channels (1− oa) becomes open with rate α and
the fraction of open channels oa becomes closed with rate β. Note that α and
5
β may be voltage-dependent for specific types of ion-channels, as in (2.1)–(2.7).
The resulting ODE for each macroscopic gating variable oa can then be written
as
doa
dt
= α(1 − oa) − βoa.
(2.10)
Under the macroscopic formulation, the ratio of open channels is approxi-
mated as
N a ≈ M(cid:89)
Oa
i=1
[oa
i ]pi,
(2.11)
where M is the number of gating particles for the particular ion channel model,
and pi is the exponent of the ith gating variable oi.
As a concrete example, the sodium channel m3h1 depicted in Figure 2.2 is
gated by gating variables m and h, which in the macroscopic formulation have
an exponent of three and one, respectively. In relation to (2.11) this means that
o1 = m with p1 = 3, and o2 = h with p2 = 1, where m and h are here just
symbols denoting the concentration of the gating variables. The ratio of open
channels is then approximated as
and the deterministic gating function for each variable obeys
≈ m3h1,
ON a
NN a
= αm(1 − m) − βmm,
= αh(1 − h) − βhh,
dm
dt
dh
dt
(2.12)
(2.13)
(2.14)
where α and β are the voltage dependent transition rates defined in (2.4)–(2.7).
Solving (2.13)–(2.14) and using (2.9) thus constitutes the classical Hodgkin-
Huxley ODE model for the ionic current. By rather evolving the Markov chain
defined implicitly in Figure 2.3 and using (2.8)–(2.9), a stochastic current is
defined. This stochastic model is the result of the microphysical assumption
of discrete ion states obeying Markovian (memoryless) transition rules. For
large enough numbers of ion channels, a transition into the corresponding de-
terministic model can be expected under rather broad conditions [14]. When
a fine enough compartmentalization of a particular neuron is made, however,
the stochastic model can be expected to be more realistic in the sense that the
effects missing in a corresponding deterministic model can not be ignored.
We now proceed to discuss the appropriate model for evolving the resulting
current, be it stochastic or deterministic, over the neuronal morphology.
2.2
Intermediate scale: current-balance and cable equa-
tion
On the intermediate scale, we solve for the current-balance equation of the
neuron, describing the relation between the membrane voltage and the different
6
current sources. We first take the ionic current sources from the microscale
into account. We next model the capacitance of the membrane, a so-called
passive neuronal property. As the neuronal morphology is divided into several
compartments we also include a term for the propagation of voltage between
the compartments, resembling the cable equation.
As before, denote the compartments by the index a ∈ {1, . . . , Mcompartments}.
For each connection between compartments, an entry is made in the adjacency
matrix A [9]. Typically, a compartment is connected to two or three neighbors.
If a compartment is connected to only a single neighbor, it represents one of the
end-points of the neuron. An example is found in Figure 2.4.
Figure 2.4: Geometry of a neuron and the associated adjacency matrix A (figure
adapted from [9]). A nonzero entry in A implies a connection between the
corresponding compartments.
The current flowing in each compartment is composed from different com-
ponents. The current source of compartment a is the total axial current with
respect to its connected compartments b ∈ B(a). Note that as A represents a
directed graph, the connected component b ∈ B(a) represents in-neighbors that
are connected via an edge in the direction to vertex a, as well as out-neighbors
which are connected by an edge outwards of vertex a. According to Ohm's law
the current equals
I a
axial =
Ga
b (V a
m − V b
m) =
Ga
b V a
Ga
b V b
m,
(2.15)
(cid:88)
b∈B(a)
(cid:88)
b∈B(a)
m − (cid:88)
b∈B(a)
where Ga
the trans-membrane potential of compartment a.
b is the conductance between compartments a and b, and where V a
m is
We then add the ionic currents computed at the microscale. For the sake of
generality we extend the previous notation of a single type of ion channel into
c types, T = {T 1, . . . , T c}, acting simultaneously. Generalizing (2.9) we arrive
7
at the total ionic current
(cid:88)
T
I a
ionic =
Ga
T (V a
m − ET ).
(2.16)
We now deviate slightly from the discussion and first consider a space-
continuous solution obtained by solving for the propagation of the potential
on a line oriented along the x-axis. Adding the axial current flux Iaxial, the to-
tal trans-membrane current flux Im, and a possibly additional external current
source Iinj gives
Cm
d
dt
Vm = Iinj − Im − Iaxial,
(2.17)
where Cm is the capacitance of the neuronal membrane. We have that the
gradient of the inter-compartmental leakage current IL is
∂IL
∂x
= −Im.
(2.18)
The leakage current accounts for the ions that diffuse out of the intracellular
space due to the natural permeability of the neuronal membrane.
Given a suitable resting potential, Ohm's law is
IL = GL
∂Vm
∂x
.
(2.19)
Substituting Im from (2.17) using (2.18)–(2.19), we get what is commonly re-
ferred to as the cable equation,
GL
∂2Vm
∂x2 = Cm
∂Vm
∂t
− Iinj + Iaxial.
(2.20)
Given the parabolic character of this equation, the Crank-Nicholson scheme [8]
is a natural choice when designing numerical methods.
Turning now to the compartment-discrete version, the leakage current in
each compartment a is given as
I a
L = Ga
L(V a
m − Ea
L),
(2.21)
where the membrane leakage conductivity Ga
are regarded as constants.
L and the leak reverse potential Ea
L
The total trans-membrane current flux now becomes
I a
m = I a
L + I a
ionic = (Ga
L +
Ga
T )V a
m − (Ga
LEa
L +
Ga
T ET ).
(2.22)
In order to solve for the membrane potential in each compartment a, we
T
T
8
(cid:88)
(cid:88)
start from a discrete version of (2.17), using (2.15) and (2.22),
T ) − I a
(cid:88)
I a
(cid:88)
inj − (I a
L +
T ET ) − (Ga
Ga
inj − I a
axial = I a
inj + (Ga
V a
m = I a
= I a
d
dt
C a
m
Ga
T )V a
m
axial
T
L +
T
m − I a
(cid:88)
(cid:88)
LEa
L +
Ga
b V a
m +
T
Ga
b V b
m.
−(cid:88)
(2.23)
(2.24)
b
b
To summarize, on the intermediate scale we insert the stochastic conductiv-
ity Ga
T computed from (2.8) into (2.16) and solve the current-balance equation
(2.23) for each of the neural compartments. We simultaneously solve for the
trans-membrane current defined in (2.22), to be coupled into the macroscopic
field - next to be discussed.
2.3 Macroscale: extracellular field potentials
To simulate spatially non-homogeneous distributions of electrical fields produced
by single neurons and neuronal networks, we use an electrostatic formulation
of Maxwell's equations to be discretized with finite elements. We solve for the
electric field intensity E in terms of the electric scalar potential V ,
E = −∇V.
(2.25)
The relevant dynamic form of the continuity equation with current sources Qj
is given by
∇ · J = − ∂ρ
∂t
+ Qj,
(2.26)
with J and ρ the current density and electric charge density, respectively. Fur-
ther constitutive relations include
and Ohm's law
D = ε0εrE,
J = σE,
(2.27)
(2.28)
in which D denotes the electric flux density. Finally, Gauss' law states that
∇ · D = ρ.
(2.29)
Upon taking the divergence of (2.28) and using the continuity equation (2.26)
we get
∇ · (σE) = − ∂ρ
∂t
+ Qj.
(2.30)
Rewriting the electric charge density using Gauss' law together with the consti-
tutive relation (2.27) and finally applying the gauge condition (2.25) twice we
arrive at the time-dependent potential formulation
= Qj.
(2.31)
(cid:18)
−∇ ·
(cid:19)
σ∇V + ε0εr
∇V
∂
∂t
9
The values for the electric conductivity σ and the relative permittivity εr were
obtained from [2]. The source currents Qj are computed from the compartmen-
tal model described in §2.2. Specifically, in compartment a, we put
Qj(t) =
I a
m(t)
Aa ,
(2.32)
where I a
area of the neuronal membrane in compartment a.
m(t) is obtained by solving (2.22) and where we recall that Aa is the
The boundary conditions here are homogeneous Neumann conditions (elec-
tric isolation) everywhere except in a single point which we take to be ground
(V = 0). In all our simulations this point was placed at the axis of rotation of the
enclosing cylindrical extracellular space, and directly underneath the neuronal
geometry. This procedure ensures that the formulation has a unique solution;
without this specification it is otherwise only specified up to a constant.
3 Numerical modeling
The processes at the three scales, that is, the microphysics of the ion channel
gating, the neuron current at the intermediate scale, and the propagation of
the electric field, all take place in continuous time. Also, all processes formally
affect eachother through two-way couplings.
In §3.1 we discuss the numerical coupling of the ion channel gating and the
cable equation via a split-step strategy. In §3.2 the propagation of the electric
potential into extracellular space is summarized and here we also explain how
the often rather complicated neuronal geometry is handled. To simplify matters,
we will disregard the usually much weaker coupling from the external electric
field back to the gating process.
3.1 Numerical coupling of firing processes
To get to grip with the details of the coupling between the ion channel gating
process and the propagation of the action potential along the neuron we need
a more compact notation as follows. In compartment a, denote by Xa(t) the
gating state, that is, the number of channels being in each of the different states
at time t. For our sodium channel example we have Xa(t) = [sa
8](t)T .
The CTMC may be written compactly as
2, . . . , sa
1, sa
dXa(t) = Sµa(Xa, V a
(3.1)
for a = 1, . . . , Mcompartments, and where S is an Mstates × Ntransitions matrix of
integer transition coefficients. The dependency on the state (Xa, V a
m) is implicit
in the random counting measure µa associated with an Ntransitions-dimensional
Poisson process. As a concrete example, the transition S3 −→ S4 in (2.1)
satisfies
m; dt),
E[µa
i (Xa, V a
m; dt)] = E[Xa
3 αm(V a
m)] dt,
(3.2)
10
with
S3,i = −S4,i = −1,
(3.3)
that is, with the understanding that (2.1) is the ith transition according to
some given ordering. Since the process is a Poisson process, (3.2) expresses
independent exponentially distributed waiting times of intensity αm(V a
m) for
each of the Xa
3 possible transitions of type S3 −→ S4.
The propagation of the action potential can similarly be written in more
compact ODE notation,
dV a
m(t) = G(Xa, V a
m,{V b
m}b∈B(a)) dt,
(3.4)
where G is just the current-balance equation (2.23) and which depends on the
state (Xa, V a
m) via (2.8), (2.16), and (2.22), respectively.
Eqs. (3.1) and (3.4) form a coupled CTMC-ODE model which falls under the
scope of Piecewise Deterministic Markov Processes (PDMPs) and for which nu-
merical methods have been investigated [31]. A highly accurate implementation
is possible through event-detection in traditional ODE-solvers. This, however,
has a performance drawback since the ODE-solver must continuously determine
to sufficient accuracy what micro-event happens when. In the experiments in
§4 we have chosen to rely on a simple split-step strategy as follows. Given a
discrete time-step ∆t, and tn+1 = tn + ∆t,
Xa
n+1 = Xa
n +
Sµa(Xa(s), V a
m,n; ds),
(cid:90) tn+1
(cid:90) tn+1
tn
tn
V a
m,n+1 = V a
m,n +
G(Xa
n+1, V a
m(s),{V b
m(s)}b∈B(a)) ds.
(3.5)
(3.6)
That is, (3.5) evolves the CTMC (3.1) keeping the voltage potential fixed at
its value from the previous time-step tn. Similarly, (3.6) evolves the ODE (3.4)
while keeping the state of the ion channels fixed at the time-step tn+1. Impor-
tantly, the usually more expensive stochastic simulation in (3.5) is fully decou-
pled as it only depends on the state of the compartment a. The global step is
achieved separately by solving the connected ODE in (3.6), and is usually quite
fast.
The approximation (3.5)–(3.6) can be understood as a split-step method and
may also be analyzed as such [6, 13]. The order of the approximation can then
be expected to be 1/2 in the root mean-square sense,
E[(cid:107)Xn − X(tn)(cid:107)2 + (cid:107)Vm,n − Vm(tn)(cid:107)2] ≤ C∆t.
(3.7)
Although it appears to be difficult to increase the stochastic order of this ap-
proximation, the accuracy is likely going to increase by turning to symmetric
Strang-type splitting methods for the stochastic part [13], possibly also adopt-
ing a higher order scheme for the ODE-part. However, the efficiency of such an
approach ultimately depends on the strength of the nonlinear feedback terms
11
and is difficult to analyze a priori. In the present proof-of-concept context we
aim at a convergent and consistent coupling for which (3.5)–(3.6) are a suitable
starting point. We thus postpone the investigation of more advanced integration
methods for another occasion.
3.2 Spatial extension of the firing process
The modeling in §2.2 did not take into account a possible free-space potential
Vext, external to the cell. The necessary modifications to incorporate this are as
follows. For x ∈ Ω ⊂ R3, let Vext(x) be a given external potential and denote
by V a
ext the value at compartment a. Then replace (2.23) with
(cid:88)
I a
m = I a
= (Ga
L + I a
L +
ionic + I a
ext
T )V a
Ga
m − (Ga
(cid:88)
(3.8)
Ga
T ET ) + Ga
mV a
ext.
LEa
L +
T
T
for Ga
m the membrane transneuronal conductivity. With this modification,
(2.23) propagates the effect of the given external field Vext(x) along the neuron.
In §2.3 we discussed how the effect of a trans-membrane current propagates
into extracellular space as an electric potential V (cf. (2.31)–(2.32)). With the
previous discussion in mind, the following two-way coupling thus emerges: the
solution of the current-balance equation (2.23) yields a current source, which
feeds into (2.32). In turn, this implies an external potential Vext := V , obtained
by solving (2.31), finally to be inserted into (3.8) above. The full model coupling
thus arrived at takes the schematic form
§2.1 −−−(cid:42)(cid:41)−−− §2.2 −−−(cid:42)(cid:41)−−− §2.3.
(3.9)
It is certainly possible to devise a numerical method using a similar split-step
strategy as in §3.1 for this full coupling. As mentioned before, to simplify we
shall assume that the feedback from the external field and onto the neuron is
weak, such that Vext ≈ 0 in (3.8) and we thus consider the simplified model (see
also Figure 2.1)
§2.1 −−−(cid:42)(cid:41)−−− §2.2 −−−→ §2.3.
(3.10)
This assumption is valid whenever the simulation consists of a small number
of neighboring neurons, but becomes inaccurate if a large number of nearly
parallel neurons is considered. See [1, 4] for a further discussion. It follows that
the solution of the field potential V can be done "offline". That is, this problem
may be solved in isolation using a pre-recorded current source I a
m(t) obtained
from the split-step method described in §3.1.
The three-dimensional neuronal geometry was constructed in Comsol Multi-
physics with the help of the interface to Matlab ("LiveLink") by morphological
additions and Boolean unification of simple geometric objects, representing neu-
ronal compartments. The TREES Toolbox [9] was used to conveniently access
the geometrical properties of single compartments over the Matlab interface.
12
More specifically, the geometry was constructed by parsing the adjacency graph
A. Starting from the initial node of the graph, this procedure makes sure that
each compartment is connected to the same neighbors as in the numerical model
for the axial current flow in (2.15).
In an initial attempt, we aimed to represent the 3D geometry as an exact
counterpart of the compartmental model, where each compartment is under-
stood as a cylinder with a certain length and diameter. If the direction of the
main axes of any two joining cylinders differed, a sphere was added in between
the cylinders, followed by a removal of all interior boundaries. Although this
created a direct volumetric representation of the neuronal compartments, the
approach is difficult to generalize to neuronal branches with a more complicated
connectivity. The reason is that the triangulation of the final object becomes
extremely difficult to achieve as the mesh engine insists on fully resolving the
curvature. See Figure 3.1 for an example of a problematic mesh, emerging at
the intersection of a cylinder and a sphere.
A second and more successful attempt was made by which three-dimensional
curves made up of line segments for each compartment was constructed. This
simplifies the meshing process immensely, since the extracellular mesh is not
constrained by high-curvature cylindrical boundaries. Implicit here is the as-
sumption that the neuron is very thin compared to the external length-scale
of practical interest. This is valid as the diameter of a dendritic structure is
≈ 1µm and the considered extracellular space is typically in the range of several
mm [12]. In Figure 3.2 we show examples of both approaches.
Figure 3.1: Example of a problematic mesh at the intersection of two cylindrical
compartments.
13
Figure 3.2: Example of a neuronal morphology created with cylindrical objects
(left), and with curves (right).
4 Experiments
We devote this section to some feasibility experiments of the method proposed.
In §4.1 we look in some detail at the coupling of the microscopic and intermediate
scales. Using the stochastic currents so computed, the induced electrical field is
propagated outside the neuron in §4.2.
4.1 Micro-meso coupling
In this section we provide numerical experiments of the coupling between the
microscopic scale, introduced in §2.1, and the mesoscopic scale, introduced in
§2.2. We simulate the classical squid model proposed by Hodgkin and Huxley
[22], which is also a part of a widely used benchmark for neuronal simulators
[3]. The model includes two types of voltage-gated ion channels, N a+ and K +
ions, which in our case are both modeled as continuous-time Markov processes
(cf. §2.1). The kinetic gating scheme for the N a+ channel is shown in Figure 2.2.
The K + channel follows a similar scheme, but has only four discrete gating states
S1, . . . , S4, of which only the state S4 is open. The transition rates in this case
are voltage-dependent as follows [22]
αn(Vm) =
0.01(Vm + 10)
e(Vm+10)/10 − 1
βn(Vm) = 0.125e(Vm/80).
,
(4.1)
(4.2)
Note the similar form of (4.1) to (2.4) and (4.2) to (2.5), respectively, which
is due to using the same empirical fitting procedure. The density of the ion
channels is K = 30 µm−2 and N a = 330 µm−2, respectively [21]. The single
channel conductance equals γK = 360/30 pS, and γN a = 1200/330 pS, while
the reversal potentials are EK = −77 mV and EN a = 50 mV [3].
The parameters for the intermediate scale are as follows. The specific mem-
brane capacitance cm = 1µF (cm)−2, resting potential Er = −65 mV , cytoplasm
resistivity c = 100 Ωcm, specific leak conductance gL = 40000−1 S(cm)−2 and
14
leak reversal potential EL = −65 mV [3]. We let the model be confined to
a cylindrical geometry with a length of 1 mm and a diameter of 1 µm. The
cylinder is compartmentalized into Mcompartments sub-cylinders of equal length
and diameter. The root node is ignited by a current injection of 0.1 nA.
In practice we use the Gillespie's "Direct Method" [17] to evolve (3.5) and
the Crank-Nicholson scheme [8] to solve (3.6), relying on the fact that the latter
is linear in V a
m.
In Figure 4.1 we show the numerical solution of the coupled model, overlaid
with the deterministic solution where ion channels are described by ODEs as in
(2.13)–(2.14). We show three traces of the membrane voltage Vm in one neuronal
compartment over time, where the injected current has been varied from a lower
value (0.045nA) to a larger value (0.1nA). The dynamics of the stochastic
model for the initial current injections clearly differ from the dynamics of the
deterministic one, where a single spike or a train of spikes can be triggered in the
stochastic representation, while no spike can be obtained in the deterministic
one. At the higher amount of current injection, we observe a trace with similar
characteristics for both model representations, but with an increasingly different
phase shift.
In Figure 4.2 we show a numerical convergence study of the coupled model,
concerning two method parameters. We show how the interspike interval (ISI)
changes as a function of the coupling time step ∆t, as well as the discretization
of the geometry ∆x. The ISI is defined as the duration between the peaks
of two spikes. As the neuronal firing is now a stochastic process, the ISI will
be a distribution, and hence we present the first and second moments. We
find that, for the study of the coupling time step, the ISI appears to be well
resolved at ∆t = 10−1 ms for the spatial discretization presented. For the spatial
discretization, we find that the ISI distributions do not significantly differ for a
voxel length of under 1µm.
15
Figure 4.1: Membrane voltage behavior around threshold current. The accuracy
of the deterministic solver was verified with the reference solution of Rallpack
3. In all cases the discretization is ∆t = 0.05 and ∆x = 0.05. Upper: injected
current is Iinj = 0.045 nA, middle: injected current is Iinj = 0.063 nA, lower:
injected current is Iinj = 0.1 nA.
16
050100150200250Time [ms]-100-50050Vm [mV]DeterministicStochastic050100150200250Time [ms]-100-50050Vm [mV]050100150200250Time [ms]-100-50050Vm [mV] Figure 4.2: Top: The mean interspike interval (ISI) ± standard error as a
function of the time step ∆t. The compartment length is here ∆x = 0.02.
Bottom: The ISI as a function of the compartment length ∆x at a time step
∆t = 0.05. The red line represents the interspike interval from the Rallpack 3
reference solution [3]. The number of trajectories in all runs is N = 40.
4.2 Three-scale coupling
For this example we took the morphological description of a pyramidal CA1
neuron from [28], which contains about 1500 compartments. We re-sampled the
geometry in order to aggregate short compartments of sizes less than 50 µm
into larger compartments, leading to a final representation consisting of approx-
imately 400 compartments. Although the reduction might alter the properties
of the model and more evolved protocols for compartment reduction could be
used [20, 26], we ignore the induced discretization error here since the purpose
of the model is to demonstrate the overall numerical method.
We solved (3.5) and (3.6) over the morphology, incorporating the active
properties described in §4.1. Next, we scaled the transmembrane currents Im(t)
(cf. (2.32)) and mapped this to the corresponding curve segment as a current
source Qj(t) [7]. It can be noted in passing that we here assume that transmem-
brane currents are the only cause of change of extracellular potential, which is
not the case in a real neuron, as for example synaptic calcium-mediated currents
17
10-310-210-1100Timestep [ms]1020304050Interspike interval [ms]10010110214151617Interspike interval [ms] Figure 4.3: Solution of the full three-scale model framework: propagation of an
extracellular action-potential into homogeneous extracellular space.
are suspected to contribute to a large fraction of the extracellular signature [4].
Equipped with the source currents (2.32), the formulation (2.31) is efficiently
solved by Comsol's "Time discrete solver", which is based on the observation
that the variable W := ∆V satisfies a simple ODE. Solving for W in an inde-
pendent manner up to some time t, it is then straightforward to solve a single
static PDE to arrive at the potential V itself. For the Time discrete solver the
time-step was set to ∆t as used in the split-step method, thus ensuring a correct
transition to the macroscopic scale.
A tetrahedral mesh [23] was applied to discretize space (using the "finest"
mesh setting; resolution of curvature 0.2, resolution of narrow regions 0.8). The
simulations were verified against coarser mesh settings in order to ensure a
practically converged solution.
The result of the simulation is visualized in Figure 4.3. We have inserted
four "point probes" at a radial distance of 1000 µm to the neuron, measuring
the extracellular voltage at single points. The electric potential thus monitored
by the probes is shown in Figure 4.4.
Figure 4.4: Electric potential (mV) in the four point probes placed around the
neuron.
18
00.050.10.150.20.250.30.350.40.45Time [ms]-0.4-0.200.20.40.60.8Electric Potential [mV]Probe at x=400Probe at x=0Probe at x=-200Probe at x=-400 5 Conclusions
In this paper we have proposed a model framework for neuronal firing processes
consisting of three layers: on the microscale, ion channel gating is modeled
as a continuous-time Markov chain. On the intermediate scale, the currents
produced by open channels are integrated into the current-balance equation
proposed by Hodgkin and Huxley. Finally, on the macroscale, the outward cur-
rent of neurons are propagated into extracellular space, simulating the emission
of an extracellular potential. We have also described a numerical approach to
the coupling of the different scales and indicated through computational results
the feasibility of the overall approach.
To date, several exact and approximate simulation methods [5, 16, 18, 25] of
stochastically gated ion channel models have been proposed. However, to our
knowledge, the coupling between the microscopic gating layer and the meso-
scopic layer describing the action potential initiation and propagation, has not
yet been rigorously studied. In this paper, we provide the formulation of the
coupling as a split-step method and moreover numerically observe the conver-
gence of the method with respect to the size of the coupling time-step as well as
the spatial discretization. We show that the distribution of interspike intervals
may be significantly affected by the choice of the coupling time-step, but does
not appear to depend strongly on the spatial discretization. Finally, we discuss
the theory and praxis of incorporating the macroscopic scale into the model,
including the spatial representation of neuronal compartments and appropri-
ate numerical procedures for the simulation of the local field potential (LFP)
propagation.
While it has often been taken for granted that ion channel gating occurs
deterministically, accumulating research evidence indicates that the presence of
stochasticity significantly influences the neuronal behavior [15, 32, 35]. In turn,
neurons respond with a high level of variability to the repeated presentation of
equal stimuli. This leads to remarkable difficulties in studying the link between
single cell biophysical properties and their function in larger neuronal networks,
both in health and disease. Additionally, a recent study [5] has shown that
stochastic ion channel gating largely differs not only between different neuronal
cell types, but also locally between different parts of the neuron. Given the
technical difficulties of assessing the signal channel properties of smaller neuronal
compartments such as axons and dendrites, developing reliable mathematical
models is essential to tackle this problem.
Several studies [5, 15, 27, 32, 33, 35] have demonstrated how incorporation
of channel noise into the Hodgkin and Huxley equations could resemble multi-
ple realistic neuronal behaviors. Although it is not the scope of this work to
mimic any particular experimental problem, but rather to provide a modeling
framework that could be used by a diverse group of scientists in the future,
we foresee multiple interesting applications. For example, it was shown previ-
ously [24, 34, 35] that the stochastic nature of voltage-gated ion channels in the
medial entorhinal cortex stellate cells are crucial for generating subthreshold
oscillations in theta (≈ 8Hz) frequency range. These intrinsic oscillations in
19
stellate cells were suggested to generate theta oscillations in vivo LFP [34]. It
is well established that theta oscillatory activity in vivo provides a temporal
window in which spatial and declarative memories are formed [19]. Thus, our
framework could be used to test the link between the stochastic nature of sin-
gle channels in specific cell types and their effect at both the cellular and the
network level.
Acknowledgment
PB and SE were supported by the Swedish Research Council within the UP-
MARC Linnaeus center of Excellence. SM was supported by the Olle Engkvist
post-doctoral fellowship.
Availability and reproducibility
The computational results can be reproduced within the upcoming release 1.4
of the URDME open-source simulation framework, available for download at
www.urdme.org.
References
[1] C. A. Anastassiou, R. Perin, H. Markram, and C. Koch. Ephaptic coupling
of cortical neurons. Nature Neuroscience, 14(2):217–223, 2011.
[2] C. B´edard, H. Kroger, and A. Destexhe. Modeling extracellular field po-
tentials and the frequency-filtering properties of extracellular space. Bio-
physical Journal, 86(3):1829–1842, 2004.
[3] U. S. Bhalla, D. H. Bilitch, and J. M. Bower. Rallpacks: a set of benchmarks
for neuronal simulators. Trends in Neurosciences, 15(11):453 – 458, 1992.
[4] G. Buzs´aki, C. A. Anastassiou, and C. Koch. The origin of extracellu-
lar fields and currents - EEG, ECoG, LFP and spikes. Nature Reviews
Neuroscience, 13(6):407–420, 2012.
[5] R. C. Cannon, C. O'Donnell, and M. F. Nolan. Stochastic Ion Channel
Gating in Dendritic Neurons: Morphology Dependence and Probabilistic
Synaptic Activation of Dendritic Spikes. PLoS Computational Biology, 6
(8), 2010.
[6] A. Chevallier and S. Engblom.
Pathwise error bounds in multi-
scale variable splitting methods for spatial stochastic kinetics, 2017.
Accepted for publication in SIAM J. Numer. Anal. Available at
https://arxiv.org/abs/1607.00805.
[7] AC/DC Module User's Guide. Comsol, 2012. Version 4.3.
20
[8] J. Crank and P. Nicolson. A practical method for numerical evaluation
of solutions of partial differential equations of the heat-conduction type.
Advances in Computational Mathematics, 6(1):207–226, 1947.
[9] H. Cuntz, F. Forstner, A. Borst, and M. Hausser. One Rule to Grow
Them All: A General Theory of Neuronal Branching and Its Practical
Application. PLoS Computational Biology, 6(8), 2010.
[10] A. D. Dorval and J. A. White. Channel noise is essential for perithreshold
oscillations in entorhinal stellate neurons. The Journal of Neuroscience:
The Official Journal of the Society for Neuroscience, 25(43):10025–10028,
2005.
[11] G. T. Einevoll, D. K. W´ojcik, and A. Destexhe. Modeling extracellular
potentials. Journal of Computational Neuroscience, 29(3):367–369, 2010.
[12] G. T. Einevoll, C. Kayser, N. K. Logothetis, and S. Panzeri. Modelling
and analysis of local field potentials for studying the function of cortical
circuits. Nature Reviews Neuroscience, 14(11):770–785, 2013.
[13] S. Engblom. Strong convergence for split-step methods in stochastic jump
kinetics. SIAM Journal of Numerical Analysis, 53(6):2655–2676, 2015.
[14] S. N. Ethier and T. G. Kurtz. Markov Processes: Characterization and
Convergence. Wiley series in Probability and Mathematical Statistics. John
Wiley & Sons, New York, 1986.
[15] A. A. Faisal, L. P. J. Selen, and D. M. Wolpert. Noise in the nervous
system. Nature Reviews Neuroscience, 9(4):292–303, 2008.
[16] R. F. Fox. Stochastic versions of the Hodgkin-Huxley equations. Biophysical
Journal, 72(5):2068–2074, May 1997.
[17] D. T. Gillespie. Exact stochastic simulation of coupled chemical reactions.
Journal of Physical Chemistry, 81(25):2340–2361, 1977.
[18] J. H. Goldwyn, N. S. Imennov, M. Famulare, and E. Shea-Brown. Stochas-
tic differential equation models for ion channel noise in Hodgkin-Huxley
neurons. Physical Review. E, Statistical, Nonlinear, and Soft Matter
Physics, 83(4 Pt 1):041908, 2011.
[19] T. Hartley, C. Lever, N. Burgess, and J. O'Keefe. Space in the brain:
how the hippocampal formation supports spatial cognition. Philosophical
Transactions of the Royal Society of London. Series B, Biological Sciences,
369(1635):20120510, 2014.
[20] E. B. Hendrickson, J. R. Edgerton, and D. Jaeger. The capabilities and
limitations of conductance-based compartmental neuron models with re-
duced branched or unbranched morphologies and active dendrites. Journal
of Computational Neuroscience, 30(2):301–321, 2011.
21
[21] B. Hille. Ionic Channels of Excitable Membranes. Sinauer Associates, 1992.
[22] A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane
current and its application to conduction and excitation in nerve. The
Journal of Physiology, 117(4):500–544, 1952.
[23] J. Jin. The Finite Element Method in Electromagnetics. John Wiley &
Sons, New York, 2nd edition, 2002.
[24] R. Klink and A. Alonso. Ionic mechanisms for the subthreshold oscillations
and differential electroresponsiveness of medial entorhinal cortex layer II
neurons. Journal of Neurophysiology, 70(1):144–157, 1993.
[25] D. Linaro, M. Storace, and M. Giugliano. Accurate and fast simulation of
channel noise in conductance-based model neurons by diffusion approxima-
tion. PLoS Computational Biology, 7(3):e1001102, 2011.
[26] A. Marasco, A. Limongiello, and M. Migliore. Using Strahler's analysis to
reduce up to 200-fold the run time of realistic neuron models. Scientific
Reports, 3:srep02934, 2013.
[27] B. Moezzi, N. Iannella, and M. D. McDonnell. Ion channel noise can ex-
plain firing correlation in auditory nerves. Journal of Computational Neu-
roscience, 41(2):193–206, 2016.
[28] T. M. Morse, N. T. Carnevale, P. G. Mutalik, M. Migliore, and G. M.
Shepherd. Abnormal Excitability of Oblique Dendrites Implicated in Early
Alzheimer's: A Computational Study. Frontiers in Neural Circuits, 4, 2010.
[29] D. Purves. Neuroscience. Sinauer Associates, 2012. 4th Edition.
[30] D. C. Rees, G. Chang, and R. H. Spencer. Crystallographic Analyses of
Ion Channels: Lessons and Challenges. Journal of Biological Chemistry,
275(2):713–716, 2000.
[31] M. G. Riedler. Almost sure convergence of numerical approximations for
piecewise deterministic Markov processes. Journal of Computational and
Applied Mathematics, 239(0):50–71, 2013.
[32] E. Schneidman, B. Freedman, and I. Segev. Ion channel stochasticity may
be critical in determining the reliability and precision of spike timing. Neu-
ral Computation, 10(7):1679–1703, 1998.
[33] K. M. Stiefel, B. Englitz, and T. J. Sejnowski. Origin of intrinsic irregular
firing in cortical interneurons. Proceedings of the National Academy of
Sciences of the United States of America, 110(19):7886–7891, 2013.
[34] J. A. White, R. Klink, A. Alonso, and A. R. Kay. Noise from voltage-gated
ion channels may influence neuronal dynamics in the entorhinal cortex.
Journal of Neurophysiology, 80(1):262–269, 1998.
22
[35] J. A. White, J. T. Rubinstein, and A. R. Kay. Channel noise in neurons.
Trends in Neurosciences, 23(3):131–137, 2000.
23
|
1710.02423 | 1 | 1710 | 2017-10-06T14:30:59 | Voltage laws for three-dimensional microdomains with cusp-shaped funnels derived from Poisson-Nernst-Planck equations | [
"q-bio.NC",
"math.AP",
"physics.bio-ph",
"q-bio.SC"
] | We study the electro-diffusion properties of a domain containing a cusp-shaped structure in three dimensions when one ionic specie is dominant. The mathematical problem consists in solving the steady-state Poisson-Nernst-Planck (PNP) equation with an integral constraint for the number of charges. A non-homogeneous Neumann boundary condition is imposed on the boundary. We construct an asymptotic approximation for certain singular limits that agree with numerical simulations. Finally, we analyse the consequences of non-homogeneous surface charge density. We conclude that the geometry of cusp-shaped domains influences the voltage profile, specifically inside the cusp structure. The main results are summarized in the form of new three-dimensional electrostatic laws for non-electroneutral electrolytes. We discuss applications to dendritic spines in neuroscience. | q-bio.NC | q-bio |
Voltage laws for three-dimensional microdomains with
cusp-shaped funnels derived from
Poisson-Nernst-Planck equations
J. Cartailler∗ and D. Holcman∗†
June 6, 2021
Abstract
We study the electro-diffusion properties of a domain containing a cusp-shaped
structure in three dimensions when one ionic specie is dominant. The mathematical
problem consists in solving the steady-state Poisson-Nernst-Planck (PNP) equation
with an integral constraint for the number of charges. A non-homogeneous Neu-
mann boundary condition is imposed on the boundary. We construct an asymptotic
approximation for certain singular limits that agree with numerical simulations. Fi-
nally, we analyse the consequences of non-homogeneous surface charge density. We
conclude that the geometry of cusp-shaped domains influences the voltage profile,
specifically inside the cusp structure. The main results are summarized in the form
of new three-dimensional electrostatic laws for non-electroneutral electrolytes. We
discuss applications to dendritic spines in neuroscience.
Keywords. Electro-diffusion, Cusp Funnel, Poisson Nernst-Planck, Non-Electro-neutrality;
Asymptotics; Nonlinear PDEs.
AMS subject classification. 35J66
1
Introduction
We study here the Poisson-Nernst-Planck (PNP) equations in three dimensional for do-
mains containing a cups-shaped funnel. These equations are used to describe electro-
diffusion processes in ionic channels [1, 2] and also in neurobiological microdomains [3, 4],
∗Ecole Normale Sup´erieure, 46 rue d'Ulm 75005 Paris, France.
†Mathematical Institute, University of Oxford, Andrew Wiles Building, Woodstock Rd, Oxford OX2
6GG, United Kingdom.
Corresponding author email: [email protected]
1
where charges are coupled though the electric field. We consider here a generic domain
formed of a ball with an attached cusp-shaped funnel on its boundary. Such geometry
is common in cellular neurobiology, for instance dentritic spines [5], where the structure
cannot be reduced to 1D geometry [6]. Phenomenological descriptions of electro-diffusion,
using the linear cable theory, RC-electric circuit representation, and even electronic de-
vices, are insufficient to describe non-cylindrical geometry [3, 6], since they assume a
simple reduced one-dimensional or an overly simplified geometry.
We present here nove results about the voltage landscape based on the electro-diffusion
model in various microdomains, when the condition of electro-neutrality is not satisfied
and one ionic specie is dominant. The boundary is impermeable to particles (ions) and the
electric field satisfies the compatibility condition resulting from Poisson's equation. Under
the non-electro-neutrality assumption and with charge distributed in bounded domains,
confined by a dielectric membrane, Debye's law of charge screening decaying exponentially
away from a charge [7] does not apply and long-range correlation are expected, leading to
a gradient of charges in a domain with no inward current. We derived a new capacitance
law for an electrolyte ball [8] and for a two-dimensional cusp [9], where the difference of
potential V (C) − V (S) between the center C and the surface S increases, first linearly
and then logarithmically when the total number of charges in the ball increases.
Our aim here is to estimate the effect of boundary curvature on three-dimensional
electrical domains such as dendritic spines. In particular, we explore the effect of boundary
curvature on the charge and field distribution at steady state. The curvature of neuronal
dendrites and axons membranes possesses many local maxima that can modulate the
channel's local electric potential [4, 10, 11]. In this article, we study the effects of local
curvature on the distribution of charges with no electro-neutrality. The effect of non-
electro-neutrality was recently studied in [8, 9] and a long-range electrostatic length,
much longer than the Debye length was found. This effect is due to the combined effects
of non-electro-neutrality and di-electric boundary, which lead to charge accumulation.
The cusp-shaped funnel was studied in the context of diffusion in [12], but we focus here
on a three-dimensional nonlinear problem with non-homogeneous Neumann boundary
condition and we further extend the matched asymptotic analysis based on conformal
mapping, different from the classical matched asymptotic methods [13, 14, 15, 16, 17].
The manuscript is composed of three parts: in sections 1 and 2, we extend the results
we have obtained in [9], that describe the voltage in a planar cusp with homogeneous
surface charge density. We then focus on an uncharged cusp for a 3D cusp-shaped funnel.
In the third section, we extend the results derived in section 1 to a non-homogeneous
surface charge density. We summarized now the new electrostatic laws we derived here
for the difference of potential V (C) − V (S) where C is the center of mass of the domain
and S is located at the bottom of funnel (Fig. 1A).
For a constant surface charge density (section 2, eq. (76)), the voltage difference is
2
given by
V (C) − V (S) =
kT
e
ln sin2
π∂ Ω
(e2/kT )N ε
(cid:16)
− ln
,
(cid:17)2 + O(1)
2 (e2/kT )2π2N 2R2
c
4∂ Ω + (e2/kT )N ε
that depends the number N of ions enclosed in the domain Ω, the thermal energy kT and
the elementary charge e of the electron (1.602 · 10−19C), the cusp-shaped funnel width at
the base ε, and its curvature radius Rc (see. Fig. 1A).
When the surface of the cusp does not carry any charges, the voltage difference (section
(cid:32)
3, eq. (97)) is
V (C) − V (S) =
kT
e
− ln
8Rc ε
π4∂ Ωε (1 + Nbulk/Nε)
− ln sin2
(cid:112)
2∂ Ωε
(e2/kT )Nε
Rc ε
(cid:33)
+ O(1)
,
,
(cid:17)2 + O(1)
which depends on the surface ∂Ωε at the end of the funnel, the number of charges Nbulk
and Nε in bulk and at the end of the funnel respectively. When the surface charge density
is non-homogeneously distributed, the potential differences (section 4, formula (103)) are
given by
V (C) − V (S) =
kT
e
π∂ Ωε
(e2/kT )Nε ε
− ln
(cid:16)
2 (e2/kT )2π2N 2
cuspR2
c
4∂ Ωcusp + (e2/kT )Ncusp ε
ln sin2
which depends on the total surface charge density Ncusp on the cusp.
These new electrostatics expressions are asymptotic formula derived in the limit ε (cid:28) 1
and for a large number of charges. There are the main results of the present study.
Figure 1: Ball with a cusp-shaped funnel and image Ωw of the domain Ω cross-
section under the conformal mapping (17) A. schematic representation of the domain Ω,
with the funnel curvature radius RC, the north pole N , the funnel tip S, and the center of mass
C. B-C The neck (B) is mapped onto the semi-annulus enclosed between the like-style arcs and
the large disk in Ω is mapped onto the small red disk. The short green segment AB (left) (of
length ε) is mapped onto the thick green segment AB (of length 2
ε + O(ε)).
√
3
BBAABACABBNSCABrz 2 The Poisson-Nernst-Planck equations
The Poisson-Nernst-Planck equations is a classical model of electro-diffusion. In a domain
Ω, the total charge in Ω results from the sum of the positive Np and negative Nm charges.
The concentration of mobile ions [3] shows an imbalance of positive negative ions Np (cid:29)
Nm, such that the charges in Ω can be approximated [9] by N identical positive ions with
an initial density q(x) in Ω. The valence is z and the total number of particles is
(cid:90)
q(s) ds = N.
(1)
The total charge in the domain Ω is
Ω
Q = zeN,
(cid:104)
where e is the electron charge. The charge density ρ(x, t) is the solution of the Nernst-
Planck equation
D
∆ρ(x, t) +
(cid:20)∂ρ(x, t)
ze
kT
D
∂n
∇ (ρ(x, t)∇φ(x, t))
∂φ(x, t)
ze
kT
ρ(x, t)
+
for x ∈ Ω
∂ρ(x, t)
∂t
=
= 0 for x ∈ ∂ Ω
∂n
ρ(x, 0) = q(x) for x ∈ Ω,
(2)
(3)
(4)
(cid:105)
(cid:21)
where kT represents the thermal energy. The electric potential φ(x, t) in Ω is the solution
of the Poisson equation
∆φ(x, t) = − zeρ(x, t)
for x ∈ Ω
= − σ(x, t) for x ∈ ∂ Ω,
εrε0
∂φ(x, t)
∂n
(5)
(6)
where εrε0 is the permitivity of the medium and σ(x, t) is the surface charge density on
the boundary ∂ Ω.
2.1 Steady solution in a three-dimensional ball with a cusp-
shaped funnel
To study the effect of a narrow funnel attached to a sphere filled with an electrolyte as
illustrated Fig. 1A, we study the solution of the steady-state equation (2)
(cid:27)
(cid:27)
ρ(x) = N
(cid:90)
Ω
exp
exp
−zeφ(x)
kT
−zeφ(s)
kT
(cid:26)
(cid:26)
4
,
ds
(7)
hence (5) results in the Poisson equation
∆φ(x) = −
and (6) gives the boundary condition
εrε0
exp
Ω
(cid:26)
(cid:26)
zeN exp
(cid:90)
(cid:27)
(cid:27)
−zeφ(x)
kT
−zeφ(s)
kT
∂φ(x)
∂n
= − Q
εrε0∂ Ω for x ∈ ∂ Ω.
.
ds
(8)
(9)
Equation (9) represents the compatibility condition obtained by integrating the Poisson's
equation (5) over the domain Ω, assuming the surface charge density is constant. Using
non-dimensional variables, we define the normalized field
(ze)2N
εrε0kT
¯u(x) =
zeφ(x)
(10)
λ =
kT
,
,
where λ generalizes the Bjerrum length lB = e2/kT . The Poisson's equation (8) reduces
to
∆¯u(x) = − λ exp{−¯u(x)}
exp{−¯u(s)} ds
(cid:90)
Ω
(11)
(12)
(13)
(14)
and the boundary condition (9) becomes
= − λ
∂ ¯u(x)
∂n
∂ Ω for x ∈ ∂ Ω.
We consider now the PNP problem (11)-(12) in the solid of revolution (Fig. 1A), ob-
tained by rotating the symmetric planar domain Fig. 1B around its z−axis of symmetry.
Consequently, Ω represents now a ball with a cusp-shaped funnel, with a radius curvature
Rc at the entrance of the funnel (blue dashed circles in Fig. 1A-B).
Ω
Rc
Using the change of variable x =
3 and u(x) = ¯u(x) +
2 and Ω =
∂ Ω
Rc
, ∂Ω =
x
Rc
(cid:19)
ln
λR2
c/
exp{−u(s)} ds
converts (11) into
(cid:18)
(cid:90)
Ω
−∆u(x) = exp{−u(x)} for x ∈ Ω
for x ∈ ∂Ω.
= − λ
∂u(x)
∂ΩRc
The non-dimensional surface charge density is
∂n
σ =
λ
∂ΩRc
.
We first consider a uniform surface charge density in (13) and then study the consequences
of a non-homogeneously distributions.
5
2.2 Poisson-Nernst-Planck solutions in a 3D cusp-shaped funnel
The cylindrical symmetry of the Neumann boundary value problem (BVP) (13) in the
(r, z, φ) coordinates (Fig. 1A) centered on the axis of symmetry, implies that u(x) is
independent of the angle φ in the domain Ω. It follows that (13) in the domain Ω can be
written as
∂2u(r, z)
∂r2 +
1
r
∂u(r, z)
∂r
+
∂2u(r, z)
∂z2
∂u(r, z)
∂n
= − exp(−u(r, z))
= −σ,
(15)
where n = [nr, nz]T is the outward normal unit vector to the surface ∂Ω and r is the
distance to the symmetry axis of Ω. The opening at the cusp funnel is small AB = ε (cid:28) 1
(green line Fig. 1B), so the funnel is a narrow passage. To remove the cusp singularity,
we use first the transformation to the rotated and translated coordinates given by r =
r − 1 − ε/2 and z = −z + 1. Setting u(r, z) = u(r, z), eq. (15) becomes,
∂2 u(r, z)
∂r2 +
∂2 u(r, z)
∂ z2 +
1
∂ u(r, z)
(r + 1 + ε/2)
∂r
= − exp(−u(r, z))
(16)
∂ u(r, z)
∂ n
= −σ.
We shall construct an asymptotic expansion of the solution u(r, z) for small ε by first
mapping the cross section in the (r, z)−plane conformally into its image under the Mobius
transformation [12]
where
w(ξ) = ρeiθ =
ξ − α
1 − αξ
,
α = −1 − √
ε + O(ε),
(17)
(18)
and ξ = r + iz. In the dimensionless domain Ω, the parameter ε is also dimensionless
and Rcε = ε. Mobius transformation maps the two osculating circles A and B (dashed
blue) into concentric circles (see Fig. 1B-C). The Mobius transformation (17) maps the
right circle B (dashed blue) into itself and Ω is mapped onto the banana-shaped domain
Ωw = w(Ω) as shown in Figure 1C.
The second order derivative for u(ξ) = v(w) is computed using (17) in (16) [18]
∂2 u
∂r2 +
In the small ε limit, we have
∂2 u
∂ z2 = w(cid:48)(ξ)2∆wv(w).
(1 − √
ε)eiθ − 1 + O(ε)4
4ε + O(ε3/2)
(19)
(20)
.
w(cid:48)(ξ)2 =
6
The 3D BVP (16) differs from the 2D problem [9] by the extra first order radial derivative.
For small ε limit, we have
r + 1 + ε/2 =
ε
1 − cos(θ)
+ O(ε3/2).
(21)
In complex coordinates we have
(22)
where (cid:60)e(·) is the real part. Under the conformal mapping (17), the gradient from (22)
transforms as follows [18]
∂r
∂u(r, z)
= (cid:60)e (∇u(ξ)) ,
∇u(ξ) = ∇wv(w) w(cid:48)(ξ).
Using polar coordinates (ρ, θ) in the mapped domain Ωw, we write
w(cid:48)(ξ) = w1(ρ, θ) + i w2(ρ, θ),
where i2 = −1. Using (17), we obtain
w1(ρ, θ) =
1 − α2ρ2 + 2αρ cos(θ)(1 + αρ cos(θ))
1 − α2
w2(ρ, θ) = −2αρ sin(θ)
1 + αρ cos(θ)
1 − α2
.
Using (22) and (25), in polar coordinates (see Appendix), it follows that
∂ u(r, z)
∂r
=
∂v(ρ, θ)
(cos(θ) w1(ρ, θ) − sin(θ) w2(ρ, θ))
∂ρ
∂v(ρ, θ)
∂θ
− 1
ρ
(sin(θ) w1(ρ, θ) + cos(θ) w2(ρ, θ)) .
(23)
(24)
(25)
(26)
To leading order, using (21) and (26), we get (Appendix)
∂r
1
r
∂ u(r, z)
(27)
In summary, using (19) in polar (ρ, θ)−coordinates, eq. (27) and (16) in Ωw, are changed
to
(1 − √
= −ρ(1 − cos(θ))2
(cid:18) ∂2v(ρ, θ)
− sin(θ)(1 − cos(θ))
∂2v(ρ, θ)
∂v(ρ, θ)
∂v(ρ, θ)
∂v(ρ, θ)
∂v(ρ, θ)
(cid:19)
ε3/2
.
∂θ
∂ρ
ε
ε)eiθ − 14
4ε
∂ρ2 +
1
ρ
∂ρ
+
1
ρ2
ε3/2
− ρ(1 − cos(θ))2
− sin(θ)(1 − cos(θ))
= − exp{−v(ρ, θ)}
= − σ
√
1 − cos(θ)
ε
ε
.
∂ρ
∂v(ρ, θ)
∂θ
(28)
∂θ2
∂v(ρ, θ)
∂n
7
2.3 Asymptotic analysis of the PNP equations in a cusp-shaped
funnel
To analyse eq. (28) in the limit of σ (cid:29) 1, ε → 0 [8], we approximate the domain Ωw by
two subregions
ε > π, ρ − 1 ≤ √
ε}
(29)
A = {(ρ, θ) ∈ Ωw : θ − √
B = {w = (1 − √
ε)eiθ : θ − π ≤ √
ε},
as illustrated in Fig. 2A. The regions B consists of a circular arc (dashed red). We
construct now the solution uA(r, θ) and uB(θ) of (13) in each subregion.
Figure 2: Decomposition of the domain Ωw into two subregions regions A and B A.
Representation of the two subregions A (blue) and B (dotted red) of Ωw. B. Solutions of (43)
ε.
(dashed blue), (54) (red dots), and the uniform approximation uunif (58) (green) for r = 1−√
Asymptotics of uA(r, θ) in region A
√
To construct the asymptotics solution uA(r, θ) in region A, we use that the radial derivative
ε) → ∞ in the regime σε3/2 = O(1) as σ → ∞ and ε → 0. Thus the angular
∂
∂r is O(σ
derivatives are negligible relative to the radial ones. It follows in a regular expansion of
equation 28 along the rays θ = θ0 = const, for ρ ∈ [1 − √
the solution, the θ derivative can be neglected relative to the ρ derivative and we will
ε, 1].
8
Setting uA(ρ, θ0) = v(ρ, θ0), to leading order in σ
ε, equation (28) reduces to
(cid:19)
+
1
ρ
∂uA(ρ, θ0)
∂ρ
(30)
√
(cid:18) ∂2uA(ρ, θ0)
∂ρ2
∂ uA(ρ, θ0)
∂ρ
(1 − √
ε)eiθ0 − 14
4ε
−ρ(1 − cos(θ0))2
√
ε3/2
ε
1 − cos(θ0)
−e−uA(ρ, θ0) =
(cid:12)(cid:12)(cid:12)(cid:12)ρ=1
(cid:12)(cid:12)(cid:12)(cid:12)ρ=1−√
ε
duA(ρ, θ0)
dρ
duA(ρ, θ0)
dρ
= −
= 0.
In the limit ε (cid:28) 1, we note that ρeiθ0(1 − √
√
ε) and using the
ε and setting uA(ρ, θ0) = vA(ρ, θ0), to leading order in ε (cid:28) 1,
ε) − 14 = eiθ0 − 14 + O(
√
change of variable ρ = ρ
eq. (30) becomes
(cid:18)
1 − 4(1 − cos(θ0))2
eiθ0 − 14
(cid:19)
.
(31)
−4ε2e−vA(ρ, θ0)
eiθ0 − 14
=
∂2vA(ρ, θ0)
∂ ρ2
− √
ε
∂vA(ρ, θ0)
∂ ρ
Using the function,
h(θ0) =
4ε2
eiθ0 − 14
and vA(ρ, θ0) = vA(ρ, θ0) − ln(h(θ0)), eq. (31) is transformed into
∂2vA(ρ, θ0)
∂ ρ2
= −e−vA(ρ, θ0) +
∂vA(ρ, θ0)
∂ ρ
√
ε
1 − (1 − cos(θ0))2
eiθ0 − 14
(cid:18)
(cid:19)
.
(32)
(33)
Figure 3: PNP solution (13) in a 3D domain with a cusp-shaped funnel A. Repre-
sentation of the domain Ω with a surface charge density σ, the north pole N , the funnel tip S,
and the center of mass C, respectively. B. Numerical (13) (solid) and analytical (58) (dashed)
solutions in the domain Ωw for several values of σ = 10, 100, 1000 and 4000 for ε = 0.01. C.
Difference u(C) − u(S) computed numerically (solid blue) from (13) and analytically (dashed
green) from (76).
9
SAy0CNBC3Dnumericsuunif3Dnumericsuunif40001000=10100CS--- Using a regular expansion in the small ε limit (in the regime σε3/2 = O(1))
√
εvA,1(ρ, θ0) + O(ε)
vA(ρ, θ0) = vA,0(ρ, θ0) +
in (33), we get
= −e−vA,0(ρ, θ0)
=
σε
1 − cos(θ0)
= 0.
∂2vA,0(ρ, θ0)
∂ ρ2
∂vA,0(ρ, θ0)
∂ ρ
∂vA,0(ρ, θ0)
(cid:12)(cid:12)(cid:12)(cid:12)ρ=0
(cid:12)(cid:12)(cid:12)(cid:12)ρ=1
(cid:18)
A direct integration of (35) is [9]
∂ ρ
vA,0(ρ, θ0) = ln
2C1(θ0)2 cos2
(cid:18) ρ + C2(θ0)
(cid:19)(cid:19)
2C1(θ0)
,
(cid:18) ρ + C2(θ0)
(cid:19)
2C1(θ0)
.
(34)
(35)
(36)
(37)
(38)
where C1(θ0) and C2(θ0) are two constants that depend on θ0. To compute these constants,
we differentiate (36)
v(cid:48)
A,0(ρ, θ0) =
−1
C1(θ0)
tan
Using the Neumann boundary condition at ρ = 1 in (35), we get
C2(θ0) = −1.
Using (38) and (37) and the boundary condition at ρ = 0 in (35), we find that C1 is
solution of the transcendental equation,
In the regime σ = O(ε−3/2), we have
σεC1(θ0)
(1 − cos(θ0))
C1(θ0) =
Using (36),(38) and (40) in (36), we obtain to leading order
2(1 − cos(θ0)) + σε
.
= tan
2C1(θ0)
(cid:18) 1
(cid:19)
(cid:18) 1
(cid:19)
(cid:19)2(cid:33)
(cid:18)2(1 − cos(θ0)) + σε
(cid:18)
πσε(ρ − 1)
+ O
σε
πσε
πσε
(cid:32)
cos2
2(2(1 − cos(θ0)) + σε)
(cid:18)
10
(39)
.
(40)
(cid:19)(cid:19)
.
(41)
2
vA,0(ρ, θ0) = ln
+ ln
Using (41), (34) and (32), we conclude
vA(ρ, θ0) = ln
(cid:32)
(cid:19)2(cid:33)
(cid:18)2(1 − cos(θ0)) + σε
(cid:18)
(cid:18)
πσε
2
πσε(ρ − 1)
(cid:18) 4ε2
(cid:19)(cid:19)
eiθ0 − 14
+ ln
2(2(1 − cos(θ0)) + σε)
+ ln
cos2
√
+O(
ε).
(cid:19)
(42)
In particular the solution at ρ = 1 − √
(cid:32)
uA(1 − √
ε, θ0) = ln
8
ε is
(cid:18)2(1 − cos(θ0)) + σε
πσeiθ0 − 12
(cid:19)2(cid:33)
√
ε).
+ O(
(43)
We note that the three dimensional solution (43) is identical to the one obtained inside a
planar cusped-shaped domain [9].
Asymptotics of uB(θ) in region B
√
The asymptotic solution uA(ρ, θ) in A does not satisfy the boundary condition (28) at
θ = π. Indeed, ∂uA(ρ, θ)/∂θθ=π = 0, while the boundary condition (28) is ∂v/∂θθ=π =
−σ
ε/2 (cid:29) 1, thus a boundary layer should develop.
The boundary layer solution uB(θ) is derived by taking into account the θ derivatives
in eq. (28):
(1 − √
In small ε limit, for ρ = 1 − √
ε)eiθ − 14
4ρ2ε
∂2uB(θ)
∂θ2 +
sin(θ)(1 − cos(θ))
ε
∂ uB(θ)
∂θ
= −e−uB(θ).
(44)
ε, we have
4ε
ρeiθ(1 − √
(45)
which is constant. Using (45) in (44) and η = π − θ, we define uB(θ) = uB(η), leading to
ε) − 14
=
,
ε
4
Since 0 ≤ η ≤ √
∂2 uB(η)
∂η2 − 1
4
sin(η)(1 + cos(η))
∂ uB(θ)
∂η
= − ε
4
e−uB(η).
(46)
ε, we shall approximate the first order term and thus eq. (46) reduces to
∂2 uB(η)
∂η2 − η
2
∂ uB(θ)
∂η
= − ε
4
e−uB(η).
Using v(η) = uB(η) − ln (4/ε), eq. (47) is transformed to
−∂2v(η)
∂η2 +
η
2
∂v(η)
∂η
11
= e−v(η).
(47)
(48)
Using the boundary condition (28), we further reduce the solution v(η) to the equation
= e−v(η) + O(λε2)
∂η
∂v(η)
∂η
The solution is
−∂2v(η)
∂η2
∂v(η)
(cid:12)(cid:12)(cid:12)(cid:12)η=0
(cid:12)(cid:12)(cid:12)(cid:12)η=
(cid:32)
√
ε
σ
=
= 0.
√
ε
2
(cid:32)
v(η) = ln
2 C 2
1 cos2
(cid:33)(cid:33)
,
η + C2
2 C1
where
C2 = −√
(cid:32)
and C1 is solution of the transcendental equation
√
ε,
2 C1√
ε
arctan
σ
ε C1
2
(cid:33)
= 1.
In the limit σ (cid:29) 1, we have
C1 =
(cid:18)√
ε
2
2
π
+
√
2
ε
σ
(cid:19)
+ O
(cid:18) 1
√
(σ
(cid:19)
.
ε)3
(49)
(50)
(51)
(52)
(53)
We note that
(36) that for θ ∈ B, the asymptotic solution is
∂η
∂v(η)
η
2
is small, justifying our simplifications. We conclude from (53)-(51)-
(cid:114)
(θ − (π − √
ε))2
ε
(cid:19)
(cid:18)
1 − 4
σε
+ C0,
(54)
uB(θ) = ln cos2 π
2
where C0 is a constant that we find in the next paragraph by matching the solution in
two regions A and B.
A uniform approximation of u(ρ, θ) in Ωw
We now construct a uniform asymptotic approximation uunif (ρ, θ) in the region A ∪ B
(Fig. 3A) using uA(ρ, θ) with uB(ρ, θ) that match for θ = π − √
ε, leading to
(cid:0)1 − √
ε(cid:1) .
ε, π − √
C0 = uA
(55)
12
Using the analytical expression (43) of uA, we get
(cid:32)
(cid:33)
.
(56)
C0 = ln
(4 + σε)2
2(πσ)2
Thus,
(cid:114)
(θ − (π − √
ε))2
ε
(cid:18)
1 − 4
σε
(cid:32)
(cid:19)
+ ln
(4 + σε)2
2(πσ)2
(cid:33)
.
(57)
uB(θ) = ln cos2 π
2
Consequently, using (43) and (57) the solution in the funnel is
(cid:32)
ln
8
(cid:18) 2(1 − cos(θ)) + σε
(cid:114)
πσeiθ − 12
(θ − (π − √
ε))2
(cid:19)2(cid:33)
(cid:18)
,
ln cos2 π
2
ε
1 − 4
σε
(cid:32)
(cid:19)
+ ln
(cid:33)
,
(4 + σε)2
2(πσ)2
for θ ∈ [0, π − √
ε]
(58)
for θ ∈ [π − √
ε, π].
uunif (ρ, θ) =
The numerical solution of eq. (13) in Ωw and the approximation uunif (ρ, θ) of (58) are
shown in Fig. 3B.
2.4 Estimating the potential drop in Ωw
The difference of potential between the center of mass C and the tip of the funnel S (see
Fig. 3A) is defined as
∆f unnelu = u(C) − u(S),
where
u(S) = u(1 − √
ε, π) and u(C) = u(1 − √
√
ε),
ε, c
(59)
(60)
u is solution of eq. 13 and the constant c depends on the domain geometry and is
defined by the conformal mapping w (relation (17)). To compute ∆f unnelu, we use the
two differences
∆uA = uA(1 − √
ε, π) − uA(1 − √
√
ε),
ε, c
and
It follows that
∆uB = uB(π) − uB(π − √
ε).
∆f unnel = ∆uA + ∆uB.
13
(61)
(62)
(63)
To compute ∆uA, we use the analytical expression (43) for ρ = 1 − √
(cid:19)2
(cid:18)
uA(1 − √
ε, θ0) = − ln
eiθ0 − 14
8(1 − √
ε)2
σπ
2(1 − cos(θ0)) + σε
ε and any θ0,
+ O(ε).
(64)
At the point S (θ0 = π),
uA(S) = − ln
2σ2π2
(4 + σε)2 + 2 ln(1 − √
√
ε, we observe that for ε (cid:28) 1 in relation (64),
ε) + O(ε).
To estimate uA(C) for which θ0 = c
eiθ0 − 14 = c4ε2 + O(ε3),
and
2(1 − cos(c
√
ε)) + σε = ε(c2 + σ) + O(ε2).
We use (66) and (67), so eq. (64) reduces to
(cid:18) σπ
(cid:19)2
c2 + σ
+ 2 ln(1 − √
ε) + O (ε) .
uA(C) = − ln
c4
8
In the large σ limit,
(cid:18)
(cid:18)
1
σ
(cid:19)
(cid:19)
.
1
σ
uA(C) = − ln
π2c4
8
+ 2 ln(1 − √
ε) + O
ε,
.
Using uA(C) and uA(S), we conclude that
∆uA = − ln
2σ2π2
(4 + σε)2 + ln
π2c4
8
+ O
ε,
For σ (cid:29) 1, to leading order, the solution of eq. (70) does not depend on σ
∆uA ∼ − ln
24
c4ε2 .
We now estimate the difference ∆uB. We have from (54) that
(65)
(66)
(67)
(68)
(69)
(70)
(71)
(72)
(73)
and
ε) = C0
uB(π − √
uB(π) = ln sin2(cid:16) π
(cid:17)
+ C0.
σε
14
Using (72) and (73) in (62), we obtain
∆uB = ln sin2(cid:16) π
(cid:17)
.
σε
For σ (cid:29) 1, eq. (74) reduces to
∆uB = −2 ln σ + 2 ln
π
ε
+ O
(cid:18) 1
(cid:19)
σ2
.
(74)
(75)
Finally, using (70), (74) and (63), we find that the difference in the funnel is
∆u = ln sin2 π
σε
− ln
2σ2π2
(4 + σε)2 + ln
π2c4
8
+ O
ε,
.
(76)
(cid:18)
(cid:19)
1
σ
(cid:18) 1
(cid:19)
σ
The results in large σ limit found in (71), (75) and leads to
∆u = − ln σ2 + 2 ln
πc2
4
+ O
.
(77)
Equation (74) shows that for σ (cid:29) 1, the potential drop in the cusp-shaped funnel is
dominant in region B. We compare (Fig. 3C) expression (76) with the numerical solution
of 13. We note that the distribution of the potential inside a 3D solid funnel is to leading
order identical to the one we obtained inside a planar cusp [9].
3 The PNP equations in a cusp-shaped domain with
non-homogeneous surface charge density
When the surface charge density is not homogeneously distributed over the surface ∂Ω,
we expect a re-organization of the potential u of (13). we subdivide the surface ∂Ω into
three regions (Fig. 4),
∂Ω = ∂Ωε ∪ ∂Ωcusp ∪ ∂Ωbulk,
(78)
where ∂Ωε is the bottom of the funnel, ∂Ωcusp the funnel area and ∂Ωbulk the bulk surface.
The Neuman boundary conditions on each sub-regions are defined by
∂u(x)
∂n
∂u(x)
∂n
∂u(x)
∂n
=
−λε
∂Ωε on ∂Ωε
= − λcusp
∂Ωcusp on ∂Ωcusp
−λbulk
∂Ωbulk on ∂Ωbulk.
=
15
(79)
Using the compatibility condition obtained by integrating the Poisson equation (13)
(cid:90)
∂u(x)
dS = −λ.
∂Ω
∂n
we obtain that
We will use the notation
where j ∈ {ε , cusp , bulk}.
λ = λε + λcusp + λbulk.
σj =
λj
∂Ωj ,
(80)
(81)
(82)
Figure 4: Schematic representation of the ∂Ω boundary subregions. Subregions of the
boundary ∂Ω: the cusp ∂Ωcusp (red), the bulk ∂Ωbulk (blue) and (as shown in the inset panel)
the funnel bottom ∂Ωε (orange). Their respective surface charge densities are σbulk, σcusp and
σε.
3.1 PNP solutions for σcusp = 0, σbulk = σε = σ in 3D
To compute the solution of (13) for an uncharged funnel (σcusp = 0), we will use the
same conformal mapping (17) as describe above with now reflecting boundary condition
on ∂Ωcusp, which are invariant under the conformal mapping. As a result the boundary
conditions on the two like-style arcs of the domain Ωw are also reflective. Consequently,
instead of searching a solution in the banana-shaped domain Ωw, we will construct it in
the circular arc as a one-dimensional solution.
The boundary value problem (28) in the conformal image Ωw becomes
exp(cid:8)−v(eiθ)(cid:9)
(83)
v(cid:48)(cid:48) − 4 sin(θ)(1 − cos(θ))
√
ε4
√
v(cid:48)(c
eiθ − 1 − eiθ
v(cid:48) = −
ε) = 0
√
ε4
4ε
eiθ − 1 − eiθ
√
v(cid:48)(π) = −σ
ε
.
2
16
y0CNS To construct an asymptotic approximation to the solution of (83) in the limits ε → 0 and
σ → ∞, we first construct the outer-solution in the form of a series in powers of ε, which
ε. After dropping the terms in
is an approximation valid away from the boundary θ = c
ε in 83, we obtain the outer solution by a direct integration
v1(θ) = −A(θ − sin(θ)) + v(0),
(84)
√
where v(0) and A are constants. The outer solution (84) cannot satisfy all boundary con-
ditions, consequently a boundary layer correction is needed at θ = π. An approximation
of the solution can be obtained by freezing the power-law term and neglecting the first
order derivatives in (83), for which the equation is for a generic parameter b > 0,
The solution is [8]
d2
dθ2 vb(θ) + be−vb(θ) = 0,
vb(θ) = ln cos2
dvb(0)
(cid:19)
dθ
(cid:18) b
2
= vb(0) = 0.
θ
.
(85)
Putting the outer and boundary layer solutions together gives the uniform asymptotic
approximation
(cid:19)
(cid:18) b
2
θ
.
(86)
(87)
(88)
The condition at θ = π gives that
The compatibility condition for (13),
yunif(θ) = −A(θ − sin(θ)) + v(0) + ln cos2
unif(π) = −2A − b tan
y(cid:48)
(cid:90)
λε + λbulk =
√
π = −σε
2
b
2
ε
.
exp{−u(x)}dSx,
gives in Ωw that
(cid:90)
Ωw
Ω
exp{−v(w)}
dw
φ(cid:48)(φ−1(w)) .
λε + λbulk =
Using the uniform approximation (86) in the compatibility condition (88), we obtain the
second condition
λε + λbulk = 8
√
π(cid:90)
ε e−v(0)
ε(cid:90)
√
c
√
π/
ε
≈ 8 e−v(0)
ε
1
cos2 b
2
θ
dθ
exp{A(θ − sin(θ))}
eiθ(1 − √
ε) − 14
√
√
εξ − sin(
1 + ξ22
exp{A(
√
1
cos2 b
2
εξ
0
17
εξ))}
dξ,
(89)
where we used the change of variable θ =
√
εξ. Integrating by parts, we get for ε (cid:28) 1
2
√
b
tan
ε
b
2
π
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)1 +
eAπ
(cid:18) π√
ε
(cid:19)2(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)2 −
ε(cid:90)
√
π/
0
, (90)
√
2
b
ε
tan
b
2
θ Ψ(θ) dθ
λε + λbulk ∼ 8 e−v(0)
ε
where
√
√
εξ − sin(
exp{A(
1 + ξ22
Thus, it remains to solve the asymptotic equation
Ψ(ξ) =
d
dξ
εξ))}
λε + λbulk ∼ 8 e−v(0)ε1/2
bπ4 tan
exp{Aπ} + O
πb
2
(cid:20) 2
.
(cid:18)
ln
(cid:12)(cid:12)(cid:12)(cid:12)cos
(cid:12)(cid:12)(cid:12)(cid:12)(cid:19)(cid:21)
.
πb
2
for A and b in the limit ε → 0. We consider the limiting case where
√
A
σε
ε
(cid:28) 1 for σε → ∞,
for which condition (87) can be simplified and gives to leading order
We conclude from expression 86 that
(91)
(92)
(93)
(94)
(95)
(96)
that is, for σε
√
ε (cid:28) 1 (94) gives
b tan
πb
2
=
It follows from (92) using (82) that
√
ε
,
σε
2
√
tan
b
2
π ∼σε
2
ε
.
A = − 1
π
yunif(θ) =
1
π
b ≈ 1 − 4
π
ln
ln
+ ln cos2
√
1
σε
,
ε
π4∂Ωε
(cid:18)
8ε
1 +
(cid:18)
8ε
1 +
λbulk
λε
π4∂Ωε
1 − 4
√
1
π
ε
σε
2
θ
18
λbulk
λε
− v(0)
.
(cid:19)
(θ − sin(θ))
− v(0)
(cid:19)
+ v(0).
We compare in Fig. 5A-D, the uniform approximation (96) with numerical simulations of
the reduced eq. (83) and the three-dimensional numerical solution (eq13). The difference
of potential ∆yunif = yunif(0) − yunif(π), can now be estimated using (96) and we obtain
ln
(cid:18)
8ε
1 +
(cid:19)
λbulk
λε
π4∂Ωε
∆yunif = −
− v(0)
− ln sin2
(cid:18) 2
√
(cid:19)
.
(97)
ε
σε
In the small ε limit, the constant v(0) = O(1) can be neglected. We compare the analytical
expression for difference of potential (97) with the result of the reduced equation (83)
computed numerically in Fig. 5E. We note that the solution in 3D differs from 2D, as
shown in Fig. 5F.
Figure 5: Numerical (13)-(83) versus analytical (96) solutions with zero Neumann
boundary conditions, except at the end of the funnel. A-D Analytical (dashed
green) obtained from (96) and numerical solutions (13) (blue) computed in 3D and the
1D reduced equation (83) (dashed red). E. Potential difference v(0) − v(π) computed
numerically from (83) (blue) and the asymptotics (97). F. Comparison of eq.
(13)
numerical solutions in 2D (red) and 3D (blue).
19
13839697=103=2500=104ABDCFE=5000v()v()v(0)-v()510x103 4 PNP solution for σε (cid:54)= σcusp and σbulk = O(1)
4.1 Analytical representation of the PNP solution
We study here the effect of the charge density σcusp located on the cusp-shaped funnel on
the solution u(x) of
−∆u(x) = exp{−u(x)} for x ∈ Ω
(98)
∂u(x)
∂n
∂u(x)
∂n
∂u(x)
∂n
= −σε on ∂Ωε
= −σcusp on ∂Ωcusp
= −σbulk on ∂Ωbulk,
√
in the small ε and large σcusp limits, such as σcusp
O(1).
√
ε (cid:29) 1, σε/σcusp = O(1) and σbulk =
ε, π −
ε] (region A, Fig. 2), the angular derivatives of uunif can be neglected. We thus use the
√
ε limit, we have shown (section 2) that for θ in the range [c
In the large σcusp
√
result of eq. (43) by changing σ by σcusp to obtain
(cid:32)
(cid:19)2(cid:33)
(cid:18)2(1 − cos(θ)) + σcuspε
(cid:18)
πσcuspε
(cid:18)
2
πσcusp(ρ − ε)
(cid:19)
(cid:18) 4ε2
(cid:19)(cid:19)
eiθ − 14
+ ln
2(2(1 − cos(θ)) + σcuspε)
ucusp(ρ, θ) = ln
(99)
For ρ = 1 − √
ε).
+ ln
cos2
√
+O(
ε, π − √
√
(cid:32)
ε, θ) = ln
8
ε and θ ∈ [c
ucusp(1 − √
ε], we get
(cid:18)2(1 − cos(θ)) + σcuspε
πσcuspeiθ − 12
(cid:19)2(cid:33)
√
+ O(
ε).
(100)
To construct a uniform solution uunif , we match to a solution uB in region
B = {(ρ, θ), θ ∈ [π − √
ε, π] and ρ = 1 − √
(cid:18)2(1 − cos(θ)) + σcuspε
ε}. We obtain the general expression
for θ ∈ [0, π − √
(cid:19)2(cid:33)
ε]
(cid:32)
πσcuspeiθ − 12
uunif (ρ, θ) =
ln
8
uB(θ)
Thus the difference of potential u(C)− u(S) between the center of mass C and the funnel
base S is then
(101)
for θ ∈ [π − √
ε, π].
(cid:18)
(cid:19)
ε,
1
σcusp
,
(102)
V (C) − V (S) = − ln
2σcuspπ2
(4 + σcuspε)2 + ln
π2c4
8
+ ∆uB + O
20
√
where c
ε is the angular coordinate of the mapped center of mass C in Ωw. We compare
in Fig. 6A-B the analytical expression (dashed) of (101) with the three-dimensional
numerical simulations (solid) of u (eq. (98)). When uB is given by expression (57) with
condition σε
ε (cid:29) 1, then the difference of potential is given by
√
(cid:18)
(cid:19)
u(C) − u(S) =
kT
e
ln sin2 π
σεε
− ln
2σcuspπ2
(4 + σcuspε)2
+ O(1).
(103)
The two conditions σε/σcusp = O(1) and σbulk = O(1) imply that the uniform solution
is not affected by the bulk or the tip of the cusp. This is in contrast with the results
computed for σcusp = 0 (section 3.1) for which the solution in the cusp is entirely defined
by the surface charge densities σcusp and σbulk (see eq. (96)). However, when the previous
conditions are not satisfied (σε/σcusp = O(1) is not verified), the numerical solution (red)
and the analytical expression (101) (dashed blue) do not agree (Fig. 6A-B).
4.2 PNP solution with reflecting boundary at the end of the
funnel
When we impose a reflecting boundary condition at the end of the cusp ∂Ωε (σε = 0), we
construct an approximation of equation (98) in the regimes ε (cid:28) 1 and σcusp (cid:29) 1 in the
following regime of parameters σcusp
To construct the approximation uunif in Ωw, we use expression in the cusp (101), where
the solution uB is constructed by extending ucusp(ρ, θ) to region B. We have
ε (cid:29) 1 and σbulk = O(1).
√
To show that ucusp satisfies the same boundary condition, we differentiate ucusp(ρ, θ), (eq.
(99)), in θ at θ = π:
∂uunif (ρ, θ)
∂θ
∂ucusp(ρ, θ)
= 0.
(cid:12)(cid:12)(cid:12)(cid:12)θ=π
(cid:12)(cid:12)(cid:12)(cid:12)θ=π
(cid:18)2(1 − cos(θ)) + σcuspε
= 0.
πσcuspeiθ − 12
∂θ
(cid:32)
(cid:19)2(cid:33)
(104)
(105)
We conclude that ucusp matches at θ = π the boundary condition satisfied by the solution
u(x) for σε = 0. Consequently,
uunif (ρ, θ) = ln
8
.
(106)
Thus the difference of potential between the funnel base S and the center of mass C is
u(C) − u(S) = − ln
2σcuspa2π2
(4 + σcuspε)2 + ln
π2c4
8
.
(107)
21
We obtain a good agreement between the analytical expression (eq. (106)) and the three
dimensional numerical solution of (98) (Fig. 6C).
The result obtained from (107) can be used to model the voltage in a domain with a
cusp-shaped funnel connecting a reservoir with a fixed electrical potential and zero electric
field at the of funnel-reservoir junction. This no field condition is satisfied when σε = 0.
This result can be applied to the electrical properties of dendritic spines with a short neck
(see [10], p.28, Fig. 3.9, spine 7), approximated by a cusp and the parent dendrite as a
reservoir.
Figure 6: Comparison of numerical and analytical solutions for non-homogeneous
surface charge density A. Numerical (eq. (98)) in 3D (solid) and analytical (eq. (101))
(dashed) solutions for σcusp = 1000 and σcusp = 1, σbulk = σε = 2500. B. Magnification of
panel A in the region of θ = π. C. 3D Numerical (solid) from eq. (98) and analytical (eq.(106))
(dashed) solutions computed for σcusp = 10, 100, 1000 and 4000, where σε = 0 and σcusp = σbulk.
Here ε = 0.01.
5 Discussion and conclusion
Based on the steady-state solution of the Poisson-Nernst-Planck equations, we derived
here electrostatic properties of non-electro-neutral electrolytes confined in a cusp-shaped
funnel geometry. We showed that the local curvature and the distribution of surface
charges shape the electrical landscape within small domains. The new electrical proper-
ties have been obtained for a dominant ionic specie, in an electrolyte having an excess of
charges in two dimensions in [9]. The new mathematical methods consist here in the con-
struction of an asymptotic expansion of the nonlinear PNP equations inside 3D domains,
with non-homogeneous Neumann boundary conditions.
Using asymptotics methods validated by numerical solutions of the PNP equations,
we found several explicit voltage drops: first, for a surface charge density homogeneously
distributed, the electrical potential distribution in 3D and 2D domains is quite similar
to leading order potential inside a planar cusp (Fig. 3). However, the voltage inside
an uncharged funnel (Fig. 5), associated to the condition σcusp = 0 varies significantly
22
------3DNumerics:Approx.(101):=2500=ABC=03DNumericsApprox.(105):=1010010004000 between a 2D and 3D domain with a cusp funnel. We summarize in table 1 the results we
have obtained in the three sections above, where we reintroduce the physical units and
used σi = σizeRc/kT (section 2).
The presence of negative ions in Ω may slightly reduces the voltage. However, as
shown in [9], accounting for negative charges carried by chloride anions present in the
cytosol at physiological concentration [3], does not alter the voltage to leading order.
Consequently the voltages summarized in table 1 provide insights for understanding the
electro-diffusion properties. The present results could be used in the design accurately
quartz nanopipettes with an optimal shape [12, 23, 24]. It would be interesting to vary
the surface charge densities [25] in some sub-regions σbulk, σcusp or σε (79).
Finally, the present analytical results can be used to predict the voltage drop in neu-
ronal microdomains such as dendritic spines [10]. The local curvature is certainly a key
factor in modulating the voltage and thus we are beginning to understand how nano-
and micrometer geometry can encode synaptic modulation, that underlyes learning and
memory in the brain. Indeed, in compartment such as dendritic spines, the high curva-
ture variation play a major role in converting injected current into voltage. This effect
may as well influence the propagation and genesis of local depolarization in excitable cells
[6, 21, 22].
(cid:32)
kT
e
ln sin2 kT π
eεσ
V (C) − V (S)
− 2 ln
√
2 eπRcσ
4kT + eεσ
(cid:33)
+ O(1)
Conditions
σ (cid:29) 1
σcusp = 0
σε (cid:29) 1
σbulk = C ste
σcusp (cid:54)= σε
ε (cid:29) 1
σcusp
σbulk = C ste
√
kT
e
(cid:18)
− ln
− ln sin2
√
2kT
eσε
Rc ε
8Rc ε
(cid:32)
π4∂ Ωε (1 + σbulk/σε)
√
2 eπRcσcusp
4kT + eεσcusp
ln sin2 kT π
eεσε
− 2 ln
kT
e
+ O(1)
(cid:19)
+ O(1)
(cid:33)
Table 1: Electrodiffusion laws for voltage drop for various surface charge densities
6 Appendix
6.1 Radial derivative under the Mobius map (17)
We shall describe in this appendix the computation step to reduce the first order radial
derivative from (16) leading to the result (28) in section 2.2. First, we note that in complex
23
coordinates, we have
where we define
∂u(r, z)
∂r
= (cid:60)e (∇u(ξ)) ,
∇u(ξ) =
∂u(r, z)
∂r
+ i
∂u(r, z)
∂ z
.
Under the conformal mapping (17), the gradient (109) is transformed as follows
∇u(ξ) = ∇wv(w) w(cid:48)(ξ).
Using the notation w = X + iY , we get
∇wv(w) =
∂v(X, Y )
∂X
+ i
∂v(X, Y )
∂Y
.
We define the real functions w1(X, Y ) and w2(X, Y ) that satisfy
w(cid:48)(ξ) = w(cid:48)(w−1(X, Y )) = w1(X, Y ) + i w2(X, Y ).
Using (17) (Mobius transformation), we get
w(cid:48)(w−1(X, Y )) =
(1 + αw)2
1 − α
.
Equations (112) and (113) lead to
w1(X, Y ) =
w2(X, Y ) = −2αY (1 + αX)
1 − α2
1 − α2
(1 + αX)2 − α2Y 2
.
From (108)-(110)-(111)-(112), we obtain
∂u
∂r
= w1(X, Y )
∂v(X, Y )
∂X
− w2(X, Y )
∂v(X, Y )
∂Y
.
(108)
(109)
(110)
(111)
(112)
(113)
(114)
(115)
Due to the round geometry of the banana-shaped domain Ωw, it is convenient to switch
from Cartesian coordinates (X, Y ) to polar coordinates (ρ, θ). Setting v(X, Y ) = v(ρ, θ),
we get
∂v(X, Y )
∂X
∂v(X, Y )
∂Y
=
=
∂v(ρ, θ)
∂ρ
∂v(ρ, θ)
∂ρ
∂ρ
∂X
∂ρ
∂Y
+
+
∂v(ρ, θ)
∂θ
∂v(ρ, θ)
∂θ
∂θ
∂X
∂θ
∂Y
24
(116)
where,
∂ρ
∂X
∂θ
∂X
= cos(θ)
= −sin(θ)
ρ
,
,
∂ρ
∂Y
∂θ
∂Y
= sin(θ)
=
cos(θ)
ρ
.
Using (116) and (117) in (115), it follows that
∂ u(r, z)
∂r
=
∂v(ρ, θ)
(cos(θ) w1(ρ, θ) − sin(θ) w2(ρ, θ))
∂ρ
∂v(ρ, θ)
∂θ
− 1
ρ
(sin(θ) w1(ρ, θ) + cos(θ) w2(ρ, θ)) ,
where we set wi(ρ, θ) = wi(X, Y ) for i ∈ {1 , 2}, such as
w1(ρ, θ) =
1 − α2ρ2 + 2αρ cos(θ)(1 + αρ cos(θ))
1 − α2
w2(ρ, θ) = −2αρ sin(θ)
1 + αρ cos(θ)
1 − α2
.
(117)
(118)
(119)
Using (119) and (118), we obtain to leading order
1
r
∂ u(r, z)
∂r
= −ρ(1 − cos(θ))2
ε3/2
∂v(ρ, θz)
∂ρ
− sin(θ)(1 − cos(θ))
ε
∂v(ρ, θz)
∂θ
.
(120)
6.2 The numerical procedure
Numerical solutions were constructed by the COMSOL Multiphysics 5.0 (BVP problems),
Maple 2015 (Shooting problems) and Matlab R2015 (Conformal mapping). The boundary
value problems in 1D, 2D, and 3D were solved by the finite elements method in the COM-
SOL 'Mathematics' package. We used an adaptive mesh refinement to ensure numerical
convergence for large value of the parameters σ, σε, σbulk and σcusp. We solved the PDEs
by the shooting procedure for boundary value problems using Runge-Kutta fourth-order
method.
25
References
[1] Schuss, Z., B. Nadler and R.S. Eisenberg, Derivation of Poisson and Nernst-
Planck equations in a bath and channel from a molecular model, Phys. Rev. E, 64
(2001).
[2] Singer, A. and J. Norbury, A Poisson-Nernst-Planck model for biological ion
channelsan asymptotic analysis in a three-dimensional narrow funnel,SIAM of Ap-
plied Mathematics, 7(3), pp.949 -- 968 (2009).
[3] Hille, B., Ion Channels of Excitable Membranes, Third Edition, Sinauer Associates,
(2001).
[4] Bezanilla, F., How membrane proteins sense voltage, Nat Rev Mol Cell Biol., 9,
pp.323 -- 332 (2008).
[5] Bourne J.N. and K.M. Harris, Balancing structure and function at hippocampal
dendritic spines, Annu. Rev. Neurosci., 31, pp.47 -- 67 (2008).
[6] Holcman, D. and R. Yuste, The new nanophysiology: regulation of ionic flow in
neuronal subcompartments, Nature Reviews Neuroscience, 16, pp.685 -- 692 (2015).
[7] Debye, P. and E. Huckel, Zur Theorie der Elektrolyte. I. Gefrierpunktserniedri-
gung und verwandte Erscheinungen Physikalische Zeitschrift, 24(9), pp.185 -- 206
(1923).
[8] Cartailler, J., Z. Schuss and D. Holcman, Analysis of the Poisson-Nernst-
Planck equation in a ball for modeling the voltage-current relation in neurobiological
microdomains, Physica D: Nonlinear Phenomena, 339, pp.39 -- 48 (2016).
[9] Cartailler, J., Z. Schuss and D. Holcman, Geometrical effects on nonlinear
electrodiffusion in cell physiology, J. Nonlin. Sci., doi:10.1007/s00332-017-9393-2, pp.
1 -- 30 (2017).
[10] Yuste, R., Dendritic Spines, The MIT Press, Cambridge, MA (2010).
[11] Cartailler J., Kwon T, Yuste R., Holcman D., Electro-diffusion modulation
of synaptic input in dendritic spines using deconvolved voltage sensor time series,
doi:10.1101/097279, (2016).
[12] Holcman, D. and Z. Schuss, Brownian motion in dire straits,Multiscale Modeling
& Simulation,10(4), pp.1204 -- 1231 (2012).
[13] Ward, MJ.; Keller, Joseph B. Nonlinear eigenvalue problems under strong localized
perturbations with applications to chemical reactors. Stud. Appl. Math. 85 (1991),
no. 1, 128.
26
[14] Ward, MJ.; Henshaw, William D.; Keller, Joseph B. Summing logarithmic expansions
for singularly perturbed eigenvalue problems. SIAM J. Appl. Math. 53 (1993), no. 3,
799828.
[15] Pillay, S., A. Peirce, T. Kolokolnikov and M. Ward,An Asymptotic Anal-
ysis of the Mean First Passage Time for Narrow Escape Problems: Part I: Two-
Dimensional Domains, SIAM Multiscale Modeling and Simulation, 8(3), pp.803-835
(2010).
[16] Delgado, M. Ward and D. Coombs, Conditional Mean First Passage Times to
Small Traps in a 3-D Domain with a Sticky Boundary, SIAM J. Multiscale Analysis
and Simulation, 13(4), pp.1224 -- 1258 (2015).
[17] Lindsay, A., A. Bernoff and M. Ward,First Passage Statistics for the Capture
of a Brownian Particle by a Structured Spherical Target with Multiple Surface Traps,
SIAM J. Multiscale Modeling and Simulation, 15(1), pp.74-109 (2016).
[18] Henricci, P., Applied and Computational Complex Analysis, Volume 1: Power
Series Integration Conformal Mapping Location of Zero, Wiley-Blackwell, Volume 1
(1997).
[19] Sylantyev S., Savtchenko L.P., Ermolyuk Y., Michaluk P., Rusakov
D.A., Spike-driven glutamate electrodiffusion triggers synaptic potentiation via a
homer-dependent mGluR-NMDAR link, Neuron, 77(3), pp.528 -- 41 (2013).
[20] Ahirwar, D.K., M.W. Nasser, T.H. Jones, E.K. Sequin, J.D. West, T.L.
Henthorne, J. Javor, A.M. Kaushik, R.K. Ganju and V.V. Subramaniam,
Non-contact method for directing electrotaxis, Scientific Reports, 5 (2015).
[21] Rall, W., C. Koch, and I. Segev, Cable Theory for Dendritic Neurons. In Meth-
ods in Neuronal Modeling: from Synapses to Networks, (eds), Cambridge, Mass., The
MIT Pres, pp.9 -- 63 (1989).
[22] Qian, N. and T.J. Sejnowski, An electro-diffusion model for computing mem-
brane potentials and ionic concentrations in branching dendrites, spines and axons,
Biol. Cybern., 62, pp.1 -- 15 (1989).
[23] Perry, D., D. Momotenko, R.A. Lazenby, M. Kang and P.R. Unwin,
Characterization of Nanopipettes, Anal. Chem., 88(10), pp.5523 -- 30 (2016).
[24] Jayant K., Hirtz J.J., Plante I.J., Tsai D.M., De Boer W.D., Semonche
A., Peterka D.S., Owen J.S., Sahin O., Shepard K.L. and Yuste R.,
Targeted intracellular voltage recordings from dendritic spines using quantum-dot-
coated nanopipettes, Nat. Nano., 12(4), pp.335 -- 342 (2017).
27
[25] Sparreboom, W., A. van den Berg and J.C.T. Eijkel, Principles and appli-
cations of nanofluidic transport, Nat. Nano., 4, pp.713 -- 720 (2009).
28
|
1911.09451 | 1 | 1911 | 2019-11-21T13:08:17 | Deep neuroethology of a virtual rodent | [
"q-bio.NC"
] | Parallel developments in neuroscience and deep learning have led to mutually productive exchanges, pushing our understanding of real and artificial neural networks in sensory and cognitive systems. However, this interaction between fields is less developed in the study of motor control. In this work, we develop a virtual rodent as a platform for the grounded study of motor activity in artificial models of embodied control. We then use this platform to study motor activity across contexts by training a model to solve four complex tasks. Using methods familiar to neuroscientists, we describe the behavioral representations and algorithms employed by different layers of the network using a neuroethological approach to characterize motor activity relative to the rodent's behavior and goals. We find that the model uses two classes of representations which respectively encode the task-specific behavioral strategies and task-invariant behavioral kinematics. These representations are reflected in the sequential activity and population dynamics of neural subpopulations. Overall, the virtual rodent facilitates grounded collaborations between deep reinforcement learning and motor neuroscience. | q-bio.NC | q-bio |
DEEP NEUROETHOLOGY OF A VIRTUAL RODENT
Josh Merel(cid:63)1, Diego Aldarondo(cid:63)2,3, Jesse Marshall(cid:63)3,4, Yuval Tassa1,
Greg Wayne1, Bence Olveczky3,4
1DeepMind, London, UK.
2Program in Neuroscience, 3Center for Brain Science, 4Department of Organismic and
Evolutionary Biology, Harvard University, Cambridge, MA 02138, USA.
[email protected], [email protected],
jesse d [email protected]
ABSTRACT
Parallel developments in neuroscience and deep learning have led to mutually
productive exchanges, pushing our understanding of real and artificial neural net-
works in sensory and cognitive systems. However, this interaction between fields
is less developed in the study of motor control. In this work, we develop a virtual
rodent as a platform for the grounded study of motor activity in artificial models
of embodied control. We then use this platform to study motor activity across
contexts by training a model to solve four complex tasks. Using methods famil-
iar to neuroscientists, we describe the behavioral representations and algorithms
employed by different layers of the network using a neuroethological approach to
characterize motor activity relative to the rodent's behavior and goals. We find
that the model uses two classes of representations which respectively encode the
task-specific behavioral strategies and task-invariant behavioral kinematics. These
representations are reflected in the sequential activity and population dynamics of
neural subpopulations. Overall, the virtual rodent facilitates grounded collabora-
tions between deep reinforcement learning and motor neuroscience.
1
INTRODUCTION
Animals have nervous systems that allow them to coordinate their movement and perform a diverse
set of complex behaviors. Mammals, in particular, are generalists in that they use the same general
neural network to solve a wide variety of tasks. This flexibility in adapting behaviors towards many
different goals far surpasses that of robots or artificial motor control systems. Hence, studies of the
neural underpinnings of flexible behavior in mammals could yield important insights into the classes
of algorithms capable of complex control across contexts and inspire algorithms for flexible control
in artificial systems.
Recent efforts at the interface of neuroscience and machine learning have sparked renewed interest in
constructive approaches in which artificial models that solve tasks similar to those solved by animals
serve as normative models of biological intelligence. Researchers have attempted to leverage these
models to gain insights into the functional transformations implemented by neurobiological circuits,
prominently in vision (Khaligh-Razavi & Kriegeskorte, 2014; Yamins et al., 2014; Kar et al., 2019),
but also increasingly in other areas, including audition (Kell et al., 2018) and navigation (Banino
et al., 2018; Cueva & Wei, 2018). Efforts to construct models of biological locomotion systems have
informed our understanding of the mechanisms and evolutionary history of bodies and behavior
(Grillner et al., 2007; Ijspeert et al., 2007; Ramdya et al., 2017; Nyakatura et al., 2019). Neural
control approaches have also been applied to the study of reaching movements, though often in
constrained behavioral paradigms (Lillicrap & Scott, 2013), where supervised training is possible
(Sussillo et al., 2015; Michaels et al., 2019).
While these approaches model parts of the interactions between animals and their environments
(Chiel & Beer, 1997), none attempt to capture the full complexity of embodied control, involving
how an animal uses its senses, body and behaviors to solve challenges in a physical environment.
(cid:63)Equal contribution.
1
The development of models of embodied control is valuable to the field of motor neuroscience,
which typically focuses on restricted behaviors in controlled experimental settings. It is also valuable
for AI research, where flexible models of embodied control could be applicable to robotics.
Here, we introduce a virtual model of a rodent to facilitate grounded investigation of embodied motor
systems. The virtual rodent affords a new opportunity to directly compare principles of artificial
control to biological data from real-world rodents, which are more experimentally accessible than
humans. We draw inspiration from emerging deep reinforcement learning algorithms which now
allow artificial agents to perform complex and adaptive movement in physical environments with
sensory information that is increasingly similar to that available to animals (Peng et al., 2016; 2017;
Heess et al., 2017; Merel et al., 2019a;b). Similarly, our virtual rodent exists in a physical world,
equipped with a set of actuators that must be coordinated for it to behave effectively. It also possesses
a sensory system that allows it to use visual input from an egocentric camera located on its head and
proprioceptive input to sense the configuration of its body in space.
There are several questions one could answer using the virtual rodent platform. Here we focus
on the problem of embodied control across multiple tasks. While some efforts have been made to
analyze neural activity in reduced systems trained to solve multiple tasks (Song et al., 2017; Yang
et al., 2019), those studies lacked the important element of motor control in a physical environment.
Our rodent platform presents the opportunity to study how representations of movements as well as
sequences of movements change as a function of goals and task contexts.
To address these questions, we trained our virtual rodent to solve four complex tasks within a physi-
cal environment, all requiring the coordinated control of its body. We then ask "Can a neuroscientist
understand a virtual rodent?" -- a more grounded take on the originally satirical "Can a biologist fix a
radio?" (Lazebnik, 2002) or the more recent "Could a neuroscientist understand a microprocessor?"
(Jonas & Kording, 2017). We take a more sanguine view of the tremendous advances that have been
made in computational neuroscience in the past decade, and posit that the supposed 'failure' of these
approaches in synthetic systems is partly a misdirection. Analysis approaches in neuroscience were
developed with the explicit purpose of understanding sensation and action in real brains, and often
implicitly rooted in the types of architectures and processing that are thought relevant in biological
control systems. With this philosophy, we use analysis approaches common in neuroscience to ex-
plore the types of representations and dynamics that the virtual rodent's neural network employs to
coordinate multiple complex movements in the service of solving motor and cognitive tasks.
2 APPROACH
2.1 VIRTUAL RODENT BODY
Figure 1: (A) Anatomical skeleton of a rodent (as reference; not part of physical simulation). (B) A
body designed around the skeleton to match the anatomy and model collisions with the environment.
(C) Purely cosmetic skin to cover the body. (D) Semi-transparent visualization of (A)-(C) overlain.
We implemented a virtual rodent body (Figure 1) in MuJoCo (Todorov et al., 2012), based on mea-
surements of laboratory rats (see Appendix A.1). The rodent body has 38 controllable degrees of
freedom. The tail, spine, and neck consist of multiple segments with joints, but are controlled by
tendons that co-activate multiple joints (spatial tendons in MuJoCo).
The virtual rodent has access to proprioceptive information as well as "raw" egocentric RGB-camera
(64×64 pixels) input from a head-mounted camera. The proprioceptive inputs include internal
joint angles and angular velocities, the positions and velocities of the tendons that provide actua-
2
tion, egocentric vectors from the root (pelvis) of the body to the positions of the head and paws, a
vestibular-like upright orientation vector, touch or contact sensors in the paws, as well as egocentric
acceleration, velocity, and 3D angular velocity of the root.
2.2 VIRTUAL RODENT TASKS
Figure 2: Visualizations of four tasks the virtual rodent was trained to solve: (A) jumping over gaps
("gaps run"), (B) foraging in a maze ("maze forage"), (C) escaping from a hilly region ("bowl es-
cape"), and (D) touching a ball twice with a forepaw with a precise timing interval between touches
("two-tap").
We implemented four tasks adapted from previous work in deep reinforcement learning and motor
neuroscience (Merel et al., 2019a; Tassa et al., 2018; Kawai et al., 2015) to encourage diverse motor
behaviors in the rodent. The tasks are as follows: (1) Run along a corridor, over "gaps", with a
reward for traveling along the corridor at a target velocity (Figure 2A). (2) Collect all the blue orbs
in a maze, with a sparse reward for each orb collected (Figure 2B). (3) Escape a bowl-shaped region
by traversing hilly terrain, with a reward proportional to distance from the center of the bowl (Figure
2C). (4) Approach orbs in an open field, activate them by touching them with a forepaw, and touch
them a second time after a precise interval of 800ms with a tolerance of ±100ms; there is a time-out
period if the touch is not within the tolerated window and rewards are provided sparsely on the first
and second touch (Figure 2D). We did not provide the agent with a cue or context indicating its task.
Rather, the agent had to infer the task from the visual input and behave appropriately.
2.3 TRAINING A MULTI-TASK POLICY
Figure 3: The virtual rodent agent architecture. Egocentric visual image inputs are encoded into
features via a small residual network (He et al., 2016) and proprioceptive state observations are
encoded via a small multi-layer perceptron. The features are passed into a recurrent LSTM module
(Hochreiter & Schmidhuber, 1997). The core module is trained by backpropogation during training
of the value function. The outputs of the core are also passed as features to the policy module (with
the dashed arrow indicating no backpropogation along this path during training) along with shortcut
paths from the proprioceptive observations as well as encoded features. The policy module consists
of one or more stacked LSTMs (with or without skip connections) which then produce the actions
via a stochastic policy.
Emboldened by recent results in which end-to-end RL produces a single terrain-adaptive policy
(Peng et al., 2016; 2017; Heess et al., 2017), we trained a single architecture on the multiple motor-
control-reliant tasks (see Figure 3). To train a single policy to perform all four tasks, we used
an IMPALA-style setup for actor-critic DeepRL (Espeholt et al., 2018); parallel workers collected
rollouts, logged them to a replay, from which a central learner sampled data to perform updates. The
3
value-function critic was trained using off-policy correction via V-trace. To update the actor, we
used a variant of MPO (Abdolmaleki et al., 2018) where the E-step is performed using advantages
determined from the empirical returns and the value-function, instead of the Q-function (Song et al.,
2019). Empirically, we found that the "escape" task was more challenging to learn during interleaved
training relative to the other tasks. Consequently, we present results arising from training a single-
task expert on the escape task and training the multi-task policies using kickstarting for that task
(Schmitt et al., 2018), with a weak coefficient (.001 or .005). Kickstarting on this task made the seeds
more reliably solve all four tasks, facilitating comparison of the multi-task policies with different
architectures (i.e. the policy having 1, 2, or 3 layers, with or without skip connections across those
layers). The procedure yields a single neural network that uses visual inputs to determine how to
behave and coordinates its body to move in ways required to solve the tasks. See video examples of
a single policy solving episodes of each task: gaps, forage, escape, and two-tap.
3 ANALYSIS
Figure 4: Ethology of the virtual rodent. (A) Example jumping sequence in gaps run task with a rep-
resentative subset of recorded behavioral features. Dashed lines denote the time of the corresponding
frames (top). (B) tSNE embedding of 60 behavioral features describing the pose and kinematics of
the virtual rodent allows identification of rodent behaviors. Points are colored by hand-labeling of
behavioral clusters identified by watershed clustering. (C) The first two principal components of
different behavioral features reveals that behaviors are more shared across tasks at short, 5-25 Hz
timescales (fast kinematics), but no longer 0.3-5 Hz timescales (slow kinematics).
We analyzed the virtual rodent's neural network activity in conjunction with its behavior to char-
acterize how it solves multiple tasks (Figure 4A). We used analyses and perturbation techniques
adapted from neuroscience, where a range of techniques have been developed to highlight the prop-
erties of real neural networks. Biological neural networks have been hypothesized to control, select,
and modulate movement through a variety of debated mechanisms, ranging from explicit neural
representations of muscle forces and behavioral primitives, to more abstract production of neural
dynamics that could underly movement (Graziano, 2006; Kalaska, 2009; Churchland et al., 2012).
A challenge with nearly all of these models however is that they have largely been inspired by find-
ings from individual behavioral tasks, making it unclear how to generalize them to a broader range
of naturalistic behaviors. To provide insight into mechanisms underlying movement in the virtual
rodent, and to potentially give insight by proxy into the mechanisms underlying behavior in real
rats, we thus systematically tested how the different network layers encoded and generated different
aspects of movement.
For all analyses we logged the virtual rodent's kinematics, joint angles, computed forces, sensory
inputs, and the cell unit activity of the LSTMs in core and policy layers during 25 trials per task
from each network architecture.
4
3.1 VIRTUAL RODENTS EXHIBIT BEHAVIORAL FLEXIBILITY.
We began our analysis by quantitatively describing the behavioral repertoire of the virtual rodent. A
challenge in understanding the neural mechanisms underlying behavior is that it can be described at
many timescales. On short timescales, one could describe rodent locomotion using a set of actuators
that produce joint-specific patterns of forces and kinematics. However on longer timescales, these
force patterns are organized into coordinated, re-used movements, such as running, jumping, and
turning. These movements can be further combined to form behavioral strategies or goal-directed
behaviors. Relating neural representations to motor behaviors therefore requires analysis methods
that span multiple timescales of behavioral description. To systematically examine the classes of
behaviors these networks learn to generate and how they are differentially deployed across tasks,
we developed sets of behavioral features that describe the kinematics of the animal on fast (5-25
Hz), intermediate (1-25 Hz) or slow (0.3-5 Hz) timescales (Appendix A.2, A.3 ). As validation that
these features reflected meaningful differences across behaviors, embedding these features using
tSNE (Maaten & Hinton, 2008) produced a behavioral map in which virtual rodent behaviors, were
segregated to different regions of the map (Figure 4B)(see video). This behavioral repertoire of the
virtual rodent consisted of many behaviors observed in rodents, such as rearing, jumping, running,
climbing and spinning. While the exact kinematics of the virtual rodent's behaviors did not exactly
match those observed in real rats, they did reproduce unexpected features. For instance the stride
frequency of the virtual rodent during galloping matches that observed in rats (Appendix A.3).
We next investigated how these behaviors were used by the virtual rodent across tasks. On short
timescales, low-level motor features like joint speed and actuator forces occupied similar regions
in principal component space (Figure 4C). In contrast, behavioral kinematics, especially on long,
0.3-5 Hz timescales, were more differentiated across tasks. Similar results held when examining
overlap in other dimensions using multidimensional scaling. Overall this suggests that the network
learned to adapt similar movements in a selective manner for different tasks, suggesting that the
agent exhibited a form of behavioral flexibility.
3.2 NETWORKS PRIMARILY REFLECT BEHAVIORS, NOT FORCES
Figure 5: Representational structure of the rodent's neural network. (A) Example similarity matrices
of neural networks and behavioral descriptors. We grouped behavioral descriptors into 50 clusters
that and we computed the average neural population vector during each cluster (AppendixA.4).
Similarity was assessed by computing the dot product of either the neural population vector or the
behavioral feature vector within each cluster. (B) Centered Kernel Alignment (CKA) index of neural
and behavioral feature similarity matrices for 3 and 1 policy layer architectures. (C) CKA index of
feature similarity matrices across all pairs of network layers. (D) Average CKA index between core
and policy layers and behavioral features, compared across architectures. Points show values from
individual network seeds. Policy values are averaged across layers.
We next examined the neural activity patterns underlying the virtual rodent's behavior to test if net-
works produced behaviors through explicit representations of forces, kinematics or behaviors. As
5
expected, core and policy units operate on distinct timescales (See Appendix A.3, Figure 9). Units
in the core typically fluctuated over timescales of 1-10 seconds, likely representing variables asso-
ciated with context and reward. In contrast, units in policy layers were more active over subsecond
timescales, potentially encoding motor and behavioral features.
To quantify which aspects of behavior were encoded in the core and policy layers, and how these pat-
terns varied across layers, we used representational similarity analysis (RSA) (Kriegeskorte et al.,
2008; Kriegeskorte & Diedrichsen, 2019). RSA provides a global measure of how well different
features are encoded in layers of a neural network by analyzing the geometries of network activity
upon exposure to several stimuli, such as objects. To apply RSA, first a representational similarity
(or equivalently, dissimilarity) matrix is computed that quantifies the similarity of neural population
responses to a set of stimuli. To test if different neural populations show similar stimulus encodings,
these similarity matricies can then be directly compared across different network layers. Multiple
metrics, such as the matrix correlation or dot product can be used to compare these neural represen-
tational similarity matricies. Here we used the linear centered kernel alignment (CKA) index, which
shows invariance to orthonormal rotations of population activity (Kornblith et al., 2019).
RSA can also be used to directly test how well a particular stimulus feature is encoded in a popula-
tion. If each stimuli can be quantitively described by one or more feature vectors, a similarity matrix
can also be computed across the set of stimuli themselves. The strength of encoding of a particular
set of features can by measured by comparing the correlation of the stimulus feature similarity ma-
trix and the neuronal similarity matrix. The correlation strength directly reflects the ability of a linear
decoder trained on the neuronal population vector to distinguish different stimuli (Kriegeskorte &
Diedrichsen, 2019). Unlike previous applications of RSA in the analysis of discrete stimuli such as
objects, (Khaligh-Razavi & Kriegeskorte, 2014; Yamins et al., 2014) behavior evolves continuously.
To adapt RSA to behavioral analysis, we partitioned time by discretizing each behavioral feature
into 50 clusters (Appendix A.4).
As expected, RSA revealed that core and policy layers encoded somewhat distinct behavioral fea-
tures. Policy layers contained greater information about fast timescale kinematics in a manner that
was largely conserved across layers, while core layers showed more moderate encoding of kinemat-
ics that was stronger for slow behavioral features (Figure 5B,C). This difference in encoding was
largely consistent across all architectures tested (Figure 5D).
The feature encoding of policy networks was somewhat consistent with the emergence of a hier-
archy of behavioral abstraction. In networks trained with three policy layers, representations were
distributed in timescales across layers, with the last layer (policy 2) showing stronger encoding of
fast behavioral features, and the first layer (policy 0) instead showing stronger encoding of slow
behavioral features. However, policy layer activity, even close to the motor periphery, did not show
strong explicit encoding of behavioral kinematics or forces.
3.3 BEHAVIORAL REPRESENTATIONS ARE SHARED ACROSS TASKS
We then investigated the degree to which the rodent's neural networks used the same neural represen-
tations to produce behaviors, such as running or spinning, that were shared across tasks. Embedding
population activity into two-dimensions using multidimensional scaling revealed that core neuron
representations were highly distinct across all tasks, while policy layers contained more overlap
(Figure 6A), suggesting that some behavioral representations were re-used. Comparison of rep-
resentational similarity matricies for behaviors that were shared across tasks revealed that policy
layers tended to possess a relatively similar encoding of behavioral features, especially fast behav-
ioral features, over tasks (Figure 6C; Appendix A.4). This was validated by inspection of neural
activity during individual behaviors shared across tasks (Appendix A.5, Figure 10). Core layer rep-
resentations across almost all behavioral categories were more variable across tasks, consistent with
encoding behavioral sequences or task variables.
Interestingly, when comparing this cross-task encoding similarity across architectures, we found that
one layer networks showed a marked increase in the similarity of behavioral encoding across tasks
(Figure 6D). This suggests that in networks with lower computational capacity, animals must rely
on a smaller, shared behavioral representation across tasks.
6
Figure 6: Policy representations are shared across tasks. (A) Two-dimensional multidimensional
scaling embeddings of core and policy activity shows that while policy representations overlap
across some tasks, core representations are largely distinct. (B) CKA index of the policy 2 and
core network representations of behavioral features during behaviors shared across different tasks
(Appendix A.4). Policy 2, but not core networks show similar encoding patterns across the across
the maze forage and two-tap tasks, as well as the gaps run and maze forage tasks, consistent with the
shared behaviors used across these tasks. (C) The similarity of behavioral feature encoding (CKA
index) across different architectures demonstrates that networks with fewer layers show greater sim-
ilarity across tasks. Points show values from individual seeds.
3.4 NEURAL POPULATION DYNAMICS ARE SYNCHRONIZED WITH BEHAVIOR
Figure 7: Neurons in core and policy networks show sequential activity during stereotyped behavior.
(A) Example video stills showing the virtual rodent engaged in the two-tap task (B) Average absolute
z-scored activity traces of all 128 neurons in each layer during performance of the two-tap sequence.
Traces are sorted by the time of peak average firing rate. Dashed lines indicate the times of first and
second taps. Sequential neural activity is present during the two-tap sequence.
While RSA described which behavioral features were represented in core and policy activity, we
were also interested in describing how neural activity changes over time to produce different behav-
iors. We began by analyzing neural activity during the production of stereotyped behaviors. Activity
7
patterns in the two-tap task showed peak activity in core and policy units that was sequentially orga-
nized (Figure 7), uniformly tiling time between both taps of the two-tap sequence. This sequential
activation was observed across tasks and behaviors in the policy network, including during running
(see video) where, consistent with policy networks encoding short-timescale kinematic features in a
task-invariant manner, neural activity sequences were largely conserved across tasks (See Appendix
A.5, Figure 10). These sequences were reliably repeated across instances of the respective behaviors,
and in the case of the two-tap sequence, showed reduced neural variability relative to surrounding
timepoints (See Appendix A.6, Figure 11).
Figure 8: Latent network dynamics within tasks reflect rodent behavior on different timescales. (A)
Vector field representation of the first two principal components of neural activity in the core and
final policy layers during the two-tap task. PC spaces show signatures of rotational dynamics. (B)
Vector field representation of first two jPC planes for the core and final policy layers during the two-
tap task. Apparent rotations within the different planes are associated with behaviors and behavioral
features of different timescales, labeled above. Columns denote layer (as in (A)), while rows denote
jPC plane. (C) Characteristic frequency of rotations within each jPC plane. Groups of three points
respectively indicate the first, second, and third jPC planes for a given layer. Rotations in the core
are slower than those in the policy. (D) Variance explained by each jPC plane.
The finding of sequential activity hints at a putative mechanism for the rodent's behavioral pro-
duction. We next hoped to systematically quantify the types of sequential and dynamical activity
present in core and policy networks without presupposing the behaviors of interest. To describe
population dynamics in relation to behavior, we first applied principal components analysis (PCA)
to the activity during the performance of single tasks, and visualized the gradient of the population
vector as a vector field. Figure 8A shows such a vector field representation of the first two principal
components of the core and final policy layer during the two-tap task. We generated vector fields
by discretizing the PC space into a two-dimensional grid and calculating the average neural activity
gradient with respect to time for each bin.
The vector fields showed strong signatures of rotational dynamics across all layers, likely a signature
of previously described sequential activity. To extract rotational patterns, we used jPCA, a dimen-
sionality reduction method that extracts latent rotational dynamics in neural activity (Churchland
et al., 2012). The resulting jPCs form an orthonormal basis that spans the same space as the first
six traditional PCs, while maximally emphasizing rotational dynamics. Figure 8B shows the vector
8
fields of the first two jPC planes for the core and final policy layers along with their characteristic
frequency. Consistent with our previous findings, jPC planes in the core have lower characteristic
frequencies than those in policy layers across tasks (Figure 8C). The jPC planes also individually
explained a large percentage of total neural variability (Figure 8D).
These rotational dynamics in the policy and core jPC planes were respectively associated with the
production of behaviors and the reward structure of the task. For example, in the two-tap task,
rotations in the fastest jPC plane in the core were concurrent with the approach to reward, while
rotations in the second fastest jPC were concurrent with long timescale transitions between running
to the orb and performing the two-tap sequence. Similarly, the fastest jPC in policy layers was corre-
lated with the phase of running, while the second fastest was correlated with the phase of the two-tap
sequence (video). This trend of core and policy neural dynamics respectively reflecting behavioral
and task-related features was also present in other tasks. For example, in the maze forage task, the
first two jPC planes in the core respectively correlated with reaching the target orb and discovering
the location of new orbs, while those in the policy were correlated with low-level locomotor features
such as running phase (video). Along with RSA, these findings support a model in which the core
layer transforms sensory information into a contextual signal in a task-specific manner. This signal
then modulates activity in the policy toward different trajectories that generate appropriate behav-
iors in a more task-independent fashion. For a more complete set of behaviors with neural dynamics
visualizations overlaid, see Appendix A.7.
3.5 NEURAL PERTURBATIONS CORROBORATE DISTINCT ROLES ACROSS LAYERS
To causally demonstrate the differing roles of core and policy units in respectively encoding task-
relevant features and movement, we performed silencing and activation of different neuronal subsets
in the two-tap task. We identified two stereotyped behaviors (rears and spinning jumps) that were
reliably used in two different seeds of the agent to reach the orb in the task. We ranked neurons
according to the degree of modulation of their z-scored activity during the performance of these
behaviors. We then inactivated subsets of neurons by clamping activity to the mean values between
the first and second taps and observed the effects of inactivation on trial success and behavior.
In both seeds analyzed, inactivation of policy units had a stronger effect on motor behavior than the
inactivation of core units. For instance, in the two-tap task, ablation of 64 neurons in the final policy
layer disrupts the performance of the spinning jump (Appendix A.8 Figure 12B video). In contrast,
ablation of behavior-modulated core units did not prevent the production of the behavior, but mildly
affected the way in which the behavior is directed toward objects in the environment. For example,
ablation of a subset of core units during the performance of a spinning jump had a limited effect, but
sometimes resulted in jumps that missed the target orbs (video; See Appendix A.8, Figure 12C).
We also performed a complementary perturbation aimed to elicit behaviors by overwriting the cell
state of neurons in each layer with the average time-varying trajectory of neural activity measured
during natural performance of a target behavior. The efficacy of stimulation was found to depend
on the gross body posture and behavioral state of an animal, but was nevertheless successful in
some cases. For example, during the two-tap sequence, we were able to elicit spinning movements
common to searching behaviors in the forage task (video; See Appendix A.8, Figure 12D, E). The
efficacy of this activation was more reliable in layers closer to the motor output (Figure 12D). In
fact, activation of core units rarely elicited spins, but rather elicited sporadic dashes reminiscent of
the searching strategy of many models during the forage task (video).
4 DISCUSSION
For many computational neuroscientists and artificial intelligence researchers, an aim is to reverse-
engineer the nervous system at an appropriate level of abstraction. In the motor system, such an
effort requires that we build embodied models of animals equipped with artificial nervous systems
capable of controlling their synthetic bodies across a range of behavior. Here we introduced a virtual
rodent capable of performing a variety of complex locomotor behaviors to solve multiple tasks using
a single policy. We then used this virtual nervous system to study principles of the neural control
of movement across contexts and described several commonalities between the neural activity of
artificial control and previous descriptions of biological control.
9
A key advantage of this approach relative to experimental approaches in neuroscience is that we can
fully observe sensory inputs, neural activity, and behavior, facilitating more comprehensive testing
of theories related to how behavior can be generated. Furthermore, we have complete knowledge
of the connectivity, sources of variance, and training objectives of each component of the model,
providing a rare ground truth to test the validity of our neural analyses. With these advantages
in mind, we evaluated our analyses based on their capacity to both describe the algorithms and
representations employed by the virtual rodent and recapitulate the known functional objectives
underlying its creation without prior knowledge.
To this end, our description of core and policy as respectively representing value and motor pro-
duction is consistent with the model's actor-critic training objectives. But beyond validation, our
analyses provide several insights into how these objectives are reached. RSA revealed that the cell
activity of core and policy layers had greater similarity with behavioral and postural features than
with short-timescale actuators. This suggests that the representation of behavior is useful in the
moment-to-moment production of motor actions in artificial control, a model that has been pre-
viously proposed in biological action selection and motor control (Mink, 1996; Graziano, 2006).
These behavioral representations were more consistent across tasks in the policy than in the core,
suggesting that task context and value activity in the core engaged task-specific behavioral strategies
through the reuse of shared motor activity in the policy.
Our analysis of neural dynamics suggests that reused motor activity patterns are often organized as
sequences. Specifically, the activity of policy units uniformly tiles time in the production of several
stereotyped behaviors like running, jumping, spinning, and the two-tap sequence. This finding is
consistent with reports linking sequential neural activity to the production of stereotyped motor and
task-oriented behavior in rodents (Berke et al., 2009; Rueda-Orozco & Robbe, 2015; Dhawale et al.,
2019), including during task delay periods (Akhlaghpour et al., 2016), as well as in singing birds
(Albert & Margoliash, 1996; Hahnloser et al., 2002). Similarly, by relating rotational dynamics
to the virtual rodent's behavior, we found that different behaviors were seemingly associated with
distinct rotations in neural activity space that evolved at different timescales. These findings are
consistent with a hierarchical control scheme in which policy layer dynamics that generate reused
behaviors are activated and modulated by sensorimotor signals from the core.
This work represents an early step toward the constructive modeling of embodied control for the
purpose of understanding the neural mechanisms behind the generation of behavior. Incrementally
and judiciously increasing the realism of the model's embodiment, behavioral repertoire, and neural
architecture is a natural path for future research. Our virtual rodent possesses far fewer actuators
and touch sensors than a real rodent, uses a vastly different sense of vision, and lacks integration
with olfactory, auditory, and whisker-based sensation (see Zhuang et al., 2017). While the virtual
rodent is capable of locomotor behaviors, an increased diversity of tasks involving decision making,
memory-based navigation, and working memory could give insight into "cognitive" behaviors of
which rodents are capable. Furthermore, biologically-inspired design of neural architectures and
training procedures should facilitate comparisons to real neural recordings and manipulations. We
expect that this comparison will help isolate residual elements of animal behavior generation that
are poorly captured by current models of motor control, and encourage the development of artificial
neural architectures that can produce increasingly realistic behavior.
AUTHOR CONTRIBUTIONS
Josh and Yuval built the rodent MuJoCo model, with measurements collected by Diego and Jesse.
Josh trained the virtual rodent model. Jesse performed behavioral and neural representation analyses.
Diego performed neural dynamics analyses. Josh, Jesse, and Diego drafted the manuscript. All
authors contributed to the conception of the project.
ACKNOWLEDGMENTS
The rodent skeleton reference model was purchased from leo3Dmodels on TurboSquid. Thanks to
Max Cant for the rodent skin, and Marcus Wainwright for the skybox and ground textures. D.A. was
supported by NSF GRFP DGE1745303. J.D.M was supported by a fellowship from the Helen Hay
Whitney foundation sponsored by Vertex and a K99/R00 award from the NINDS.
10
REFERENCES
Abbas Abdolmaleki, Jost Tobias Springenberg, Yuval Tassa, Remi Munos, Nicolas Heess, and Mar-
tin Riedmiller. Maximum a posteriori policy optimisation. In International Conference on Learn-
ing Representations, 2018.
Hessameddin Akhlaghpour, Joost Wiskerke, Jung Yoon Choi, Joshua P Taliaferro, Jennifer Au, and
Ilana B Witten. Dissociated sequential activity and stimulus encoding in the dorsomedial striatum
during spatial working memory. Elife, 5:e19507, 2016.
C Yu Albert and Daniel Margoliash. Temporal hierarchical control of singing in birds. Science, 273
(5283):1871 -- 1875, 1996.
Andrea Banino, Caswell Barry, Benigno Uria, Charles Blundell, Timothy Lillicrap, Piotr Mirowski,
Alexander Pritzel, Martin J Chadwick, Thomas Degris, Joseph Modayil, et al. Vector-based
navigation using grid-like representations in artificial agents. Nature, 557(7705):429, 2018.
Joshua D Berke, Jason T Breck, and Howard Eichenbaum. Striatal versus hippocampal representa-
tions during win-stay maze performance. Journal of neurophysiology, 101(3):1575 -- 1587, 2009.
Gordon J Berman, Daniel M Choi, William Bialek, and Joshua W Shaevitz. Mapping the stereotyped
behaviour of freely moving fruit flies. Journal of The Royal Society Interface, 11(99):20140672,
2014.
Hillel J Chiel and Randall D Beer. The brain has a body: adaptive behavior emerges from inter-
actions of nervous system, body and environment. Trends in neurosciences, 20(12):553 -- 557,
1997.
Mark M Churchland, M Yu Byron, John P Cunningham, Leo P Sugrue, Marlene R Cohen, Greg S
Corrado, William T Newsome, Andrew M Clark, Paymon Hosseini, Benjamin B Scott, et al. Stim-
ulus onset quenches neural variability: a widespread cortical phenomenon. Nature neuroscience,
13(3):369, 2010.
Mark M Churchland, John P Cunningham, Matthew T Kaufman, Justin D Foster, Paul Nuyujukian,
Stephen I Ryu, and Krishna V Shenoy. Neural population dynamics during reaching. Nature, 487
(7405):51, 2012.
Christopher J Cueva and Xue-Xin Wei. Emergence of grid-like representations by training recurrent
neural networks to perform spatial localization. arXiv preprint arXiv:1803.07770, 2018.
Ashesh K Dhawale, Steffen BE Wolff, Raymond Ko, and Bence P Olveczky. The basal ganglia can
control learned motor sequences independently of motor cortex. bioRxiv, pp. 827261, 2019.
Lasse Espeholt, Hubert Soyer, Remi Munos, Karen Simonyan, Volodymyr Mnih, Tom Ward, Yotam
IMPALA: Scalable distributed deep-rl
Doron, Vlad Firoiu, Tim Harley, Iain Dunning, et al.
with importance weighted actor-learner architectures. In International Conference on Machine
Learning, pp. 1406 -- 1415, 2018.
Michael Graziano. The organization of behavioral repertoire in motor cortex. Annu. Rev. Neurosci.,
29:105 -- 134, 2006.
Sten Grillner, Alexander Kozlov, Paolo Dario, Cesare Stefanini, Arianna Menciassi, Anders Lansner,
and Jeanette Hellgren Kotaleski. Modeling a vertebrate motor system: pattern generation, steering
and control of body orientation. Progress in brain research, 165:221 -- 234, 2007.
Richard HR Hahnloser, Alexay A Kozhevnikov, and Michale S Fee. An ultra-sparse code underli-
esthe generation of neural sequences in a songbird. Nature, 419(6902):65, 2002.
Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recog-
nition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.
770 -- 778, 2016.
Nicolas Heess, TB Dhruva, Srinivasan Sriram, Jay Lemmon, Josh Merel, Greg Wayne, Yuval Tassa,
Tom Erez, Ziyu Wang, Ali Eslami, et al. Emergence of locomotion behaviours in rich environ-
ments. arXiv preprint arXiv:1707.02286, 2017.
11
Norman C Heglund and C Richard Taylor. Speed, stride frequency and energy cost per stride: how
do they change with body size and gait? Journal of Experimental Biology, 138(1):301 -- 318, 1988.
Sepp Hochreiter and Jurgen Schmidhuber. Long short-term memory. Neural computation, 9(8):
1735 -- 1780, 1997.
Auke Jan Ijspeert, Alessandro Crespi, Dimitri Ryczko, and Jean-Marie Cabelguen. From swimming
to walking with a salamander robot driven by a spinal cord model. Science, 315(5817):1416 --
1420, 2007.
Eric Jonas and Konrad Paul Kording. Could a neuroscientist understand a microprocessor? PLoS
computational biology, 13(1):e1005268, 2017.
John F Kalaska. From intention to action: motor cortex and the control of reaching movements. In
Progress in Motor Control, pp. 139 -- 178. Springer, 2009.
Kohitij Kar, Jonas Kubilius, Kailyn Schmidt, Elias B Issa, and James J DiCarlo. Evidence that
recurrent circuits are critical to the ventral streams execution of core object recognition behavior.
Nature neuroscience, 22(6):974, 2019.
Risa Kawai, Timothy Markman, Rajesh Poddar, Raymond Ko, Antoniu L Fantana, Ashesh K
Dhawale, Adam R Kampff, and Bence P Olveczky. Motor cortex is required for learning but
not for executing a motor skill. Neuron, 86(3):800 -- 812, 2015.
Alexander JE Kell, Daniel LK Yamins, Erica N Shook, Sam V Norman-Haignere, and Josh H Mc-
Dermott. A task-optimized neural network replicates human auditory behavior, predicts brain
responses, and reveals a cortical processing hierarchy. Neuron, 98(3):630 -- 644, 2018.
Seyed-Mahdi Khaligh-Razavi and Nikolaus Kriegeskorte. Deep supervised, but not unsupervised,
models may explain it cortical representation. PLoS computational biology, 10(11):e1003915,
2014.
Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey E. Hinton. Similarity of neural
network representations revisited. CoRR, abs/1905.00414, 2019.
Nikolaus Kriegeskorte and Jorn Diedrichsen. Peeling the onion of brain representations. Annual
Review of Neuroscience, 42(1):407 -- 432, 2019.
Nikolaus Kriegeskorte, Marieke Mur, and Peter A Bandettini. Representational similarity analysis-
connecting the branches of systems neuroscience. Frontiers in systems neuroscience, 2:4, 2008.
Yuri Lazebnik. Can a biologist fix a radio? Or, what I learned while studying apoptosis. Cancer
cell, 2(3):179 -- 182, 2002.
Timothy P Lillicrap and Stephen H Scott. Preference distributions of primary motor cortex neurons
reflect control solutions optimized for limb biomechanics. Neuron, 77(1):168 -- 179, 2013.
Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. Journal of machine
learning research, 9(Nov):2579 -- 2605, 2008.
Josh Merel, Arun Ahuja, Vu Pham, Saran Tunyasuvunakool, Siqi Liu, Dhruva Tirumala, Nicolas
Heess, and Greg Wayne. Hierarchical visuomotor control of humanoids. In International Con-
ference on Learning Representations, 2019a.
Josh Merel, Leonard Hasenclever, Alexandre Galashov, Arun Ahuja, Vu Pham, Greg Wayne,
Yee Whye Teh, and Nicolas Heess. Neural probabilistic motor primitives for humanoid control.
In International Conference on Learning Representations, 2019b.
Jonathan A Michaels, Stefan Schaffelhofer, Andres Agudelo-Toro, and Hansjorg Scherberger. A
neural network model of flexible grasp movement generation. bioRxiv, pp. 742189, 2019.
Jonathan W Mink. The basal ganglia: focused selection and inhibition of competing motor pro-
grams. Progress in neurobiology, 50(4):381 -- 425, 1996.
12
John A Nyakatura, Kamilo Melo, Tomislav Horvat, Kostas Karakasiliotis, Vivian R Allen, Amir
Andikfar, Emanuel Andrada, Patrick Arnold, Jonas Laustroer, John R Hutchinson, et al. Reverse-
engineering the locomotion of a stem amniote. Nature, 565(7739):351, 2019.
Xue Bin Peng, Glen Berseth, and Michiel Van de Panne. Terrain-adaptive locomotion skills using
deep reinforcement learning. ACM Transactions on Graphics (TOG), 35(4):81, 2016.
Xue Bin Peng, Glen Berseth, KangKang Yin, and Michiel Van De Panne. DeepLoco: Dynamic
locomotion skills using hierarchical deep reinforcement learning. ACM Transactions on Graphics
(TOG), 36(4):41, 2017.
Pavan Ramdya, Robin Thandiackal, Raphael Cherney, Thibault Asselborn, Richard Benton,
Auke Jan Ijspeert, and Dario Floreano. Climbing favours the tripod gait over alternative faster
insect gaits. Nature communications, 8:14494, 2017.
Alfonso Renart and Christian K Machens. Variability in neural activity and behavior. Current
opinion in neurobiology, 25:211 -- 220, 2014.
Pavel E Rueda-Orozco and David Robbe. The striatum multiplexes contextual and kinematic infor-
mation to constrain motor habits execution. Nature neuroscience, 18(3):453, 2015.
Simon Schmitt, Jonathan J Hudson, Augustin Zidek, Simon Osindero, Carl Doersch, Wojciech M
Czarnecki, Joel Z Leibo, Heinrich Kuttler, Andrew Zisserman, Karen Simonyan, et al. Kickstart-
ing deep reinforcement learning. arXiv preprint arXiv:1803.03835, 2018.
H Francis Song, Guangyu R Yang, and Xiao-Jing Wang. Reward-based training of recurrent neural
networks for cognitive and value-based tasks. Elife, 6:e21492, 2017.
H Francis Song, Abbas Abdolmaleki, Jost Tobias Springenberg, Aidan Clark, Hubert Soyer, Jack W
Rae, Seb Noury, Arun Ahuja, Siqi Liu, Dhruva Tirumala, et al. V-MPO: On-Policy Maxi-
mum a Posteriori Policy Optimization for Discrete and Continuous Control. arXiv preprint
arXiv:1909.12238, 2019.
Greg J Stephens, Bethany Johnson-Kerner, William Bialek, and William S Ryu. Dimensionality and
dynamics in the behavior of c. elegans. PLoS computational biology, 4(4):e1000028, 2008.
David Sussillo, Mark M Churchland, Matthew T Kaufman, and Krishna V Shenoy. A neural network
that finds a naturalistic solution for the production of muscle activity. Nature neuroscience, 18(7):
1025, 2015.
Yuval Tassa, Yotam Doron, Alistair Muldal, Tom Erez, Yazhe Li, Diego de Las Casas, David Bud-
den, Abbas Abdolmaleki, Josh Merel, Andrew Lefrancq, Timothy Lillicrap, and Martin Ried-
miller. DeepMind control suite. arXiv preprint arXiv:1801.00690, 2018.
Emanuel Todorov, Tom Erez, and Yuval Tassa. MuJoCo: A physics engine for model-based control.
In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pp. 5026 -- 5033.
IEEE, 2012.
Daniel LK Yamins, Ha Hong, Charles F Cadieu, Ethan A Solomon, Darren Seibert, and James J
DiCarlo. Performance-optimized hierarchical models predict neural responses in higher visual
cortex. Proceedings of the National Academy of Sciences, 111(23):8619 -- 8624, 2014.
Guangyu Robert Yang, Madhura R Joglekar, H Francis Song, William T Newsome, and Xiao-Jing
Wang. Task representations in neural networks trained to perform many cognitive tasks. Nature
neuroscience, 22(2):297, 2019.
Chengxu Zhuang, Jonas Kubilius, Mitra JZ Hartmann, and Daniel L Yamins. Toward goal-driven
neural network models for the rodent whisker-trigeminal system. In Advances in Neural Informa-
tion Processing Systems, pp. 2555 -- 2565, 2017.
13
A APPENDIX
A.1 RAT MEASUREMENTS
To construct the virtual rodent model, we obtained the mass and lengths of the largest body segments
that influence the physical properties of the virtual rodent. First, we dissected cadavers of two
female Long-Evans rats, and measured the mass of relevant limb segments and organs. Next, we
measured the lengths of body segments over the skin of animals anesthetized with 2% v/v isoflurane
anesthesia in oxygen. We confirmed that these skin based measurements approximated bone lengths
by measuring bone lengths in a third cadaver. The care and experimental manipulation of all animals
were reviewed and approved by the appropriate Institutional Animal Care and Use Committee.
Animal (#)
63
64
Body part
Mass (g)
Average mass (g)
Hindlimb L
Hindlimb R
Tail
Forelimb R
Forelimb L
Full torso
Head
Upper torso
Lower torso
Torso without organs
Intestines and stomach
Liver
Pelvis and kidneys
Jaw
Skull
Tail (base to mid)
Tail (mid to tip)
Scapula L
Humerus L
Radius/ulna L
Forepaw L
Scapula R
Humerus R
Radius/ulna R
Forepaw R
Hindpaw L
Tibia L
Femur L
Hindpaw R
Tibia R
Femur R
21
21
8
11
12
176
26
78
98
54
22
26
74
2.43
23
5.92
1.78
3.19
6.25
2.61
0.53
2.23
6.08
2.17
0.53
1.66
9
13
1.81
5
13
26
26
10
14
13
187
26
71
114
58
32
17
80
4.70
21
7.20
2.30
4.70
4.70
2.8
0.5
3.9
6.7
3.3
0.5
1.7
9
16
1.6
6
18
23.5
23.5
9
12.5
12.5
181.5
26
74.5
106
56
27
21.5
77
3.57
22
6.56
2.04
3.94
5.48
2.70
0.52
3.07
6.39
2.74
0.52
1.68
9
14.5
1.71
5.5
15.5
Total
281
301
291
Table 1: Before weighing, limb segments were divided at their respective joints. Mass of all seg-
ments includes all bones, skin, muscle, fascia and adipose layers. L and R refer to the left and right
sides of the animal. Precision of measurements listed without decimal places is ±0.5g
14
Age (days)
Mass (g)
Body part
Ankle to claw L
Ankle to toe L
Ankle to pad L
Ankle to claw R
Ankle to toe R
Ankle to pad R
Tibia L
Femur L
Tibia R
Femur R
Pelvis
Wrist to claw L
Wrist to finger L
Wrist to pad L
Wrist to olecranon L
Humerus L
Scapula L
Wrist to claw R
Wrist to finger R
Wrist to pad R
Wrist to olecranon R
Humerus R
Scapula R
Headcap width
Headcap length
Skull width
Skull length
Skull height
Head to thoracic
Thoracic to sacral
Head to sacral
Head width
Ear
Eye
40.2
38.4
23.4
50
44.5
25.8
15
29.1
31.9
22.7
39
30
38.8
57
145
53.4
18
7.2
55
330
389
39.7
37.7
23
40.4
38.7
23.3
38.5
32.1
39.1
32.1
31.7
17.6
15.8
6.4
32.5
31
26.4
17
13
5.8
33.5
33.5
29.5
39.5
38.12
22.2
38.2
37
22.4
36.3
31.6
36.7
32.9
32
18.8
16
6
34
29.52
24
16.8
14.1
5.6
30.6
28.2
23.8
330
348
37.8
35.6
22.12
38.3
36.3
21.9
49.2
37.9
37.9
38.7
30.2
18.6
17.4
8.34
31.7
28.2
29.3
17.8
15.6
6.55
31.2
28.8
25.9
23.35
51.1
23
61
48.6
73.1
126
71.4
73.6
145.5
21.8
56.48
68.68
62.9
127.05
17.55
8.25
19.3
8.6
17.9
8.8
22.8
53.16
21.59
65
65.04
127.2
19.2
8.2
48
382
325
62
82
273
Animal (#)
56
64
83
283
63
83
269
62*
83
273
Average ± std
309 ± 47
Length (mm)
39.9
36.6
22.5
39.3
37.7
21.8
35.8
33.4
35.1
31.9
26.7
16
15.5
4.9
33.9
27
25.9
15.9
15.6
5.2
30.4
25
26.2
39.8
38
6.4
38.3
36.2
24.1
34.1
32.4
36.18
32.6
19.2
17.6
6.4
29.9
25.4
26.2
18.1
16.9
5.8
29.9
25.2
24.4
22.2
48
21
71.2
68.8
140.9
41.5
39.3
23.3
39.6
38.6
23.1
38.7
35.35
38.4
32.1
27.2
19.02
17.07
6.1
32.1
31.2
29.1
16.3
15.3
5
31.8
28.2
28.8
23.9
58.13
21.5
60.4
64.7
123.7
18.8
8.3
39.8 ± 1.1
37.7 ± 1.2
20.4 ± 6.2
39.0 ± 0.9
37.4 ± 1.1
22.8 ± 0.9
40.4 ± 6.5
35.3 ± 4.6
37.2 ± 1.5
33.4 ± 2.6
28.9 ± 2.7
17.7 ± 1.6
16.6 ± 0.9
6.4 ± 1.1
31.9 ± 1.9
29.2 ± 2.4
26.2 ± 2.4
17.0 ± 0.8
15.1 ± 1.4
5.7 ± 0.5
31.2 ± 1.3
28.1 ± 3.1
26.4 ± 2.3
39
30
25.1 ± 6.1
55.0 ± 4.5
21.4 ± 0.3
64.2 ± 8.7
68.0 ± 4.6
133.6 ± 9.7
53.4
18.5 ± 0.7
8.2 ± 0.6
Table 2: Length measurements of limb segments used to construct the virtual rodent model from 7
female Long-Evans rats. Measurements were performed using calipers either over the skin or over
dissected bones (*). Thoracic and sacral refer to vertebral segments. L and R refer to the left and
right sides of the animal's body.
15
A.2 BEHAVIORAL ANALYSIS
We generated features describing the whole-body pose and kinematics of the virtual rodent on fast,
intermediate, and slow temporal scales. To describe the whole-body pose, we took the top 15 prin-
cipal components of the virtual rodent's joint angles and joint positions to yield two 15 dimensional
sets of eigenpostures (Stephens et al., 2008). We combined these into a 30 dimensional set of postu-
ral features. To describe the animal's whole-body kinematics, we computed the continuous wavelet
transform of each eigenposture using a Morlet wavelet spanning 25 scales. For each set of eigen-
postures this yielded a 375 dimensional time-frequency representation of the underlying kinematics.
We then computed the top 15 principal components of each 375 dimensional time-frequency repre-
sentation and combined them to yield a 30 dimensional representational description of the animal's
behavioral kinematics. To facilitate comparison of kinematics to neural representations on differ-
ent timescales, we used three sets of wavelet frequencies on 1 to 25 Hz (intermediate), 0.3 to 5 Hz
(slow) or 5-25 Hz (fast) timescales. In separate work, we have found that combining postural and
kinematic information improves separation of animal behaviors in behavioral embeddings. There-
fore, we combined postural and dynamical features, the later on intermediate timescales, to yield a
60 dimensional set of 'behavioral features' that we used to map the animal's behavior using tSNE
(Figure 4C) (Berman et al., 2014). tSNEs were made using the Barnes-Hut approximation with a
perplexity of 30.
A.3 POWER SPECTRAL DENSITY OF BEHAVIOR AND NETWORK ACTIVITY
Figure 9: (A) Power spectral density estimates of four different features describing animal behavior,
computed by averaging the spectral density of the top ten principal components of each feature,
weighted by the variance they explain. (B) Power spectral density estimates of four different net-
work layers, computed by averaging the spectral density of the top ten principal components of each
matrix of activations, weighted by the variance they explain. Notice that policy layers have more
power in high frequency bands than core layers. Arrows mark peaks in the power spectra corre-
sponding to locomotion. Notably, the 4-5 Hz frequency of galloping in the virtual rat matches that
measured in laboratory rats (Heglund & Taylor, 1988). Power spectral density was computed using
Welch's method using a 10 s window size and 5 s overlap.
16
A.4 REPRESENTATIONAL SIMILARITY ANALYSIS
We used representational similarity analysis to compare population representations across different
network layers and to compute the encoding strength of different features describing animal behavior
in the population. Representational similarity analysis has in the past been used to compare neural
population responses in tasks where behavioral stimuli are discrete, for instance corpuses of objects
or faces (Kriegeskorte et al., 2008; Kriegeskorte & Diedrichsen, 2019). A challenge in scaling such
approaches to neural analysis in the context of behavior is that behavior unfolds continuously in
time. It is thus a priori unclear how to discretize behavior into discrete chunks in which to compare
representations.
Formally, we defined eight sets of features Bi=1...8 describing the behavior of the animal on different
timescales. These included features such as joint angles, the angular speed of the joint angles,
eigenposture coefficients, and actuator forces that vary on short timescales, as well as behavioral
kinematics, which vary on longer timescales and 'behavioral features', which consisted of both
kinematics and eigenpostures. Each feature set is a matrix Bi ∈ RM xqi where M is the number of
timepoints in the experiment and qi is the number of features in the set. We discretized each set Bi
using k-means clustering with k = 50 to yield a partition of the timepoints in the experiment Pi.
Using the discretization defined in Pi, we can perform representational similarity analysis to com-
pare the structure of population responses across neural network layers Lm and Ln or between a
given network layer and features of the behavior Bi. Following notation in (Kornblith et al., 2019)
we let X ∈ Rkxp be a matrix of population responses across p neurons and the k behavioral cat-
egories in Pi. We let Y ∈ Rkxq be either the matrix of population responses from q neurons in a
distinct network layer, or a set of q features describing the behavior of the animal in the feature set
Bi.
After computing the response matricies in a given behavioral partition, we compared the representa-
tional structure of the matricies XX T and Y Y T . To do so, we compute the similarity between these
matricies using the linear Centered Kernel Alignment index, which is invariant under orthonormal
rotations of the population activity. Following (Kornblith et al., 2019), the CKA coeffient is:
(cid:107)XY T(cid:107)F
CKA(XX T , Y Y T ) =
(cid:107)XX T(cid:107)F(cid:107)Y Y T(cid:107)F
(1)
Where (cid:107) · (cid:107)F is the Frobenius norm. For centered X and Y , the numerator is equivalent to the
dot-product between the vectorized responses (cid:107)XY T(cid:107)F = (cid:104)vec(XX T ), vec(Y Y T )(cid:105).
For a given network layer Lm, and a behavioral partition Pi, we can denote XX T = DLm
Pi
Similarly, for a given feature set Bi, let DBi
Pi
CKA (Dm
i .
= Dm
i. Thus we are interested in characterizing both
i , Dn
i )
= Di
(2)
and
CKA(cid:0)Dm
(cid:1) .
i , Di
i
(3)
The former equation describes the similarity across two layers of the network, and the later describes
the similarity of the network activity to a set of behavioral descriptors.
An additional challenge comes when restricting this analysis to comparing the neural representations
of behavioral across different tasks Ta, Tb, where not all behaviors are necessarily used in each task.
To make such a comparison, we denote Bi(Ta) to be the set of behavioral clusters observed in task
= Bi(Ta) ∩ Bi(Ta) to be the set of behaviors used in each of the two tasks. We
Ta, and BTaTb
can then define a restricted partition of timepoints for each task P Ta,Tb
that includes only
these behaviors, and compute the representational similarity between the same layer across tasks:
or P Tb,Ta
i
i
i
CKA(cid:0)Dm
, Dm
i,Tb
i,Ta
(cid:1) .
(4)
We have presented a means of performing representational similarity analysis across continuous
time domains, where the natural units of discretization are unclear and likely manifold. While we
focused on analyzing responses on the population level, it is likely that different subspaces of the
population may encode information about distinct behavioral features at different timescales, which
is still an emerging domain in representational similarity analysis techniques.
17
A.5 NEURAL POPULATION ACTIVITY ACROSS TASKS DURING RUNNING
Figure 10: Average activity in the final policy layer (policy 2) during running cycles across different
tasks.
In each heatmap, rows correspond to the absolute averaged z-scored activity for individ-
ual neurons, while columns denote time relative to the mid stance of the running phase. Across
heatmaps, neurons are sorted by the time of peak activity in the tasks denoted on the left, such
that each column of heatmaps contains the same average activity information with rearranged rows.
Aligned running bouts were acquired by manually segmenting the the principal component space of
policy 2 activity to find instances of mid-stance running and analyzing the surrounding 200 ms.
18
A.6 STEREOTYPED BEHAVIOR INITIATION AND NEURAL VARIABILITY
During the execution of stereotyped behaviors, neural variability was reduced (Figure 11). Recall
that in our setting, neurons have no intrinsic noise, but inherit motor noise through observations of
the state (i.e. via sensory reafference). This effect loosely resembles, and perhaps informs one line
of interpretation of the widely reported phenomenon of neural variability reducing with stimulus or
task onset (Churchland et al., 2010). Our reproduction of this effect, which simply emerges from
training, suggests that variance modulation may partly arise from moments in a task that benefit
from increased behavioral precision (Renart & Machens, 2014).
Figure 11: Quantification of neural variability in inter-tap interval of two-tap task relative to the
second tap. (A) Example normalized activity traces of ten randomly selected neurons in the final
policy layer. Lines indicate mean normalized activity whiles shaded regions range from the 20th
percentile to the 80th percentile. Dashed lines indicate the times of first and second taps.
(B)
Standard deviation of normalized activity across all neurons in the final policy layer as a function
of time relative to the second tap. Lines indicate the mean standard deviation while shaded regions
range from the 20th percentile to the 80th percentile. Observe that variability is reduced during the
two-tap interval.
A.7 NEURAL DYNAMICS VISUALIZED DURING TASK BEHAVIOR
For completeness, we provide links to videos of a few variants of neural dynamics for each task.
Network
Visualization Task (link)
1-layer policy
3-layer policy
3-layer policy
PCA
PCA
PCA
PCA
PCA
PCA
PCA
PCA
jPCA
jPCA
jPCA
jPCA
gaps
forage
escape
two-tap
gaps
forage
escape
two-tap
gaps
forage
escape
two-tap
Table 3: Links to representative visualizations of neural dynamics and behavior
19
A.8 PERTURBATION RESULTS
Figure 12: Causal manipulations reveal distinct roles for core and policy layers in the production
of behavior. (A) Two-tap accuracy during the inactivation of units modulated by idiosyncratic be-
haviors within the two-tap sequence. Core inactivation has a weaker negative effect on trial success
than policy inactivation for several levels of inactivation. (B) Representative example of a failed
trial during inactivation of the final policy layer in a model that performs a spinning jump during the
two-tap sequence. The model is incapable of producing the spinning jump behavior while inacti-
vated. (C) Representative example of a failed trial during core inactivation in a model that performs
a spinning jump during the two-tap sequence. The model is still able to perform the spinning jump
behavior, but misses the orb. (D) Proportion of attempts at stimulation that successfully elicited spin
behavior during the two-tap task. The efficacy of this activation was more reliable in layers closer
to the motor output. (E) Representative example of a single trial in which an extra spin occurs after
policy 2 activation.
20
|
1704.08306 | 1 | 1704 | 2017-03-21T16:12:31 | A Digital Neuromorphic Architecture Efficiently Facilitating Complex Synaptic Response Functions Applied to Liquid State Machines | [
"q-bio.NC",
"cs.NE",
"stat.ML"
] | Information in neural networks is represented as weighted connections, or synapses, between neurons. This poses a problem as the primary computational bottleneck for neural networks is the vector-matrix multiply when inputs are multiplied by the neural network weights. Conventional processing architectures are not well suited for simulating neural networks, often requiring large amounts of energy and time. Additionally, synapses in biological neural networks are not binary connections, but exhibit a nonlinear response function as neurotransmitters are emitted and diffuse between neurons. Inspired by neuroscience principles, we present a digital neuromorphic architecture, the Spiking Temporal Processing Unit (STPU), capable of modeling arbitrary complex synaptic response functions without requiring additional hardware components. We consider the paradigm of spiking neurons with temporally coded information as opposed to non-spiking rate coded neurons used in most neural networks. In this paradigm we examine liquid state machines applied to speech recognition and show how a liquid state machine with temporal dynamics maps onto the STPU-demonstrating the flexibility and efficiency of the STPU for instantiating neural algorithms. | q-bio.NC | q-bio |
A Digital Neuromorphic Architecture Efficiently
Facilitating Complex Synaptic Response Functions
Applied to Liquid State Machines
Michael R. Smith∗, Aaron J. Hill∗, Kristofor D. Carlson∗, Craig M. Vineyard∗, Jonathon Donaldson∗,
David R. Follett†, Pamela L. Follett†‡, John H. Naegle∗, Conrad D. James∗ and James B. Aimone∗
∗Sandia National Laboratories, Albuquerque, NM 87185 USA
Email:{msmith4, ajhill, kdcarls, cmviney, jwdonal, jhnaegl, cdjame, jbaimon}@sandia.gov
†Lewis Rhodes Labs, Concord, MA 01742 USA
‡Tufts University, Medford, MA 02155 USA
Email:{drfollett, plfollett}@earthlink.net
Abstract-Information in neural networks is represented as
weighted connections, or synapses, between neurons. This poses
a problem as the primary computational bottleneck for neural
networks is the vector-matrix multiply when inputs are multiplied
by the neural network weights. Conventional processing architec-
tures are not well suited for simulating neural networks, often re-
quiring large amounts of energy and time. Additionally, synapses
in biological neural networks are not binary connections, but
exhibit a nonlinear response function as neurotransmitters are
emitted and diffuse between neurons. Inspired by neuroscience
principles, we present a digital neuromorphic architecture, the
Spiking Temporal Processing Unit (STPU), capable of modeling
arbitrary complex synaptic response functions without requiring
additional hardware components. We consider the paradigm of
spiking neurons with temporally coded information as opposed to
non-spiking rate coded neurons used in most neural networks. In
this paradigm we examine liquid state machines applied to speech
recognition and show how a liquid state machine with temporal
dynamics maps onto the STPU-demonstrating the flexibility and
efficiency of the STPU for instantiating neural algorithms.
I. INTRODUCTION
Neural-inspired learning algorithms are achieving state of
the art performance in many application areas such as speech
recognition [1], image recognition [2], and natural language
processing [3]. Information and concepts, such as a dog or
a person in an image, are represented in the synapses, or
weighted connections, between the neurons. The success of a
neural network is dependent on training the weights between
the neurons in the network. However, training the weights in a
neural network is non-trivial and often has high computational
complexity with large data sets requiring long training times.
One of the contributing factors to the computational com-
plexity of neural networks is the vector-matrix multiplications
This work was supported by Sandia National Laboratories Laboratory
Directed Research and Development (LDRD) Program under the Hardware
Acceleration of Adaptive Neural Algorithms (HAANA) Grand Challenge
project. Sandia National Laboratories is a multi-mission laboratory managed
and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed
Martin Corporation, for the U. S. Department of Energys National Nuclear
Security Administration under Contract DE-AC04-94AL85000.
(the input vector multiplied by the synapse or weight matrix).
Conventional computer processors are not designed to process
information in the manner that a neural algorithm requires
(such as the vector-matrix multiply). Recently, major advances
in neural networks and deep learning have coincided with
advances in processing power and data access. However, we
are reaching the limits of Moore's law in terms of how much
more efficiency can be gained from conventional processing
architectures. In addition to reaching the limits of Moore's
law, conventional processing architectures also incur the von
Neumann bottleneck [4] where the processing unit's program
and data memory exist in a single memory with only one
shared data bus between them.
In contrast to conventional processing architectures which
consist of a powerful centralized processing unit(s) that oper-
ate(s) in a mostly serialized manner, the brain is composed of
many simple distributed processing units (neurons) that are
sparsely connected and operate in parallel. Communication
between neurons occurs at
the synaptic connection which
operate independently of the other neurons that are not in-
volved in the connection. Thus, vector-matrix multiplications
are implemented more efficiently facilitated by parallel oper-
ations. Additionally, the synaptic connections in the brain are
generally sparse and information is encoded in a combination
of the synaptic weights and the temporal latencies of a spike
on the synapse [5]. Biological synapses are not simply a
weighted binary connection but rather exhibit a non-linear
synaptic response function due to the release and dispersion
of neurotransmitters in the space between neurons.
Biological neurons communicate using simple "data pack-
ets," that are generally accepted as binary spikes. This is in
contrast to the neuron models used in traditional artificial neu-
ral networks (ANN) which are commonly rate coded neurons.
Rate coded neurons encode information between neurons as
a real-valued magnitude of the output of a neuron-a larger
output represents a higher firing rate. The use of rate coded
neurons stems from the assumption that the firing rate of a
Fig. 1. High level overview of the STPU. The STPU is composed of a set of leaky integrate and fire neurons. Each neuron has an associated temporal buffer
such that inputs can be mapped to a neuron with a time delay. W(t) is the neuronal encoding transformation which addresses connectivity, efficacy and
temporal shift. The functionality of the STPU mimics the of functionality of biological neurons.
neuron is the most important piece of information, whereas
temporally coded neurons encode information based on when
a spike from one neuron arrives at another neuron. Temporally
coded information has been shown to be more powerful than
rate coded information and more biologically accurate [6] .
Based on these neuroscience principles, we present the Spik-
ing Temporal Processing Unit (STPU), a novel neuromorphic
hardware architecture designed to mimic neuronal functional-
ity and alleviate the computational restraints inherent in con-
ventional processors. Other neuromorphic architectures have
shown very strong energy efficiency [7], powerful scalability
[8], and aggressive speed-up [9] by utilizing the principles ob-
served in the brain. We build upon these efforts leveraging the
benefits of low energy consumption, scalability, and run time
speed ups and include an efficient implementation of arbitrarily
complex synaptic response functions in a digital architecture.
This is important as the synaptic response function has strong
implications in spiking recurrent neural networks [10].
We also examine liquid state machines (LSMs) [11] to show
how the constructs available in the STPU facilitate complex
dynamical neuronal systems. While we examine the STPU in
the context of LSMs, the STPU is a general neuromorphic
architecture. Other spiked-based algorithms have been imple-
mented on the STPU [12], [13].
In Section II, we present the STPU. A high level comparison
with other neuromorphic architectures is presented in Section
III. We present LSMs in Section IV. In Section V, we examine
how LSMs map onto the STPU and show results from running
the LSM on the STPU. We conclude in Section VI.
II. THE SPIKING TEMPORAL PROCESSING UNIT
In this section, we describe the Spiking Temporal Processing
Unit (STPU) and how the components in the STPU map
to functionality in biological neurons. The design of the
STPU is based on the following three neuroscience principles
observed in the brain: 1) the brain is composed of simple
processing units (neurons) that operate in parallel and are
sparsely connected, 2) each neuron has its own local mem-
ory for maintaining temporal state, and 3) information is
encoded in the connectivity, efficacy, and signal propagation
characteristics between neurons. A high-level overview of a
biological neuron and how its components map onto the STPU
are shown in Figure 1. The STPU derives its dynamics from
the leaky integrate and fire (LIF) neuron model [14]. Each
LIF neuron j maintains a membrane potential state variable,
vj, that tracks its stimulation at each time step based on the
following differential equation [10]:
(cid:88)
(cid:88)
wkj · s(t − tkl − ∆kl).
(1)
dvj
dt
= − vj
τj
+
k
l
The variable τj is the time constant of the first-order dynamics,
k is the index of the presynaptic neuron, wkj is the weight
connecting neuron j to neuron k, tkl is the time of the lth spike
from neuron k, ∆kl is the synaptic delay from neuron k on the
lth spike, and s(·) is the dynamic synaptic response function
to an input spike. In the LIF model, neuron j will fire if vj
exceeds a threshold θj. The synapses between input neurons
to destination neurons are defined in the weight matrix W(t)
for a given time t as the weights between inputs and neurons
can change over time.
Unique to the STPU, each LIF neuron has a local temporal
memory buffer R composed of D memory cells to model
synaptic delays. When a biological neuron fires, there is a
latency associated with the arrival of the spike at the soma of
the postsynaptic neuron due to the time required to propagate
down the axon of the presynaptic neuron and the time to
propagate from the dendrite to the soma of the postsynaptic
neuron (∆kl). The temporal buffer represents different synaptic
junctions in the dendrites where a lower index value in the
temporal buffer constitutes a dendritic connection closer to the
soma and/or a shorter axon length than one with a larger index
value. Thus, synapses in the STPU are specified as a weight
wkjd from a source input neuron k, to a destination neuron j in
the dth cell of the temporal buffer, d ∈ {0, 1, 2, . . . , D − 1}.
This allows multiple connections between neurons with dif-
ferent synaptic delays. At each time step a summation of
the product of the inputs i(t) and synaptic weights W(t)
occurs and is added to the current value in that position of
k ik(t)wkjd(t) where
R(t) is a temporary state of the temporal buffer. The value
in each cell of the temporal buffer is then shifted down one
position, that is Rd(t+1) = Rd−1(t). The values at the bottom
of the buffer are fed into the LIF neuron.
the temporal buffer Rd(t) = Rd(t) +(cid:80)
In biological neurons, when a neuron fires a (near) binary
spike is propagated down the axon to the synapse, which
defines a connection between neurons. The purpose of the
synapse is to transfer the electric activity or information from
one neuron to another neuron. Direct electrical communication
does not take place, rather a chemical mediator is used. In the
presynaptic terminal, an action potential from the emitted spike
causes the release of neurotransmitters into the synaptic cleft
(space between the pre and postsynaptic neurons) from the
synaptic vescles. The neurotransmitters cross the synaptic cleft
and attach to receptors on the postsynaptic neuron injecting
a positive or negative current into the postsynaptic neuron.
Through a chemical reaction, the neurotransmitters are broken
down in receptors on the postsynaptic neuron and are released
back into the synaptic cleft where the presynaptic neuron
reabsorbs the broken down molecules to synthesize new neu-
rotransmitters. In terms of electrical signals, the propagation
of activation potentials on the axon is a digital signal as shown
in Figure 2. However, the chemical reactions that occur at the
synapse to release and reabsorb neurotransmitters are modeled
as an analog signal.
The behavior of the synapse propagating spikes between
neurons has important ramifications on the dynamics of the
liquid. In Equation 1, the synaptic response function is rep-
resented by s(·). Following Zhang et al. [10], the Dirac delta
function δ(·) can be used as the synaptic response function
and is convenient for implementation on digital hardware.
However, the Dirac delta function exhibits static behavior.
Zhang et al. show that dynamical behavior can be modeled in
the synapse by using the first-order response to a presynaptic
spike:
(2)
where τ s is the time constant of the first-order response, H(·)
is the Heaviside step function, and 1/τ s normalizes the first-
order response function. The dynamical behavior can also be
implemented using a second-order dynamic model for s(·):
) · H(t − tkl − ∆kl) (3)
− t−tkl−∆kl
− t−tkl−∆kl
− e
(e
τ s
2
τ s
1
1
τ s e− t−tkl−∆kl
τ s
· H(t − tkl − ∆kl)
1
1 − τ s
τ s
2
Fig. 2. Spike propagation along the axon and across the synapse. The spike
propagated on the axon is generally accepted as a binary spike. Upon arrival
at the synapse, the spike initiates a chemical reaction in the synaptic cleft
which stimulates the postsynaptic neuron. This chemical reaction produces
an analog response that is fed into the soma of the postsynaptic neuron. In
the STPU, arbitrary synaptic response functions are modeled efficiently using
the temporal buffer. The synaptic response function is discretely sampled and
encoded into the weights connecting one neuron to another and mapped to
the corresponding cells in the temporal buffer.
1 and τ s
2 are the time constants for the second
where τ s
order response and 1/(τ s
2 ) normalizes the second-order
dynamical response function. Zhang et al. showed significant
improvements in accuracy and the dynamics of the liquid when
using these dynamical response functions.
1 − τ s
Implementing exponential functions in hardware is expen-
sive in terms of the resources needed to implement ex-
ponentiation. Considering that
the STPU is composed of
individual parallel neuronal processing units, each neuron
would need its own exponentiation functionality. Including the
hardware mechanisms for each neuron to do exponentiation
would significantly reduce the number of neurons by orders
of magnitude as there are limited resources on an FPGA.
Rather than explicitly implement the exponential functions in
hardware, we use the temporal buffer associated with each
neuron. The exponential function is discretely sampled and
the value at each sample is assigned a connection weight
wkjd from the presynaptic neuron k to the corresponding
cell d in the temporal buffer of the postsynaptic neuron j.
Thus, a single weighted connection between two neurons is
expanded to multiple weighted connections between the same
two neurons. This is shown graphically in Figure 2. The use
of the temporal buffer allows for an efficient implementation
of the digital signal propagation down the axon of a neuron or
HIGH-LEVEL COMPARISON OF THE STPU WITH TRUE NORTH AND SPINNAKER.
TABLE I
Platform:
Interconnect:
Neuron Model:
Synapse Model:
STPU
3D mesh multicast1
LIF
Programmable4
TrueNorth
2D mesh unicast
LIF2
Binary
SpiNNaker
2D mesh multicast
Programmable3
Programmable5
1The 3D mesh is enabled due to the temporal buffer available for each neuron in STPU.
2TrueNorth provides a highly programmable LIF to facilitate additional neural dynamics.
3SpiNNaker provides flexibility for the neuron model, however more complex biological models are more computationally expensive.
4The synapse model is programmable in the STPU via the temporal buffer by discretely sampling an arbitrary synapse model.
5As with the neuron model, SpiNNaker is optimized for simpler synaptic models. More complex synaptic models incur a cost in computational complexity.
the analog signal propagation between neurons at the synapse.
III. COMPARISON WITH OTHER NEUROMORPHIC
ARCHITECTURES
The STPU is not the first neuromorphic architecture. Four
prominent neuromorphic architectures are IBM's TrueNorth
chip [7], the Stanford Neurogrid [15], the Heidelberg Brain-
ScaleS machine [16] and the Manchester Spiking Neural Net-
work Architecture (SpiNNaker) [17]. The Stanford Neurogrid
and the Heidelberg BrainScaleS are analog circuits while
TrueNorth and SpiNNaker are digital circuits. As the STPU
is also a digital system, we will focus on a comparison with
TrueNorth and SpiNNaker.
The TrueNorth chip leverages a highly distributed cross-
bar based architecture designed for high energy-efficiency
composed of 4096 cores. The base-level neuron is a highly
parametrized LIF neuron. A TrueNorth core is a 256 × 256
binary crossbar where the existence of the synapse is encoded
at each junction, and individual neurons assign weights to
particular sets of input axons. The crossbar architecture allows
for efficient vector-matrix multiplication. TrueNorth only al-
lows for point-to-point routing. Each of the 256 neurons on
a core is programmed with a spike destination addressed to a
single row on a particular core which could be the same core,
enabling recurrence, or a different core. The crossbar inputs
are coupled via delay buffers to insert axonal delays. A neuron
is not natively able to connect to multiple cores or to connect
to a single neuron with different temporal delays. As a work
around, a neuron is to be replicated within the same core and
mapped to the different cores. For multiple temporal delays
between two neurons (such as those in the STPU), there is no
obvious mechanism for an implementation [18].
SpiNNaker is a massively parallel digital computer com-
posed of simple ARM cores with an emphasis on flexibility.
Unlike the STPU and TrueNorth, SpiNNaker is able to model
arbitrary neuron models via an instruction set that is provided
to the ARM core. SpiNNAker is designed for sending large
numbers of small data packages to many destination neurons.
While SpiNNaker was designed for modeling neural networks,
it could potentially be used more generally due to its flexibility.
The STPU architecture falls in between the TrueNorth
and SpiNNaker architectures. The STPU implements a less
parameterized LIF neuron than TrueNorth, however, its routing
of neural spikes is more flexible and allows a multicast similar
to SpiNNaker rather than the unicast used in TrueNorth. A
key distinguishing feature of the STPU is the temporal buffer
associated with each neuron, giving the STPU 3-dimensional
routing. A high-level summary of the comparison of STPU
with TrueNorth and SpiNNaker is shown in Table I.
IV. LIQUID STATE MACHINES
The liquid state machine (LSM) [11] is a neuro-inspired
algorithm that mimics the cortical columns in the brain.
It is conjectured that the cortical microcircuits nonlinearly
project
input streams into a high-dimensional state space.
This high-dimension representation is then used as input to
other areas in the brain where learning can be achieved. The
cortical microcircuits have a sparse representation and fading
memory-the state of the microcircuit "forgets" over time.
While LSMs may be able to mimic certain functionality in
the brain, it should be noted that LSMs do not try to explain
how or why the brain operates as it does.
In machine learning, LSMs are a variation of recurrent neu-
ral networks that fall into the category of reservoir computing
(RC) [19] along with echo state networks [20]. LSMs differ
from echo state machines in the type of neuron model used.
LSMs use spiking neurons while echo state machines use rate
coded neurons with a non-linear transfer function.
LSMs operate on temporal data composed of multiple
related time steps. LSMs are composed of three general
components: 1) input neurons, 2) randomly connected leaky-
integrate and fire spiking neurons called the liquid, and 3)
readout nodes that read the state of liquid. A diagram of an
LSM is shown in Figure 3. Input neurons are connected to
a random subset of the liquid neurons. The readout neurons
may be connected to all the neurons in the liquid or a subset
of them. Connections between neurons in the liquid are based
on probabilistic models of brain connectivity [11]:
r2
− E(N1,N2 )
Pconnection(N1, N2) = q · e
(4)
where N1 and N2 represent two neurons and E(N1, N2) is
the Euclidean distance between N1 and N2. The variables q
and r are two chosen constants. In this paper, we use a 3-
dimensional grid to define the positions of neurons on the
liquid. The liquid functions as a temporal kernel, casting the
input data into a higher dimension. The LIF neurons allow
for temporal state to be carried from one time step to another.
LSMs avoid the problem of training recurrent neural models
V. MAPPING THE LSM ONTO THE STPU
In this section, we implement the LSM on the STPU. There
have been previous implementations of LSMs on hardware,
however, in most cases an FPGA or VLSI chip has been
designed specifically for a hardware implementation of an
LSM. Roy et al. [25] and also Zhang et al. [10] present
a low-powered VLSI hardware implementation of an LSM.
Schrauwen et al. [26] implement an LSM on an FPGA chip.
In contrast to other work, the STPU has been developed to
be a general neuromorphic architecture. Other neuroscience
work and algorithms are being developed against the STPU
such as spike sorting and using spikes for median filtering
[13]. Currently, we have an STPU simulator implemented in
MATLAB as well as an implementation on an FGPA chip. The
MATLAB simulator has a one-to-one correspondence with the
hardware implementation.
Given the constructs provided by the STPU, the LSM with
a liquid composed of LIF neurons maps naturally onto the
STPU. We use the second-order synaptic response function of
Equation 3 that is based on the work of Zhang et al. [10].
They found that the second-order response function produced
more dynamics in the liquid allowing the neural signals to
persist longer after the input sequence had finished. This lead
to improved classification results. Following Zhang et al., the
synaptic properties of the liquid, including parameters for the
connection probabilities between the liquid neurons defined in
Equation 4 and the synaptic weights, are given in Table II.
There are two types of neurons: excitatory (E) and inhibitory
(I). As has been observed in the brain [11], the liquid is
made up of an 80/20 network where 80% of the nuerons
are excitatory and 20% of the neurons are inhibitory. The
probability of a synapse existing between two neurons and
the weights between the neurons is dependent on the types of
the considered neurons. E/I → E/I denotes the presynaptic
and postsynaptic neurons being connected by the synapse. For
example, E → I denotes the connection between an excitatory
presynaptic neuron with an inhibitory postsynaptic neuron.
Excitatory neurons increase the action potential at a target
neurons (positive synaptic weights) while the inhibitory neu-
rons decrease the action potential (negative synaptic weights).
When the connections are generated between neurons in the
liquid,
the neurons are randomly connected according to
Equation 4 with the parameters given in Table II. Each input
neuron is randomly connected to a subset of 30% of the
neurons in the liquid with a weight of 8 or -8 chosen uniformly
at random.
To implement the second-order synaptic response function,
Equation 3 is sampled at discrete time steps and multiplied by
the synaptic weight value between the neurons as specified in
Table II. The discretely sampled weights are then encoded via
multiple weights at corresponding cells in the temporal buffer
for the postsynaptic neuron. In this implementation, there is
2 is set
no synaptic delay (∆kl = 0) and τ s
1 is set to 4 and τ s
to 8 for excitatory neurons. For inhibitory neurons, τ s
1 and τ s
2
are set to 4 and 2 respectively. For all neurons, τj is set to 32.
Fig. 3. A liquid state machine, composed of three components: 1) a set of
input neurons, 2) the liquid-a set of recurrent spiking neurons, and 3) a set
of readout neurons with plastic synapses that can read the state of the neurons
in the liquid.
PARAMETERS FOR THE SYNAPSES (OR CONNECTIONS BETWEEN
NEURONS) IN THE LIQUID.
TABLE II
Parameter
r from Equation 4
q from Equation 4
Synaptic weight
type
ALL
E → E
E → I
I → E
I → I
E → E
E → I
I → E
I → I
value
2
0.45
0.30
0.60
0.15
3
6
-2
-2
by only training the synaptic weights from the liquid to the
readout nodes, similar to extreme machine learning that use
a random non-recurrent neural network for non-temporal data
[21]. It is assumed that all temporal integration is encompassed
in the liquid. Thus, the liquid in an LSM acts similarly to
the kernel in a support vector machine on streaming data
by employing a temporal kernel. In general, the weights and
connections in the liquid do not change, although some studies
have looked at plasticity in the liquid [22].
The readout neurons are the only neurons that have plastic
synapses, allowing for synaptic weight updates via training.
Using each neurons firing state from the liquid, the temporal
aspect of learning on temporal data is transformed to a static
(non-temporal) learning problem. As all temporal integration
is done in the liquid, no additional mechanisms are needed to
train the readout neurons. Any classifier can be used, but often
a linear classifier is sufficient. Training of the readout neurons
can be done in a batch or on-line manner [10].
LSMs have been successfully applied to several applications
including speech recognition [10], vision [23], and cognitive
neuroscience [11], [24]. Practical applications suffer from the
fact that traditional LSMs take input in the form of spike trains.
Transforming numerical input data into spike data, such that
the non-temporal data is represented temporally, is nontrivial.
The plastic readout neurons are connected to all of the
neurons in the liquid. Training is done off-line using a linear
classifier on the average firing rate of the neurons in the liquid.
We examine the effect of the various linear classifiers below.
A. Experiments
To evaluate the effect of different parameters for the liquid
state machine, we use a data set for spoken digit recognition
of Arabic digits from 0 to 9 [27]. The dataset is composed of
the time series Mel-frequency cepstral coefficients (MFCCs)
of 8800 utterances of each digit from 88 speakers on with 10
repetitions per digit (10 × 10 × 88). The MFCCs were taken
from 44 male and 44 female native Arabic speakers between
the ages of 18 and 40. The dataset is partitioned into a training
set from 66 speakers and a test set from the other 22 speakers.
We scale all variables between 0 and 1.
To evaluate the performance of the LSM, we examine
the classification accuracy on the test set, and measure the
separation in the liquid from the training set. If there is good
separation within the liquid, then the state vectors from the
trajectories for each class should be distinguishable from each
other. To measure the separability of a liquid Ψ on a set of
state vectors O from the liquid perturbed by a given input
sequence, we follow the definition from Norton and Ventura
cv+1 where cd is the inter-class distance
[22]: Sep(Ψ, O) = cd
and cv is the intra-class variance. Separation is the ratio of the
distance between the classes divided by the class variance.
ok∈Ol
l=1
(cid:80)n
(cid:80)n
l=1
m=1
n2
(cid:80)
(cid:80)n
(cid:107)µ(Ol)−µ(Om)(cid:107)2
the center of mass
The inter-class difference is
the mean difference of
for every pair of classes: cd =
where (cid:107)·(cid:107)2 is the L2 norm, n is
the number of classes, and µ(Ol) is the center of mass for each
class. For a given class, the intra-class variance is the mean
variance of the state vectors from the inputs from the center
of mass for that class: cv = 1
n
(cid:107)µ(Ol)−ok(cid:107)2
Ol
We investigate various properties of the liquid state ma-
chine, namely the synaptic response function, the input en-
coding scheme, the liquid topology, and the readout train-
ing algorithm. We also consider the impact of θj on the
liquid. A neuron j will spike if vj exceeds θj. Thus, θj
can have a significant impact on the dynamics of the liq-
uid. Beginning with a base value of 20 (as was used by
Zhang et al.) we consider the effects of decreasing values
of θj ∈ {20, 17.5, 15, 12.5, 11, 10, 9, 7.5, 5, 3, 2.5, 2, 1}. For
default parameters, we use a reservoir of size 3 × 3 × 15, we
feed the the magnitude of the inputs into the input neurons
(current injection) and a linear SVM to train the synapses for
the readout neurons.
1) Synaptic Response Functions: We first investigate the
effect of the synaptic response function using default param-
eters. Using θj = 20, the average separation values, average
spiking rates and the classification accuracy of a linear SVM
are given in Table III. As highlighted in bold, the second-order
synaptic function (Equation 3) achieves the largest separation
values for training and testing, the lowest average spike rate,
and the highest classification accuracy. The average spike rate
is significantly higher for the first-order response function
SEPARATION VALUES, AVERAGE SPIKING RATES, AND CLASSIFICATION
ACCURACY FROM DIFFERENT SYNAPTIC RESPONSE FUNCTIONS.
TABLE III
Synaptic Res
Dirac Delta
First-Order
Second-Order
First-Order 30
First-Order 40
First-Order 50
TrainSep
TrainRate
0.129
0.251
0.263
0.352
0.293
0.129
0.931
0.845
0.261
0.689
0.314
0.138
TestSep
0.139
0.277
0.290
0.389
0.337
0.134
TestRate
0.931
0.845
0.255
0.688
0.314
0.138
SVM
0.650
0.797
0.868
0.811
0.817
0.725
l
a
v
e
s
n
o
p
s
e
r
0.2
0.1
0
First-Order
Second-Order
0
5
10
15
time
20
25
30
35
Fig. 4. Visualization of the first- and second-order response functions.
than for the second-order response function, which is counter-
intuitive since the second-order response function perpetuates
the signal through the liquid longer. However, examining the
first-order and second-order response functions, as shown in
Figure 4, shows that the first-order response function has a
larger initial magnitude and then quickly subsides. The second-
order response function has a lower initial magnitude, but is
slower to decay giving a more consistent propagation of the
spike through time.
Adjusting the value of θj to 30, 40, and 50 accommodate
this behavior for the first-order response function (the bottom
three rows int he Table III) shows that an improvement can be
made in the separation values, spiking rate, and classification
accuracy. Despite this improvement, the first-order response
function does not achieve a better performance than the
second-order response function for the classification accuracy.
The first-order response function does get a better separation
score, but this does not translate into better accuracy.
2) Input Encoding Schemes: In traditional LSMs, the input
is temporally encoded in the form of a spike train. Un-
fortunately, most datasets are not temporally encoded, but
rather are numerically encoded. A spike train input aligns
with neuroscience, but practically it is non-trivial to encode
all information temporally as the brain does. Therefore, we
examine three possible encoding schemes: 1) rate encoding
where the magnitude of the numeric value is converted into
a rate and a spike train at that rate is fed into the liquid 2)
bit encoding where the magnitude of the numeric value is
converted into its bit representation at a given precision, and
3) current injection. Rate encoding requires n time steps to
encode a single input by converting a magnitude to a rate.
This is similar to binning and has some information loss. Bit
encoding, only requires one time step, however, it requires
m inputs per standard input to convert the magnitude into its
.
THE SEPARATION VALUES, AVERAGE SPIKING RATES OF THE LIQUID, AND
TABLE IV
CLASSIFICATION ACCURACY EXAMINING LIQUID TOPOLOGIES WITH
DIFFERENT INPUT ENCODING SCHEMES AND VALUES FOR θj . THE
LARGEST SEPARATION VALUES AND ACCURACIES FOR EACH ENCODING
SCHEME ARE IN BOLD
Encoding Scheme
20
Current Injection
Bit Encoding
Rate Encoding
0.263
0.261
0.868
0.271
0.434
0.741
0.164
0.146
0.747
θj
10
0.378
0.750
0.894
0.338
0.544
0.735
0.622
0.594
0.643
5.5
0.334
0.843
0.873
0.350
0.592
0.755
0.197
0.952
0.601
15
0.409
0.580
0.905
0.310
0.497
0.741
0.364
0.199
0.733
3
0.324
0.873
0.868
0.353
0.634
0.764
0.047
0.985
0.548
2
0.324
0.878
0.866
0.357
0.620
0.761
0.048
0.985
0.558
m bit-precise representation. We set m to 10. Compared to
current injection, the execution time increases linearly in the
number of time steps for rate encoding.
Table IV shows the separation values (first row for each
encoding scheme), average spiking rates (second row), and the
accuracy from a linear SVM on the test set (third row) for the
input encoding schemes with various values of θj. The average
spiking rate gives the percentage of neurons that were firing
in the liquid through the time series and provides insight into
how sparse the spikes are within the liquid. Table IV shows
a representative subset of the values that were used for θj.
The bold values represent the highest separation value and
classification accuracy for each encoding scheme.
The results show that the value of θj has a significant effect
on the separation of the liquid as well as the classification
accuracy from the SVM. This is expected as the dynamics of
the liquid are dictated by when neurons fire. A lower threshold
allows for more spikes as is indicated in the increasing values
of the average spiking rates as the values for θj decrease.
Overall, using rate encoding produces the greatest values in
separation. However,
there is significant variability as the
values for θj change. For rate encoding, the greatest accuracy
from the SVM is achieved with a low separation value. In the
other encoding schemes, separation and classification accuracy
appear to be correlated. The greatest classification accuracy is
achieved from current injection.
3) Liquid Topology: The topology of the liquid in an LSM
determines the size of the liquid and influences the connections
within the liquid as the distance between the neurons impacts
the connections made between neurons. A more cubic liquid
(e.g. 5×5×5) should be more densely connected compared to
a column of liquid (e.g. 2×2×20). In this section, we examine
using liquids with grids of 3× 3× 15, 2× 2× 20, 4× 5× 20,
and 5 × 5 × 5. As before, we consider different values for θj.
The separation values, average spike rates, and the accuracy
from a linear SVM are given in Table V for the values of θj
that provided the largest separation values for each topology
configuration and encoding scheme combination.
THE SEPARATION VALUES AND AVERAGE SPIKE RATE OF THE LIQUID
USING DIFFERENT LIQUID TOPOLOGIES. THE LARGEST SEPARATION
TABLE V
VALUES AND ACCURACIES FOR EACH TOPOLOGY ARE IN BOLD
Topology/
Num Neurons
θj:
3x3x15/135
5x5x5/125
θj:
θj:
4x5x10/200
θj:
2x2x20/80
Current Injection
15
0.409
0.580
0.905
12.5
0.467
0.534
0.889
15
0.418
0.698
0.879
11
0.397
0.608
0.903
12.5
0.405
0.693
0.884
11
0.432
0.682
0.900
12.5
0.384
0.760
0.876
10
0.389
0.658
0.912
Encoding Scheme
Bit Encoded
3
2
0.353
0.634
0.764
3
0.384
0.596
0.765
15
0.506
0.576
0.830
3
0.264
0.537
0.672
0.357
0.620
0.761
2
0.380
0.586
0.755
12.5
0.501
0.603
0.835
2
0.247
0.550
0.666
Rate Encoded
10
11
0.622
0.622
0.594
0.563
0.658
0.643
10
0.384
0.473
0.637
15
0.479
0.289
0.644
10
0.305
0.609
0.695
9
0.451
0.600
0.610
12.5
0.352
0.577
0.612
9
0.255
0.846
0.678
Again, the value for θj has a significant impact on the
separation of the liquid and the classification accuracy. The
greatest separation values and classification accuracies for each
topology is highlighted in bold. For all of the topologies,
current injection achieves the highest classification accuracy.
Interestingly, the separation values across encoding schemes
and topologies do not correlate with accuracies. Within the
same encoding scheme and topology, however, the accuracy
generally improves as the separation increases.
For current injection, the different topologies do not appear
to have a significant impact on the classification accuracy
except for the 4 × 5 × 10 topology which has a decrease in
accuracy. This may be due to increased number of liquid nodes
that are used as input to the SVM. The converse is true for
bit encoding as the 4 × 5 × 10 topology achieves the highest
accuracy possibly due to the increased number of inputs due
to the bit representation of the input.
4) Readout Training Algorithms: How the plastic synapses
are trained will have a significant effect on the performance of
the LSM. Traditionally, LSMs use a linear classifier based on
the assumption that the liquid has transformed the state space
such that the problem is linearly separable. Linear models are
represented as a set of weights and a threshold-which can
be implemented in neuromorphic hardware. By using a linear
model, the liquid and the classification can all be done on
the STPU avoiding the overhead of going off chip to make a
prediction. We consider four linear classifiers: 1) linear SVM,
2) linear discriminant analysis (LDA), 3) ridge regression, and
4) logistic regression. With each of these algorithms, we use
the default parameters as they are set in the Statistics and
Machine Learning Toolbox in MATLAB.
We examine the classification of each of the linear classifiers
CLASSIFICATION ACCURACY ON THE TEST SET FROM DIFFERENT LINEAR
CLASSIFIERS. THE GREATEST ACCURACY FOR EACH TOPOLOGY IS IN
TABLE VI
Linear Model:
Linear SVM
LDA
Ridge Regress
Logistic Regress
3 × 3 × 15
θj = 15
0.906
0.921
0.745
0.431
BOLD
5 × 5 × 5
θj = 11
0.900
0.922
0.717
0.254
4 × 5 × 10
θj = 15
0.900
0.922
0.717
0.254
2 × 2 × 20
θj = 10
0.914
0.946
0.897
0.815
on the topologies and θj values that achieved the highest clas-
sification accuracy on the linear SVM in previous experiments.
We also limit ourselves to examining current injection for the
input scheme as current injection consistently achieved the
highest classification accuracy. The results are shown in Table
VI. LDA consistently achieves the highest classification ac-
curacy of the considered classifiers. The highest classification
accuracy achieved is 0.946.
VI. CONCLUSION AND FUTURE WORK
In this paper, we presented the Spiking Temporal Processing
Unit or STPU-a novel neuromorphic processing architecture.
It is well suited for efficiently implementing neural networks
and synaptic response functions of arbitrary complexity. This
is facilitated by using the temporal buffers associated with
each neuron in the architecture. The capabilities of the STPU,
including complex synaptic response functions, were demon-
strated through implementing the functional mapping and
implementation of an LSM onto the STPU architecture.
As neural algorithms grow in scale and conventional pro-
cessing units reach the limits of Moore's law, neuromorphic
computing architectures, such as the STPU, allow efficient
implementations of neural algorithms. However, neuromorphic
hardware is based on spiking neural networks to achieve low
energy. Thus, more research is needed to understand and
develop spiking-based algorithms.
REFERENCES
[1] L. Deng, G. E. Hinton, and B. Kingsbury, "New types of deep neural
network learning for speech recognition and related applications: an
overview," in Proceedings of International Conference on Acoustics,
Speech, and Signal Processing (ICASSP), 2013, pp. 8599–8603.
[2] D. Ciresan, U. Meier, and J. Schmidhuber, "Multi-column deep neural
networks for image classification," in Proceedings of the 2012 IEEE
Conference on Computer Vision and Pattern Recognition.
IEEE
Computer Society, 2012, pp. 3642–3649.
[3] R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Ng,
and C. Potts, "Recursive deep models for semantic compositionality
over a sentiment treebank," in Proceedings of the 2013 Conference on
Empirical Methods in Natural Language Processing. Association for
Computational Linguistics, October 2013, pp. 1631–1642.
[4] J. Backus, "Can programming be liberated from the von neumann style?:
A functional style and its algebra of programs," Communications of the
ACM, vol. 21, no. 8, pp. 613–641, Aug. 1978.
[5] P. L. Follett, C. Roth, D. Follett, and O. Dammann, "White matter
damage impairs adaptive recovery more than cortical damage in an in
silico model of activity-dependent plasticity," Journal of Child Neurol-
ogy, vol. 24, no. 9, pp. 1205–1211, 2009.
[6] T. J. Sejnowski, "Time for a new neural code?" Nature, vol. 376, no.
6535, pp. 21–22, Jul. 1995.
[7] P. Merolla, J. Arthur, R. Alvarez-Icaza, A. Cassidy, J. Sawada,
F. Akopyan, B. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo,
I. Vo, S. Esser, R. Appuswamy, B. Taba, A. Amir, M. Flickner, W. Risk,
R. Manohar, and D. Modha, "A million spiking-neuron integrated circuit
with a scalable communication network and interface," Science, pp. 668–
673, August 2014.
[8] S. B. Furber, D. R. Lester, L. A. Plana, J. D. Garside, E. Painkras,
S. Temple, and A. D. Brown, "Overview of the spinnaker system
architecture," IEEE Transactions on Computers, vol. 62, no. 12, pp.
2454–2467, 2013.
[9] J. Schemmel, J. Fieres, and K. Meier, "Wafer-scale integration of analog
neural networks," in IEEE International Joint Conference on Neural
Networks, june 2008, pp. 431 –438.
[10] Y. Zhang, P. Li, Y. Jin, and Y. Choe, "A digital liquid state machine with
biologically inspired learning and its application to speech recognition,"
IEEE Transactions on Neural Networks and Learning Systems, vol. 26,
no. 11, pp. 2635–2649, 2015.
[11] W. Maass, T. Natschlager, and H. Markram, "Real-time computing
without stable states: A new framework for neural computation based
on perturbations," Neural Computation, vol. 14, no. 11, pp. 2531–2560,
Nov. 2002.
[12] S. J. Verzi, C. M. Vineyard, E. D. Vugrin, M. Galiardi, C. D. James, and
J. B. Aimone, "Optimization-based computation with spiking neurons,"
in Proceedings of the IEEE International Joint Conference on Neural
Network, 2017, p. Accepted.
[13] S. J. Verzi, F. Rothganger, O. D. Parekh, T.-T. Quach, N. E. Miner, C. D.
James, and J. B. Aimone, "Computing with spikes: The advantage of
fine-grained timing," submitted.
[14] P. Dayan and L. F. Abbott, Theoretical neuroscience : computational
and mathematical modeling of neural systems, ser. Computational neu-
roscience. Cambridge (Mass.), London: MIT Press, 2001.
[15] B. V. Benjamin, P. Gao, E. McQuinn, S. Choudhary, A. Chandrasekaran,
J.-M. Bussat, R. Alvarez-Icaza, J. Arthur, P. Merolla, and K. Boahen,
"Neurogrid: A mixed-analog-digital multichip system for large-scale
neural simulations," Proceedings of the IEEE, vol. 102, no. 5, pp. 699–
716, 2014.
[16] J. Schemmel, D. Briiderle, A. Griibl, M. Hock, K. Meier, and S. Millner,
"A wafer-scale neuromorphic hardware system for large-scale neural
modeling," in Proceedings of 2010 IEEE International Symposium on
Circuits and Systems, May 2010, pp. 1947–1950.
[17] S. Furber, F. Galluppi, S. Temple, and L. A. Plana, "The spinnaker
project," Proceedings of the IEEE, vol. 102, no. 5, pp. 652–665, 2014.
[18] W. Severa, K. D. Carlson, O. Parekh, C. M. Vineyard, and J. B. Aimone,
"Can we be formal in assessing the strengths and weaknesses of neural
architectures? A case study using a spiking cross-correlation algorithm,"
in NIPS Workshop Computing with Spikes, 2016.
[19] M. Lukosevicius and H. Jaeger, "Reservoir computing approaches to
recurrent neural network training," Computer Science Review, vol. 3,
no. 3, pp. 127–149, August 2009.
[20] H. Jaeger, "Adaptive nonlinear system identification with echo state
networks," in Advances in Neural Information Processing Systems 15.
MIT Press, 2003, pp. 609–616.
[21] G. B. Huang, Q. Y. Zhu, and C. K. Siew, "Extreme learning machine:
theory and applications," Neurocomputing, vol. 70, no. 1-3, pp. 489–501,
2006.
[22] D. Norton and D. Ventura, "Improving liquid state machines through
iterative refinement of the reservoir," Neurocomputing, vol. 73, no. 16-
18, pp. 2893–2904, 2010.
[23] H. Burgsteiner, M. Kroll, A. Leopold, and G. Steinbauer, "Movement
prediction from real-world images using a liquid state machine," Applied
Intelligence, vol. 26, no. 2, pp. 99–109, 2007.
[24] D. Buonomano and W. Maass, "State-dependent computations: Spa-
tiotemporal processing in cortical networks," Nature Reviews Neuro-
science, vol. 10, no. 2, pp. 113–125, 2009.
[25] S. Roy, A. Banerjee, and A. Basu, "Liquid state machine with dendrit-
ically enhanced readout for low-power, neuromorphic VLSI implemen-
tations," IEEE Transactions on Biomedical Circuits and Systems, vol. 8,
no. 5, 2014.
[26] B. Schrauwen, M. D'Haene, D. Verstraeten, and D. Stroobandt, "Com-
pact hardware liquid state machines on fpga for real-time speech
recognition," Neural Networks, no. 21, pp. 511–523, 1 2008.
[27] N. Hammami and M. Bedda, "Improved tree model for arabic speech
recognition," in Proceddings of the IEEE International Conference on
Computer Science and Information Technology, 2010.
|
1505.01228 | 2 | 1505 | 2015-05-07T02:41:07 | Superchords: the atoms of thought | [
"q-bio.NC",
"q-bio.QM"
] | Electroencephalography (EEG) signals' interpretation is based on waveform analysis, where meaningful information should emerge from a plethora of data. Nonetheless, the continuous increase in computational power and the development of new data processing algorithms in the recent years have put into reach the possibility of analysing raw EEG signals. Bearing that motivation, the authors propose a new approach using raw data EEG signals and deep learning neural networks, for the classification of motor activities (executed and imagery). The hypothesis to be presented here is: each instantaneous measurement of the raw signal of all EEG channels (superchord) is unique per motor activity regardless the moment of measurement. This study has confirmed the hypothesis (results with accuracy over 80%, mean for 109 subjects), reinforcing the need of further research for the understanding of mental processes. | q-bio.NC | q-bio | Superchords: the atoms of thought1
NORMAND, Rogério & FERREIRA, Hugo Alexandre
Institute of Biophysics and Biomedical Engineering (IBEB)
Faculty of Science, University of Lisbon
Abstract
Electroencephalography (EEG) signals’ interpretation is based on waveform analysis, where
meaningful information should emerge from a plethora of data. Nonetheless, the continuous
increase in computational power and the development of new data processing algorithms in the
recent years have put into reach the possibility of analysing raw EEG signals. Bearing that
motivation, the authors propose a new approach using raw data EEG signals and deep learning
neural networks, for the classification of motor activities (executed and imagery). The hypothesis to
be presented here is: each instantaneous measurement of the raw signal of all EEG channels
(superchord) is unique per motor activity regardless the moment of measurement. This study has
confirmed the hypothesis (results with accuracy over 80%, mean for 109 subjects), reinforcing the
need of further research for the understanding of mental processes.
Subject Terms: Artificial Neural Networks, Brainprint, Deep Learning, EEG, Flatcharts,
Neuroscience, Superchords.
Introduction
Electroencephalography (EEG) records the scalp voltage of brain activities with a high temporal
resolution. Microstates are defined as "quasi-stable spatial distributions (landscapes) of electric
potential that are connected by quick changes in landscapes” (Lehmann et al, 2009). Our brains
present a quasi-stable brain topography for sub-second periods of time intercalated with abrupt
changes. Those processes are believed to be connected with brain functional states during the
execution of activities and are generated by changes on the brain electric field.
The authors would like to propose a new approach potentially useful for better understanding of
mental processes through EEG. In the Brain Orchestra Approach (BOA), a superchord is defined
as the instantaneous measurement of the raw signal of all EEG channels, analog to a musical
chord .3
The hypothesis tested here is that for any single motor activity, each superchord is unique and it
can be used to determine that activity, regardless the moment of measurement.
2
1
A small homage to Dr. Dietrich Lehmann (Dec. 3, 1929 - Jun. 16, 2014) - Lehmann, 1990.
One neurone is equivalent to a note; a group of neurones equals to an instrument; a group of similar
2
instruments is like a brain region; so, the brain is like an orchestra. Electrodes work like microphones,
capturing the electric signals from brain processes with “louder" signals from regions close to them.
Superchords are represented in a straight horizontal line of pixels, where each one represents a channel
3
and the colour scale varies from bright green to bright red, respectively for higher and lower than mean raw
voltage values. Black colour represents the mean of all channels in the superchord.
5
of
1
In order to facilitate visualisation, authors have developed the concept of a flatchart , where it is
possible to identify the brain signature/pattern for a specific task, the activity brainprint . Fig. 1
pictures brainprints of making-a-fist activity with left and right hand.
5
4
z-score
4σ
3σ
2σ
1σ
0
-1σ
-2σ
-3σ
-4σ
Figure 1: On the left, the brainprint of a left hand make-a-fist, displaying 100 superchords (y-
axis) and 64 channels (x-axis), for subject 84, where red/green scale represents negative/
positive values in terms of standard deviations for each channel. On the right side, the same for
the right hand. The period of each superchord is 6.25 ms (1/160 Hz).
6
Methodology
Using a public dataset (Goldberger, 2000; Schalk, 2004), R language (R Core Team, 2014;
RStudio Team, 2012) and H2O Deep Learning algorithms (Ambati, Candel, Click et al., 2015), it
was possible to test the hypothesis.
The dataset was produced from a visual stimulus/motor response paradigm with 5 possible
reactions (movements with left hand, right hand, both hands, both feet or resting), also segregated
into real actions or imagery (R/I), for 109 subjects with a sampling rate of 160 Hz (period = 1/160 =
6.25 ms).
The methodology applied was to read the datasets for each subject, adding proper label for each
performed task, having columns as channels/task label and measurements per period for each
variable (channels/task label) in rows (superchords). Considering that the mind focus is higher just
after the stimulus, the complete dataset is sliced to consider only the first 10, 20, 30, 40, 60, 80,
100, 120, 140 and 160 first superchords (time range from 62.5 to 1000 ms) for each stimulus,
originating ten new sliced datasets.
Each subject sliced dataset is divided in two equal parts after a H2O built-in shuffle per row
(randomisation), being the first part used for training and the second used for validation purposes.
7
A flatchart is the result of chronologically juxtaposing superchords, representing in a 2D chart a 4D process
4
(3D spatial one for the brain and one for time). Through flatcharts it is possible to observe the whole event at
once. Flatcharts have superchords in rows and channels in columns.
A brainprint is the specific signature of a brain activity displayed in a flatchart. It is built through
5
superimpositions of same activity flatcharts, previously normalised by channel (mean equals to 0 and divided
by standard deviation). The final flatchart is normalised once more to be independent of the number of
superimposed flatcharts. Red/Green colour scale is used to emphasise lower/higher values (voltage).
H2O is the leading open source machine learning project from the H2O.ai community.
6
Each dataset was composed by 66 columns and a different number of superchords/rows (according to time
7
slices). The first column (timestamp) was not used. Columns 2 to 65 represent each channel raw data and
column 66 is the label for one of the 5 possible tasks.
5
of
2
8
A programming script in R/RStudio was developed to feed the training dataset into the H2O
classification algorithm in order to generate a model to be tested against the validation dataset.
The classification error was calculated as the number of wrong predictions divided by the number
of total predictions and the mean error was defined as the sum of each case subject error divided
by the total number of subjects (109).
Results & Discussion
In the fig. 2 below it is possible to compare the worse and the best case scenarios for this study.
Figure 2: On the left, the error measurement for each subject is shown considering only the first 10
superchords after the stimulus, being red for the real actions and blue for the imagery. Horizontal
lines correspond to the mean error for each set (Real ≈ 18.24%; Imagery ≈ 19.79%). On the right
side, the error results for the same subject are shown, but now considering the first 100
superchords (Real ≈ 10.71%; Imagery ≈ 11.74%).
In the worse case, with only 10 superchords after the stimulus has been presented, the results
have a mean error of ≈18.24% for real movements and ≈19.79% for imagery. By itself, this is a
strong encouragement towards the correctness of the hypothesis. The best case, with 100
superchords, the errors, on average, were ≈10.71% and ≈11.74%, respectively for real/imagery
datasets, another strong evidence to support the correctness of the proposed hypothesis. A chart
of the mean error per superchord time slice is displayed in fig. 3 - left side.
Figure 3: On the left, the mean error for each superchord time slice (10, 20, 30, 40, 60, 80, 100,
120, 140 and 160) is shown. Y-axis is defined between 0.10-0.20 to better display the differences
among the results. On the right side, the error per subject for the first 100 superchords is shown in a
density histogram measured in percentage and for real actions.
8
R scripts could be provided on a case-by-case basis. Please contact the authors.
3
of
5
It is noteworthy to mention that, with 100 superchords datasets for real actions, 39 subjects
presented errors below 1% (69 lower than 5%, 74 lower than 10%), meaning that for those
individuals, with a single superchord, regardless the moment of measurement, it would be possible
to determine their task/mental activity in each of the 5 possibilities with great precision. In fig. 3 -
right side shows the density distribution of errors for real actions and first 100 superchords (≈67%
of subjects with less than 7% error) . 9
Nevertheless, some considerations deserve attention:
1. Based on the paradigm protocol, the experiment was not totally random as each action task
should be followed by a resting task. At each round, only 3 task options were available for each
subject, no matter if real or imagery. Due to that, there is the possibility of subject anticipation
(Chavarriaga, 2012), which allowed an individual to be ready for the next task prior to the
stimulus;
2. The quality of the measurements, despite the best efforts employed: there is also a
considerable chance of electrode misplacement, saline bridges, low signal-to-noise ratio, etc.;
3. Sample representativity: there is no information regarding how subjects were selected and if
they are representative of the population, as well as to their mental health, use of drugs or any
other factor that may have contributed to measurement differences;
4. Paradigm type: the protocol involved only real or imagery motor reactions to stimuli.
On the other hand, if 1) is true, it would reinforce the proposed hypothesis, as anticipation would
create specific superchords for a mental task. The others aspects should be objectively considered
after new tests with specific protocols to avoid possible biases. Although the dataset referred motor
activities, the time slices considered, as well the presence of imagery rounds, points towards the
possibility of the correctness of the hypothesis for other mental activities.
Conclusions
Despite more work is required on this subject, there are positive indications that superchords are
singular to each motor task present on the dataset. The authors do believe it would be possible to
expand the concept of singular superchord to other mental activities, like reactions to faces or
words (another paper, covering this possibility, is being prepared).
More than to provide definitive conclusions, the authors would like to collaborate with those
interested to apply this methodology to other paradigms, experiments and cases in order to
promote advancements in neuroscience.
Acknowledgements
The authors thank Fundação para a Ciência e Tecnologia and Ministério da Ciência e Educação
Portugal (PIDDAC) for financial support under the grant UID/BIO/00645/2013. Authors also would
like to convey their thanks to IBEB/FCUL, the Portuguese National Distributed Computing
Infrastructure and H2O team for valuable contributions and support during all phases of the study.
The hardware employed to obtain the results was a cluster of 16 CPUs with 64 GB RAM running on Linux.
9
The use of lower amount of RAM can reduce the accuracy in up to 1 percent point. Due to the nature of
Deep Learning neural networks, the results can present slight variations (up to ±0.5 percent point).
5
of
4
References
Ambati SriSatish, Candel Arno, Click Cliff, Aboyoun Patrick, Aiello Spencer, Fu Anqi, Kraljevic Tom
et al., H2O, the leading open source machine learning project from the H2O.ai community, http://
h2o.ai, cf. https://github.com/h2oai/h2o for full list of contributors, 2015.
Chavarriaga, Ricardo, et al. "Anticipation-and error-related EEG signals during realistic human-
machine interaction: A study on visual and tactile feedback."Engineering in Medicine and Biology
Society (EMBC), 2012 Annual International Conference of the IEEE. pp. 6723-6726; (2012).
Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PCh, Mark RG, Mietus JE, Moody GB,
Peng C-K, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New
Research Resource for Complex Physiologic Signals. Circulation 101(23):e215-e220 [Circulation
Electronic Pages; http://circ.ahajournals.org/cgi/content/full/101/23/e215]; 2000 (June 13). - http://
www.physionet.org/pn4/eegmmidb/
Gregory R. Warnes, Ben Bolker, Lodewijk Bonebakker, Robert Gentleman, Wolfgang Huber Andy
Liaw, Thomas Lumley, Martin Maechler, Arni Magnusson, Steffen Moeller, Marc Schwartz and Bill
Venables (2015). gplots: Various R Programming Tools for Plotting Data. R package version
2.16.0. http://CRAN.R-project.org/package=gplots
Lehmann, Dietrich. et al. (2009) EEG microstates. Scholarpedia, 4(3):7632.
Lehmann, Dietrich. "Brain electric microstates and cognition: the atoms of thought." Machinery of
the Mind. Birkhäuser Boston, 1990. 209-224.
R Core Team (2014). R: A language and environment for statistical computing. R Foundation for
Statistical Computing, Vienna, Austria. URL http://www.R-project.org/.
RStudio Team (2012). RStudio: Integrated Development for R. RStudio, Inc., Boston, MA URL
http://www.rstudio.com/.
Schalk, G., McFarland, D.J., Hinterberger, T., Birbaumer, N., Wolpaw, J.R. BCI2000: A General-
Purpose Brain-Computer Interface (BCI) System. IEEE Transactions on Biomedical Engineering
51(6):1034-1043, 2004.
5
of
5
|
1903.10446 | 1 | 1903 | 2019-03-25T16:27:07 | The functional role of cue-driven feature-based feedback in object recognition | [
"q-bio.NC",
"cs.CV",
"cs.LG"
] | Visual object recognition is not a trivial task, especially when the objects are degraded or surrounded by clutter or presented briefly. External cues (such as verbal cues or visual context) can boost recognition performance in such conditions. In this work, we build an artificial neural network to model the interaction between the object processing stream (OPS) and the cue. We study the effects of varying neural and representational capacities of the OPS on the performance boost provided by cue-driven feature-based feedback in the OPS. We observe that the feedback provides performance boosts only if the category-specific features about the objects cannot be fully represented in the OPS. This representational limit is more dependent on task demands than neural capacity. We also observe that the feedback scheme trained to maximise recognition performance boost is not the same as tuning-based feedback, and actually performs better than tuning-based feedback. | q-bio.NC | q-bio |
The functional role of cue-driven feature-based feedback in object recognition
Sushrut Thorat
Marcel van Gerven
Marius Peelen
[email protected], [email protected], [email protected]
Donders Institute for Brain, Cognition and Behaviour, Radboud University
Abstract
Visual object recognition is not a trivial task, especially
when the objects are degraded or surrounded by clutter
or presented briefly. External cues (such as verbal cues
or visual context) can boost recognition performance in
such conditions. In this work, we build an artificial neural
network to model the interaction between the object pro-
cessing stream (OPS) and the cue. We study the effects of
varying neural and representational capacities of the OPS
on the performance boost provided by cue-driven feature-
based feedback in the OPS. We observe that the feed-
back provides performance boosts only if the category-
specific features about the objects cannot be fully rep-
resented in the OPS. This representational limit is more
dependent on task demands than neural capacity. We
also observe that the feedback scheme trained to max-
imise recognition performance boost is not the same as
tuning-based feedback, and actually performs better than
tuning-based feedback.
Keywords: feature-based feedback; vision; neural networks
Introduction
Visual object recognition is a non-trivial task, especially when
the objects are degraded, surrounded by clutter or presented
briefly. The introduction of external cues (such as verbal cues
or visual context) can constrain the space of the possible ob-
ject features and/or categories and improve recognition per-
formance (Carrasco, 2011; Bar, 2004).
External cues can interact with the object processing
stream in two ways. They can either modulate the information
transformations in the object processing stream (through feed-
back) and/or get combined with the object evidence present
at the end of the stream, to improve the overall decision about
the category of the object. Intuitively, the interaction involving
feedback would help with object recognition especially when
the feature information required to recognise the object cannot
be extracted by the object processing stream. This can hap-
pen either due to a capacity limit or due to a lack of information
present in the input.
Such a capacity limit can arise due to two reasons. One,
due to a limit on the number of neurons available in the object
processing stream, which would reduce the object information
that can be extracted from the image. We term this the neu-
ral capacity limit. Two, due to the limits imposed by the task
for which the stream is trained. For example, if the stream is
trained to represent one object, it will not perform well if two
objects are presented in the same image unless feature se-
lection is employed in the early stages of the network. We
term this the representational capacity limit. In this work, we
aim to understand how the feedback-driven performance gain
due to the cue depends on the capacity limits of the visual
processing stream.
Cue-driven feedback can affect
the object processing
stream in a feature-specific and/or a location-specific manner.
These interactions also account for feature-based and spa-
tial attention (Carrasco, 2011). In this work, we focus on the
feature-based feedback interaction.
We developed an artificial neural network (ANN) to model
the object processing stream, probing the stream output for an
object's presence in the image, and the feedback-based inter-
action between an external cue and the stream. The param-
eters of the ANN let us manipulate the neural and represen-
tational capacities of the object processing stream. We train
the ANN to maximise the difference (termed as the recogni-
tion performance) between correct (true positives) and incor-
rect (false positives) identification of the objects in the image.
The external cue and the probe contain category-level infor-
mation. For example, the external cue could correspond to
'Look for a Shoe' and the probe could correspond to 'Was
there a Shoe?'. Then the category-level information would be
'Shoe'. We show that the external cue substantially boosts
recognition performance when the object processing stream
cannot represent (low representational capacity) the informa-
tion required for categorising the target object (given by the
probe). We then comment on the nature of the feedback that
maximises these performance boosts.
Methods
Stimuli
We use the dataset Fashion-MNIST (Xiao, Rasul, & Vollgraf,
2017), which contains 28×28 images of 70,000 fashion prod-
ucts from 10 categories1.
We want to assess the effects of cue-driven feature-based
feedback on two object-feature manipulations. One, reducing
the feature information through blurring. Two, introducing fea-
ture competition by adding more objects to the image. To do
so, we construct 2 × 2 grid (40px × 40px) images, in which
we can place 1 to 4 objects (category overlaps are allowed)
and blur them with a Gaussian kernel with standard devia-
tions varying uniformly from 0 to 4 pixels. Example images
are shown in Figure 1.
Network Architecture
The artificial neural network (ANN) accepts three inputs - the
image, the cue, and the probe, and outputs whether the probe
1We split the dataset into 55k, 5k, and 10k images as train, valida-
tion, and test sets (equally split over the 10 categories). For testing,
we generate 10k grid images with the test image set.
Figure 1: Examples of stimuli used (with the number of ob-
jects, N, and the Gaussian blur standard deviations, SD).
category is present in the image, as shown in Figure 2. We
will now describe the sub-networks corresponding to the indi-
vidual inputs, and their interactions.
Nature of the object processing stream The object pro-
cessing stream (OPS, in green in Figure 2) is a fully-connected
ANN with one hidden layer. The input layer consists of 1600
units, representing the 40px x 40px images. The output layer
consists of 11 units, 10 of which give the probabilities of the
categories present in the image. The 11th unit is the out-
of-sample detector which gives the probability that the input
image does not belong to the space of images from the train-
ing set. This is done to prevent the OPS from make high-
confidence errors on out-of-sample images.
The hidden layer contains either 8, 32, or 3072 rectified lin-
ear units (ReLU). An increase in the number of hidden layer
units corresponds to an increase in neural capacity.
Nature of the probe The probe is a one-hot encoding2 (10
units) of the category of interest.
It is fed into another ANN
with the output of the OPS (the category probabilities). This
feedforward query ANN (fully-connected, in orange in Fig-
ure 2) has 200 rectified linear units (ReLU) in its hidden layer.
It has 2 output units (Yes/No) which give the probability of the
presence of the probe category in the image.
Nature of the cue The cue consists of 11 units, 10 of which
correspond to the 'informative' cue as they correspond to the
object categories. The 11th unit corresponds to the 'unin-
formative' cue ("Ready?" as seen in Figure 2). The infor-
mative cue is the same as the probe. This cue, after being
transformed into feedback templates, interacts with the OPS
through bias and gain modulation, as explained next. This cue
network is shown in blue in Figure 2.
Cue -- OPS interaction The responses O of the units in the
hidden layer of the OPS are given by Oh = [gh(IWh + bh)]+,
where I are the inputs, Wh are the input weights, and bh and
gh are the biases and gains of the units.
[x]+ = x if x > 0,
& [x]+ = 0 if x ≤ 0. The feedback templates bc and gc are
linear transformations of the cue, given by bc = zcWb and
gc = zcWg, where zc is the one-hot encoding of the cue cat-
egory. These templates are added to bh and gh respectively,
causing either an additive or multiplicative boost in the units'
responses. This interaction between the cue and the OPS
2Given 5 categories, and 1,2,5 being the categories of interest,
an n-hot encoding (n= 3 here) would be [1,1,0,0,1]
Figure 2: The artificial neural network (ANN) designed
to gauge the dependence of feedback-driven performance
boosts due to the cue on the capacity limits of the object pro-
cessing stream (OPS). The ANN outputs whether the probe
category is present in the input image or not. The cue inter-
acts with the OPS through bias (b) and/or gain (g) modulation
of the hidden units of the OPS.
was adapted from Lindsay & Miller (2017).
Network Training
We now describe the input-output maps used in training the
ANN shown in Figure 2.
In each case, we learn the maps
by minimising the cross-entropy between the network out-
put and target probability distributions. We do so by using
stochastic gradient descent (SGD) with Dropout regularisa-
tion (Srivastava, Hinton, Krizhevsky, Sutskever, & Salakhutdi-
nov, 2014). The training is done in three steps.
Training the OPS First, we train the parameters of the net-
work marked in green in Figure 2. The inputs are the images
mentioned in the Stimuli section. For each image, the tar-
get distribution, at the end of the OPS, is an n-hot encoding
normalised to a unit vector, representing the n unique object
categories in the image. The 11th unit of the OPS output is
associated with random images3. To manipulate the repre-
sentational capacity of the OPS, we train the OPS either with
images containing a single object which is not blurred or with
the full range of feature manipulations as shown in the Stimuli
section. The representational capacity for the full range of fea-
ture manipulations is lower in the former case, which we refer
to as low representational capacity here.
Training the probe Second, we train the parameters of the
network marked in orange in Figure 2. The OPS parameters
are frozen. The inputs are the images with the full range of
feature manipulations. The images are paired with correct or
3Uniformly random intensities are generated for all pixels. Then
the image is scaled, blurred, and occluded to cover subspaces of
interest better.
Figure 3: Cue-driven recognition performance boosts as a
function of the neural capacity (NC) and representational ca-
pacity (RC). The values of the boosts (∆avg) given by the infor-
mative cue beyond the uninformative cue, are mentioned for
NC/RC pair. As the neural capacity is reduced, the informative
cue provides lower performance boosts. Given a fixed neu-
ral capacity, the informative cue provides higher performance
boosts when the representational capacity is reduced.
incorrect probe categories equally. The output of the ANN is
a one-hot encoding of the correctness of the probe.
Training the cue Third, we train the parameters of the net-
work marked in blue in Figure 2. All other ANN parameters
are frozen.
In the case of informative cues, where the cue
category is the same as the probe category, the maps used
in training the probe are paired with the respective cues.
In
the case of the uninformative cue, all the maps used in train-
ing the probe are paired with the cue. We train bias and gain
modulation together, allowing for interactions between them.
Evaluation metric
Recognition performance is defined as the difference between
the proportion of correct and incorrect assessment that the
probe category exists in the input image. We assess the ef-
fects of imposing the two capacity limits on the recognition per-
formance (True positives (TP) - False positives (FP)) boosts
provided by the cues. If category information in the informa-
tive cue adds any functional (in terms of object categorisation)
value, it should boost performance beyond the performance
given by the uninformative cue (this boost is denoted by ∆).
Results and Discussion
We evaluate the recognition performance on the full range of
feature manipulations (number of objects and the strength of
blurring), and in the case of joint training of the gain and bias
modulation (in the cue ANN). The accuracies for recognis-
ing single objects in the image when the object processing
stream (OPS) is trained on single objects in the image (num-
ber of hidden units mentioned in brackets) are 82.4% (3072),
75.7% (32), and 54.8% (8). So, the representational capacity
Figure 4: The effect of cuing on the representation of the rele-
vant object in the OPS hidden layer (3072 units and low RC).
The standard deviation of blurring is 4px. The category-level
RSMs for the activities in the hidden layer for the mentioned
cases are shown. The categoricality indices for each RSM are
shown. The informative cue makes the representations in the
hidden layer selectively more distinct for the relevant object.
for single objects reduces with a reduction in neural capacity.
The recognition performance of the probe-only, the unin-
formative cue, and the informative cue cases are mentioned
in Figure 3. As seen in the figure, the informative cue pro-
vides higher recognition performance than the uninformative
cue and the probe-only case when the representational ca-
pacity is reduced. The uninformative cue boosts performance
over the probe-only case when the representational capacity
is low. This performance boost could be a result of boosting
the overall activity of the hidden units (through bias/gain) that
provide reliable differences in activity for the object categories,
in the case of the images with feature manipulations.
Trends observed in Figure 3 are preserved if we vary
only the number of objects (given 3072 OPS hidden units,
∆avg,RC↑ = 1.5%, ∆avg,RC↓ = 11.7%; if nob j = 4 , ∆RC↑ = 2.3%,
∆RC↓ = 15.5%) or the strength of blurring in the test images
(given 3072 OPS hidden units, ∆avg,RC↑ = 1.2%, ∆avg,RC↓ =
5.4%; if blur SD = 4px, ∆RC↑ = 2.3%, ∆RC↓ = 21.4%).
So, cue-driven feature-based feedback seems to be use-
ful for recognising objects subject to feature blurring and/or
competition, only when the features required to classify those
objects cannot be fully represented in the object processing
stream. Intuitively, if the OPS can represent all the category-
specific information about an object given the implicit repre-
sentational limits imposed by the neural capacity, such feed-
back should not be able to add to the recognition performance.
Influence of the trained feedback
How does the external cue influence the representation of the
relevant object in the hidden layer?
To assess this influence, we generate category-level rep-
resentational similarity matrices (RSM, based on Kendall's τ
correlation) for the case of a single object in the grid, which
is either presented intact, with blurring (with SD = 4px), with
blurring and the uninformative cue, or with blurring and the in-
formative cue. We define the categoricality index of the RSM
as the difference between the mean values of the diagonal
and off-diagonal elements. As seen in Figure 4, the informa-
tive cue makes the representation of the relevant object more
distinct, making it accessible to the output of the OPS.
In Abdelhack & Kamitani (2018), it was shown that neural
representations of blurred objects in the human visual cortex
are more similar to the corresponding intact object represen-
tations in a feedforward neural network than the blurred object
representations in that neural network. They attributed this ef-
fect to top-down information interacting with the stimulus rep-
resentations. This effect became stronger when a category
cue was introduced. However, in our case, the neural repre-
sentations of blurred objects (with either the informative or the
uninformative cue in effect) are equally similar to the corre-
sponding intact object representations (with no cue) and the
blurred object representations (with no cue) (∆τ < 0.04). This
inconsistency will be probed in further work by using more
complex and more biologically-plausible networks (such as
convolutional neural networks and recurrent neural networks)
for the object processing stream, which would make our net-
work a better model of the human visual system.
Comparison with tuning-based feedback
A popular model to describe the effects of the cue on neu-
ronal responses in the brain is the feature similarity gain model
(FSGM) (Martinez-Trujillo & Treue, 2004).
It claims that the
neuronal response is multiplicatively scaled according to its
preference to the properties of the attended (or task-relevant)
stimuli. Such cue-driven 'tuning-based' feedback was shown
to boost object recognition performance (with multiple objects
in a grid or overlaid) in Lindsay & Miller (2017). We deployed
tuning-based bias and gain modulation according to the math-
ematical framework outlined in Lindsay & Miller (2017)4. We
ran a grid search to compute the parameters to maximise
the recognition performance boosts provided by tuning-based
feedback over the probe-only case when representational ca-
pacity is low. Across the three neural capacities, the maximum
performance boost observed was 3%. This is small compared
to the boosts observed with the feedback trained with SGD as
seen in Figure 3. This implies the trained feedback is not the
same as tuning-based feedback.
Lindsay & Miller did observe a higher performance using
gradient-based feedback (of which feedback trained with SGD
is the natural extension) than with tuning-based feedback. As
also mentioned in their paper, this is not surprising as neu-
ronal tuning is not necessarily a measure of neuronal func-
tion. It has been shown that category-selective responses of
hidden units in ANNs do not imply that those units are rel-
atively more important to the recognition of objects of those
categories (Morcos, Barrett, Rabinowitz, & Botvinick, 2018).
In fact, the greater the category-selectivity of the hidden units,
the harder it is for the ANN to generalise to new data.
4To implement the tuning-based modulations, the following steps
are taken in Lindsay & Miller (2017). Compute category-specific (av-
eraged across multiple images) hidden layer activations. Mean- and
variance-normalise the category values for each hidden unit to gen-
erate the feedback templates. Tune (multiplicative scaling only) these
templates for bias (additive) or gain (multiplicative) modulation.
Conclusions
In this work, we investigated the nature and usefulness of cue-
driven feature-based feedback in recognising objects suffering
from feature blurring and/or competition. We built an artificial
neural network and asked how feature-based feedback can
be deployed, and how its recognition performance boosts are
dependent on neural and representational capacities of the
object processing stream. We found that the feedback boosts
performance only if the category-specific features about the
objects cannot be fully represented in the base ANN. These
representational limits are not dependent on the neural capac-
ity but on the task demands on the object processing stream.
The trained feedback does not resemble (but performs bet-
ter than) tuning-based feedback which is based on the feature
similarity gain model (Martinez-Trujillo & Treue, 2004).
To gauge the robustness of our observations, in subsequent
work we will run these analyses on different datasets and ar-
chitectures (such as convolutional neural networks). We shall
also assess these effects for location-based feedback.
Acknowledgements
We thank Giacomo Aldegheri, Surya Gayet, and Nadine Dijk-
stra for their comments and suggestions.
This work was funded by the European Research Coun-
cil (ERC) under the European Union's Horizon 2020 research
and innovation program (Grant Agreement No. 725970). This
manuscript reflects only the authors' view, and the Agency is
not responsible for any use that may be made of the informa-
tion it contains.
References
Abdelhack, M., & Kamitani, Y.
(2018). Sharpening of hier-
archical visual feature representations of blurred images.
eNeuro, 5(3), ENEURO -- 0443.
Bar, M.
(2004). Visual objects in context. Nature Reviews
Neuroscience, 5(8), 617.
Carrasco, M.
(2011). Visual attention: The past 25 years.
Vision research, 51(13), 1484 -- 1525.
Lindsay, G. W., & Miller, K. D.
(2017). Understanding bio-
logical visual attention using convolutional neural networks.
bioRxiv, 233338.
Martinez-Trujillo, J. C., & Treue, S.
(2004). Feature-based
attention increases the selectivity of population responses
in primate visual cortex. Current Biology, 14(9), 744 -- 751.
Morcos, A. S., Barrett, D. G., Rabinowitz, N. C., & Botvinick,
M. (2018). On the importance of single directions for gen-
eralization. arXiv preprint arXiv:1803.06959.
Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &
Salakhutdinov, R. (2014). Dropout: A simple way to prevent
neural networks from overfitting. The Journal of Machine
Learning Research, 15(1), 1929 -- 1958.
Xiao, H., Rasul, K., & Vollgraf, R. (2017). Fashion-mnist: a
novel image dataset for benchmarking machine learning al-
gorithms. arXiv preprint arXiv:1708.07747.
|
1809.06199 | 1 | 1809 | 2018-08-29T06:42:41 | Noninvasive Blockade of Action Potential by Electromagnetic Induction | [
"q-bio.NC",
"eess.SP",
"q-bio.TO"
] | Conventional anesthesia methods such as injective anesthetic agents may cause various side effects such as injuries, allergies, and infections. We aim to investigate a noninvasive scheme of an electromagnetic radiator system to block action potential (AP) in neuron fibers. We achieved a high-gradient and unipolar tangential electric field by designing circular geometric coils on an electric rectifier filter layer. An asymmetric sawtooth pulse shape supplied the coils in order to create an effective blockage. The entire setup was placed 5 cm above 50 motor and sensory neurons of the spinal cord. A validated time-domain full-wave analysis code Based on cable model of the neurons and the electric and magnetic potentials is used to simulate and investigate the proposed scheme. We observed action potential blockage on both motor and sensory neurons. In addition, the introduced approach shows promising potential for AP manipulation in the spinal cord. | q-bio.NC | q-bio | Noninvasive Blockade of Action Potential by Electromagnetic Induction
Soheil Hashemi 1, Amirhossein Hajiaghajani 2, and Ali Abdolali 1*
1Applied Electromagnetics Laboratory, School of Electrical Engineering, Iran University of Science and Technology, Tehran, 1684613114, Iran
2 Department of Electrical Engineering and Computer Science, University of California, Irvine, 92697, USA
Conventional anesthesia methods such as injective anesthetic agents may cause various side effects such as injuries,
allergies, and infections. We aim to investigate a noninvasive scheme of an electromagnetic radiator system to block
action potential (AP) in neuron fibers. We achieved a high-gradient and unipolar tangential electric field by designing
circular geometric coils on an electric rectifier filter layer. An asymmetric sawtooth pulse shape supplied the coils in
order to create an effective blockage. The entire setup was placed 5 cm above 50 motor and sensory neurons of the
spinal cord. A validated time-domain full-wave analysis code Based on cable model of the neurons and the electric
and magnetic potentials is used to simulate and investigate the proposed scheme. We observed action potential
blockage on both motor and sensory neurons. In addition, the introduced approach shows promising potential for AP
manipulation in the spinal cord.
1. Introduction
Patients who suffer from obesity, diabetes, and potent
anticoagulants run the risk of injuring their neurons by
traditional methods of general or regional anesthesia [1].
Human sensing receptors communicate with the central
neural system by propagating action potentials (AP) [2].
AP blockage of a sensory neuron by static magnetic fields
is reported [3] -- [5]. However, such a blockade required a
long time to emerge and occurred in less than 70% of all
AP propagations in the experiments [6].
The external stimulation of neurons by electric (E)
fields has been studied in the past decades [7], [8], using
the Hodgkin-Huxley cable equations to model neuron
stimulation based on the equations derived by the cable
theory. In these studies, the E fields that are tangential to
the neuron's axis have been considered as the cause of
neuron stimulation [9], [10]. Besides, optimization in
stimulus pulse shape was conducted, leading researchers
into designing an optimum stimulator [11] -- [13]. While,
we use the theory for investigating the blocking the AP in
this paper.
We introduced a new setup that blocks the propagating
AP in the neuron fiber and provides anesthesia in the
desired parts of the body by inducing a low-frequency
tangential E on the spinal cord. This method can instantly
affect the nervous system and offers interesting potentials
to avoid all the disadvantages of conventional methods
including injuries, infections, and other side effects.
2. Theory
In order to induce anesthesia in the spinal cord, it
is necessary that all neurons block the AP
propagation. Incident electromagnetic fields (EM)
which block AP along the neuron are likely to
reduce the electric potential of the neuron's
membrane in one point [8].
To investigate the idea and design the required
setup, we have used cable theory. From the modified
cable equations mentioned in [8], we obtain
𝜆2 𝜕2𝑉𝑚
𝜕𝑥2 − 𝜆2 𝜕𝐸𝑥
𝜕𝑥
− 𝜏
𝜕𝑉𝑚
𝜕𝑡
− 𝑉𝑚 = 0
(1)
where Vm, λ and τ are respectively the membrane
potential, space, and time constants, which are defined in
[8]. We appoint the x direction as the neuron axis and
refer to the tangential electric fields (E) throughout the
present letter by the notation Ex. According to (1), Ex must
follow
𝜕𝐸𝑥
𝜕𝑥
< 0 can reduce the membrane potential. This
reduction can cancel propagated AP in the very point if
the potential reduction is enough. Creating an EM field
* Corresponding author: Tel: +98 21 7322 5726, Fax: +98 21 7322 5777; E-mail address: [email protected]
with
𝜕𝐸𝑥
𝜕𝑥
< 0 in whole space is not possible. Thus, we
proposed a spatial Ex formation as shown in Fig. 1a.
coordinates. We have used (2) to design the coils and
calculate electric fields.
The coils were placed on the lumbar spine where the
skin-muscle tissue is considered as an effective half-space
confined by 𝑧 < 0. Frequency should not exceed 10 kHz
to avoid thermal side effects in the tissue according to
exposure restrictions [15] of E. In 10 kHz, the muscles
surrounding a neuron have a relative permittivity, relative
permeability, and conductivity of 25909, 1, and 0.34083
S/m, respectively [16], [17].
3. Method
Fig. 1. Theoretical schematic forms of induced tangential E for
the AP blockade. (a): Required spatial form according to the
cable theory-based equivalent for neuron fiber. (b): Required
coil current pulse with shortened degenerative interval and
constant slope that gives constant E at its positive cycle. (c,e):
Without filtering the degenerative cycle. The steep-decreasing
E along the x-axis can block the AP. Nevertheless, the steep-
increasing E degenerates the blockade. (d,f): After filtering the
degenerative cycle. This is to create a biphasic stimulus pulse
with one long weak phase and one brief strong phase. At a
moment that diodes turn off, magnetic potential follows the coil
currents and rise abruptly. This causes a spike in filtered E. By
this filtering procedure, the degenerative interval changes to a
spike, whose effect on nerve will be evaluated.
According to (1), the steep descending Ex region (50-52
cm) causes the membrane potential to decrease, such that
the AP barrier will be created at the point. On the other
hand, an Ex that ascends weakly will not affect the AP
significantly. The low-gradient region (0-50 cm) is
considered spatially wide in order not to noticeably
change the membrane potential. As a result, no AP will be
produced or propagated toward the central neural system.
To create this spatial shape of Electric field (E), we used
multi-turn coils to radiate an incident EM wave for the
production of E along the neuron. Electrical fields is
obtained from the Maxwell's equations in general case
[14]:
∇ × ∇ × 𝐸⃗ (𝑟 ) + 𝜇𝜀(𝑟 )
∂2𝐸⃗ (𝑟 )
∂𝑡2 + 𝜇𝜎(𝑟 )
∂𝐸⃗ (𝑟 )
∂𝑡
= −𝜇
(2)
⃗⃗⃗⃗⃗⃗⃗⃗⃗⃗
∂𝐽(𝑟′⃗⃗⃗⃗ )
∂𝑡
where ε, μ, σ and J respectively represent permittivity,
permeability, conductivity and the coil's current density.
Also, vectors r and r' denote observation and source
We aimed to hyperpolarize the neuron at the blockage
point incessantly using an external radiation system.
Therefore, the realization of the proposed system revolves
around three main phases; first, to appoint Ex as a function
of space; second, to filter the degenerative interval in
pulse shape; third, to validate the entire setup by the
modified full wave analysis achieved from cable model of
neurons [18], while exerting the proposed Ex.
3.1. Developing the spatial form of E
First, the spatial form of Ex must follow the required
form for blockade (Fig. 1a). We employed circular coils
including 30 turns, placed in the xy plane with a radius of
R. A typical coil with the center of (0,R,z0) exhibits
sinusoidal current distribution characteristic in which the
x component is very similar to "𝑐𝑜𝑠 (
𝜋
2𝑅
𝑥) " for 𝑥 ∈
(−𝑅, +𝑅). We observed
that, with a very good
approximation, the spatial form of Ex (that is itself
analogous to the x component of the circular coils' current
𝑥) " are quite similar on
[8]) and the sine lobes of "𝑠𝑖𝑛 (
𝜋
2𝑅
the neuron line (y=z+5cm=0). The required Ex which can
be denoted by spatial Fourier sine series can be created by
a series of circular coils with determined parameters such
as radius (relevant to the spatial frequency), current
(relevant to the series' coefficient), and coils' center
coordinates. Subsequently, a total form of Ex which
involves several sine lobes can be created by a setup
composed of individual coils that work together (see Fig.
2a). Thus, to form the low-gradient region of the desired
Ex (space between 0-50 cm in Fig. 1a), we used three coils
(labels 1-3 in Fig. 3).
explicitly mentioned that in a controlled environment it is
acceptable to exceed the limit.
Fig. 3. The entire setup that makes noninvasive anesthesia: 1.
Coil No. one; 2. Coil No. two; 3. Coil No. three; 4. Gold rods
connected to the external biasing; 5. GaAs layers, equivalent to
diodes that work in the x direction; 6. Air; 7. Metallic plate; 8.
Effective skin-muscle layer; 9. Neuron axon located 5 cm deep
inside the tissue. Coils run currents in clockwise direction.
Fig. 2. Results validated by the finite element method. (a)
Comparison between analytic sine lobes and simulation of
tangential E with PEC plate. (b) Asymmetric sawtooth current
pulse is applied to the coils in which the rise time is shortened.
The externally controlled diodes filter the E induced only in this
interval. Although E remains constant at positive cycles, a spike
is engendered by abrupt increase in magnetic potential.
We found a very good agreement between sine
functions and simulated Ex depicted in Fig. 2a. Still, to
realize the steep region of Ex (space between 50-52 cm in
Fig. 1a) without using any additional coil, we employed a
thin metallic (perfect electric conductor-PEC like steel)
plate under the coils and over the skin. Coils induce E on
the plate (label 7 in Fig. 3) instead of neuron, leading to
an eddy current distribution.
The PEC dissipates the magnetic field's energy; hence,
E drops dramatically where the plate is placed. Also, the
high current density on the plate's edge (at x=50 cm)
causes Ex to reach its maximum beneath the plate's edge
and creates a region with steep Ex. Due to IEEE Std
C95.6-2002, long term exposure limit of electric fields in
very low frequency (VLF) and ultra low frequency (ULF)
are 1842 V/m and 10000 V/m consequently though it
3.2. Filtering degenerative intervals in pulse shape
In addition to (2), from the Biot-Savart law we know:
𝐸𝑥(𝑡) = −
𝑑𝐴𝑥(𝐽(𝑡))
𝑑𝑡
∝ −
𝑑𝐼𝑐𝑜𝑖𝑙(𝑡)
𝑑𝑡
(3)
and
coil
where A, J and Icoil represent magnetic potential, current
distribution,
respectively.
Inevitably, a periodic pulse for the coils' current includes
both rise and fall intervals which, regarding (3), creates
both negative and positive Ex, respectively. Therefore, it
ampere-turn,
affects the polarity of
𝜕𝐸𝑥
𝜕𝑥
. Thus, in order to shorten the
degenerative interval, we used a current pulse shape
consisting of a long fall time relative to the rise time (see
Fig. 1a & b).
It was shown that the ascending interval in the current
𝜕𝐸𝑥
𝜕𝑥
pulse (Fig. 1b and, consequently, the positive steep
shown in Fig. 1e) leads to stimulating additional AP since
it induces an Ex with negative polarity; hence, E must
vanish in this period. We designed a thin filter to be
placed beside the metallic plate, involving a repeated
sequence of externally biased gold-gallium-arsenide
(GaAs) PN junction layers (labels 4 and 5 in Fig. 3).
Physically, this means placing many thin diodes whose
on/off state is controlled by external biasing. The external
biasing of gold layers enables our filter to turn the diodes
fiber inside the spinal cord. We developed the validated
full wave analysis including the effect of external
stimulus electric fields [8] in Matlab interface using the
finite difference method (voltage/time dependencies of
ion channels of the membrane are available in [18]). The
fiber was exposed to the Ex 0.5 ms after simulation (see
Fig. 4). The exposure reduced the membrane's potential
at the steep point to create blockade. However, in the
weak side, the membrane's potential did not change due
to the low gradient of induced Ex.
off in the current's fall time interval (which was creating
positive Ex polarity). To specify the method of controlling
the induced Ex's polarity, it should be noted that all coils
were supplied by synced in-phase current waveform.
When diodes are switched off, they act like a dielectric
with zero electric conductivity and an effective relative
permittivity of 12.9. Thereby, the filter is transparent to
the incident E and is allowed to be induced on the neuron
behind the filter (see Fig. 2b). In the current rise time,
diodes are switched on and act like a good electric
conductor with the conductivity of 1 MS/m. This allows
E to be induced on the filter instead of the neurons and
produce eddy currents, dissipating the magnetic field's
energy. This method allows for filtering Ex as shown in
Fig. 1f. It should be noted that the Biot-Savart law
mentioned in (3) only can be used in linear media. To
obtain E in such a non-linear medium, (2) must be
employed.
the
study,
Although the manufacturing considerations of this
diode filter have to be well considered and merit
additional
simulations by COMSOL
Multiphysics shows there is an enormous potential for
using doped GaAs semiconductors
to achieve a
conductivity-controllable surface. Other semiconductors
with better mobility may be used in the future. It should
be mentioned that the thickness of the proposed gold-PN
junction is determined such that the spatial form of Ex
does not change significantly and the incident wave
would not bias the diodes. Table 1 presents the setup
parameters.
Table 1. Specification of the final setup. All coils consist of
30 turns.
Coil
No.
Radius
[cm]
Current
(CW)
[kA]
40
15
15
1
2
3
Iron dimensions
80.8
69.3
69.3
x
[cm]
y
[cm]
z
[cm]
0
-15
15
(20,60)
30
-15
15
(-60,60)
1
2.4
2.4
(0,0.2)
Diode layer boundaries
(-60,20)
(-60,60)
(0,0.2)
3.3. Validation by the Full Wave Analysis
In the third phase, to investigate our demonstration, we
simulated the neuron behavior under the designed setup.
As an example of several neurons that were being
simulated, a motor neuron was appointed as the neuron
Fig. 4. AP propagation along the motor neuron axis inside the
spinal cord. At the beginning, neuron reaches its steady state.
Simulation of the blockade setup starts at 0.5 ms. Voltage clamp
excites the marked point at 1 ms. Then, the AP propagates
toward both ends of the cell, and its blockage can be seen at 6.5
ms when it meets the steep point. The second AP propagates
toward the weak region without perceived change. Some of
monitored moments in which the electric field's spikes change
the membrane potential (at 1.7, 4.1 and 10.4 ms) are starred. It
is observed that the negative spikes in Ex result in membrane
potential to increase by 5 mV. This voltage fluctuation does not
affect the blockage.
The fiber under exposure was excited by a voltage
clamp [19] 1 ms after starting the simulation. Afterwards,
the AP propagated toward the blockage point (steep
region of Ex) and the other side of the neuron (weak
region). No interference was observed in the membrane
voltage while the AP propagated toward the weak region
of the induced field. In this case, the neuron was
hyperpolarized and went back to rest immediately.
However, we directly enforced an artificial incessant
hyperpolarization [20] by radiating external E fields such
that the AP that was reaching the blockage point could not
stimulate the blockage area and the propagation stopped
in this region. Because at the point, hyperpolarization
avoided neuron stimulation, no AP could continue
propagating and consequently, anesthesia was induced. It
was observed that neuron remained hyperpolarized
despite the short intervals that the diode filter eliminated
E (Fig. 1f). As we validated the AP blockage on motor
and sensory neurons, we claim that the blockade region
can be developed by the proposed setup on every fiber in
the spinal cord, causing the AP to stop moving from signal
receptors toward the central neural system or vice versa.
4. Conclusion
We reported a new form of electric field as a function
of space and time, to cancel the AP that was propagating
toward the central neural system and induce anesthesia
with the use of full wave analysis of neurons. Afterwards,
we described a new radiative setup in order to realize the
proposed Ex on the neuron fibers located 5 cm deep inside
the tissue close by the spinal cord. The setup consisted of
three multiturn coils which were controlled by the signal
generator, a metallic plate that created the steep Ex, and a
conductivity-controlled surface to convert the bipolar E to
unipolar on neurons. Finally, the validated full wave
analysis based on the cable model of neurons was used to
successfully investigate and validate the operation of the
entire system. Unlike conventional anesthesia methods,
this technique eliminates all injuries, contagion, and side
effects; it also takes effect immediately after exposure.
Therefore, we hope the future researches would illustrate
remarkable potentialities of this non-invasive technique
by experimenting the proposed setup.
References
1995.
[4] B. J. Roth et al., "In vitro evaluation of a 4-leaf coil
design for magnetic stimulation of peripheral nerve.,"
Electroencephalogr. Clin. Neurophysiol., vol. 93, no. 1,
pp. 68 -- 74, 1994.
[5] S. Hashemi and A. Abdolali, "Three-dimensional
analysis, modeling, and simulation of the effect of static
magnetic fields on neurons," Bioelectromagnetics, vol.
38, no. 2, pp. 128 -- 136, 2017.
[6] A. D. Rosen, "Mechanism of Action of Moderate-
Intensity Static Magnetic Fields on Biological Systems,"
Cell Biochem. Biophys., vol. 39, no. 2, pp. 163 -- 174,
2003.
[7] E. Dayan, N. Censor, E. R. Buch, M. Sandrini, and L. G.
Cohen, "Noninvasive brain stimulation: from physiology
to network dynamics and back.," Nat. Neurosci., vol. 16,
no. May 2016, pp. 838 -- 44, 2013.
[8] B. J. Roth and P. J. Basser, "A Model of the Stimulation
of a Nerve Fiber by Electromagnetic Induction," IEEE
Trans. Biomed. Eng., vol. 37, no. 6, pp. 588 -- 597, 1990.
[9] S. Silva, P. J. Basser, and P. C. Miranda, "Elucidating
the mechanisms and loci of neuronal excitation by
transcranial magnetic stimulation using a finite element
model of a cortical sulcus," Clin. Neurophysiol., vol.
119, no. 10, pp. 2405 -- 2413, 2008.
[10] J. Platkiewicz and R. Brette, "A threshold equation for
action potential initiation," PLoS Comput. Biol., vol. 6,
no. 7, p. 25, 2010.
[11] S. M. Goetz, C. N. Truong, M. G. Gerhofer, A. V.
Peterchev, H. G. Herzog, and T. Weyh, "Analysis and
Optimization of Pulse Dynamics for Magnetic
Stimulation," PLoS One, vol. 8, no. 3, 2013.
[12] K. D'Ostilio et al., "Effect of coil orientation on
strength -- duration time constant and I-wave activation
with controllable pulse parameter transcranial magnetic
stimulation," Clin. Neurophysiol., vol. 127, no. 1, pp. 1 --
9, 2015.
[13] S. M. Goetz et al., "Enhancement of Neuromodulation
with Novel Pulse Shapes Generated by Controllable
Pulse Parameter Transcranial Magnetic Stimulation,"
Brain Stimul., 2015.
[1] R. Brull, C. J. L. McCartney, V. W. S. Chan, and H. El-
[14] C. A. Balanis, Advanced Engineering electromagnetics.
Beheiry, "Neurological complications after regional
anesthesia: Contemporary estimates of risk," Anesth.
Analg., vol. 104, no. 4, pp. 965 -- 974, 2007.
[2] M. L. DuBois, Action Potential: Biophysical and
Cellular Context, Initiation, Phases and Propagation.
New York: Nova Science, 2010.
[3] A. V Cavopol, A. W. Wamil, R. R. Holcomb, and M. J.
McLean, "Measurement and analysis of static magnetic
fields that block action potentials in cultured neurons.,"
Bioelectromagnetics, vol. 16, no. 1 995, pp. 197 -- 206,
Hoboken, NJ: John Wiley & Sons, 2012.
[15] Institute of Electrical and Electronics Engineers, IEEE
Standard for Safety Levels with Respect to Human
Exposure to Electromagnetic Fields, 0 -- 3 kHz, no.
October. 2002.
[16] "Calculation of the Dielectric Properties of Body Tissues
in the frequency range 10 Hz - 100 GHz." [Online].
Available: http://niremf.ifac.cnr.it/tissprop/. [Accessed:
01-Feb-2017].
[17] C. Gabriel, C. Gabriel, S. Gabriel, S. Gabriel, E.
Corthout, and E. Corthout, "The dielectric properties of
biological tissues: I. Literature survey.," Phys. Med.
Biol., vol. 41, no. 11, pp. 2231 -- 49, 1996.
[19] David Ogden, D. Ogden, and Plymouth Workshop.,
Microelectrode techniques: the Plymouth Workshop
handbook, 2, illustr ed. Company of Biologists Limited,
1994.
[18] C. C. McIntyre and W. M. Grill, "Extracellular
[20] B. J. Roth and N. P. Charteris, "How hyperpolarization
stimulation of central neurons: influence of stimulus
waveform and frequency on neuronal output.," J.
Neurophysiol., vol. 88, no. 4, pp. 1592 -- 1604, 2002.
and the recovery of excitability affect propagation
through a virtual anode in the heart," Comput. Math.
Methods Med., vol. 2011, 2011.
|
1902.03309 | 1 | 1902 | 2019-02-08T22:21:24 | Linear Dynamics & Control of Brain Networks | [
"q-bio.NC",
"math.OC",
"physics.bio-ph"
] | The brain is an intricately structured organ responsible for the rich emergent dynamics that support the complex cognitive functions we enjoy as humans. With around $10^{11}$ neurons and $10^{15}$ synapses, understanding how the human brain works has proven to be a daunting endeavor, requiring concerted collaboration across traditional disciplinary boundaries. In some cases, that collaboration has occurred between experimentalists and technicians, who offer new physical tools to measure and manipulate neural function. In other contexts, that collaboration has occurred between experimentalists and theorists, who offer new conceptual tools to explain existing data and inform new directions for empirical research. In this chapter, we offer an example of the latter. Specifically, we focus on the simple but powerful framework of linear systems theory as a useful tool both for capturing biophysically relevant parameters of neural activity and connectivity, and for analytical and numerical study. We begin with a brief overview of state-space representations and linearization of neural models for non-linear dynamical systems. We then derive core concepts in the theory of linear systems such as the impulse and controlled responses to external stimuli, achieving desired state transitions, controllability, and minimum energy control. Afterwards, we discuss recent advances in the application of linear systems theory to structural and functional brain data across multiple spatial and temporal scales, along with methodological considerations and limitations. We close with a brief discussion of open frontiers and our vision for the future. | q-bio.NC | q-bio |
Linear Dynamics & Control of Brain Networks
Jason Z. Kim and Danielle S. Bassett
Abstract The brain is an intricately structured organ responsible for the rich emer-
gent dynamics that support the complex cognitive functions we enjoy as humans.
With around 1011 neurons and 1015 synapses, understanding how the human brain
works has proven to be a daunting endeavor, requiring concerted collaboration across
traditional disciplinary boundaries. In some cases, that collaboration has occurred
between experimentalists and technicians, who offer new physical tools to measure
and manipulate neural function. In other contexts, that collaboration has occurred
between experimentalists and theorists, who offer new conceptual tools to explain
existing data and inform new directions for empirical research. In this chapter, we
offer an example of the latter. Specifically, we focus on the simple but powerful
framework of linear systems theory as a useful tool both for capturing biophysically
relevant parameters of neural activity and connectivity, and for analytical and nu-
merical study. We begin with a brief overview of state-space representations and
linearization of neural models for non-linear dynamical systems. We then derive
core concepts in the theory of linear systems such as the impulse and controlled
responses to external stimuli, achieving desired state transitions, controllability, and
minimum energy control. Afterwards, we discuss recent advances in the application
of linear systems theory to structural and functional brain data across multiple spatial
and temporal scales, along with methodological considerations and limitations. We
close with a brief discussion of open frontiers and our vision for the future.
Jason Z. Kim
Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104, USA
e-mail: [email protected]
Danielle S. Bassett
Departments of Bioengineering, Electrical & Systems Engineering, Physics & Astronomy, Neurol-
ogy, & Psychiatry, University of Pennsylvania, Philadelphia, PA 19104, USA
e-mail: [email protected]
1
2
Jason Z. Kim and Danielle S. Bassett
1 Emergence in the Structure and Function of Complex Systems
In the observable world, some of the most beautiful and most puzzling phenomena
arise in physical and biological systems characterized by heterogeneous interactions
between constituent elements. For example in materials physics, heterogeneous inter-
actions between particles in granular matter (such as a sand pile) constrain whether
the matter acts as a liquid (flowing with gravity) or a solid (supporting load-bearing)
[1, 2]. In sociology, heterogeneous interactions between humans in a society are
thought to be responsible for surges in online activity, peaks in book sales, traf-
fic jams, and correlated spikes in demand for emergency services [3]. In biology,
heterogeneous interactions between computational units in the brain are thought to
support a divergence of the correlation length, an anomalous scaling of correlation
fluctuations, and the manifestation of mesoscale structure in patterns of functional
coupling between units, all features that allow for a diversity of dynamics underlying
a diversity of cognitive functions [4, 5]. The feature of these systems that often drives
our fascination is the capacity for heterogeneous interactions to produce suprising
dynamics, in the form of drastic state transitions, spikes of collective activity, and
multiple accessible dynamical regimes.
Because element-element interactions are heterogeneous in such systems, tra-
ditional approaches from statistical mechanics -- such as continuum models and
mean-field approximations -- fail to offer satisfying explanations for system function.
There exists a critical need to develop alternative approaches to understand how
interactions map to emergent behavior. The need is particularly salient in the context
of neural systems, where such an understanding could directly inform models of
neurological disease and psychiatric disorders [6, 7]. Moreover, gaining such an un-
derstanding is a prerequisite for the well-reasoned development of interventions [8],
whether in the form of brain stimulation [9, 10], pharmacological agents [11, 12],
or other therapies [13]. Technically, such interventions in systems characterized by
heterogeneous interactions can be parsimoniously considered as forms of network
control, thus motivating extensive recent interest in the utility of network control
theory for neural systems [8].
Despite the generic importance of understanding how interactions map to emer-
gent properties, and the specific importance of understanding that mapping in the
human brain, progress towards that understanding has remained surprisingly slow.
Some efforts have sought to develop detailed multiscale computational models [14].
Yet such efforts are faced with the ever-present quandary that, in point of fact, "The
best material model of a cat is another, or preferably the same, cat" [15]. Detailed
models are difficult to construct, intractable to analytic approaches, require exten-
sive time to simulate, contain parameters that are frequently underconstrained by
experimental data, and in the end produce dynamics that are themselves difficult
to understand or to explain from any specific choices in the model. In contrast, ap-
proaches from physics consider natural phenomena as if dynamics at macroscopic
length scales were almost independent of the underlying, shorter length scale de-
tails [16]. A hallmark of effective physical theories is a marked compression of the
full parameter space into a few governing variables that are sufficient to describe
Linear Dynamics & Control of Brain Networks
3
the observables of interest at the scale of interest. Interestingly, recent theoretical
work demonstrates that such simple models are the natural culmination of processes
maximizing the information learned from finite data [17].
Here we embrace simplicity by considering the utility of linear systems theory for
the understanding and control of neural systems comprised of computational units
coupled by heterogeneous interactions. We begin by placing our remarks within the
context of quantitative dynamical models of neurons and their interactions, as well
as the spatial and temporal considerations inherent in choosing such models. We
will then turn to a discussion of approximations to those dynamical models, the
incorporation of exogeneous control input, and model linearization. Our treatment
then naturally brings us to a discussion of the theory of linear systems, as well as
their response to perturbative impulses, and to explicit control strategies. We lay out
the formalism for probing state transitions, controllabilty, and the minumum control
energy needed for a given state transition. After completing our formal treatment,
we discuss the application of linear systems theory to neural systems, and efforts
to map network architecture to control properties. We close with a description of
several particularly pertinent methodological considerations and limitations, before
outlining emerging frontiers.
2 Quantitative Dynamical Models of Neural Systems &
Interactions
Historically, many neural behaviors and mechanisms have been successfully modeled
quantitatively. Here we briefly describe several illustrative examples of such models.
The classic fundamental biophysical model of a single neuron (Fig. 1, left) was
developed by Alan Hodgkin and Andrew Huxley in 1952 (see [18] for details). The
model is now known as the Hodgkin-Huxley model. It treats a segment of a neuron as
an electrical circuit, where the membrane (capacitor) and voltage-gated ion channels
(resistors) are parallel circuit elements. The time-evolution of membrane voltage,
Vm, between the inside and the outside of the neuron is given by
Cm
(cid:219)Vm(t) = ¯gK n4(t)(VK − Vm) + ¯gN am3(t)h(t)(VN a − Vm) + ¯gl(Vl − Vm) + I(t),
where Cm is the membrane capacitance, ¯gK, ¯gN a, ¯gl are maximum ion conductances
for potassium, sodium, and passive leaking ions, and I is an external stimulus current,
all per unit area. In addition, VK, VN a, Vl represent the reversal potential of these ions.
The variables n, m, h vary between 0 and 1, and model the ion channel gate kinetics
to determine the fraction of open sodium (m, h) and potassium (n) channels
(cid:219)n(t) = αn(Vm(t))(1 − n(t)) − βn(Vm(t))n(t)
(cid:219)m(t) = αm(Vm(t))(1 − m(t)) − βm(Vm(t))m(t)
(cid:219)h(t) = αh(Vm(t))(1 − h(t)) − βh(Vm(t))h(t),
4
Jason Z. Kim and Danielle S. Bassett
where the functions αi(Vm) and βi(Vm) are empirically determined. These segments
are then spatially connected together, such that the propagation of an action potential
across a neuron is modeled by a set of partial differential equations. Due to the
biophysical realism of variables and parameters, this model can make powerful and
accurate predictions of neuron activity in different environments and stimulation
regimes [19, 20, 21]. Simplified versions of this model, such as the FitzHugh-
Nagumo model [22], can also produce many of the same neuronal dynamics.
Fig. 1 Schematic of neural models and controlling perturbations at different scales. Here,
the Hodgkin-Huxley model describes the biophysical behavior of single neurons (left) that may be
excitatory (blue) or inhibitory (gray). The artificial neuron models describe the simplified weighted
connections and binary states of many neurons (center). The Wilson-Cowan model describes the
activity of large neural populations in a region (right) or in a cortical column by modeling the
excitatory and inhibitory connections of each population. In each case, a controlling perturbation
(yellow) can affect the neural system at different scales.
However, many complex behaviors of neural systems arise from interactions
between multiple neurons. With four variables (membrane voltage, gates) and even
more parameters to model the behavior of a single neuron, the space of models
to explore interacting neurons quickly becomes intractable to both analytical and
numerical interrogation. An alternative approach is to capture the simplest aspects
of neural interactions that are crucial for the phenomenon of interest. Such was the
approach taken by Warren McCulloch and Walter Pitts [23], who developed what
would later become a canonical model of an artificial neuron. In this model, each
neuron i at any point in time t exists in one of two states: firing xi(t) = 1 or not
firing xi(t) = 0. The state of the neuron is determined by a weighted sum of inputs
from connected neurons at the previous time step. Then, neuron i in a system of N
neurons evolves in time as
where wi j is the strength of excitation (wi j > 0) or inhibition (wi j < 0) from neuron
j to neuron i, and function fi is typically a thresholding function (Fig. 1, center).
Instantiations and extensions of this model are used to study associative memory
(Hopfield [24]), machine learning (perceptron [25]), and cellular automata [26].
In many cases, the sheer number of neurons and interactions renders even these
simple models difficult to study. A typical solution is to instead model the average
activity of a population of neurons. This is the approach taken by Hugh Wilson and
xi(t + 1) = fi(cid:169)(cid:173)(cid:171) N
j=1
wi j xj(t)(cid:170)(cid:174)(cid:172) ,
Linear Dynamics & Control of Brain Networks
5
Jack Cowan [28] in the Wilson-Cowan model. Here, a group of neurons is separated
into excitatory and inhibitory populations, where the fraction of cells firing at time
t in each population is E(t) and I(t), respectively, that evolve in time as
(cid:219)E(t) = −E(t) + (ke − reE(t))Se(c1E(t) − c2I(t) + P(t))
τe
(cid:219)I(t) = −I(t) + (ki − ri I(t))Si(c3E(t) − c4I(t) + Q(t)).
τi
Here, c1, c2 > 0 represent connection strength into the excitatory population, and
c3, c4 > 0 represent connection strength into the inhibitory population, re, ri are the
refractory periods, and Se, Si are sigmoid functions from the distribution of neuron
input thresholds for firing. Such models produce oscillations such as those observed
in non-invasive measurements of large-scale brain activity (Fig. 1, right) in patients
with epilepsy [29].
In these and many other models, a common theme is the tradeoff between realism
and tractability. We desire sufficient realism to study crucial features of neural
systems such as the activity of each unit, the interaction strength between units, the
connection topology, and the effect of external stimulation. We also desire sufficient
tractability (either to analytical or numerical interrogation) to make consistent and
meaningful predictions about our neural system by understanding relations between
the model parameters and the model behavior. In this chapter, we will discuss one
such model from the theory of linear dynamical systems.
2.1 Spatial and Temporal Considerations
When modeling neural systems, an immediately salient consideration is the vast
range of spatial and temporal scales at which nontrivial -- and thus quite interesting
-- dynamics occur. It stands to reason that the most relevant type of model for
understanding a given phenomenon depends on the spatiotemporal scale at which
that phenomenon is observed. For example, consider the fact that while it is generally
known that certain sensory regions such as the visual cortex are both anatomically
linked to and functionally responsible for sensory inputs, it is more difficult to
assign a set of neurons that are necessary for distributed cognitive processes such
as attention and cognitive control. Thus, biophysical models at the level of single
neurons may be viable for simulating receptive fields in visual processing, but may
be less useful for studies of task-switching or gating. Similarly, consider the fact that
a single neuron may fire every few milliseconds, while human reaction times are on
the order of hundreds of milliseconds, and brain-wide fluctuations in activity on the
order of seconds. Thus, the form of the model considered should match the temporal
scales of the behavior to be studied.
From a modeling perspective, balancing these considerations of spatial and tem-
poral scales with model realism impacts the category of model that has the greatest
utility. If one wishes to consider small spatial scales, then a rather simplistic neuron-
level model such as the McCulloch-Pitts may be particularly useful, where each
6
Jason Z. Kim and Danielle S. Bassett
neural unit has discrete states such that each neuron i is either firing xi(t) = 1 or not
xi(t) = 0. In contrast, if one wishes to consider larger spatial scales characteristic of
distributed cognitive processes, it may be more appropriate to consider models in
which each neural unit reflects the average population activity of a brain region as
a continuous state, where xi(t) is a real number. Similar considerations are relevant
and important in the time domain. For models that assume fairly uniform delays
in neuronal interactions such as the McCulloch-Pitts, a discrete time model where
time evolves in integer increments may be appropriate. In contrast, if the timing of
interactions between neural units such as myelinated versus unmyelinated axons is
heterogeneous, a continuous time model may be more suitable, where time t is a real
number.
In addition to affecting the definition of neural activity and the nature of its propa-
gation, these considerations also affect the meaning of interactions between units. In
a neuron-level model whose units reflect neurons, the unit-to-unit interactions may
represent structural synapses between neurons. In contrast, in a population model
whose units reflect average neural activity of a brain region, unit-to-unit interactions
may represent a summary measure of the collective strength or extent of structural
connections between regions. Both types of connections can be empirically mea-
sured using either invasive (staining, flourescence imaging, tract tracing [30]) and
non-invasive (tractography [31]) methods. The specific type of interaction studied
constrains the sorts of inferences that one can draw from the subsequent model, as
well as the types of model-generated hypotheses that one can test in new experiments.
2.2 Dynamical Model Approximations
Both here and in the following sections, we will consider systems with both con-
tinuous state and time. However, we note that the theory of linear systems extends
naturally to discrete time systems as well. We begin our formulation with a set of N
neural units, where each unit has an associated level of activity xi(t) that is a real
number at some time t ≥ 0 that is also a real number. Then the collection of activity
for all units into column vector x(t) = [x1(t); x2(t); · · · ; xN(t)] is called the state of
our system at time t. For example, in the Hodgkin-Huxley equations, our state vector
is x = [V; n; m; h]. In many models including Hodgkin-Huxley, the time evolution
of the system states can be written as a vector derivative
(cid:219)x1(t)
(cid:219)x2(t)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:125)
...(cid:219)xN(t)
(cid:219)x(t)
=
,
f1(x(t))
f2(x(t))
...
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
fN(x(t))
f (x(t))
Linear Dynamics & Control of Brain Networks
7
where f , the vector of functions fi, determines how the system states change, (cid:219)x, at
every particular state x. We can think of these equations as generating a vector field,
where at each point x, we draw an arrow with magnitude and direction equal to
f (x). As an example, consider the following two neuron system x1, x2 that evolves
in time as
(cid:219)x1(t) = 2x2(t) − sin(x1(t))
(cid:219)x2(t) = x2
1(t) − x2(t),
where the vector field and example trajectory from initial state x(0) = [−0.3;−0.4]
is shown (Fig. 2, top). Note how at every point x1, x2, the above equation determines
a vector of motion (cid:219)x that the system traces from the initial point. This quantitative
modeling of neural dynamics allows us to study and predict the response of our
neural system to changes in interaction strength or external stimulation.
Fig. 2 Vector fields and trajectories, with and without control inputs. Example simple vector
field of two states with a particular trajectory from initial condition x(0) = [−0.3;−0.4] (top left) in
state space, with the corresponding plot of each state over time (top right), and the corresponding
vector field and trajectory with control input u(t) = 0.5 (bottom left) with corresponding states
over time (bottom right).
-0.50.5-0.50.502-0.50.5-0.50.5-0.50.502-0.50.5 8
2.3 Incorporating Exogenous Control
Jason Z. Kim and Danielle S. Bassett
While modeling intrinsic system behavior is already a broad topic of current research,
there is an increasing need for the principled study of therapeutic interventions to
correct dysfunctional neural activity. These interventions may take the form of
targeted invasive (deep bran stimulation) or non-invasive (transcranial magnetic
stimulation) inputs, or more diffusive drug treatments. Hence, in our modeling efforts
we also often desire to incorporate the effect of some external stimuli u1(t), · · · , uk(t).
We collect these stimuli into a vector u(t) = [u1(t); u2(t); · · · ; uk(t)], and include
their effect on the rates of change of system states in our function
(cid:219)x1(t)
(cid:219)x2(t)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:125)
...(cid:219)xN(t)
(cid:219)x(t)
f1(x(t), u(t))
f2(x(t), u(t))
=
.
...
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
fN(x(t), u(t))
f (x(t),u(t))
As an example in our two unit system, we can apply an input to the first unit
(cid:219)x1(t) = 2x2(t) − sin(x1(t)) + u(t)
(cid:219)x2(t) = x2
1(t) − x2(t),
thereby changing our system of equations. We plot the vector field and trajectory of
our system under some constant input u(t) = 0.5 (Fig. 2, bottom). Notice how the
control input changes the trajectory and final state of our system by modifying the
vector field. Also notice that our input only shifts the x1 component of our vectors
because we only stimulate x1. These abilities to map neural interactions f to the full
trajectory of activity x(t), and to find control inputs u(t) that drive our neural system
to a desired final state x(T) are among the core contributions of linear systems theory.
2.4 Model Linearization
While we have a quantitative framework for the evolution of a controlled neural
system, there are no general principles for determining the full trajectory x(t) or
control input u(t) to reach a desired final state for a general nonlinear system. In
systems of only a few neural units, there exist several powerful numerical and analytic
tools. However, the study and control of large neural systems is made difficult by our
inability to know how a stimulus will affect our system without first simulating the
full trajectory. Further, for multiple stimuli, the number of possible stimulus patterns
grows exponentially.
A special class of simplified systems called linear systems circumvents this issue.
In our state representation, a linear system is described by
9
(1)
Linear Dynamics & Control of Brain Networks
(cid:219)x1(t)
(cid:219)x2(t)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:125)
...(cid:219)xN(t)
=
...
a11 a12 · · · a1N
a21 a22 · · · a2N
...
...
aN1 aN2 · · · aN N
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
j=1 ai j xj(t) and external inputsk
x1(t)
x2(t)
...
xN(t)
x(t)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:125)
. . .
+
A
b11 b12 · · · b1k
b21 b22 · · · b2k
...
...
bN1 bN2 · · · bN k
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
. . .
...
B
,
u1(t)
u2(t)
...
uk(t)
u(t)
(cid:124)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:125)
(cid:219)x(t)
of current statesN
that is characterized by the time evolution of any state (cid:219)xi(t) being a weighted sum
j=1 bi juj(t). We see that our
two-unit system is not linear, because the first state (cid:219)x1(t) depends on sin(x1(t)), and
the second state (cid:219)x2(t) depends on x2
To transform the nonlinear system (cid:219)x = f (x, u), into a linear system (cid:219)x = Ax + Bu,
we can create an approximate model of our vector field about a particular operating
state x∗ and input u∗. We first evaluate the dynamics at this operating point, f (x∗, u∗).
Then we approximate the vector field along small deviations from this point by
computing the derivative of f (x, u) with respect to the states to get matrix A, and
with respect to control inputs to get matrix B
1(t), and is therefore a non-linear system.
B =
∂ f1
∂u1
∂ f2
∂u1
...
∂ f1
∂u2
∂ f2
∂u2
...
∂ fN
∂u1
∂ fN
∂u2
∂uk
· · · ∂ f1
· · · ∂ f2
...
. . .
· · · ∂ fN
∂uk
∂uk
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)x=x∗,u=u∗
.
A =
∂ f1
∂x1
∂ f2
∂x1
...
∂ f1
∂x2
∂ f2
∂x2
...
∂ fN
∂x1
∂ fN
∂x2
∂x N
· · · ∂ f1
· · · ∂ f2
...
. . .
· · · ∂ fN
∂x N
∂x N
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)x=x∗,u=u∗
Then, for states near x∗ and inputs near u∗, the vector field is approximately
∗
, u
∗) + A(x(t) − x
∗) + B(u(t) − u
(cid:219)x(t) = f (x, u) ≈ f (x
(2)
A typical operating point for the input is u∗ = 0 corresponding to no input, because
neural stimulation is viewed as a perturbation to the natural and unstimulated dynam-
ics. A typical operating point for the state x∗ is a fixed point where f (x∗, u∗) = 0,
because then the evolution of our system Eq. 2 only depends on deviations from the
point, and not on its actual value. Finally, we can write the linearized equation explic-
itly as a function of these deviations through a change of variables y(t) = x(t) − x∗,
∗).
(cid:219)y(t) = (cid:219)x(t) ≈ Ay(t) + Bu(t).
We will continue to use variable x instead of y with the understanding that it
represents deviations from the fixed point. For example, in our two-unit system, we
can linearize about x∗
2 = 0, u∗ = 0 to yield
1 = 0, x∗
(cid:21)
(cid:20) (cid:219)x1(t)
(cid:219)x2(t)
(cid:20)−1 2
0 −1
(cid:21)
(cid:21)(cid:20)x1(t)
x2(t)
≈
(cid:20)1
(cid:21)
0
+
u(t).
We show the vector fields and trajectories for both the nonlinear and linear equations
without control where u(t) = 0 (Fig. 3, top), and with control where u(t) = 0.5 (Fig. 3,
10
Jason Z. Kim and Danielle S. Bassett
bottom) from the same initial condition, and we notice that in the neighborhood of
x∗
1 = 0, x∗
2 = 0, the field and trajectories are similar. Hence, by linearizing our neural
dynamics about x∗, u∗, we can preserve the behavior of our neural system at state x(t)
and inputs u(t) near this point, while enabling the use of powerful tools developed
in the next section.
Fig. 3 Vector fields and trajectories for a nonlinear system and its linearized form. Example
vector field of two states with a particular trajectory from initial condition x(0) = [−0.3;−0.4]
for the uncontrolled nonlinear system (top left), the uncontrolled linear system (top right), the
controlled nonlinear system (bottom left) and the controlled linear system (bottom right).
3 Theory of Linear Systems
A useful model for therapeutic intervention in a neural system should capture both
how the activity over time depends on the connections between neural units, and
how to change the activity in a desired way through stimulation. Now that we have
a model that captures features of neural activity and connectivity in a linearized
form, we will develop equations that yield precisely these features. Specifically, we
will first determine the system's response to control through mathematical relations
-0.50.5-0.50.5-0.50.5-0.50.5-0.50.5-0.50.5-0.50.5-0.50.5 Linear Dynamics & Control of Brain Networks
11
as opposed to simulations. Then we will use these principles to design stimuli that
optimally guide our system from some initial state x(0) to some final state x(T).
3.1 Impulse Response
First, we find the natural evolution of system states from some initial neural state
x(0) without any external input. This task amounts to finding the state trajectory x(t)
that solves our dynamic equation (cid:219)x(t) = Ax(t). For scalar systems where x(t) is not
a vector, we are reminded of the solution to (cid:219)x = ax:
= ax
dx
dt
dx = adt
Þ
Þ 1
1
x
adt + c
dx =
x
ln x = at + c
x = Ceat,
k=0
(cid:19)
1 +
d
dt
eat =
d
dt
= 0 +
satisfies (cid:219)x = ax by using a Taylor series of the exponential function eat =∞
where the constant is the initial condition C = x(0). We can prove that this solution
(at)k
.
k!
Taking the time derivative of x(t) = eat, we see (cid:219)x = ax
a3t2
3! + · · · +
3! + · · · + k
aktk
(cid:18)
a2t2
at
1! +
2! +
2! + 3 a3t2
1! + 2 a2t
a2t2
at
2! + · · · +
1! +
k! + · · ·
(cid:19)
aktk−1
k! + · · ·
= aeat .
A matrix exponential is defined exactly the same as above with e At =∞
(cid:19)
(At)k
,
k!
and we again show that the time derivative satisfies the vector relation (cid:219)x(t) = Ax(t)
k! + · · ·
aktk
1 +
= a
(cid:18)
k=0
a
1 +
(cid:18)
A2t2
At
1! +
2! +
1! + 2 A2t
2! + 3 A3t2
A2t2
At
2! + · · · +
1! +
1 +
A
A3t2
3! + · · · +
3! + · · · + k
Aktk
k!
k! + · · ·
Aktk−1
Aktk
k! + · · ·
(cid:19)
+ · · ·
d
dt
e At =
d
dt
= 0 +
(cid:18)
= A
= Ae At .
12
Jason Z. Kim and Danielle S. Bassett
Hence, we see that the following solution
x(t) = e At x(0),
(3)
satisfies our dynamic equation. Here, the matrix exponential e At is called the state
transition matrix, and Eq. 3 is called the impulse response of our system. Hence, we
can find the state at any time T without solving for intermediate states 0 < t < T.
As an example in our two unit model, to find the state of our system at T = 2 given
an initial start at x(0) = [−0.3;−0.4], we can use software to numerically compute
the matrix exponential at time t = 2, and multiply by our initial state Eq. 3
(cid:20)0.1353 0.5413
(cid:21)(cid:20)−0.3
(cid:21)
(cid:21)
(cid:20)−0.2571
−0.0541
,
x(2) = e2Ax(0) =
0.1353
which agrees with the simulation results (Fig. 3).
0
−0.4
=
3.2 Control Response
Next, we derive the system response from an initial state x(0) to some control-
ling input u(t) through some algebraic manipulation and calculus. We begin with
our system equations (cid:219)x(t) − Ax(t) = Bu(t), and multiply both sides by a matrix
exponential
e−At (cid:219)x(t) − e−At Ax(t) = e−At Bu(t).
dt(e−At x(t)) =
Next, we see that the left-hand side is the result of a product rule where d
e−At (cid:219)x(t) − Ae−At x(t), recalling that functions of matrices can switch orders of
multiplication, such that Ae−At = e−At A. Hence, we can write our equation as
(e−At x(t)) = e−At Bu(t),
d
dt
and integrate both sides from t = 0 to t = T to yield
e−AT x(T) − x(0) =
e−At Bu(t)dt.
Þ T
0
Þ T
0
We note the matrix exponential at t = 0 becomes e−A0 = I from the Taylor series.
Next, we move the initial state x(0) to the right hand side, and multiply by e AT
e AT e−AT x(T) = e AT x(0) + e AT
e−At Bu(t)dt.
Finally we use the fact that e AT and e−AT are inverses of each other where e AT e−AT =
I, and we bring e AT into the integral to derive the system's response to control input
Linear Dynamics & Control of Brain Networks
(cid:124)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:125)
x(T) = e AT x(0)
natural
+
Þ T
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
e A(T−t)Bu(t)dt
0
controlled
13
(4)
.
Intuitively, we see that the first part of the response, e AT x(0), is just the natural
evolution of our system from an initial state, and that the second part of the response
is a convolution of our mapped inputs, Bu(t), with the impulse response. We will next
take advantage of the convolution's property of linearity to draw powerful relations
between the state evolution, control input, and system structure.
3.3 Linear Relation Between the Convolution and Control Input
Previously, we focused on the evolution of a neural system in response to a known
control input u(t) in Eq. 4. However, our goal is to design a control input that drives
our neural system to some desired final state that may stabilize an epileptic seizure
[80], or aid in memory recall [27]. In this scenario, we fix the initial state x(0) = x0
and the final state x(T) = xT as constants, and we search for an input u(t) that
satisfies
Þ T
0
e A(T−t)B u(t)(cid:124)(cid:123)(cid:122)(cid:125)
dt = x(T) − e AT x(0)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
.
variable
constant
This formulation is a linear equation with a structure that is similar to a typical
system of linear equations used in regression, M v = b, where v is the variable, b is
a constant vector, and matrix M is the linear function acting on v. Here, the control
input u(t) is the variable, x(T) − e AT x(0) is the constant vector, and the convolution
Þ T
0
L(u(t)) =
e A(T−t)Bu(t)dt,
is the linear function acting on our control inputs. By linear function, we mean that
for two control inputs u1(t), u2(t), if L(u1(t)) = c1, and L(u2(t)) = c2, then a
weighted sum of inputs yields the same weighted sum of outputs, such that
L(au1(t) + bu2(t)) = ac1 + bc2.
(5)
This linearity allows us to treat solutions to our control function problem the same as
solutions to our linear system of equations. Specifically, suppose control input u∗(t)
was a particular solution to our control problem such that L(u∗(t)) = xT − e AT x0,
and u1(t), u2(t), · · · were homogeneous solutions such that L(ui(t)) = 0. Then the
set of all valid control inputs is given by u(t) = u∗(t) +
i ai ui(t), because
14
L(u(t)) = L(u
∗(t)) +
= xT − e AT x0 +
i
ai0
Jason Z. Kim and Danielle S. Bassett
L(ai ui(t))
= xT − e AT x0.
i
3.4 Controllability
For any system, we would first like to know if a particular solution exists to the
control problem described above. A system is controllable if there is a control input
that brings our system from any initial state to any final state in finite time. For
nonlinear systems, if we know that the input u∗(t) brings our system from the initial
state 0 to some final state xT , there is in general no way to know what input will take
our system to a scaled final state axT .
In contrast, due to the linearity of our convolution operator, we know that a
scaled input au∗(t) will produce a scaled output L(au∗(t)) = axT . Further, any
N-dimensional vector can be written as a weighted sum of N linearly independent
vectors v1, v2, · · · , vN. Here, linear independence means that no vector vi in the set
can be written as a weighted sum of the remaining vectors v j(cid:44)i. For example, a
column vector a = [a1; a2; · · · ; aN] can be written as the weighted sum
(cid:124)(cid:123)(cid:122)(cid:125)
a1
a2
...
aN
a
(cid:124)(cid:123)(cid:122)(cid:125)
1
0
...
0
v1
(cid:124)(cid:123)(cid:122)(cid:125)
0
1
...
0
v2
= a1
+a2
+ · · · + aN
(cid:124)(cid:123)(cid:122)(cid:125)
0
0
...
1
v N
,
where none of the vectors vi can be written as a weighted sum of remaining vectors
v j(cid:44)i. Hence, our system is controllable if we can find input functions u1(t), · · · , u N(t)
that reach N linearly independent vectors L(u1(t)), · · · , L(u N(t)), because then we
can always reach any final state from any initial state through the weighted sum
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
xT − e AT x0
a
+ · · · + aN L(u N(t))
,
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
v N
v1
= a1 L(u1(t))
+a2 L(u2(t))
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
C =(cid:2)B, AB, A2B, · · · , AN−1B(cid:3) ,
v2
through the control input u(t) = a1u1(t) + a2u2(t) +· · · + aN u N(t). This information
of reachable states is encoded in the controllability matrix
(6)
where the rank of this matrix (given by the number of linearly independent columns
of C) tells us how many of these N independent vectors can be reached using control
Linear Dynamics & Control of Brain Networks
15
input. If this rank = N, then the system is controllable and can reach all states.
Further, if some vector xT − e AT x0 can be written as a weighted sum of the columns
of C, then there exists a control input that drives the system from x0 to xT . This set
of vectors spanned by the columns of C is called the controllable subspace.
As an example in our two unit system, the controllability matrix is written as
(cid:20)1 −1
(cid:21)
0 0
,
C =
which is not controllable, because the rank of C is 1. To consider the controllable
subspace, notice that the columns of C only have non-zero entry in the first row.
Hence, the controllable subspace contains any desired value of x1(T), but excludes
all values of x2(T). Intuitively, this loss of controllability arises because x2 does not
receive an input, nor is it affected by x1. Hence, there is no way to influence the
activity of x2 in a desired way.
3.5 Minimum Energy Control
Once we know a system is controllable, we would like to determine the control input
function u(t) that transitions our system from initial x0 to final xT states. However,
there are often limitations on the input magnitude such as electrical and thermal
damage of neural tissue, or battery life of chronic implanted stimulators. Due to the
system's linearity, we can not only find an input function, but an optimal one u∗(t)
that minimizes input cost.
First, we must define a measure of the size of our control input functions u(t). In
many applications of electrical and electromagnetic stimulation, the cost of control
scales quadratically with the input such as resistive heating with electrical current.
This quadratic measure of size is mathematically and intuitively defined using the
inner product. For N-dimensional column vectors of numbers, a, the inner product
is the well known dot product
< a, a >= a2
1 + a2
2 + · · · + a2
N
(cid:48)
a,
= a
where a(cid:48) is the transpose that turns column vector a into a row vector. We see that
doubling a will quadruple the inner product. For k-dimensional column vectors of
functions, a(t), the inner product is similarly defined as
< a(t), a(t) >=
1(t) + a2
a2
2(t) + · · · + a2
N(t)dt =
(cid:48)(t)a(t)dt,
a
that has the same quadratic relation. Hence, we define the control energy as
E =< u(t), u(t) > .
(7)
Þ
Þ
16
Jason Z. Kim and Danielle S. Bassett
Now that we have a measure of how large an input is, we wish to find a minimal
input u∗(t) that minimizes the control energy. This task is analogous to a typical
linear system of equations, M v = b, where we want to find v∗ that solves the equation
with the smallest cost < v∗, v∗ >. Here, if M has full row rank where the rows of
M are linearly independent, then the minimum solution is given by the equation for
least squares v∗ = M(cid:48)(M M(cid:48))−1
This same principle holds for our linear system L(u(t)) = xT − e AT x0, where we
want to find u∗(t) that solves the equation with the smallest cost < u∗(t), u∗(t) >.
However, while matrix M inputs a vector of numbers v and outputs a vector of
numbers b, our linear function L inputs a vector of functions and outputs a vector
of numbers. Hence, we need to carefully define the transpose, or adjoint L(cid:48). In the
case of matrix M, the adjoint preserves the inner product between inputs and outputs
such that
b. Here, M(cid:48) is the transpose, or adjoint of M.
< M v, b > =< v, M(cid:48)
b >
(cid:48)(M(cid:48)
b).
(M v)(cid:48)
b = v
Identically, for state transition x = e AT x0 − xT , the adjoint of L preserves the inner
product between the vectors of input functions u(t), and output numbers x as
(cid:18)Þ T
0
(cid:19)(cid:48)
Þ T
0
< L(u(t)), x > =< u(t), L(cid:48)(x) >
e A(T−t)Bu(t)dt
x =
(cid:48)(t)(B(cid:48)e A(cid:48)(T−t)
x)dt.
u
Notice that the inner product on the left is over vectors of numbers, while the inner
product on the right is over vectors of functions. Then, we see that our adjoint is
L(cid:48)(x) = B(cid:48)e A(cid:48)(T−t)
x,
and takes as input a vector of numbers, and outputs a vector of functions. Then, just
as our system M v = b, the minimum input u∗(t) is given by
∗(t) = L(cid:48)(LL(cid:48))−1(xT − e AT x0).
(8)
u
Finally, through substitution into Eq. 7, we can write the minimum control energy as
Emin = (xT − e AT x0)(cid:48)(LL(cid:48))−1(xT − e AT x0).
(9)
In conclusion, we point out the crucially important term of the minimum energy,
LL(cid:48), as the controllability Gramian written as
Wc(T) = LL(cid:48) =
e A(T−t)BB(cid:48)e A(cid:48)(T−t)dt.
(10)
First, we notice that this Gramian is only a function of the underlying neural relation-
ships, A, the matrix determining where the inputs are placed, B, and time T. Next,
Þ T
0
Linear Dynamics & Control of Brain Networks
17
we notice that Wc(T) is actually an N × N matrix, and can therefore be numerically
evaluated and analytically studied. Finally, we see that if our system begins at an
initial state of x0 = 0, then the minimum energy can be written
Emin = x
TW−1
(cid:48)
c (T)xT,
where the role of neural interactions and stimulation parameters on our ability to
control the system is fully encapsulated in the Gramian. This ability to decouple
the states xT from the neural interactions and stimulation parameters A, B, T is a
powerful tool for studying and designing control properties of neural systems.
4 Mapping Network Architecture to Control Properties
By formulating our neural system in a linear way, we can solve difficult problems
such as predicting the system's response to control, finding the set of states that the
system can reach, and designing efficient input stimuli, without the need to try every
control input and simulate every trajectory. Further, by directly mapping control
properties to neural activity and network architecture in an algebraic way, we can
study how features of interaction patterns impact our ability to control neural activity
[8]. As an active area of research, the variety of questions being asked and systems
being studied is very large, and require simultaneous innovations in experiment,
computation, and theory. In this section, we will describe a few recent applications
and advances.
4.1 Neuronal Control in Model Organisms
While most neural systems are too large to empirically measure activity and con-
nectivity or to analyze numerically, there do exist a few sufficiently simple model
organisms. Among these is the worm Caenorhabditis elegans [38] with several hun-
dred neurons that can be recorded from simultaneously [59]. Even for such a small
system, it is difficult to map the functional form of how activity in neuron i affects
the activity in neuron j. However, the presence or absence of connections between
neurons in this organism, and by consequence the presence or absence of elements
in the connectivity matrix A, is well known.
Advances in the study of structural controllability [34] allow us to ask questions
about our ability to control a system given only the binary presence or absence
of edges. Colloquially, this framework focuses on connectivity matrices A where
non-zero entries can only exist in the presence of binary edges, and can be used
to determine whether the system is controllable for most values where an edge is
present. Using this framework, recent work has sought to determine whether the
removal of certain neurons in C. elegans will reduce structural controllability [33].
18
Jason Z. Kim and Danielle S. Bassett
Specifically, the modeling involves input to the sensory receptor neurons as the
control input that is mapped to the system through a matrix B, and the connectivity
between neurons and muscle cells through a matrix A. Further, instead of recording
the activity of each neuron, the motion of muscles was recorded. This framework
involves the appended control framework
(cid:219)x(t) = Ax(t) + Bu(t)
y(t) = C x(t),
where y(t) represents the states (muscles) that are measured, and C is the map from
neurons and muscles x(t) to the measured output [35]. Here, the authors find that the
ablation of a neuron not previously implicated in motion, PDB, decreased structural
controllability, significantly reducing ventral bias in deep body bends in C. elegans.
4.2 State Transitions in the Human Brain
While neuron-level structural synapses map most directly to functional relationships
between neurons, there are also well-characterized structural connections between
larger-scale brain regions. These connections contain thick bundles of myelinated
axonal fibers that run throughout the brain, and are thought to play a crucial role
in coupling the activity of distant brain regions [39]. These fibers are resolved by
measuring water diffusion throughout the brain using magnetic resonance [40], and
tracing fibers along this diffusion field using computational algorithms [31]. The
whole brain is typically divided into hundreds to thousands of discrete brain regions
using a variety of parcellation schemes [41, 42], and the strength of fibers between
these regions comprise the connectivity matrix A [32].
Such region-level study of brain dynamics has led to the discovery of macroscopic
functional organization in the human brain at rest [36] and during various cognitively
demanding tasks [37]. Here, brain activity can be empirically measured through
methods such as magnetic resonance imaging (blood oxygen level dependent) or
electrophysiology (aggregate electrical activity). Of particular interest are large-
scale functional brain networks that display stereotyped changes in activity patterns
during tasks that demand certain cognitive or sensorimotor processes [43]. Here, it is
thought that the brain uses underlying structural connections to support circuit-level
coordination, as well as to guide itself to specific patterns of activity using cognitive
control [44, 45].
Recent work has begun formulating cognitive control as a linear systems problem
[44, 46, 47, 49, 48], where matrix A is the network of white matter connections
between brain regions, B represents the regions that were chosen to be responsible for
control, and x(t) represents the activity of each region over time. Specifically in [46,
49], the authors quantify cognitive states as vectors corresponding to activity in the
brain regions during cognitive tasks, and compute the minimum control energy Eq. 8
to transition between cognitive states for various sets of control regions. Colloquially,
Linear Dynamics & Control of Brain Networks
19
if a set of regions requires less input energy to transition between cognitive states,
then those regions may easily transition the whole brain between these states along
an optimal trajectory given they are responsible for cognitive control. Moreover,
individual differences in the minimal control energy are correlated with individual
differences in performance on cognitive control tasks [50]. In complementary studies,
individual differences in controllability statistics calculated for distinct regions of
the brain are correlated with individual differences in measures of cognitive control
assessed with common neuropsychological test batteries [47, 48].
5 Methodological Considerations and Limitations
While the theory of linear systems is a powerful quantitative framework for studying
and controlling dynamical neural systems, there are several important caveats. Here
we mention three: dimensionality and numerical stability, model validation and
experimental data, and the assumption of linearity.
5.1 Dimensionality and Numerical Stability
The benefit of studying linear systems is that we take difficult and largely intractable
questions of controllability and control input design, and greatly simplify them into
algebraic problems of computing objects like the controllability matrix Eq. 6 and the
controllability Gramian Eq. 10. However, these matrices scale quadratically with the
number of neural units, and numerical calculations and manipulations using these
matrices quickly face computational issues.
Most viable approaches to dealing with these issues involve numerically repre-
senting the elements of our matrices, and performing algebraic operations. However,
these representations are imperfect, as it is impossible to completely represent irra-
tional numbers such as π. Hence, the matrices are truncated to numerical precision,
and this truncation error propagates with each computation. Further, the propagation
of error tends to scale faster than the number of dimensions. This issue is prevalent
in the computation of the state-transition matrix [51], as well as in the calculation
of the controllability Gramian and its inverse. With the application of this theory
to high dimensional neural systems, the study of useful controllability metrics is an
active area of research [52].
5.2 Model Validation and Experimental Data
A fundamental limitation for modeling any neural system is the ability to empirically
and accurately measure model parameters and variables. A crucial parameter is the
20
Jason Z. Kim and Danielle S. Bassett
network of connectivity encoded by our adjacency matrix A, where the element in
the i-th column and j-th row models the effect of unit i on the rate of change of unit
j. While we typically use the structural connections in synapses between neurons,
or bundles of axons between brain regions as a proxy for A, it is very difficult
to measure the true functional effect that activity in unit i has on activity in unit j,
particularly for large systems. This problem is exacerbated by further methodological
limitations such as the inability to resolve directionality of connections in diffusion
tractography. Along these lines, many statistical and autoregressive methods have
been developed to infer functional relationships from recordings of neural activity
[53, 54, 55, 56, 57], and to use that inferred activity to better understand control [58].
However, the degree of causality in these methods as measured by true response to
external stimuli remains controversial.
Another such fundamental limitation is our inability to fully measure every state
of the system. The state-space representation of our model requires that every state
is observed. However, it is impossible to simultaneously record the activity of every
neuron in almost all biological systems, although this recording has been achieved
in sufficiently simple organisms [59]. As a result of only being able to observe a
small subset of the full state-space, these models of interactions may become largely
descriptive and phenomenological in nature. In response, there is a continuing effort
to improve the spatial and temporal resolution of neuroimaging methods [60].
5.3 Assumption of Linearity
An inherent limitation is the lack of generality in our linear approximation of the
full nonlinear neural dynamics. In response, there is a sizable quantity of research
studying the control properties of nonlinear dynamical systems [61]. An interesting
bridge between these two disciplines exists in the theory of the Koopman or compo-
sition operator [62]. The underlying benefit of this theory is that, while our system
of equations may evolve nonlinearly in time given the current set of N states, there
may exist a higher-dimensional set of M > N state variables in which the dynam-
ical system does evolve linearly [63]. While the extension of linear systems theory
to actually controlling this higher-dimensional system may be limited, it remains a
promising future area of research.
6 Open Frontiers
Many exciting and open frontiers exist in the study of brain network dynamics using
linear systems theory. Here we constrain our remarks to three main topic areas, but
freely admit that this discussion is far from comprehensive. First, we describe oppor-
tunities in the further development of useful controllability statistics as well as in the
development of foundational theory linking control profiles to the system's underly-
Linear Dynamics & Control of Brain Networks
21
ing network architecture. Second, we underscore the need for a better understanding
of how control is implemented in the brain, how control strategies might depend
on context, and how control processes could facilitate the effective manipulation
of information. Third, we describe the relevance of the modeling efforts we dis-
cussed here for our understanding of neurological disease and psychiatric disorders
as well as the development of personalized and targeted therapeautic interventions
for alterations in mental health.
6.1 Theory and Statistics
Linear systems theory has its basis in a rich literature stemming from now well-
developed areas of mathematics, physics, and engineering [64]. Yet, much is still
unknown about exactly how the network topology of a given unit-to-unit interaction
pattern impacts the capacity for control, the trajectories accessible to the systems,
and the minimum control energy. Some preliminary efforts have begun to make
headway by using linear network control theory to derive accurate closed-form
expressions that relate the connectivity of a subset of structural connections (those
linking driver nodes to non-driver nodes) to the minimum energy required to control
networked systems [65]. Further work is needed gain an intuition for the role of
higher order structures (e.g., cycles) in the control of the networked system, and
any dependence on edge directionality [67]. Moreover, it would be fruitful in the
future to further develop a broader set of controllablity statistics, extending beyond
node controllability [52], and edge controllability [70], to the control of motifs [71].
Finally, throughout such investigations it will be useful to understand which features
of control are shared across networks with various topologies, versus those features
which are specific to networks with a particular topology [66, 68, 69].
6.2 Context, Computations, and Information Processing
Despite the emerging appreciation that linear systems theory has considerable utility
in the study of cognitive function, we still know very little about exactly how control
is implemented in the brain, across spatial scales, and capitalizing on the unit-to-
unit interaction patterns at each of those scales. Some initial evidence suggests that
features of synaptic connectivity -- and particularly autaptic connections -- can serve
to tune the excitability of the neural circuit, altering its controllability profile and
propensity to display synchronous bursts of activity [72]. Complementary evidence
also at the cellular scale demonstrates how intrinsic network structure and exogeneous
stimulus patterns together determine the manner in which a stimulus propagates
through the network, with important implications for cognitive faculties that require
persistent activation of neuronal patterns such as working memory and attention [73].
There are interesting similarities between these observations and evidence at larger
22
Jason Z. Kim and Danielle S. Bassett
spatial scales, which suggests that the architecture of white matter tracts connecting
brain areas can be used to infer the probability with which the brain persists in
certain states [74]. Such conceptual similarities motivate concerted efforts to better
understand how the architecture of brain networks across spatial scales supports
information processing and cognitive computations, and how those processes and
computations might depend on the context in which the brain is placed. Formally,
it would be interesting to consider context as a form of exogeneous input to the
system, in a manner reminiscent of how we currently consider brain stimulation [8].
We speculate that such a formulation of the problem could help to explain a range
of observations, such as the ability of cognitive effort to suppress epileptic activity
[75].
6.3 Disease and Intervention
The fact that controllability can depend on network topology [66, 65] and can be
altered by edge pruning [76], suggests that it might also be a useful biomarker in some
neurological diseases and psychiatric disorders, many of which are associated with
changes in the structural topology of neural circuitry at various spatial scales [6, 7].
Indeed, recent studies have reported differences in controllability statistics estimated
in brain networks of patients with bipolar disorder [77], temporal lobe epilepsy [78],
and mild traumatic brain injury [49]. In a complementary line of work, studies are
beginning to ask whether the altered controllability profiles of brain networks in
these patients could help to inform the development of more targeted interventions
for their illness, in the form of brain stimulation [79, 80], pharmacological agents,
or cognitive behavioral therapy. Other efforts have begun to consider symptoms of a
given disease as a network, and to identify symptoms predicted to have high impulse
response in the patient's daily life [81]. It would be interesting in future to determine
whether the linear systems approach could be useful in more carefully formalizing
that problem as a network control problem, which in turn could be used to determine
which symptom to treat in order to move the entire symptom network towards a
healthier state [82].
7 Acknowledgements
We gratefully acknowledge comments and feedback from Arian Ashourvan, Ann E.
Sizemore, Melody X. Lim, Jennifer A. Stiso, Erin G. Teich, Teresa Karrer, Zhixin
Lu, Harang Ju, and Eli J. Cornblath. We also thank Ann E. Sizemore for generous
assistance with and input on schematic figure construction. JZK acknowledges sup-
port from the NIH T32-EB020087, PD: Felix W. Wehrli, and the National Science
Foundation Graduate Research Fellowship No. DGE-1321851. DSB acknowledges
support from the John D. and Catherine T. MacArthur Foundation, the Alfred P.
Linear Dynamics & Control of Brain Networks
23
Sloan Foundation, the Paul G. Allen Foundation, the Army Research Laboratory
through contract number W911NF-10-2-0022, the Army Research Office through
contract numbers W911NF-14-1-0679 and W911NF-16-1-0474, the National Insti-
tute of Health (2-R01-DC-009209-11, 1R01HD086888-01, R01-MH107235, R01-
MH107703, R01MH109520, 1R01NS099348 and R21-M MH-106799), the Office
of Naval Research, and the National Science Foundation (BCS-1441502, CAREER
PHY-1554488, BCS-1631550, and CNS-1626008).The content is solely the respon-
sibility of the authors and does not necessarily represent the official views of any of
the funding agencies.
24
8 Problems
Jason Z. Kim and Danielle S. Bassett
• Problem 1: Linearize the following system about point x∗
1(t) − 2x2(t) + x3(t) − 1
2x1(t) − 2x2
2(t) + 2x3(t)
x1(t)x2(t) − x3(t) + 1
(cid:219)x1(t)
(cid:219)x2(t)
(cid:219)x3(t)
=
−x2
and demonstrate that this point is a fixed point where (cid:219)x1 = (cid:219)x2 = (cid:219)x3 = 0.
1 = 1, x∗
2 = −1, x∗
3 = 0,
• Problem 2: Prove that the matrix exponential of A =
is
.
(cid:21)
(cid:20)a 0
0 b
(cid:21)
(cid:20)ea 0
0 eb
,
e A =
Þ t
0
Þ T
0
using the Taylor series of the scalar and matrix exponentials.
• Problem 3: Prove that the system response to control
x(t) = e At x(0) +
e A(t−τ)Bu(τ)dτ,
satisfies the dynamical equation (cid:219)x(t) = Ax(t) + Bu(t) by substitution.
• Problem 4: Prove that the convolution operator
L(u(t)) =
e A(T−τ)Bu(τ)dτ
is linear according to Eq. 5; that is, if L(u1(t)) = c1, and L(u2(t)) = c2, then
demonstrate that L(au1(t) + bu2(t)) = ac1 + bc2.
• Problem 5: Determine if the following system is controllable
(cid:219)x1(t)
(cid:219)x2(t)
(cid:219)x3(t)
=
0 1 0
0 0 1
1 0 0
x1(t)
+
x2(t)
x3(t)
1
0
0
u(t),
by constructing the controllability matrix.
• Problem 6: Determine for what value of a the system is not controllable
Linear Dynamics & Control of Brain Networks
(cid:219)x1(t)
(cid:219)x2(t)
(cid:219)x3(t)
=
0 0 0
1 1 0
1 0 a
x1(t)
x2(t)
x3(t)
25
+
u(t),
1
0
0
by constructing the controllability matrix.
• Problem 7: Derive the minimum energy equation Eq. 9
Emin = (xT − e AT x0)(cid:48)(LL(cid:48))−1(xT − e AT x0),
by substituting the minimum input u∗(t) into the control energy Eq. 7
E =< u(t), u(t) > .
• Problem 8: Show that the controllability Gramian can be written
WC(T) =
e A(T−t)BBT e AT (T−t)dt =
e Aτ BBT e AT τdτ,
Þ T
0
using the substitution τ = T − t.
Þ T
0
0 1
(cid:21)
(cid:20)1 0
(cid:0)e2bT − 1(cid:1)(cid:21)
(cid:20)1 0
(cid:21)
0
• Problem 9: Show that the controllability Gramian for system
is
A =
WC(T) =
(cid:21)
(cid:20)a 0
(cid:20) 1
(cid:0)e2aT − 1(cid:1)
0 b
,
2a
0
B =
1
2b
• Problem 10: Compute the minimum energy required for the system
to transition from initial state x(0) =
to final state x(T) =
B =
0 1
,
(cid:21)
(cid:20)1
2
in time T = 1.
(cid:20) 1
(cid:21)
A =
2 0
0 2
,
(cid:21)
(cid:20)0
0
26
References
Jason Z. Kim and Danielle S. Bassett
1. Maier, M., Zippelius, A., Fuchs, M. Emergence of Long-Ranged Stress Correlations at the
Liquid to Glass Transition. Phys Rev Lett. 2017 Dec 29;119(26):265701. doi: 10.1103/Phys-
RevLett.119.265701.
2. Kivelson, S., Kivelson, S. A. Defining emergence in physics. Quantum Materials. 1:16024
(2016). doi: 10.1038/npjquantmats.2016.24.
3. Lynn, C. W., Papadopoulos, L., Lee, D., Bassett, D. S. Surges of collective human activity
emerge from simple pairwise correlations. Phys. Rev. X. (2018) In Press.
4. Bassett, D. S., Gazzaniga, M. S. Understanding complexity in the human brain. Trends Cogn
Sci. 2011 May;15(5):200-9. doi: 10.1016/j.tics.2011.03.006.
5. Haimovici, A., Tagliazucchi, E., Balenzuela, P., Chialvo, D. R. Brain organization into resting
state networks emerges at criticality on a model of the human connectome. Phys Rev Lett.
2013 Apr 26;110(17):178101. doi: 10.1103/PhysRevLett.110.178101.
6. Braun, U., Schaefer, A., Betzel, R. F., Tost, H., Meyer-Lindenberg, A., Bassett, D. S. From
Maps to Multi-dimensional Network Mechanisms of Mental Disorders. Neuron. 2018 Jan
3;97(1):14-31. doi: 10.1016/j.neuron.2017.11.007.
7. Stam, C. J. Modern network science of neurological disorders. Nat Rev Neurosci. 2014
Oct;15(10):683-95. doi: 10.1038/nrn3801.
8. Tang, E., Bassett, D. S. Control of dynamics in brain networks. Rev. Mod. Phys. 2018.
90:031003. doi: 10.1103/RevModPhys.90.031003.
9. Downar, J., Geraci, J., Salomons, T. V., Dunlop, K., Wheeler, S., McAndrews, M. P., Bakker,
N., Blumberger, D. M., Daskalakis, Z. J., Kennedy, S. H., Flint, A. J., Giacobbe, P. Anhedonia
and reward-circuit connectivity distinguish nonresponders from responders to dorsomedial
prefrontal repetitive transcranial magnetic stimulation in major depression. Biol Psychiatry.
2014 Aug 1;76(3):176-85. doi: 10.1016/j.biopsych.2013.10.026.
10. Medaglia, J. D., Harvey, D. Y., White, N., Kelkar, A., Zimmerman, J., Bassett, D. S., Hamil-
ton, R. H. Network Controllability in the Inferior Frontal Gyrus Relates to Controlled Lan-
guage Variability and Susceptibility to TMS. J Neurosci. 2018 Jul 11;38(28):6399-6410. doi:
10.1523/JNEUROSCI.0092-17.2018.
11. Gass, N., Becker, R., Sack, M., Schwarz, A. J., Reinwald, J., Cosa-Linan, A., Zheng, L., von
Hohenberg, C. C., Inta, D., Meyer-Lindenberg, A., Weber-Fahr, W., Gass, P., Sartorius, A.
Antagonism at the NR2B subunit of NMDA receptors induces increased connectivity of the
prefrontal and subcortical regions regulating reward behavior. Psychopharmacology (Berl).
2018 Apr;235(4):1055-1068. doi: 10.1007/s00213-017-4823-2
12. Braun, U., Schafer, A., Bassett, D. S., Rausch, F., Schweiger, J. I., Bilek, E., Erk, S.,
Romanczuk-Seiferth, N., Grimm, O., Geiger, L. S., Haddad, L., Otto, K., Mohnke, S., Heinz,
A., Zink, M., Walter, H., Schwarz, E., Meyer-Lindenberg, A., Tost, H. Dynamic brain net-
work reconfiguration as a potential schizophrenia genetic risk mechanism modulated by
NMDA receptor function. Proc Natl Acad Sci U S A. 2016 Nov 1;113(44):12568-12573.
doi: 10.1073/pnas.1608819113
13. Yang, Z., Gu, S., Honnorat, N., Linn, K. A., Shinohara, R. T., Aselcioglu, I., Bruce, S.,
Oathes, D. J., Davatzikos, C., Satterthwaite, T. D., Bassett, D. S., Sheline, Y. I. Network
changes associated with transdiagnostic depressive symptom improvement following cognitive
behavioral therapy in MDD and PTSD. Mol Psychiatry. 2018 Dec;23(12):2314-2323. doi:
10.1038/s41380-018-0201-7.
14. Markram, H., Muller, E., Ramaswamy, S., Reimann, M. W., Abdellah, M., Sanchez, C.
A., Ailamaki, A., Alonso-Nanclares, L., Antille, N., Arsever, S., Kahou, G. A., Berger, T.
K., Bilgili, A., Buncic, N., Chalimourda, A., Chindemi, G., Courcol, J. D., Delalondre, F.,
Delattre, V., Druckmann, S., Dumusc, R., Dynes, J., Eilemann, S., Gal, E., Gevaert, M. E.,
Ghobril, J. P., Gidon, A., Graham, J. W., Gupta, A., Haenel, V., Hay, E., Heinis, T., Hernando,
J. B., Hines, M., Kanari, L., Keller, D., Kenyon, J., Khazen, G., Kim, Y., King, J. G., Kisvarday,
Z., Kumbhar, P., Lasserre, S., Le, Be, J. V., Magalhaes, B. R., Merchan-Perez, A., Meystre,
J., Morrice, B. R., Muller, J., Munoz-Cespedes, A., Muralidhar, S., Muthurasa, K., Nachbaur,
Linear Dynamics & Control of Brain Networks
27
D., Newton, T. H., Nolte, M., Ovcharenko, A., Palacios, J., Pastor, L., Perin, R., Ranjan, R.,
Riachi, I., Rodriguez, J. R., Riquelme, J. L., Rossert, C., Sfyrakis, K., Shi, Y., Shillcock, J. C.,
Silberberg, G., Silva, R., Tauheed, F., Telefont, M., Toledo-Rodriguez, M., Trankler, T., Van
Geit, W., Diaz, J. V., Walker, R., Wang, Y., Zaninetta, S. M., DeFelipe, J., Hill, S. L., Segev, I.,
Schurmann, F. Reconstruction and Simulation of Neocortical Microcircuitry. Cell. 2015 Oct
8;163(2):456-92. doi: 10.1016/j.cell.2015.09.029.
15. Rosenblueth, A., Wiener, N. The role of models in science. Philosophy of Science. (1945) 12
(4):316-321.
16. Machta, B. B., Chachra, R., Transtrum, M. K., Sethna, J. P. Parameter space compression
underlies emergent theories and predictive models. Science. 2013 Nov 1;342(6158):604-7.
doi: 10.1126/science.1238723.
17. Mattingly, H. H., Transtrum, M. K., Abbott, M. C., Machta, B. B. Maximizing the informa-
tion learned from finite data selects a simple model. Proc Natl Acad Sci U S A. 2018 Feb
20;115(8):1760-1765.
18. Hodgkin, A. L., Huxley, A. F.: A quantitative description of membrane current and its
application to conduction and excitation in nerve. J. Physiol. (1952) doi: 10.1113/jphys-
iol.1952.sp004764
19. Cano, G., Dilao, R. Intermittency in the Hodgkin-Huxley model. J Comput Neurosci. 2017
Oct;43(2):115-125. doi: 10.1007/s10827-017-0653-9.
20. Goldwyn, J. H., Shea-Brown, E. The what and where of adding channel noise to the
Hodgkin-Huxley equations. PLoS Comput Biol. 2011 Nov;7(11):e1002247. doi: 10.1371/jour-
nal.pcbi.1002247.
21. Teka, W., Stockton, D., Santamaria, F. Power-Law Dynamics of Membrane Conductances
Increase Spiking Diversity in a Hodgkin-Huxley Model. PLoS Comput Biol. 2016 Mar
3;12(3):e1004776. doi: 10.1371/journal.pcbi.1004776.
22. FitzHugh, R.: Impulse and physiological states in theoretical models of nerve membrane.
Biophys. J. (1961) doi: 10.1016/S0006-3495(61)86902-6
23. McCulloch, W. S., Pitts, W.: A logical calculus of the ideas immanent in nervous activity.
Bull. Math. Sci. (1943) doi: 10.1007/BF02478259
24. Hopfield, J. J.: Neural networks and physical systems with emergent collective computational
abilities. Proc. Natl. Acad. Sci. (1982) doi: 10.1073/pnas.79.8.2554
25. Rosenblatt, F.: The perceptron: a probabilistic model for information storage and organization
in the brain. Psychol. Rev. (1958) doi: 10.1037/h0042519
26. Hedlund, G.A. Math. Systems Theory (1969) 3: 320. https://doi.org/10.1007/BF01691062
27. Ezzyat, Y. et al.: Direct brain stimulation modulates encoding states and memory performance
in humans. Curr. Biol. (2017) doi: 10.1016/j.cub.2017.03.028
28. Wilson, H. R., Cowan, J. D.: Excitatory and inhibitory interactions in localized populations
of model neurons. Biophys. J. (1972) doi: 10.1016/S0006-3495(72)86068-5
29. Shusterman, V., Troy, W. C.: From baseline to epileptiform activity: A path to synchronized
rhythmicity in large-scale neural networks. Phys. Rev. E. Stat. Nonlin. Soft. Matter. Phys.
(2008) doi: 10.1103/PhysRevE.77.061911
30. Oh, S. W. et al.: A mesoscale connectome of the mouse brain. Nature. (2014) doi: 10.1038/na-
ture13186
31. Basser, P. J., Pajevic, S., Pierpaoli, C., Duda, J., Aldroubi, A.: In vivo fiber tractography using
DT-MRI data. Magn. Reson. Med. (2000) doi: 10.1002/1522-2594(200010)44:4<625::AID-
MRM17>3.0.CO;2-O
32. Bassett, D. S., Zurn, P., Gold, J. I. On the nature and use of models in network neuroscience.
Nat Rev Neurosci. 2018 Sep;19(9):566-578. doi: 10.1038/s41583-018-0038-8
33. Yan, G. et al.: Network control principles predict neuron function in the Caenorhabditis
elegans connectome. Nature. (2017) doi: 10.1038/nature24056
34. Lin, C. T.: Structural Controllability.
IEEE T. Automat. Contr.
(1974)
doi:
10.1109/TAC.1974.1100557
35. Towlson, E. K. et al.: Caenorhabditis elegans and the network control framework -- FAQs.
Philos. Trans. Royal Soc. B. (2018) doi: 10.1098/rstb.2017.0372
28
Jason Z. Kim and Danielle S. Bassett
36. Raichle, M. E. et al.: A default mode of brain function. PNAS.
(2001) doi:
10.1073/pnas.98.2.676
37. Sporns, O., Betzel, R. F. Modular Brain Networks. Annu Rev Psychol. 2016;67:613-40. doi:
10.1146/annurev-psych-122414-033634.
38. White, J. G., Southgate, E., Thomson, J. N., Brenner, S.: The structure of the nervous system
of the nematode Caenorhabditis elegans. Philos. Trans. R. Soc. Lond. B. Biol. Sci. (1986) doi:
10.1098/rstb.1986.0056
39. Avena-Koenigsberger, A., Misic, B., Sporns, O. Communication dynamics in complex brain
networks. Nat Rev Neurosci. 2017 Dec 14;19(1):17-33. doi: 10.1038/nrn.2017.149
40. Taylor, D. G., Bushell, M. C.: The spatial mapping of translational diffusion coefficients by
the NMR imaging technique. Phys. Med. Biol. (1985) doi: 10.1088/0031-9155/30/4/009
41. Hagmann, P. et al.: Mapping the structural core of human cerebral cortex. PLoS Biol. (2008)
42. Power, J. D. et al.: Functional network organization of the human brain. Neuron. (2011) doi:
doi: 10.1371/journal.pbio.0060159
10.1016/j.neuron.2011.09.006
43. Bressler, S. L., Menon, V.: Large-scale brain networks in cognition: emerging methods and
principles. Trends. Cogn. Sci. (2010) doi: 10.1016/j.tics.2010.04.004
44. Gu, S., Pasqualetti, F., Cieslak, M., Telesford, Q. K., Yu, A. B., Kahn, A. E., Medaglia, J. D.,
Vettel, J. M., Miller, M. B., Grafton, S. T., Bassett, D. S. Controllability of structural brain
networks. Nat Commun. 2015 Oct 1;6:8414. doi: 10.1038/ncomms9414.
45. Medaglia, J. D., Huang, W., Karuza, E. A., Kelkar, A., Thompson-Schill, S. L., Ribeiro, A.,
Bassett, D. S. Functional Alignment with Anatomical Networks is Associated with Cognitive
Flexibility. Nat Hum Behav. 2018;2(2):156-164. doi: 10.1038/s41562-017-0260-9.
46. Betzel, R. F. et al.: Optimally controlling the human connectome: the role of network topology.
Sci. Rep. (2016) doi: 10.1038/srep30770
47. Tang, E., Giusti, C., Baum, G. L., Gu, S., Pollock, E., Kahn, A. E., Roalf, D. R., Moore, T. M.,
Ruparel, K., Gur, R. C., Gur, R. E., Satterthwaite, T. D., Bassett, D. S. Developmental increases
in white matter network controllability support a growing diversity of brain dynamics. Nat
Commun. 2017 Nov 1;8(1):1252.
48. Cornblath, E. J., Tang, E., Baum, G. L., Moore, T. M., Adebimpe, A., Roalf, D. R., Gur,
R. C., Gur, R. E., Pasqualetti, F., Satterthwaite, T. D., Bassett, D. S. Sex differences in
network controllability as a predictor of executive function in youth. Neuroimage. 2018 Dec
1;188:122-134. doi: 10.1016/j.neuroimage.2018.11.048.
49. Gu, S., Betzel, R. F., Mattar, M. G., Cieslak, M., Delio, P. R., Grafton, S. T., Pasqualetti, F.,
Bassett, D. S. Optimal trajectories of brain state transitions. Neuroimage. 2017 Mar 1;148:305-
317. doi: 10.1016/j.neuroimage.2017.01.003.
50. Cui, Z., Stiso, J., Baum, G. L., Kim, J. Z., Roalf, D. R., Betzel, R. F., Gu, S., Lu, Z., Xia, C.
H., Ciric, R., Moore, T. M., Shinohara, R. T., Ruparel, K., Davatzikos, C., Pasqualetti, F., Gur,
R. E., Gur, R. C., Bassett, D. S., Satterthwaite, T. D. Optimization of Energy State Transition
Trajectory Supports the Development of Executive Function During Youth. bioRxiv 424929;
doi: https://doi.org/10.1101/424929
51. Moler, C., Loan, C. V.: Nineteen Dubious Ways to Compute the Exponential of a Matrix,
Twenty-Five Years Later. SIAM Rev. (2003) doi: 10.1137/S00361445024180
52. Pasqualetti, F., Zampiere, S., Bullo, F.: Controllability metrics,
for complex networks. 2014 American Control Conference.
limitations and al-
(2014) doi:
gorithms
10.1109/ACC.2014.6858621
53. Granger, C. W. J.: Investigating Causal Relations by Econometric Models and Cross-spectral
Methods. Econometrica. (1969) doi: 10.2307/1912791
54. Seth, A. K., Barrett, A. B., Barnett, L. Granger causality analysis in neuroscience and neu-
roimaging. J Neurosci. 2015 Feb 25;35(8):3293-7. doi: 10.1523/JNEUROSCI.4399-14.2015.
55. Barnett, L., Barrett, A. B., Seth, A. K. Misunderstandings regarding the application of Granger
causality in neuroscience. Proc Natl Acad Sci U S A. 2018 Jul 17;115(29):E6676-E6677. doi:
10.1073/pnas.1714497115
56. Friston, K. J. Functional and effective connectivity: a review. Brain Connect. 2011;1(1):13-36.
doi: 10.1089/brain.2011.0008.
Linear Dynamics & Control of Brain Networks
29
57. McIntosh, A. R. Tracing the route to path analysis in neuroimaging. Neuroimage. 2012 Aug
15;62(2):887-90. doi: 10.1016/j.neuroimage.2011.09.068.
58. Becker, C. O., Bassett, D. S., Preciado, V. M. Large-scale dynamic modeling of task-
fMRI signals via subspace system identification. J Neural Eng. 2018 Dec;15(6):066016. doi:
10.1088/1741-2552/aad8c7.
59. Nguyen, J. P. et al: Whole-brain calcium imaging with cellular resolution in freely behaving
Caenorhabditis elegans. Proc. Natl. Acad. Sci. (2016) doi: 10.1073/pnas.1507110112
60. Stosiek, C., Garaschuk, O., Holthoff, K., Konnerth, A.: In vivo two-photon calcium imaging
of neuronal networks. Proc. Natl. Acad. Sci. (2003) doi: 10.1073/pnas.1232232100
61. Motter, A. E. Networkcontrology. Chaos. 2015 Sep;25(9):097621. doi: 10.1063/1.4931570.
62. Koopman, B. O.: Hamiltonian systems and transformations in Hilbert space. Proc. Natl. Acad.
Sci. (1931) doi: 10.1073/pnas.17.5.315
63. Brunton, S. L., Brunton, B. W., Proctor, J. L., Kutz, J. N.: Koopman Invariant Subspaces
and Finite Linear Representations of Nonlinear Dynamical Systems for Control. PLoS One.
(2016) doi: 10.1371/journal.pone.0150171
64. Kailath, T. Linear Systems. Prentice-Hall, 1980.
65. Kim, J. Z., Soffer, J. M., Kahn, A. E., Vettel, J. M., Pasqualetti, F., Bassett, D. S. Role of Graph
Architecture in Controlling Dynamical Networks with Applications to Neural Systems. Nat
Phys. 2018;14:91-98. doi: 10.1038/nphys4268
66. Wu-Yan, E., and Betzel, R. F., Tang, E., Gu, S., Pasqualetti, F., Bassett, D. S. Benchmarking
Measures of Network Controllability on Canonical Graph Models. Journal of Nonlinear
Science. 2018. 1-39
67. Xiao, Y., Lao, S., Hou, L., Small, M., Bai, L. Effects of Edge Directions on the Struc-
tural Controllability of Complex Networks. PLoS One. 2015 Aug 17;10(8):e0135282. doi:
10.1371/journal.pone.0135282.
68. Tu, C., Rocha, R. P., Corbetta, M., Zampieri, S., Zorzi, M., Suweis, S. Warn-
ings and caveats in brain controllability. Neuroimage. 2018 Aug 1;176:83-91. doi:
10.1016/j.neuroimage.2018.04.010
69. Menara, T., Gu, S., Bassett, D. S., Pasqualetti, F. On Structural Controllability of Symmetric
(Brain) Networks. 2017. arXiv:1706.05120.
70. Pang, S. P., Wang, W. X., Hao, F., Lai, Y. C. Universal framework for edge controllability of
complex networks. Sci Rep. 2017 Jun 26;7(1):4224. doi: 10.1038/s41598-017-04463-5.
71. Whalen, A. J., Brennan, S. N., Sauer, T. D., and Schiff, S. J. Observability and Controllability
of Nonlinear Networks: The Role of Symmetry. Phys. Rev. X. 2015, 5, 011005.
72. Wiles, L., Gu, S., Pasqualetti, F., Parvesse, B., Gabrieli, D., Bassett, D. S., Meaney, D. F.
Autaptic Connections Shift Network Excitability and Bursting. Sci Rep. 2017 Mar 7;7:44006.
doi: 10.1038/srep44006.
73. Ju, H., Kim, J. Z., Bassett, D. S. Network topology of neural systems supporting avalanche
dynamics predicts stimulus propagation and recovery. 2018. arXiv:1812.09361.
74. Cornblath, E. J., Ashourvan, A., Kim, J. Z., Betzel, R. F., Ciric, R., Baum, G. L., He, X.,
Ruparel, K., Moore, T. M., Gur, R. C., Gur, R. E., Shinohara, R. T., Roalf, D. R., Satterthwaite,
T. D., Bassett, D. S. Context-dependent architecture of brain state dynamics is explained by
white matter connectivity and theories of network control. 2018. arXiv:1809.02849.
75. Muldoon, S. F., Costantini, J., Webber, W. R. S., Lesser, R., Bassett, D. S. Locally stable brain
states predict suppression of epileptic activity by enhanced cognitive effort. Neuroimage Clin.
2018 Feb 27;18:599-607. doi: 10.1016/j.nicl.2018.02.027.
76. Mengiste, S. A., Aertsen, A., Kumar, A. Effect of edge pruning on structural controllability
and observability of complex networks. Sci Rep. 2015 Dec 17;5:18145.
77. Jeganathan, J., Perry, A., Bassett, D. S., Roberts, G., Mitchell, P. B., Breakspear, M. Fronto-
limbic dysconnectivity leads to impaired brain network controllability in young people with
bipolar disorder and those at high genetic risk. Neuroimage Clin. 2018 Mar 27;19:71-81. doi:
10.1016/j.nicl.2018.03.032.
78. Bernhardt, B. C., Liu, M., Vos de Wael, R., Smallwood, J., Jefferies, E., Gu, S., Bassett, D. S.,
Bernasconi, A., Bernasconi, N. Hippocampal pathology modulates white matter connectome
topology and controllability in temporal lobe epilepsy. Neurology. 2018. In Press.
30
Jason Z. Kim and Danielle S. Bassett
79. Muldoon, S. F., Pasqualetti, F., Gu, S., Cieslak, M., Grafton, S. T., Vettel, J. M., Bassett, D.
S. Stimulation-Based Control of Dynamic Brain Networks. PLoS Comput Biol. 2016 Sep
9;12(9):e1005076. doi: 10.1371/journal.pcbi.1005076
80. Taylor, P. N., Thomas, J., Sinha, N., Dauwels, J., Kaiser, M., Thesen, T., Ruths, J. Optimal
control based seizure abatement using patient derived connectivity. Front Neurosci. 2015 Jun
3;9:202. doi: 10.3389/fnins.2015.00202
81. Yang, X., Ram, N., Gest, S. D., Lydon-Staley, D. M., Conroy, D. E., Pincus, A L., Mole-
naar, P. C. M. Socioemotional Dynamics of Emotion Regulation and Depressive Symp-
toms: A Person-Specific Network Approach. Complexity. 2018;2018. pii: 5094179. doi:
10.1155/2018/5094179.
82. Lydon-Staley, D. M., Barnett, I., Satterthwaite, T. D., Bassett, D. S. Digital phenotyping for
psychiatry: Accommodating data and theory with network science methodologies. Current
Opinions in Biomedical Engineering. 2019. In Press.
|
1803.00338 | 2 | 1803 | 2018-04-17T08:26:53 | Synthesizing realistic neural population activity patterns using Generative Adversarial Networks | [
"q-bio.NC",
"cs.NE"
] | The ability to synthesize realistic patterns of neural activity is crucial for studying neural information processing. Here we used the Generative Adversarial Networks (GANs) framework to simulate the concerted activity of a population of neurons. We adapted the Wasserstein-GAN variant to facilitate the generation of unconstrained neural population activity patterns while still benefiting from parameter sharing in the temporal domain. We demonstrate that our proposed GAN, which we termed Spike-GAN, generates spike trains that match accurately the first- and second-order statistics of datasets of tens of neurons and also approximates well their higher-order statistics. We applied Spike-GAN to a real dataset recorded from salamander retina and showed that it performs as well as state-of-the-art approaches based on the maximum entropy and the dichotomized Gaussian frameworks. Importantly, Spike-GAN does not require to specify a priori the statistics to be matched by the model, and so constitutes a more flexible method than these alternative approaches. Finally, we show how to exploit a trained Spike-GAN to construct 'importance maps' to detect the most relevant statistical structures present in a spike train. Spike-GAN provides a powerful, easy-to-use technique for generating realistic spiking neural activity and for describing the most relevant features of the large-scale neural population recordings studied in modern systems neuroscience. | q-bio.NC | q-bio |
Published as a conference paper at ICLR 2018
SYNTHESIZING REALISTIC NEURAL
POPULATION ACTIVITY PATTERNS USING
GENERATIVE ADVERSARIAL NETWORKS
Manuel Molano-Mazon1,+, Arno Onken1,2, Eugenio Piasini1,3,∗, Stefano Panzeri1,∗
1Laboratory of Neural Computation, Istituto Italiano di Tecnologia, 38068 Rovereto (TN), Italy
2University of Edinburgh, Edinburgh EH8 9AB, UK
3University of Pennsylvania, Philadelphia, PA 19104
+Corresponding author
∗Equal contribution
[email protected], [email protected], [email protected],
[email protected]
ABSTRACT
The ability to synthesize realistic patterns of neural activity is crucial for studying
neural information processing. Here we used the Generative Adversarial Net-
works (GANs) framework to simulate the concerted activity of a population of
neurons. We adapted the Wasserstein-GAN variant to facilitate the generation of
unconstrained neural population activity patterns while still benefiting from pa-
rameter sharing in the temporal domain. We demonstrate that our proposed GAN,
which we termed Spike-GAN, generates spike trains that match accurately the
first- and second-order statistics of datasets of tens of neurons and also approxi-
mates well their higher-order statistics. We applied Spike-GAN to a real dataset
recorded from salamander retina and showed that it performs as well as state-of-
the-art approaches based on the maximum entropy and the dichotomized Gaus-
sian frameworks. Importantly, Spike-GAN does not require to specify a priori the
statistics to be matched by the model, and so constitutes a more flexible method
than these alternative approaches. Finally, we show how to exploit a trained Spike-
GAN to construct 'importance maps' to detect the most relevant statistical struc-
tures present in a spike train. Spike-GAN provides a powerful, easy-to-use tech-
nique for generating realistic spiking neural activity and for describing the most
relevant features of the large-scale neural population recordings studied in modern
systems neuroscience.
1
INTRODUCTION
Understanding how to generate synthetic spike trains simulating the activity of a population of neu-
rons is crucial for systems neuroscience. In computational neuroscience, important uses of faithfully
generated spike trains include creating biologically consistent inputs needed for the simulation of
realistic neural networks, generating large datasets to be used for the development and validation
of new spike train analysis techniques, and estimating the probabilities of neural responses in order
to extrapolate the information coding capacity of neurons beyond what can be computed from the
neural data obtained experimentally (Ince et al., 2013; Moreno-Bote et al., 2014). In experimental
systems neuroscience, the ability to develop models that produce realistic neural population pat-
terns and that identify the key sets of features in these patterns is fundamental to disentangling the
encoding strategies used by neurons for sensation or behavior (Panzeri et al., 2017) and to design
closed-loop experiments (Kim et al., 2017) in which synthetic patterns, representing salient features
of neural information, are fed to systems of electrical micro-stimulation (Tehovnik et al., 2006) or
patterned light optogenetics (Panzeri et al., 2017; Bovetti & Fellin, 2015) for naturalistic intervention
on neural circuits.
One successful way to generate realistic spike trains is that of using a bottom-up approach, focusing
explicitly on replicating selected low-level aspects of spike trains statistics. Popular methods include
1
Published as a conference paper at ICLR 2018
renewal processes (Stein (1965); Gerstner & Kistler (2002)), latent variable models (Macke et al.,
2009; Lyamzin et al., 2010) and maximum entropy approaches (Tang et al., 2008; Schneidman et al.,
2006; Savin & Tkacik, 2017), which typically model the spiking activity under the assumption that
only first and second-order correlations play a relevant role in neural coding (but see Cayco-Gajic
et al. (2015); Koster et al. (2014); Ohiorhenuan et al. (2010)). Other methods model spike train
responses assuming linear stimulus selectivity and generating single trial spike trains using simple
models of input-output neural nonlinearities and neural noise (Keat et al., 2001; Pillow et al., 2008;
Lawhern et al., 2010). These methods have had a considerable success in modeling the activity
of populations of neurons in response to sensory stimuli (Pillow et al., 2008). Nevertheless, these
models are not completely general and may fail to faithfully represent spike trains in many situations.
This is because neural variability changes wildly across different cortical areas (Maimon & Assad,
2006) due to the fact that responses, especially in higher-order areas and in behaving animals, have
complex non-linear tuning to many parameters and are affected by many behavioral variables (e.g.
the level of attention (Fries et al., 2001)).
An alternative approach is to apply deep-learning methods to model neural activity in response to
a given set of stimuli using supervised learning techniques (McIntosh et al., 2016). The potential
advantage of this type of approach is that it does not require to explicitly specify any aspect of the
spike train statistics. However, applications of deep networks to generate faithful spike patterns
have been rare. Here, we explore the applicability of the Generative Adversarial Networks (GANs)
framework (Goodfellow et al., 2014) to this problem. Three aspects of GANs make this technique
a good candidate to model neural activity. First, GANs are an unsupervised learning technique
and therefore do not need labeled data (although they can make use of labels (Odena et al., 2016b;
Chen et al., 2016)). This greatly increases the amount of neural data available to train them. Sec-
ond, recently proposed modifications of the original GANs make them good at fitting distributions
presenting multiple modes (Arjovsky et al., 2017; Gulrajani et al., 2017). This is an aspect that is
crucial for neural data because the presentation of even a single stimulus can elicit very different
spatio-temporal patterns of population activity (Churchland et al., 2007; Morcos & Harvey, 2016).
We thus need a method that generates sharp realistic samples instead of producing samples that are
a compromise between two modes (which is typical, for instance, of methods seeking to minimize
the mean squared error between the desired output and the model's prediction (Goodfellow, 2016;
Lotter et al., 2016)). Finally, using as their main building block deep neural networks, GANs in-
herit the capacity of scaling up to large amounts of data and therefore constitute a good candidate
to model the ever growing datasets provided by experimental methods like chronic multi-electrode
and optical recording techniques.
In the present work we extend the GAN framework to synthesize realistic neural activity. We adapt
the recently proposed Wasserstein-GAN (WGAN) (Arjovsky et al., 2017) which has been proven
to stabilize training, by modifying the network architecture to model invariance in the temporal
dimension while keeping dense connectivity across the modeled neurons. We show that the proposed
GAN, which we called Spike-GAN, is able to produce highly realistic spike trains matching the first
and second-order statistics of a population of neurons. We further demonstrate the applicability
of Spike-GAN by applying it to a real dataset recorded from the salamander retina (Marre et al.,
2014) and comparing the activity patterns the model generates to those obtained with a maximum
entropy model (Tkacik et al., 2014) and with a dichotomized Gaussian method (Lyamzin et al.,
2010). Finally, we describe a new procedure to detect, in a given activity pattern, those spikes
participating in a specific feature characteristic of the probability distribution underlying the training
dataset.
2 METHODS
2.1 NETWORK ARCHITECTURE
We adapted the Generative Adversarial Networks described by Goodfellow et al. (2014) to produce
samples that simulate the spiking activity of a population of N neurons as binary vectors of length
T (spike trains; Fig. S3). In their original form, GANs proved to be difficult to train, prompting
several subsequent studies that focused on making them more stable (Radford et al., 2015; Chin-
tala et al., 2016). In the present work we used the Wasserstein-GAN variant described by Arjovsky
et al. (2017). Wasserstein-GANs (WGAN) minimize the Earth-Mover (or Wasserstein-1) distance
2
Published as a conference paper at ICLR 2018
(EM) between the original distribution Pdata and the distribution defined by the generator, PG. Ar-
jovsky et al. (2017) showed that the EM distance has desirable properties in terms of continuity
and differentiability that ensure that the loss function provides a meaningful gradient at all stages
of training, which boosts considerably its stability. A further improvement was later introduced by
Gulrajani et al. (2017), who provided an alternative procedure to ensure that the critic is Lipschitz
(via gradient penalization), which is required in the WGAN framework.
Here we adapted the WGAN-GP architecture (Gulrajani et al., 2017) to simulate realistic neural pop-
ulation activity patterns. Our samples are matrices of size N × T , where N is the number of neurons
and T the number of time bins, each bin usually corresponding to a few milliseconds (Fig. S3).
Importantly, while samples present a high degree of invariance along the time dimension, they are
usually not spatially structured (i.e. across neurons) and thus we cannot expect any invariance along
the dimension spanning the different neurons. For this reason, in order to take advantage of the
temporal invariance while being maximally agnostic about the neural correlation structure underly-
ing the population activity, we modified a standard 1D-DCGAN (1 dimensional deep convolutional
GAN) architecture (Radford et al., 2015) by transposing the samples so as to make the spatial di-
mension correspond to the channel dimension (Fig. 1). Therefore our proposed GAN can be seen as
performing a semi-convolution, where the spatial dimension is densely connected while weights are
shared across the temporal dimension thus improving training, efficiency and the interpretability of
the trained networks.
The main modifications we have introduced to the WGAN-GP are:
1. The responses of different neurons are fed into different channels.
2. Following Chintala et al. (2016) we made all units LeakyReLU (the slope of the leak was
set to 0.2) except for the last layer of the generator where we used sigmoid units.
3. The critic consists of two 1D convolutional layers with 256 and 512 features, respectively,
followed by a linear layer (Fig. 1). The generator samples from a 128-dimension uniform
distribution and its architecture is the mirror image of that of the critic.
4. To avoid the checkerboard issue described by Odena et al. (2016a), we divided all genera-
tor's fractional-strided convolutions (i.e. deconvolutions) into two separate steps: upsam-
pling and convolving. The upsampling step is done using a nearest neighbor procedure, as
suggested by Odena et al. (2016a).
We called the network described above Spike-GAN. As in Arjovsky et al. (2017), Spike-GAN
was trained with mini-batch stochastic gradient descent (we used a mini-batch size of 64). All
weights were initialized from a zero-centered normal distribution with standard deviation 0.02. We
used the Adam optimizer (Kingma & Ba, 2014) with learning rate = 0.0001 and hyperparameters
β1 = 0 and β2 = 0.9. The parameter λ, used for gradient penalization, was set to 10. The critic
was updated 5 times for each generator update. All code and hyperparameters may be found at
https://github.com/manuelmolano/Spike-GAN.
2.2 SPIKE TRAIN ANALYSIS
To compare the statistics of the generated samples to the ones contained in the ground truth dataset,
we first discretized the continuously-valued samples produced by the generator and then, for each
bin with activation h, we drew the final value from a Bernoulli distribution with probability h.
Note that the last layer of the generator contains a sigmoid function and thus the h values can be
interpreted as probabilities.
We assessed the performance of the model by measuring several spike train statistics commonly used
in neuroscience: 1) Average number of spikes (spike-count) per neuron. 2) Average time course,
which corresponds to the probability of firing in each bin, divided by the bin duration (measured
in seconds). 3) Covariance between pairs of neurons. 4) Lag-covariance between pairs of neurons:
for each pair of neurons, we shift the activity of one of the neurons by one bin and compute the
covariance between the resulting activities. This quantity thus indicates how strongly the activity of
one of the neurons is related to the future activity of the other neuron. 5) Distribution of synchrony
(or k-statistic), which corresponds to the probability PN (k) that k out of the N neurons spike at
the same time. 6) Spike autocorrelogram, computed by counting, for each spike, the number of
3
Published as a conference paper at ICLR 2018
Figure 1: Critic's architecture. Samples are transposed so as to input the neurons' activities into
different channels. The convolutional filters (red box) span all neurons but share weights across
the time dimension. The critic consists of two 1D convolutional layers with 256 and 512 features.
Stride=2; all units are LeakyReLU (slope=0.2). The architecture of the generator is the same as that
of the critic, used in the opposite direction.
spikes preceding and following the given spike in a predefined time window. The obtained trace is
normalized to the peak (which is by construction at 0 ms) and the peak is then zeroed in order to
help comparisons.
3 RESULTS
3.1 FITTING THE STATISTICS OF SIMULATED SPIKE TRAINS
We first tested Spike-GAN with samples coming from the simulated activity of a population of
16 neurons whose firing probability followed a uniform distribution across the whole duration
(T=128 ms) of the samples (bin size=1 ms, average firing rate around 100 Hz, Fig. 2D). In order
to test whether Spike-GAN can approximate second-order statistics, the neurons' activities present
two extra features that are commonly found in neural recordings. First, using the method described
in Mikula & Niebur (2003), we introduced correlations between randomly selected pairs of neurons
(8 pairs of correlated neurons; correlation coefficient values around 0.3). Second, we imposed a
common form of temporal correlations arising from neuronal biophysics (refractory period): fol-
lowing an action potential, a neuron typically remains silent for a few milliseconds before it is able
to spike again. This phenomenon has a clear effect on the spike autocorrelogram that shows a pro-
nounced drop in the number of spikes present at less than 2 ms (see Fig. 2E). We trained Spike-GAN
on 8192 samples for 500000 iterations (Fig. S4 shows the critic's loss function across training).
A representative sample produced by a trained Spike-GAN together with the resulting patterns (after
binarizing the samples, see Section 2.2) is shown in Fig. 2A. Note that the sample (black traces) is
mostly binary, with only a small fraction of bins having intermediate values between 0 and 1. We
evaluated the performance of Spike-GAN by measuring several spike train statistics commonly used
in neuroscience (see Section 2.2). For comparison, we also trained a generative adversarial network
in which both the generator and the critic are a 4-layer multi-layer perceptron (MLP) and the number
of units per layer is adjusted so both models present comparable numbers of trainable variables (490
units per layer which results in ≈ 3.5M trainable variables). As Fig. 2 shows, while both models
fit fairly well the first three statistics (mean spike-count, covariances and k-statistics), the Spike-
GAN's approximation of the features involving time (average time course, autocorrelogram and
lag-covariance) is considerably better than that of the MLP GAN. This is most likely due to the
weight sharing performed by Spike-GAN along the temporal dimension, that allows it to easily
learn temporally invariant features.
In Supp. Section A.1 we further show that Spike-GAN is not only memorizing the samples present
in the training dataset but it is able to effectively mimic their underlying distribution.
4
Critic's Architecture neurons (N)time bins (T)T/2T/4 features (256) features (512)filterNx5Convolution (stride 2)+Leaky ReLUFlattenLinear projectionConvolution (stride 2)+Leaky ReLUCritic's output Published as a conference paper at ICLR 2018
Figure 2: Fitting the statistics of simulated population activity patterns. A) Representative sample
generated by Spike-GAN (black lines) and the resulting spike trains after binarizing (red lines). B-
D) Fitting of the average spike-count, pairwise covariances and k-statistics done by Spike-GAN (red
dots) and by a MLP GAN (green dots). Line indicates identity. E) Average time courses corre-
sponding to the ground truth dataset and to the data obtained with Spike-GAN and the MLP GAN.
F-G) Fitting of the autocorrelogram and the lag-covariances done by Spike-GAN (red line/dots) and
a MLP GAN (green line/dots). Black line corresponds to the autocorrelogram resulting from the
ground truth distribution.
5
Published as a conference paper at ICLR 2018
3.2 COMPARING TO STATE-OF-THE-ART METHODS
We next tested the Spike-GAN model on real recordings coming from the retinal ganglion cells
(RGCs) of the salamander retina (Marre et al., 2014; Tkacik et al., 2014). The dataset contains the
response of 160 RGCs to natural stimuli (297 repetitions of a 19-second movie clip of swimming
fish and water plants in a fish tank) discretized into bins of 20 ms. We randomly selected 50 neurons
out of the total 160 and partitioned their activity into non-overlapping samples of 640 ms (32 time
bins) which yielded a total of 8817 training samples (using overlapping samples and thus increasing
their number did not improve the results shown below). We obtained almost identical results for a
different set of 50 randomly selected neurons (data not shown).
In order to provide a comparison between Spike-GAN and existing state-of-the-art methods, we fit
the same dataset with a maximum entropy approach developed by Tkacik et al. (2014), the so-called
k-pairwise model, and a dichotomized Gaussian method proposed by Lyamzin et al. (2010). Briefly,
maximum entropy (MaxEnt) models provide a way of fitting a predefined set of statistics charac-
terizing a probability distribution while being maximally agnostic about any other aspect of such
distribution, i.e. maximizing the entropy of the probability distribution given the constraints in the
statistics (Press´e et al., 2013). In neuroscience applications, the most common approach has been
to design MaxEnt models fitting the first and second-order statistics, i.e. the average firing rates and
pairwise correlations between neurons (Tang et al., 2008; Schneidman et al., 2006; Shlens et al.,
2006). The k-pairwise model extends this approach to further constrain the activity of the neural
population by fitting the k-statistics of the dataset of interest, which provides a measure of the neu-
ral population synchrony (see Section 2.2). Dichotomized Gaussian (DG) methods, on the other
hand, model the neural activity by thresholding a correlated multivariate normal distribution with
mean and covariance chosen such that the generated samples have the desired first- and second-order
statistics. The method developed by Lyamzin et al. (2010) is an extension of previous approaches
(see e.g. Macke et al. (2009)) in which signal and noise correlations are modeled separately. Im-
portantly, unlike the k-pairwise model (and most MaxEnt models (Savin & Tkacik, 2017)), the DG
model can fit temporal correlations.
We first checked for signs of overfitting by plotting, for each method, randomly selected gener-
ated samples together with their closest sample (in terms of L1 distance) in the training dataset.
Although the generator in a GAN never 'sees' the training dataset directly but instead obtains infor-
mation about the dataset only through the critic, it is still possible that the generator obtains enough
information about the real samples to memorize them. Fig. S5 shows that this is not the case, with
the closest samples in the training dataset being very different from the generated ones.
As shown in Fig. 3, all methods provide a good approximation of the average firing rate, the covari-
ance and the k-statistics, but the fit performed by the MaxEnt (green dots) and the DG (blue pluses)
models is somewhat tighter than that produced by Spike-GAN (red dots). This is not surprising,
as these are the aspects of the population activity distribution that these models are specifically de-
signed to fit. By contrast, Spike-GAN does remarkably well without any need for these statistical
structures to be manually specified as features of the model.
As mentioned above, the k-pairwise model does not take into account the temporal dynamics of
the population and therefore ignores well-known neural features that are very likely to play a rele-
vant role in the processing of incoming information (e.g. refractory period, burst or lagged cross-
correlation between pairs of neurons). Fig. 3 shows that both Spike-GAN and the DG model ap-
proximate well the ground truth autocorrelogram and lag-covariances while the k-pairwise model,
as expected, entirely fails to do so.
Importantly, while its performance in terms of reproducing positive correlations is remarkable, the
DG method struggles to approximate the statistics of neural activity associated with negative corre-
lations (Lyamzin et al., 2010). Fig. S6 shows how the k-pairwise and the DG methods fit the dataset
described in Fig. 2. As can be seen, the DG model, while matching perfectly the (positive) correla-
tions between neurons, fails to approximate the negative correlations present in the autocorrelogram
that are caused by the refractory period (Fig. S6E).
The above results demonstrate that Spike-GAN generates samples comparable to those produced by
state-of-the-art methods without the need of defining a priori which statistical structures constitute
important features of the probability distribution underlying the modeled dataset.
6
Published as a conference paper at ICLR 2018
Figure 3: Fitting the statistics of real population activity patterns obtained in the retinal salamander.
A-C) Fitting of the average spike-count, pairwise covariances and k-statistics done by Spike-GAN
(red dots), the k-pairwise model (green crosses) and the DG model (blue pluses). Line indicates
identity. D) Average time courses corresponding to the ground truth data and to the data obtained
with Spike-GAN, the k-pairwise model and the DG model. E-F) Fitting of the autocorrelogram and
the lag-covariances done by Spike-GAN (red line/dots), the k-pairwise model (green line/crosses)
and the DG model (blue dashed line/pluses). Black line corresponds to the autocorrelogram resulting
from the ground truth distribution.
7
Published as a conference paper at ICLR 2018
3.3 USING THE TRAINED CRITIC TO INFER RELEVANT NEURAL FEATURES
We then investigated what a trained critic can tell us about the population activity patterns that
compose the original dataset. In order to do so, we designed an alternative dataset in which neural
samples contain stereotyped activation patterns each involving a small set of neurons (Fig. 4A). This
type of activation patterns, also called packets, have been found in different brain areas and have been
suggested to be fundamental for cortical coding, forming the basic symbols used by populations of
neurons to process and communicate information about incoming stimuli (Luczak et al., 2015).
Thus, besides being a good test for the capability of Spike-GAN to approximate more intricate
statistical structures, analyzing simulated samples presenting packets constitutes an excellent way of
demonstrating the applicability of the model to a highly relevant topic in neuroscience. We trained
Spike-GAN on a dataset composed of neural patterns of 32 neurons by 64 ms that present four
different packets involving non-overlapping sets of 8 neurons each (Fig. 4A). Importantly, only few
neurons out of all the recorded ones typically participate in a given packet and, moreover, neurons
are usually not sorted by the packet to which they belong. Therefore, real neural population activity
is extremely difficult to interpret and packets are cluttered by many other 'noisy' spikes (Fig. 4B).
In order to assess the applicability of Spike-GAN to real neuroscience experiments, we trained it on
these type of realistic patterns of activity (see caption of Fig. 4 for more details on the simulated
dataset and the training).
Visual inspection of the filters learned by the first layer of the critic suggests that Spike-GAN is able
to learn the particular structure of the packets described above: many of the filters display spatial
distributions that are ideally suited for packet detection (Fig. S7; note that filters have been sorted in
the neurons' dimension to help visualization).
Recently, Zeiler & Fergus (2014) developed a procedure to investigate which aspects of a given
sample are most relevant for a neural network. They proposed to systematically alter different parts
of the input and evaluate the change each alteration produces in the output of different layers of the
network. Here we have adapted this idea to investigate which are the most relevant features of a
given neural activity pattern. We first compute the output produced by the critic for a real sample.
Then, for a given neuron and a given temporal window of several milliseconds, we shuffle across
time the spikes emitted by the neuron during that period of time and compute the output of the critic
when using as input the altered sample. The absolute difference between the two outputs gives us an
idea of how important is the structure of the spike train we have disrupted. We can then proceed in
the same fashion for all neurons and for several time windows and obtain a map of the importance
of each particular spike train emitted by each neuron (importance maps, see Fig. 4C, heatmaps).
To highlight the usefulness of the procedure explained above, we produced a separate dataset in
which the same population of neurons encodes the information about a particular stimulus by emit-
ting one of the packet types shown in Fig. 4A around 16 ms after the stimulus presentation 1. Fig. 4C
(gray scale panels) shows 5 representative example patterns (see also Fig. S8). The packets are high-
lighted for visualization, but it is clear that patterns containing packets are almost indistinguishable
from those without them. Noticeably, the importance maps (heatmaps) are able to pinpoint the
spikes belonging to a packet (note that this does not require re-training of Spike-GAN). Further, by
averaging the importance maps across time and space, we can obtain unambiguous results regarding
the relevance of each neuron and time period (Fig. 4D-E; in Fig. 4E the neurons presenting higher
importance values are those participating in the packet).
The importance-map analysis thus constitutes a very useful procedure to detect the most relevant
aspects of a given neural population activity pattern. In Fig. S2 we describe a potential application
of the importance maps to the study of how a population of neurons encode the information about a
given set of stimuli.
4 DISCUSSION
We explored the application of the Generative Adversarial Networks framework (Goodfellow et al.,
2014) to synthesize neural responses that approximate the statistics of the activity patterns of a
1It has been shown that, in the sensory cortex, activity packets in response to external stimuli are very similar
to those recorded when no stimulation is applied (Luczak et al., 2015).
8
Published as a conference paper at ICLR 2018
Figure 4: A) An example pattern showing the different packets highlighted with different colors and
sorted to help visualization. The probability of each type of packet to occur was set to 0.1. Packets
of the same type do not overlap in time. B) Realistic neural population pattern (gray spikes do not
participate in any packet). C) Examples of activity patterns (grayscale panels) in which only one type
of packet is usually present (one or two times) during a period of time from 16 to 32 ms. Packets
are highlighted as white spikes. Heatmaps: importance maps showing the change that disrupting
specific spikes has on the critic's output. Note that packet spikes normally show higher values.
We used a sliding window of 8 ms (with a step size of 2 ms) to selectively shuffle the activity of
each neuron at different time periods. The Spike-GAN used to obtain these importance maps was
trained for 50000 iterations on 8192 samples. D) Average of 200 randomly selected importance
maps across the neurons dimension, yielding importance as a function of time. E) Average of the
same 200 randomly selected importance maps across the time dimension, yielding importance as a
function of neurons. Errorbars correspond to standard error.
population of neurons. For this purpose, we put forward Spike-GAN, by adapting the WGAN
variant proposed by Arjovsky et al. (2017) to allow sharing weights across time while maintaining
a densely connected structure across neurons. We found that our method reproduced to an excellent
approximation the spatio-temporal statistics of neural activity on which it was trained. Importantly,
it does so without the need for these statistics to be handcrafted in advance, which avoids making a
priori assumptions about which features of the external world make neurons fire.
Recently, Pandarinath et al. (2017) have proposed a deep learning method, LFADS (Latent Factor
Analysis via Dynamical Systems), to model the activity of a population of neurons using a varia-
tional autoencoder (in which the encoder and decoder are recurrent neural networks). LFADS allows
inferring the trial-by-trial population dynamics underlying the modeled spike train patterns and thus
9
Published as a conference paper at ICLR 2018
can be seen as a complementary method to Spike-GAN, which does not explicitly provide the latent
factors governing the response of the neurons. Regarding the application of the GANs framework to
the field of neuroscience, Arakaki et al. (2017) proposed a GAN-based approach for fitting network
models to experimental data consisting of a set of tuning curves extracted from a population of neu-
rons. However, to the best of our knowledge our work is the first to use GANs to directly produce
realistic neural patterns simulating the activity of populations of tenths of neurons.
Building on the work by Zeiler & Fergus (2014), we showed how to use Spike-GAN to visualize
the particular features that characterize the training dataset. Specifically, Spike-GAN can be used to
obtain importance maps that highlight the spikes that participate in generating activity motifs that
are most salient in the spike trains. This can be useful for unsupervised identification of highly
salient low-dimensional representations of neural activity, which can then be used to describe and
interpret experimental results and discover the key units of neural information used for functions
such as sensation and behavior.
A further and promising application of importance maps is that of designing realistic patterns of
stimulation that can be used to perturb populations of neurons using electrical or optical neural stim-
ulation techniques (Panzeri et al., 2017; Tehovnik et al., 2006; Emiliani et al., 2015). The ability of
Spike-GAN to generate realistic neural activity including its temporal dynamics and to identify its
most salient features suggests that it may become a very relevant tool to design perturbations. In
Fig. S2 we provide a more detailed description of a potential application of Spike-GAN, in which
importance maps may allow inferring the set of neurons participating in the encoding of the informa-
tion about a given set of stimuli (Fig. S2F) and the spatio-temporal structure of the packets elicited
by each stimulus (Fig. S2E).
We have compared Spike-GAN with two alternative methods based on the maximum entropy and the
dichotomized Gaussian frameworks. These methods offer the possibility of computing the sample
probabilities (MaxEnt model) and separately specifying the signal and noise correlations present in
the generated samples (DG model). Spike-GAN does not have these features; nevertheless, it does
have important advantages over the mentioned methods. First, Spike-GAN is more flexible than
the MaxEnt and DG models, being able to fit any type of spatio-temporal structure present in the
data. Further, it does not require making a priori assumptions about which statistical properties of
a dataset are relevant and thus need to be matched. Finally, Spike-GAN is based on the deep neural
network framework, and is therefore able to directly benefit from the engineering advances emerging
in this rapidly-growing field. Conceivably, this will enable Spike-GAN, or methods derived from it,
to make in the future better and better use of the datasets of ever increasing size that are produced
by the experimental neuroscience community.
ACKNOWLEDGMENTS
This work has received funding from the European Union's Horizon 2020 research and innovation
programme under the Marie Sklodowska-Curie grant agreement No 699829 (ETIC) and under the
Marie Sklodowska-Curie grant agreement No 659227 (STOMMAC). EP thanks Vittal Premachan-
dran for discussions at the Brains, Minds and Machines summer course.
REFERENCES
Alireza Alemi-Neissi, Federica Bianca Rosselli, and Davide Zoccolan. Multifeatural shape pro-
cessing in rats engaged in invariant visual object recognition. Journal of Neuroscience, 33(14):
5939–5956, 2013.
Takafumi Arakaki, Gregory Barello, and Yashar Ahmadian. Capturing the diversity of biological
tuning curves using generative adversarial networks. arXiv preprint arXiv:1707.04582, 2017.
Martin Arjovsky, Soumith Chintala, and L´eon Bottou. Wasserstein GAN.
arXiv:1701.07875, 2017.
arXiv preprint
Serena Bovetti and Tommaso Fellin. Optical dissection of brain circuits with patterned illumination
through the phase modulation of light. Journal of Neuroscience Methods, 241:66–77, 2015.
Natasha A. Cayco-Gajic, Joel Zylberberg, and Eric Shea-Brown. Triplet correlations among
similarly tuned cells impact population coding. Frontiers in Computational Neuroscience, 9:
10
Published as a conference paper at ICLR 2018
57, 2015.
frontiersin.org/article/10.3389/fncom.2015.00057.
ISSN 1662-5188. doi: 10.3389/fncom.2015.00057. URL http://journal.
Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel.
Info-
gan: Interpretable representation learning by information maximizing generative adversarial nets.
CoRR, abs/1606.03657, 2016. URL http://arxiv.org/abs/1606.03657.
Soumith Chintala, Emily Denton, Martin Arjovsky, and Michael Mathieu. How to train a GAN?
Tips and tricks to make GANs work. GitHub, 2016.
Mark M Churchland, M Yu Byron, Maneesh Sahani, and Krishna V Shenoy. Techniques for extract-
ing single-trial activity patterns from large-scale neural recordings. Current Opinion in Neurobi-
ology, 17(5):609–618, 2007.
Valentina Emiliani, Adam E Cohen, Karl Deisseroth, and Michael Hausser. All-optical interrogation
of neural circuits. Journal of Neuroscience, 35(41):13917–13926, 2015.
Pascal Fries, John H Reynolds, Alan E Rorie, and Robert Desimone. Modulation of oscillatory
neuronal synchronization by selective visual attention. Science, 291(5508):1560–1563, 2001.
Wulfram Gerstner and Werner M Kistler. Spiking neuron models: Single neurons, populations, plas-
ticity. Cambridge University Press, 2002. URL http://icwww.epfl.ch/gerstner/
BUCH.html.
Ian Goodfellow.
NIPS 2016 tutorial: Generative adversarial networks.
arXiv:1701.00160, 2016.
arXiv preprint
Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair,
Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In Advances in Neural Infor-
mation Processing Systems, pp. 2672–2680, 2014.
Ishaan Gulrajani, Faruk Ahmed, Martin Arjovsky, Vincent Dumoulin, and Aaron Courville. Im-
proved training of Wasserstein GANs. arXiv preprint arXiv:1704.00028, 2017.
Robin AA Ince, Stefano Panzeri, and Christoph Kayser. Neural codes formed by small and tem-
porally precise populations in auditory cortex. Journal of Neuroscience, 33(46):18277–18287,
2013.
Justin Keat, Pamela Reinagel, R Clay Reid, and Markus Meister. Predicting every spike: a model
for the responses of visual neurons. Neuron, 30(3):803–817, 2001.
Christina K Kim, Avishek Adhikari, and Karl Deisseroth. Integration of optogenetics with comple-
mentary methodologies in systems neuroscience. Nature Reviews Neuroscience, 18(4):222–235,
2017.
Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
Urs Koster, Jascha Sohl-Dickstein, Charles M Gray, and Bruno A Olshausen. Modeling higher-
order correlations within cortical microcolumns. PLoS Computational Biology, 10(7):e1003684,
2014.
Vernon Lawhern, Wei Wu, Nicholas Hatsopoulos, and Liam Paninski. Population decoding of motor
cortical activity using a generalized linear model with hidden states. Journal of Neuroscience
Methods, 189(2):267–280, 2010.
William Lotter, Gabriel Kreiman, and David Cox. Unsupervised Learning of Visual Structure us-
ing Predictive Generative Networks. ICLR, 2016. URL http://arxiv.org/abs/1511.
06380.
Artur Luczak, Bruce L McNaughton, and Kenneth D Harris. Packet-based communication in the
cortex. Nature Reviews Neuroscience, 16(12):745–755, 2015.
11
Published as a conference paper at ICLR 2018
Dmitry R Lyamzin, Jakob H Macke, and Nicholas A Lesica. Modeling population spike trains
with specified time-varying spike rates, trial-to-trial variability, and pairwise signal and noise
correlations. Frontiers in computational neuroscience, 4, 2010.
Jakob H Macke, Philipp Berens, Alexander S Ecker, Andreas S Tolias, and Matthias Bethge. Gen-
erating spike trains with specified correlation coefficients. Neural Computation, 21(2):397–423,
2009.
Gaby Maimon and John A Assad. A cognitive signal for the proactive timing of action in macaque
lip. Nature Neuroscience, 9(7):948–955, 2006.
Olivier Marre, Gasper Tkacik, Dario Amodei, Elad Schneidman, William Bialek, and II Berry,
Michael J. Multi-electrode array recording from salamander retinal ganglion cells. IST Austria,
2014. doi: 10.15479/AT:ISTA:61. URL https://datarep.app.ist.ac.at/61/.
Lane McIntosh, Niru Maheswaranathan, Aran Nayebi, Surya Ganguli, and Stephen Baccus. Deep
learning models of the retinal response to natural scenes.
In D. D. Lee, M. Sugiyama, U. V.
Luxburg, I. Guyon, and R. Garnett (eds.), Advances in Neural Information Processing Systems 29,
pp. 1369–1377. Curran Associates, Inc., 2016. URL http://papers.nips.cc/paper/
6388-deep-learning-models-of-the-retinal-response-to-natural-scenes.
pdf.
Shawn Mikula and Ernst Niebur. The effects of input rate and synchrony on a coincidence detector:
analytical solution. Neural Computation, 15(3):539–547, 2003.
Ari S Morcos and Christopher D Harvey. History-dependent variability in population dynamics
during evidence accumulation in cortex. Nature Neuroscience, 19(12):1672–1681, 2016.
Rub´en Moreno-Bote, Jeffrey Beck, Ingmar Kanitscheider, Xaq Pitkow, Peter Latham, and Alexandre
Pouget. Information-limiting correlations. Nature Neuroscience, 17(10):1410–1417, 2014.
Augustus Odena, Vincent Dumoulin, and Chris Olah. Deconvolution and checkerboard artifacts.
Distill, 1(10):e3, 2016a.
Augustus Odena, Christopher Olah, and Jonathon Shlens. Conditional image synthesis with auxil-
iary classifier GANs. arXiv preprint arXiv:1610.09585, 2016b.
Ifije E Ohiorhenuan, Ferenc Mechler, Keith P Purpura, Anita M Schmid, Qin Hu, and Jonathan D
Victor. Sparse coding and high-order correlations in fine-scale cortical networks. Nature, 466
(7306):617–621, 2010.
Chethan Pandarinath, Daniel J O'Shea, Jasmine Collins, Rafal Jozefowicz, Sergey D Stavisky,
Jonathan C Kao, Eric M Trautmann, Matthew T Kaufman, Stephen I Ryu, Leigh R Hochberg,
et al. Inferring single-trial neural population dynamics using sequential auto-encoders. bioRxiv,
pp. 152884, 2017.
Stefano Panzeri, Christopher D Harvey, Eugenio Piasini, Peter E Latham, and Tommaso Fellin.
Cracking the neural code for sensory perception by combining statistics, intervention, and behav-
ior. Neuron, 93(3):491–507, 2017.
Jonathan W Pillow,
Jonathon Shlens, Liam Paninski, Alexander Sher, Alan M Litke,
EJ Chichilnisky, and Eero P Simoncelli. Spatio-temporal correlations and visual signalling in
a complete neuronal population. Nature, 454(7207):995–999, 2008.
Steve Press´e, Kingshuk Ghosh, Julian Lee, and Ken A. Dill. Principles of maximum entropy and
maximum caliber in statistical physics. Rev. Mod. Phys., 85:1115–1141, Jul 2013. doi: 10.1103/
RevModPhys.85.1115. URL https://link.aps.org/doi/10.1103/RevModPhys.
85.1115.
Alec Radford, Luke Metz, and Soumith Chintala. Unsupervised representation learning with deep
convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
Caroline A Runyan, Eugenio Piasini, Stefano Panzeri, and Christopher D Harvey. Distinct
timescales of population coding across cortex. Nature, 548(7665):92–96, 2017.
12
Published as a conference paper at ICLR 2018
Cristina Savin and Gasper Tkacik. Maximum entropy models as a tool for building precise neural
controls. Current Opinion in Neurobiology, 46:120–126, 2017.
Elad Schneidman, Michael J Berry, Ronen Segev, and William Bialek. Weak pairwise correlations
imply strongly correlated network states in a neural population. Nature, 440(7087):1007–1012,
2006.
Jonathon Shlens, Greg D. Field, Jeffrey L. Gauthier, Matthew I. Grivich, Dumitru Petrusca, Alexan-
der Sher, Alan M. Litke, and E. J. Chichilnisky. The structure of multi-neuron firing patterns in
primate retina. Journal of Neuroscience, 26(32):8254–8266, 2006.
Richard B. Stein. A theoretical analysis of neuronal variability. Biophysical Journal, 5(2):173 –
ISSN 0006-3495. doi: 10.1016/S0006-3495(65)86709-1. URL http://www.
194, 1965.
sciencedirect.com/science/article/pii/S0006349565867091.
Aonan Tang, David Jackson, Jon Hobbs, Wei Chen, Jodi L Smith, Hema Patel, Anita Prieto, Dumitru
Petrusca, Matthew I Grivich, Alexander Sher, et al. A maximum entropy model applied to spatial
and temporal correlations from cortical networks in vitro. Journal of Neuroscience, 28(2):505–
518, 2008.
EJ Tehovnik, AS Tolias, F Sultan, WM Slocum, and NK Logothetis. Direct and indirect activation
of cortical neurons by electrical microstimulation. Journal of Neurophysiology, 96(2):512–521,
2006.
Kai Ming Ting. Precision and recall. In Encyclopedia of machine learning, pp. 781–781. Springer,
2011.
Gasper Tkacik, Olivier Marre, Dario Amodei, Elad Schneidman, William Bialek, and Michael J
Berry II. Searching for collective behavior in a large network of sensory neurons. PLoS Compu-
tational Biology, 10(1):e1003408, 2014.
Matthew D Zeiler and Rob Fergus. Visualizing and understanding convolutional networks.
European Conference on Computer Vision, pp. 818–833. Springer, 2014.
In
13
Published as a conference paper at ICLR 2018
A APPENDIX
A.1 GENERATING NEW, REALISTIC PATTERNS OF NEURAL ACTIVITY
In this section, we investigated how well Spike-GAN fits the whole probability density function
from which the population activity patterns present in the training dataset are drawn, following an
approach inspired by Macke et al. (2009).
We started by producing a ground truth dataset (2· 106 samples) for a small-size problem (2 neurons
x 12 time bins) so as to reduce the dimensionality of the samples and obtain a good approximation of
the underlying probability density function. We will call the probabilities computed from the ground
truth dataset numerical probabilities. We then obtained a second dataset (2 · 106 samples) drawn
from the same probability distribution as the ground truth dataset, which we will call surrogate
dataset. This dataset will provide us with a reference for the results we obtain when comparing the
distribution generated by Spike-GAN with the original ground truth distribution.
A third small dataset (training dataset, 8192 samples) coming from the same probability distribution
as the ground truth and surrogate datasets, was used to train Spike-GAN. Finally, a 2 · 106-sample
dataset (generated dataset) was obtained from the trained Spike-GAN.
The distributions corresponding to the three different datasets presented very similar entropies: the
ground truth and surrogate datasets had both 14.6 bits while the generated dataset had 14.3 bits.
Thus, the dataset generated by Spike-GAN was not affected by any evident mode collapse. We then
plotted the sample probabilities with respect to both the generated and the surrogate dataset against
the numerical probabilities. By comparing the densities in Fig. S1, we deduce that the surrogate
probability distribution deviates from the ground truth distribution (the identity line) in the same
way as the generated distribution does. Hence, this deviation can be attributed to finite sampling
effects rather than poor performance of Spike-GAN.
Importantly, the percentage of samples generated by Spike-GAN that were originally present in the
training dataset was 44% (45% for the surrogate dataset). This implies that 56% percent of the
samples produced by Spike-GAN were generated de novo. Finally, the percentage of generated
samples that had numerical probability equal to 0 was 3.2%, which is comparable to the 3.8% of
samples in the surrogate dataset that also had numerical probability 0.
Taken together, the above results strongly suggest that Spike-GAN has learned the probability dis-
tribution underlying the training dataset.
A.2 USING IMPORTANCE MAPS IN A REAL EXPERIMENT
When recording large-scale spike trains it is difficult to make hypotheses about what are the key
features of population activity that allow the animal to discriminate between the stimuli (for exam-
ple, the patterns of firing rates or the pattern of response latencies of specific subsets of neurons).
One possible use of Spike-GAN is to interpret the features of neural activity that are prominent in
the importance maps as possible candidates for being units of information relevant for behavioral
discrimination.
To illustrate the approach, we simulated a hypothetical experiment (Fig. S2A) in which we consider
N repetitions of a behavioural task, where a mouse has to discriminate two different stimuli (verti-
cal/horizontal stripes). For each repetition of the task, we assume to record several patterns of neural
activity such as those in Fig. 4 (main paper). By means of two-photon calcium imaging the activity
of a population of V1 neurons in the visual cortex of the mouse is recorded in response to the two
stimuli, which are associated with two distinct behavioral outcomes (e.g. drink from a left/right
reward port (Alemi-Neissi et al., 2013)). The mouse is assumed to have been previously trained on
the task. [Note that one possible difficulty in applying Spike-GAN to calcium imaging is that, unlike
spikes directly extracted from electrophysiology experiments, two-photon imaging signals are not
binary-valued. Nevertheless, to a first approximation, calcium signals can be binarized, and this has
been shown not to constitute an obstacle to the study of sensory encoding or decision-making in
certain preparations (Runyan et al., 2017).]
Area V1 is known to encode information about the stimulus orientation and therefore the two stimuli
will evoke two distinct activity patterns (Fig. S2B). However, these patterns are usually difficult to
14
Published as a conference paper at ICLR 2018
identify due to background activity and the lack of information about the ground truth correlation
structure of the recorded population of neurons. Fig. S2C shows the actual 'recorded' population
responses.
As described in Section 3.3, Spike-GAN can then be trained on the samples shown in Fig. S2C in or-
der to compute the importance map associated to each sample (Fig. S2D). We can then approximate
the structure of the original packets by thresholding and aligning (to the first spike above threshold)
each importance map and averaging all aligned maps for each stimulus. Importantly, the approx-
imated packets (gray maps) and the original ones (green pluses) are highly correlated (Fig. S2E;
r ≥ 0.8, yellow values) and the distribution of importance values corresponding to bins participat-
ing in a packet (those marked with a green plus) is clearly different from that corresponding to the
rest of the bins (Fig. S10A). Furthermore, we can easily identify those neurons participating in the
encoding of each stimulus by averaging the importance maps across time and trials for each stimulus
(see caption to Fig. S2 for details). As shown in Fig. S2F, the neurons that (by design) participate
in the encoding of a particular stimulus (indicated by red pluses and blue crosses for stim1 and
stim2, respectively) are those presenting the highest importance values. Given the clear bimodal
distribution of the importance values shown in Fig. S2F, we can easily set a threshold to select those
neurons encoding each stimulus. The recall and precision values (Ting, 2011) corresponding to this
threshold is also shown in Fig. S2F.
Finally, we have evaluated the importance maps in noisier scenarios, in which the spikes forming the
packets present noise in their timing (Fig. S9) or in which the number of samples available to train
Spike-GAN is smaller (Fig. S10). In both cases we have found that the importance maps analysis is
able to provide highly valuable information about the neurons participating in the encoding of the
stimulus information and their relative timing.
The features of neural activity individuated as important by the spike GAN importance maps may
or may not be used by the animal to make a decision. To test this causally, one can replace, during a
perceptual discrimination task, the presentation of a sensory stimulus with the presentation of a 2P-
optogenetic activation pattern. Importantly, the experimenter can, in each stimulation trial, either
keep or shuffle the features previously highlighted by the importance maps before imposing the
pattern. Comparing, between the altered and original patterns, the behavioral report of the animal
about which stimulus was presented can then be used to find out whether the information carried by
each specific feature is used to produce a consistent behavioral percept. For instance, a stimulation
protocol in which only part of the neurons participating in a particular feature is stimulated would
provide information about the capacity of the mouse brain to 'fill in the gaps'. Alternatively, the
time at which each neuron is stimulated could be altered in order to study the role spike timing plays
in the encoding of the stimulus.
15
Published as a conference paper at ICLR 2018
Figure S1: Fitting the whole probability distribution. Numerical probabilities obtained from
the ground truth dataset (2 correlated neurons, samples duration=12 ms, firing rate≈160 Hz,
correlation≈0.3, refractory period=2 ms) vs probabilities inferred from the surrogate and the gen-
erated datasets. Gray line is the identity. Blue tones: probabilities computed with respect to the
surrogate dataset. Red tones: probabilities computed with respect to the generated dataset. Both dis-
tributions are obtained by kernel density estimation, with a 2D Gaussian kernel with bandwidth=0.1.
16
6.56.05.55.04.5log probability in surrogate and generated distr.6.56.05.55.04.5log numerical probability Published as a conference paper at ICLR 2018
Figure S2: Using importance maps in a real experiment. A) Two-photon microscopy can be used for
both imaging and manipulating the activity of a population of neurons. A trained mouse is presented
with two different stimuli (stim1 and stim2) and makes a decision (e.g. drink from right/left reward
port) accordingly. At the same time the response of a population of neurons in V1 is recorded. B)
Simulated data presenting two different packets (18 neurons per packet, 4 of which participate in
both packets; each stimulus presentation evokes a single packet) in response to the two different
stimuli. The packets are only evident when the neurons are sorted. C) Actual recorded patterns of
activity. D) Importance maps obtained after training Spike-GAN with 4096 samples like the one
shown in panel C. E) Approximated packets are obtained by thresholding the importance maps (we
used as threshold the median of the total distribution of importance values) and aligning each of
them to the first spike above the threshold. The aligned maps are then averaged for each stimulus
(gray maps). Green pluses show the structure of the ground truth packets (in yellow the correlation
between the ground truth and the template packets). F) The importance of each neuron can be
inferred by averaging importance maps across time and trials. Blue crosses and red pluses indicate
the neurons that by design participate in the encoding of each stimulus, respectively. The computed
importance for these neurons is more than three times that of the neurons that do not participate.
Dashed gray line indicate the threshold used to select the participating neurons and compute the
precision and recall values shown in red and blue for each stimulus, respectively.
Figure S3: A) Example of neural recording in which the activity of 8 neurons is measured during an
arbitrary time period. B) Example of a sample extracted from the activity shown in A (green box).
17
stim 1stim 2in vivo two-photonimaging GCaMP6optogeneticstimulationneuronstime000000000011000000100101000001010011100110001000000110101000000101000100000000101010011010000000neuronstimesampleneural recordingAB Published as a conference paper at ICLR 2018
Figure S4: Negative critic loss corresponding to the training of Spike-GAN on samples coming
from the simulated activity of a population of 16 neurons whose firing probability follows a uniform
distribution across the whole duration (T=128 ms) of the samples (see Section 3.1).
18
Published as a conference paper at ICLR 2018
Figure S5: Ten generated samples are shown together with their closest sample in the training dataset
for Spike-GAN, the k-pairwise and the DG method. Note that we are measuring sample similarity
in terms of L1 distance. This implies that sometimes the closest sample is the one presenting no
spikes since non matching spikes are penalized more. For comparison, 10 samples contained in the
training dataset are shown together with their closest sample in the same training dataset (excluding
the sample itself).
19
Published as a conference paper at ICLR 2018
Figure S6: Comparison of Spike-GAN with the k-pairwise and the DG models in the presence of
negative correlations (refractory period).
20
Published as a conference paper at ICLR 2018
Figure S7: Filters learned by the first layer of Spike-GAN when trained on the dataset described in
Section 3.3.
21
Published as a conference paper at ICLR 2018
Figure S8: Randomly selected samples and their corresponding importance maps.
22
Published as a conference paper at ICLR 2018
Figure S9: In order to test the approach discussed in Section A.2 with less reliable responses, we
introduced noise in the packets in the form of discretized Gaussian noise (std=0.5) added to the
time of each spike (panel A shows 8 randomly selected samples for each stimulus (red and blue
boxes, respectively)). B) Inferred packets are, as expected, noisier but the correlation with the
ground truth packets is still high (yellow values). Note that the correlation is computed between
the inferred packet and the average of all noisy packets present in the training dataset, so as to
take into account the inherent uncertainty of the responses. Green pluses show the structure of
the ground truth, noise-free packets. F) Neurons importance value each stimulus (red and blue,
respectively). Blue crosses and red pluses indicate the neurons that by design participate in the
encoding of each stimulus, respectively. Dashed gray line indicate the threshold used to select the
participating neurons and compute the precision and recall values shown in red and blue for each
stimulus, respectively. Most of the relevant neurons are still being detected (recall=0.89) with just a
slight decrease in the precision (precision ≥ 0.94).
23
Published as a conference paper at ICLR 2018
Figure S10: Approximated packets for different training dataset sizes. A) Distribution of the im-
portance values (normalized to the maximum value) shown in Fig. S2E, for the bins forming the
packets (i.e. bins marked with a green plus in Fig. S2E) (black line) and those not participating in
the packets (gray line). The two distributions are clearly separated, as confirmed by the ROC curve
in panel D (solid blue line). B-C) Same as in panel A in the case where 2048 and 1024 samples have
been used to train Spike-GAN, respectively. The corresponding ROC curves are shown in green and
red, respectively, in panel D.
24
|
1702.01825 | 1 | 1702 | 2017-02-06T23:48:19 | Deep Learning Models of the Retinal Response to Natural Scenes | [
"q-bio.NC",
"stat.ML"
] | A central challenge in neuroscience is to understand neural computations and circuit mechanisms that underlie the encoding of ethologically relevant, natural stimuli. In multilayered neural circuits, nonlinear processes such as synaptic transmission and spiking dynamics present a significant obstacle to the creation of accurate computational models of responses to natural stimuli. Here we demonstrate that deep convolutional neural networks (CNNs) capture retinal responses to natural scenes nearly to within the variability of a cell's response, and are markedly more accurate than linear-nonlinear (LN) models and Generalized Linear Models (GLMs). Moreover, we find two additional surprising properties of CNNs: they are less susceptible to overfitting than their LN counterparts when trained on small amounts of data, and generalize better when tested on stimuli drawn from a different distribution (e.g. between natural scenes and white noise). Examination of trained CNNs reveals several properties. First, a richer set of feature maps is necessary for predicting the responses to natural scenes compared to white noise. Second, temporally precise responses to slowly varying inputs originate from feedforward inhibition, similar to known retinal mechanisms. Third, the injection of latent noise sources in intermediate layers enables our model to capture the sub-Poisson spiking variability observed in retinal ganglion cells. Fourth, augmenting our CNNs with recurrent lateral connections enables them to capture contrast adaptation as an emergent property of accurately describing retinal responses to natural scenes. These methods can be readily generalized to other sensory modalities and stimulus ensembles. Overall, this work demonstrates that CNNs not only accurately capture sensory circuit responses to natural scenes, but also yield information about the circuit's internal structure and function. | q-bio.NC | q-bio |
Deep Learning Models of the Retinal Response to
Natural Scenes
Lane T. McIntosh∗1, Niru Maheswaranathan∗1, Aran Nayebi1,
Surya Ganguli2,3, Stephen A. Baccus3
1Neurosciences PhD Program, 2Department of Applied Physics, 3Neurobiology Department
{lmcintosh, nirum, anayebi, sganguli, baccus}@stanford.edu
Stanford University
Abstract
A central challenge in sensory neuroscience is to understand neural computations
and circuit mechanisms that underlie the encoding of ethologically relevant, natu-
ral stimuli. In multilayered neural circuits, nonlinear processes such as synaptic
transmission and spiking dynamics present a significant obstacle to the creation of
accurate computational models of responses to natural stimuli. Here we demon-
strate that deep convolutional neural networks (CNNs) capture retinal responses to
natural scenes nearly to within the variability of a cell's response, and are markedly
more accurate than linear-nonlinear (LN) models and Generalized Linear Mod-
els (GLMs). Moreover, we find two additional surprising properties of CNNs:
they are less susceptible to overfitting than their LN counterparts when trained
on small amounts of data, and generalize better when tested on stimuli drawn
from a different distribution (e.g. between natural scenes and white noise). An
examination of the learned CNNs reveals several properties. First, a richer set
of feature maps is necessary for predicting the responses to natural scenes com-
pared to white noise. Second, temporally precise responses to slowly varying
inputs originate from feedforward inhibition, similar to known retinal mechanisms.
Third, the injection of latent noise sources in intermediate layers enables our model
to capture the sub-Poisson spiking variability observed in retinal ganglion cells.
Fourth, augmenting our CNNs with recurrent lateral connections enables them to
capture contrast adaptation as an emergent property of accurately describing retinal
responses to natural scenes. These methods can be readily generalized to other
sensory modalities and stimulus ensembles. Overall, this work demonstrates that
CNNs not only accurately capture sensory circuit responses to natural scenes, but
also can yield information about the circuit's internal structure and function.
1
Introduction
A fundamental goal of sensory neuroscience involves building accurate neural encoding models that
predict the response of a sensory area to a stimulus of interest. These models have been used to shed
light on circuit computations [1, 2, 3, 4], uncover novel mechanisms [5, 6], highlight gaps in our
understanding [7], and quantify theoretical predictions [8, 9].
A commonly used model for retinal responses is a linear-nonlinear (LN) model that combines a linear
spatiotemporal filter with a single static nonlinearity. Although LN models have been used to describe
responses to artificial stimuli such as spatiotemporal white noise [10, 2], they fail to generalize to
natural stimuli [7]. Furthermore, the white noise stimuli used in previous studies are often low
resolution or spatially uniform and therefore fail to differentially activate nonlinear subunits in the
∗These authors contributed equally to this work.
30th Conference on Neural Information Processing Systems (NIPS 2016), Barcelona, Spain.
Figure 1: A schematic of the model architecture. The stimulus was convolved with 8 learned
spatiotemporal filters whose activations were rectified. The second convolutional layer then projected
the activity of these subunits through spatial filters onto 16 subunit types, whose activity was linearly
combined and passed through a final soft rectifying nonlinearity to yield the predicted response.
retina, potentially simplifying the retinal response to such stimuli [11, 12, 2, 10, 13]. In contrast to
the perceived linearity of the retinal response to coarse stimuli, the retina performs a wide variety of
nonlinear computations including object motion detection [6], adaptation to complex spatiotemporal
patterns [14], encoding spatial structure as spike latency [15], and anticipation of periodic stimuli
[16], to name a few. However it is unclear what role these nonlinear computational mechanisms have
in generating responses to more general natural stimuli.
To better understand the visual code for natural stimuli, we modeled retinal responses to natural image
sequences with convolutional neural networks (CNNs). CNNs have been successful at many pattern
recognition and function approximation tasks [17]. In addition, these models cascade multiple layers
of spatiotemporal filtering and rectification–exactly the elementary computational building blocks
thought to underlie complex functional responses of sensory circuits. Previous work utilized CNNs
to gain insight into the neural computations of inferotemporal cortex [18], but these models have
not been applied to early sensory areas where knowledge of neural circuitry can provide important
validation for such models.
We find that deep neural network models markedly outperform previous models in predicting retinal
responses both for white noise and natural scenes. Moreover, these models generalize better to unseen
stimulus classes, and learn internal features consistent with known retinal properties, including
sub-Poisson variability, feedforward inhibition, and contrast adaptation. Our findings indicate that
CNNs can reveal both neural computations and mechanisms within a multilayered neural circuit
under natural stimulation.
2 Methods
The spiking activity of a population of tiger salamander retinal ganglion cells was recorded in response
to both sequences of natural images jittered with the statistics of eye movements and high resolution
spatiotemporal white noise. Convolutional neural networks were trained to predict ganglion cell
responses to each stimulus class, simultaneously for all cells in the recorded population of a given
retina. For a comparison baseline, we also trained linear-nonlinear models [19] and generalized
linear models (GLMs) with spike history feedback [2]. More details on the stimuli, retinal recordings,
experimental structure, and division of data for training, validation, and testing are given in the
Supplemental Material.
2.1 Architecture and optimization
The convolutional neural network architecture is shown in Figure 2.1. Model parameters were
optimized to minimize a loss function corresponding to the negative log-likelihood under Poisson
spike generation. Optimization was performed using ADAM [20] via the Keras and Theano software
libraries [21]. The networks were regularized with an (cid:96)2 weight penalty at each layer and an (cid:96)1
activity penalty at the final layer, which helped maintain a baseline firing rate near 0 Hz.
2
……time8 subunits16 subunitsconvolutionconvolutiondenseresponses We explored a variety of architectures for the CNN, varying the number of layers, number of filters per
layer, the type of layer (convolutional or dense), and the size of the convolutional filters. Increasing
the number of layers increased prediction accuracy on held-out data up to three layers, after which
performance saturated. One implication of this architecture search is that LN-LN cascade models –
which are equivalent to a 2-layer CNN – would also underperform 3 layer CNN models.
Contrary to the increasingly small filter sizes used by many state-of-the-art object recognition
networks, our networks had better performance using filter sizes in excess of 15x15 checkers. Models
were trained over the course of 100 epochs, with early-stopping guided by a validation set. See
Supplementary Materials for details on the baseline models we used for comparison.
3 Results
We found that convolutional neural networks were substantially better at predicting retinal responses
than either linear-nonlinear (LN) models or generalized linear models (GLMs) on both white noise
and natural scene stimuli (Figure 2).
3.1 Performance
Figure 2: Model performance. (A,B) Correlation coefficients between the data and CNN, GLM or
LN models for white noise and natural scenes. Dotted line indicates a measure of retinal reliability
(See Methods). (C) Receiver Operating Characteristic (ROC) curve for spike events for CNN, GLM
and LN models. (D) Spike rasters of one example retinal ganglion cell responding to 6 repeated
trials of the same randomly selected segment of the natural scenes stimulus (black) compared to the
predictions of the LN (red), GLM (green), or CNN (blue) model with Poisson spike generation used
to generate model rasters. (E) Peristimulus time histogram (PSTH) of the spike rasters in (D).
3
EABWhite noiseNatural scenesCNNLNDRetinal reliabilityCData6 trialsGLMCNNLNGLMCNNLNGLMFiring Rate (Hz)Time (seconds)ROC Curve for Natural Scenes LN models and GLMs failed to capture retinal responses to natural scenes (Figure 2B) consistent with
previous results [7]. In addition, we also found that LN models only captured a small fraction of the
response to high resolution spatiotemporal white noise, presumably because of the finer resolution that
were used (Figure 2A). In contrast, CNNs approach the reliability of the retina for both white noise
and natural scenes. Using other metrics, including fraction of explained variance, log-likelihood, and
mean squared error, CNNs showed a robust gain in performance over previously described sensory
encoding models.
We investigated how model performance varied as a function of training data, and found that LN
models were more susceptible to overfitting than CNNs, despite having fewer parameters (Figure 4A).
In particular, a CNN model trained using just 25 minutes of data had better held out performance
than an LN model fit using the full 60 minute recording. We expect that both depth and convolutional
filters act as implicit regularizers for CNN models, thereby increasing generalization performance.
3.2 CNN model parameters
Figure 3 shows a visualization of the model parameters learned when a convolutional network is
trained to predict responses to either white noise or natural scenes. We visualized the average feature
represented by a model unit by computing a response-weighted average for that unit. Models trained
on white noise learned first layer features with small (∼200 µm) receptive field widths (top left box
in Figure 3), whereas the natural scene model learns spatiotemporal filters with overall lower spatial
and temporal frequencies. This is likely in part due to the abundance of low spatial frequencies
present in natural images [22]. We see a greater diversity of spatiotemporal features in the second
layer receptive fields compared to the first (bottom panels in Figure 3). Additionally, we see more
diversity in models trained on natural scenes, compared to white noise.
Figure 3: Model parameters visualized by computing a response-weighted average for different
model units, computed for models trained on spatiotemporal white noise stimuli (left) or natural
image sequences (right). Top panel (purple box): visualization of units in the first layer. Each 3D
spatiotemporal receptive field is displayed via a rank-one decomposition consisting of a spatial filter
(top) and temporal kernel (black traces, bottom). Bottom panel (green box): receptive fields for the
second layer units, again visualized using a rank-one decomposition. Natural scenes models required
more active second layer units, displaying a greater diversity of spatiotemporal features. Receptive
fields are cropped to the region of space where the subunits have non-zero sensitivity.
3.3 Generalization across stimulus distributions
Historically, much of our understanding of the retina comes from fitting models to responses to
artificial stimuli and then generalizing this understanding to cases where the stimulus distribution is
more natural. Due to the large difference between artificial and natural stimulus distributions, it is
unknown what retinal encoding properties generalize to a new stimulus.
4
Figure 4: CNNs overfit less and generalize better across stimulus class as compared to simpler models.
(A) Held-out performance curves for CNN (∼150,000 parameters) and GLM/LN models cropped
around the cell's receptive field (∼4,000 parameters) as a function of the amount of training data. (B)
Correlation coefficients between responses to natural scenes and models trained on white noise but
tested on natural scenes. See text for discussion.
We explored what portion of CNN, GLM, and LN model performance is specific to a particular
stimulus distribution (white noise or natural scenes), versus what portion describes characteristics
of the retinal response that generalize to another stimulus class. We found that CNNs trained
on responses to one stimulus class generalized better to a stimulus distribution that the model
was not trained on (Figure 4B). Despite LN models having fewer parameters, they nonetheless
underperform larger convolutional neural network models when predicting responses to stimuli not
drawn from the training distribution. GLMs faired particularly poorly when generalizing to natural
scene responses, likely because changes in mean luminance result in pathological firing rates after
the GLM's exponential nonlinearity. Compared to standard models, CNNs provide a more accurate
description of sensory responses to natural stimuli even when trained on artificial stimuli (Figure 4B).
3.4 Capturing uncertainty of the neural response
In addition to describing the average response to a particular stimulus, an accurate model should also
capture the variability about the mean response. Typical noise models assume i.i.d. Poisson noise
drawn from a deterministic mean firing rate. However, the variability in retinal spiking is actually
sub-Poisson, that is, the variability scales with the mean but then increases sublinearly at higher mean
rates [23, 24]. By training models with injected noise [25], we provided a latent noise source in the
network that models the unobserved internal variability in the retinal population. Surprisingly, the
model learned to shape injected Gaussian noise to qualitatively match the shape of the true retinal
noise distribution, increasing with the mean response but growing sublinearly at higher mean rates
(Figure 5). Notably, this relationship only arises when noise is injected during optimization–injecting
Gaussian noise in a pre-trained network simply produced a linear scaling of the noise variance as a
function of the mean.
3.5 Feedforward inhibition shapes temporal responses in the model
To understand how a particular model response arises, we visualized the flow of signals through the
network. One prominent aspect of the difference between CNN and LN model responses is that CNNs
but not LN models captured the precise timing and short duration of firing events. By examining
the responses to the internal units of CNNs in time and averaged over space (Figure 6 A-C), we
found that in both convolutional layers, different units had either positive or negative responses to the
same stimuli, equivalent to excitation and inhibition as found in the retina. Precise timing in CNNs
arises by a timed combination of positive and negative responses, analogous to feedforward inhibition
that is thought to generate precise timing in the retina [26, 27]. To examine network responses in
5
Figure 5: Training with added noise recovers retinal sub-Poisson noise scaling property. (A) Variance
versus mean spike count for CNNs with various strengths of injected noise (from 0.1 to 10 standard
deviations), as compared to retinal data (black) and a Poisson distribution (dotted red). (B) The same
plot as A but with each curve normalized by the maximum variance. (C) Variance versus mean spike
count for CNN models with noise injection at test time but not during training.
space, we selected a particular time in the experiment and visualized the activation maps in the first
(purple) and second (green) convolutional layers (Figure 6D). A given image is shown decomposed
through multiple parallel channels in this manner. Finally, Figure 6E highlights how the temporal
autocorrelation in the signals at different layers varies. There is a progressive sharpening of the
response, such that by the time it reaches the model output the predicted responses are able to mimic
the statistics of the real firing events (Figure 6C).
3.6 Feedback over long timescales
Retinal dynamics are known to exceed the duration of the filters that we used (400 ms). In particular,
changes in stimulus statistics such as luminance, contrast and spatio-temporal correlations can
generate adaptation lasting seconds to tens of seconds [5, 28, 14]. Therefore, we additionally
explored adding feedback over longer timescales to the convolutional network.
To do this, we added a recurrent neural network (RNN) layer with a history of 10s after the fully
connected layer prior to the output layer. We experimented with different recurrent architectures
(LSTMs [29], GRUs [30], and MUTs [31]) and found that they all had similar performance to the
CNN at predicting natural scene responses. Despite the similar performance, we found that the
recurrent network learned to adapt its response over the timescale of a few seconds in response to
step changes in stimulus contrast (Figure 7). This suggests that RNNs are a promising way forward
to capture dynamical processes such as adaptation over longer timescales in an unbiased, data-driven
manner.
4 Discussion
In the retina, simple models of retinal responses to spatiotemporal white noise have greatly influenced
our understanding of early sensory function. However, surprisingly few studies have addressed
whether or not these simple models can capture responses to natural stimuli. Our work applies
models with rich computational capacity to bear on the problem of understanding natural scene
responses. We find that convolutional neural network (CNN) models, sometimes augmented with
lateral recurrent connections, well exceed the performance of other standard retinal models including
LN and GLMs. In addition, CNNs are better at generalizing both to held-out stimuli and to entirely
different stimulus classes, indicating that they are learning general features of the retinal response.
Moreover, CNNs capture several key features about retinal responses to natural stimuli where LN
models fail. In particular, they capture: (1) the temporal precision of firing events despite employing
filters with slower temporal frequencies, (2) adaptive responses during changing stimulus statistics,
and (3) biologically realistic sub-Poisson variability in retinal responses. In this fashion, this work
provides the first application of deep learning to understanding early sensory systems under natural
conditions.
6
Variance in Spike CountMean Spike CountAMean Spike CountNormalized Variancein Spike CountMean Spike CountBVariance in Spike CountCDataPoisson0.10.11.02.04.010.0DataPoisson1.02.04.010.0 Figure 6: Visualizing the internal activity of a CNN in response to a natural scene stimulus. (A-C)
Time series of the CNN activity (averaged over space) for the first convolutional layer (8 units, A),
the second convolutional layer (16 units, B), and the final predicted response for an example cell (C,
cyan trace). The recorded (true) response is shown below the model prediction (C, gray trace) for
comparison. (D) Spatial activation of example CNN filters at a particular time point. The selected
stimulus frame (top, grayscale) is represented by parallel pathways encoding spatial information in
the first (purple) and second (green) convolutional layers (a subset of the activation maps is shown
for brevity). (E) Autocorrelation of the temporal activity in (A-C). The correlation in the recorded
firing rates is shown in gray.
Figure 7: Recurrent neural network (RNN) layers capture response features occurring over multiple
seconds. (A) A schematic of how the architecture from Figure 2.1 was modified to incorporate the
RNN at the last layer of the CNN. (B) Response of an RNN trained on natural scenes, showing a
slowly adapting firing rate in response to a step change in contrast.
To date, modeling efforts in sensory neuroscience have been most useful in the context of carefully
designed parametric stimuli, chosen to illuminate a computation or mechanism of interest [32]. In
part, this is due to the complexities of using generic natural stimuli. It is both difficult to describe
the distribution of natural stimuli mathematically (unlike white or pink noise), and difficult to fit
models to stimuli with non-stationary statistics when those statistics influence response properties.
7
0Firing Rate(spikes/s)04242Time (s)StimulusIntensityRNN6LSTMABFull Field Flicker We believe the approach taken in this paper provides a way forward for understanding general natural
scene responses. We leverage the computational power and flexibility of CNNs to provide us with a
tractable, accurate model that we can then dissect, probe, and analyze to understand what that model
captures about the retinal response. This strategy of casting a wide computational net to capture
neural circuit function and then constraining it to better understand that function will likely be useful
in a variety of neural systems in response to many complex stimuli.
Acknowledgments
The authors would like to thank Ben Poole and EJ Chichilnisky for helpful discussions related to this
work. Thanks also goes to the following institutions for providing funding and hardware grants, LM:
NSF, NVIDIA Titan X Award, NM: NSF, AN and SB: NEI grants, SG: Burroughs Wellcome, Sloan,
McKnight, Simons, James S. McDonnell Foundations and the ONR.
References
[1] Tim Gollisch and Markus Meister. Eye smarter than scientists believed: neural computations in
circuits of the retina. Neuron, 65(2):150–164, 2010.
[2] Jonathan W Pillow, Jonathon Shlens, Liam Paninski, Alexander Sher, Alan M Litke,
EJ Chichilnisky, and Eero P Simoncelli. Spatio-temporal correlations and visual signalling in a
complete neuronal population. Nature, 454(7207):995–999, 2008.
[3] Nicole C Rust, Odelia Schwartz, J Anthony Movshon, and Eero P Simoncelli. Spatiotemporal
elements of macaque v1 receptive fields. Neuron, 46(6):945–956, 2005.
[4] David B Kastner and Stephen A Baccus. Coordinated dynamic encoding in the retina using
opposing forms of plasticity. Nature neuroscience, 14(10):1317–1322, 2011.
[5] Stephen A Baccus and Markus Meister. Fast and slow contrast adaptation in retinal circuitry.
Neuron, 36(5):909–919, 2002.
[6] Bence P Ölveczky, Stephen A Baccus, and Markus Meister. Segregation of object and back-
ground motion in the retina. Nature, 423(6938):401–408, 2003.
[7] Alexander Heitman, Nora Brackbill, Martin Greschner, Alexander Sher, Alan M Litke, and
EJ Chichilnisky. Testing pseudo-linear models of responses to natural scenes in primate retina.
bioRxiv, page 045336, 2016.
[8] Joseph J Atick and A Norman Redlich. Towards a theory of early visual processing. Neural
Computation, 2(3):308–320, 1990.
[9] Xaq Pitkow and Markus Meister. Decorrelation and efficient coding by retinal ganglion cells.
Nature neuroscience, 15(4):628–635, 2012.
[10] Jonathan W Pillow, Liam Paninski, Valerie J Uzzell, Eero P Simoncelli, and EJ Chichilnisky.
Prediction and decoding of retinal ganglion cell responses with a probabilistic spiking model.
The Journal of Neuroscience, 25(47):11003–11013, 2005.
[11] S Hochstein and RM Shapley. Linear and nonlinear spatial subunits in y cat retinal ganglion
cells. The Journal of Physiology, 262(2):265, 1976.
[12] Tim Gollisch. Features and functions of nonlinear spatial integration by retinal ganglion cells.
Journal of Physiology-Paris, 107(5):338–348, 2013.
[13] Adrienne L Fairhall, C Andrew Burlingame, Ramesh Narasimhan, Robert A Harris, Jason L
Puchalla, and Michael J Berry. Selectivity for multiple stimulus features in retinal ganglion
cells. Journal of neurophysiology, 96(5):2724–2738, 2006.
[14] Toshihiko Hosoya, Stephen A Baccus, and Markus Meister. Dynamic predictive coding by the
retina. Nature, 436(7047):71–77, 2005.
[15] Tim Gollisch and Markus Meister. Rapid neural coding in the retina with relative spike latencies.
science, 319(5866):1108–1111, 2008.
[16] Greg Schwartz, Rob Harris, David Shrom, and Michael J Berry. Detection and prediction of
periodic patterns by the retina. Nature neuroscience, 10(5):552–554, 2007.
[17] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444,
2015.
8
[18] Daniel L Yamins, Ha Hong, Charles Cadieu, and James J DiCarlo. Hierarchical modular
optimization of convolutional networks achieves representations similar to macaque it and
human ventral stream. In Advances in neural information processing systems, pages 3093–3101,
2013.
[19] EJ Chichilnisky. A simple white noise analysis of neuronal light responses. Network: Computa-
tion in Neural Systems, 12(2):199–213, 2001.
[20] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint
arXiv:1412.6980, 2014.
[21] Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, James Bergstra, Ian Goodfellow, Arnaud
Bergeron, Nicolas Bouchard, David Warde-Farley, and Yoshua Bengio. Theano: new features
and speed improvements. arXiv preprint arXiv:1211.5590, 2012.
[22] Aapo Hyvärinen, Jarmo Hurri, and Patrick O Hoyer. Natural Image Statistics: A Probabilistic
Approach to Early Computational Vision., volume 39. Springer Science & Business Media,
2009.
[23] Michael J Berry, David K Warland, and Markus Meister. The structure and precision of retinal
spike trains. Proceedings of the National Academy of Sciences, 94(10):5411–5416, 1997.
[24] Rob R de Ruyter van Steveninck, Geoffrey D Lewen, Steven P Strong, Roland Koberle, and
William Bialek. Reproducibility and variability in neural spike trains. Science, 275(5307):1805–
1808, 1997.
[25] Ben Poole, Jascha Sohl-Dickstein, and Surya Ganguli. Analyzing noise in autoencoders and
deep networks. arXiv preprint arXiv:1406.1831, 2014.
[26] Botond Roska and Frank Werblin. Vertical interactions across ten parallel, stacked representa-
tions in the mammalian retina. Nature, 410(6828):583–587, 2001.
[27] Botond Roska and Frank Werblin. Rapid global shifts in natural scenes block spiking in specific
ganglion cell types. Nature neuroscience, 6(6):600–608, 2003.
[28] Peter D Calvert, Victor I Govardovskii, Vadim Y Arshavsky, and Clint L Makino. Two temporal
phases of light adaptation in retinal rods. The Journal of general physiology, 119(2):129–146,
2002.
[29] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation,
9(8):1735–1780, 1997.
[30] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the
properties of neural machine translation: Encoder–decoder approaches. Eighth Workshop on
Syntax, Semantics and Structure in Statistical Translation (SSST-8), pages 103–111, 2014.
[31] Rafal Jozefowicz, Wojciech Zaremba, and Ilya Sutskever. An empirical exploration of recurrent
network architectures. Proceedings of the 32nd International Conference on Machine Learning,
37:2342–2350, 2015.
[32] Nicole C Rust and J Anthony Movshon. In praise of artifice. Nature neuroscience, 8(12):1647–
1650, 2005.
[33] Stelios M Smirnakis, Michael J Berry, David K Warland, William Bialek, and Markus Meister.
Adaptation of retinal processing to image contrast and spatial scale. Nature, 386(6620):69–73,
1997.
[34] Gašper Tkacik, Patrick Garrigan, Charles Ratliff, Grega Milcinski, Jennifer M Klein, Lucia H
Seyfarth, Peter Sterling, David H Brainard, and Vijay Balasubramanian. Natural images from
the birthplace of the human eye. PLoS One, 6(6):e20409, 2011.
9
Supplementary methods
Retinal recordings
The responses of tiger salamander retinal ganglion cells from 3 animals were recorded using a 60
channel multielectrode array. Further experimental details are described in detail elsewhere [4].
We analyzed the reliability of all recorded cells over the course of each experiment by computing this
correlation coefficient between a cell's average response to the same stimulus on different blocks of
trials and analyzed only those cells with a correlation exceeding 0.3. Thirty-seven cells exceeded the
criterion for reliability. Of these, 70.3% were fast OFF-type cells, 10.8% were medium OFF, and
16.7% were slow OFF. Our original dataset also included ON cells, however none of them passed our
retinal reliability criterion.
We interleaved natural scenes and white noise stimuli to average over any experimental drift. However,
these transitions generated contrast adaptation over tens of seconds [33, 5] that could not be captured
by the short duration of spatiotemporal filters (400 ms) in the CNN. Therefore, we focused our
analysis on steady state responses by excluding one minute of data after each transition. Spiking
responses were binned using 10 ms bins and smoothed using a 10 ms Gaussian filter.
The training dataset was divided randomly according to a 90%/10% train/validation split, and the test
set consisted of averaged repeated trials to 1 minute of novel stimuli.
Stimulus
The white noise stimulus consisted of binary checkers at 35% contrast, and the natural scene stimulus
was a sequence of jittered natural images sampled from a natural image database [34].
Of particular note is the spatial resolution of this dataset, which is considerably higher than stimuli
used in pre-existing attempts to model retinal responses. Our stimuli consisted of 50 x 50 spatial
checkers, each of which spanned 55 µm x 55 µm on the retina. At this resolution, they can
differentially activate nonlinear subunits [11, 12] within the ∼250 µm salamander ganglion cell
receptive field center. Previous stimuli are 120 µm pixels, covering roughly the entire RF center in
primate peripheral retina [2] or spatially uniform [10, 13]. Since coarser stimuli will not differentially
activate subunits, this higher resolution dataset provides a unique challenge for capturing nonlinear
retinal responses.
Comparison with other models
To compare the performance of other models, we fit GLMs using spatiotemporal stimulus filters and
temporal spike history filters, and LN models using spatiotemporal filters and a parameterized soft
rectifying nonlinearity. We found that we needed to regularize the LN model parameters in order to
prevent overfitting and offset the large number of parameters (Figure 4A). We tried (1) using a convo-
lutional filter instead of a fully-connected one, (2) only using stimuli centered around the receptive
field, and (3) using various levels of (cid:96)1 and/or (cid:96)2 penalties on the filter coefficients as regularization
techniques. Cutting out the stimulus around the receptive field (using a window size of 11 x 11
checkers, or 605 µm), in combination with (cid:96)2 regularization, led to the best held-out performance
(Figure 4A). We found the same to be true of GLMs. This cropping regularization procedure resulted
in LN models requiring only 4843 parameters, and GLMs requiring 4861 parameters. Therefore, we
report LN and GLM performance results using this regularization scheme.
The GLM parameters consist of weights, biases, and spike-history filters, however we did not include
cell coupling filters, since [2] reported that adding coupling did not improve predictions of the average
firing rate, although they improved single trial predictions. In this study we report performance as the
comparison between the models' predictions and averaged responses to repeated stimuli, thus [2]
suggests coupling filters would not improve the performance of GLMs.
10
|
1904.10508 | 2 | 1904 | 2019-11-13T02:28:54 | Quantum-Inspired Computing: Can it be a Microscopic Computing Model of the Brain? | [
"q-bio.NC",
"cs.ET",
"cs.NE"
] | Quantum computing and the workings of the brain have many aspects in common and have been attracting increasing attention in academia and industry. The computation in both is parallel and non-discrete. Though the underlying physical dynamics (e.g., equation of motion) may be deterministic, the observed or interpreted outcomes are often probabilistic. Consequently, various investigations have been undertaken to understand and reproduce the brain on the basis of quantum physics and computing. However, there have been arguments on whether the brain can and have to take advantage of quantum phenomena that need to survive in the macroscopic space-time region at room temperature. This paper presents a unique microscopic computational model for the brain based on an ansatz that the brain computes in a manner similar to quantum computing, but with classical waves. Log-scale encoding of information in the context of computing with waves is shown to play a critical role in bridging the computing models with classical and quantum waves. Our quantum-inspired computing model opens up a possibility of unifying the computing framework of artificial intelligence and quantum computing beyond quantum machine learning approaches. | q-bio.NC | q-bio |
Quantum-Inspired Computing: Can it be a Microscopic
Computing Model of the Brain?
Yasunao Katayama
IBM Research - Tokyo
email: [email protected]
Abstract: Quantum computing and the workings of the brain have many aspects in common
and have been attracting increasing attention in academia and industry. The computation in
both is parallel and non-discrete. Though the underlying physical dynamics (e.g., equation of
motion) may be deterministic, the observed or interpreted outcomes are often probabilistic.
Consequently, various investigations have been undertaken to understand and reproduce the
brain on the basis of quantum physics and computing 1 -- 9. However, there have been argu-
ments on whether the brain can and have to take advantage of quantum phenomena that
need to survive in the macroscopic space-time region at room temperature 10 -- 12. This paper
presents a unique microscopic computational model for the brain based on an ansatz that
the brain computes in a manner similar to quantum computing, but with classical waves.
Log-scale encoding of information 13 in the context of computing with waves 14 is shown to
play a critical role in bridging the computing models with classical and quantum waves. Our
quantum-inspired computing model opens up a possibility of unifying the computing frame-
work of artificial intelligence and quantum computing beyond quantum machine learning
approaches.
1
3
5
9
10
14
17
Contents
1 Introduction
2 Encoding cubits into classical waves
3 Operations for cubits
4 Log-scale encoding to bridge QC and QIC
5 Discussion
6 Conclusion
2
1 Introduction
Artificial intelligence (AI) and quantum computing (QC) are two rapidly evolving technologies
that are redefining computing. The capability of computers to handle narrowly defined AI tasks is
surpassing human capability, and the research focus is shifting toward giving computers broader
and more general AI capabilities 16. Quantum computers are expected to solve certain types of
problems that conventional computers are hard to solve 17. Since QC and the workings of the brain
share many aspects in common, it is not surprising to see that there have been various research
activities aimed at understanding the computing model of the brain on the basis of quantum physics
and QC as early as the 1970s 2 -- 6. As a result of rapid advancement of QC, quantum machine
learning has been attracting a lot of attention in both AI and QC research communities 7 -- 9 Indeed,
the computation in both cases is executed in parallel with non-discrete variables. Though the
underlying physical dynamics (e.g., equation of motion) may be deterministic, the observed or
interpreted outcomes are more probabilistic.
However, at the same time there have been controversial arguments from physics (i.e., ma-
terial) and cognitive science (i.e., mental) points of view on whether the brain, both biological and
artificial, can and have to take advantage of quantum features, such as Tensor-product statespace
and entanglement, that need to survive in macroscopic space-time regions at room temperature
10 -- 12. Though macroscopic quantum phenomena are increasingly being observed in the lab 18, 19,
they are seldom perceived in ordinary life. In addition, from a computing point of view, though
there is much less doubt on inherent QC advantages in quantum information processing 20, 21, AI
3
workloads in general process huge data sets and irreversible algorithms with noise in computing
environment and data, which may not always be effectively translated into tensor product entan-
gled statespaces with a limited number of qubits and their interconnects.
In other words, the
statespace advantage of QC may quickly diminish unless the symmetry of the problem cleanly fit
to exponentially-large Hilbert state spaces with exclusively linear operators on them. QC-unique
constraints such as no cloning, strict reversibility, measurement complications, may pose additional
challenges in application coverage, system scalability, and fault and error tolerances, if we dare to
pursue strict QC approaches universally for classical and data-intensive computing problems of
AI workloads. On the other hand, parallel and non-discrete dynamics is not unique to quantum
physics but is something natively observed in ordinary classical physics, such as wave dynamics
14.
Here in this paper, we propose quantum-inspired computing (QIC) as energy and computa-
tionally efficient microscopic computing model for classical workloads, in particular AI. As shown
in Fig. 1, our approach is unique compared with existing QC-based approaches, since quantum
physics is not prerequisite in the computing model.
Instead, parallel computing advantage of
QC is incorporated in a more general natural computing context with less QC-unique constraints.
More specifically, our model is based in as a network of elastic wave processing of spikes. Wave
superposition and thresholding can smoothly and precisely execute, without being influenced by
collision and congestion, weighted sum and ReLU operations, which are key arithmetic in modern
AI processing. Our quantum-inspired modeling approach with natively incorporating precise spike
correlations in continuous time allows for the temporal computing more rigorously based on origi-
4
nal spiking neural network (SNN) model 22 and should differentiate it from existing neuromorphic
approaches 23 -- 25. In order for physically directive and nonblocking wave nature to prevail at a small
scale for energy efficiency and integration, it is essential to reduce the wavelength by reducing the
spike signal velocity 14. Quantum-inspired approach has been improving classical counterpart at
different levels 15. The present work is quantum-inspired approach at a microscopic computing
level.
2 Encoding cubits into classical waves
The wave nature of matters and fields plays a critical role in both classical and quantum physics.
In classical physics, a wave is a phenomenon that can elastically transfer energy and informa-
tion, without transporting matter, via the interplay between displacement and a linearly responsive
restoring force, such as between a displacement current and an electrical and magnetic field in the
case of electromagnetic waves. Waves carry electrical signals, not electrons. In quantum physics,
a wave function is an essential means to describe quantum phenomena in both ground and excited
states. Though classical waves do not exhibit quantum-mechanical coherence, such as macroscop-
ically observed in lasers and superconductors, they have their own wave characteristics such as
superposition and interference. Classically coherent waves can be generated by externally exciting
the system in phase with forced vibrations. Such classical waves are in accordance with classical
dynamics with commutable variables and contiguous energy spectra. Interestingly, the coherence
of such macroscopic classical waves can often last for a substantial amount of time even at room
temperature, as observed in radio waves in wireless communication or sound waves in a music
5
hall.
In this paper, we will construct a computing model with classical waves. Let us start the
discussion on how to take advantage of the classical wave features in computing by defining,
in analogy to the qubit in quantum computing, the cubit, which is an abbreviation of classical
universal bit. The notation used for cubits is similar to the standard Dirac notation for qubits but
with double bras and kets. Let 0ii and 1ii be normalized orthogonal basis vectors for cubits:
hh00ii = hh11ii = 1,
hh01ii = hh10ii = 0.
(1)
(2)
The scalar product is defined as a conjugate integral in a predefined volume V , which can be
flying or standing as discussed later. States with higher indices and harmonics for cubits can be
considered similar to qubits 26, but we will not go further for simpler comparison. Using
hhx1x2ii = δ(x1 − x2)
enables the actual waveform in space time to be given as
f0(x) = hhx0ii ,
f1(x) = hhx1ii .
Therefore, the scalar product can be expressed as an integral in space coordinates:
hh01ii = Zx∈V
f ∗
0 (x)f1(x)dx.
6
(3)
(4)
(5)
(6)
The difference between bra or ket may not matter much if the waves are defined in R. Complex
conjugate ∗ should be used when dealing with complex numbers. An arbitrary cubit state
aii = ¯a0ii + a1ii
can have a wave form in space time as
a(x) = ¯af0 + af1.
(7)
(8)
Cubits can be flying or standing depends on whether the basis vector 0ii and 1ii and the
associated volume V is flying or standing. The difference in the mathematical formulation is not
as large as that in the implementation, given that it is only a matter of which coordinate to select, as
illustrated in Fig. 2. In other words, when a flying cubit and associated V are moving at a constant
velocity v ideally with little dispersion, the flying cubit can be considered standing when viewed
from the coordinate attached to it. Assuming v ≪ c (i.e., nonrelativistic) with no dispersion, an
arbitrary flying cubit state a(x, t)iif can be related to a standing one a(x, t)iis as
a(x, t)iif = Z a(x − x′, t)iis δ(x′ − vt)dx′.
(9)
The use of slow v is essential to exploit wave nature with energy efficiency and integration 14.
In QC systems, qubits are often standing. Flying qubits are diffusive except for massless qubits,
i.e., those flying at the speed of light. On the other hand, spikes in the brain are more consistent
with flying cubit picture. Encoding information into standing cubits may be related to digital
recording techniques, such as partial response maximum likelihood (PRML) encoding 27. While
those techniques are considered as taking advantage of standing wave nature of recording signals
7
with the finite frequency response of the media for memory and storage, the brain is considered as
using flying wave nature of spike signals for computing in the present model.
There are multiple types of logical cubits:
Normalized full cubit
Normalized half cubit
Unnormalized full cubit
Unnormalized half cubit.
aii := ¯a0ii + a1ii ,¯a2 + a2 = 1 ∈ U(1) or SO(2)
aii := a1ii , 0 ≤ a2 ≤ 1
aii := ¯a0ii + a1ii
aii := a1ii
∈ U(1) ∩ R
∈ R2 or C
∈ R
Cubits are defined not in the Bloch sphere (SU(2) or SO(3)) like qubits. The 0ii amplitude
for the half cubits is implicit and may result in more efficient gate implementation if 0ii can be
regarded as the ground state (i.e., no excitation).
There are several ways to encode logical cubits into physical waves. Specific examples are
shown in Figs. 2 (a) -- (d). Examples (a) and (c) are for half cubits, and examples (b) and (d)
are for full cubits. Examples (a) -- (c) are with single wires, while example (d) is with a spatially
distinguishable wire pair. In other words, examples (a) -- (c) require a single lane per cubit, while
example (d) requires dual-rail encoding, i.e., two physical lanes to encode a single logical cubit.
The unnormalized half qubit could be ill defined without dual rail coding, since the amplitude
of 0ii cannot be estimated from that of 1ii. Example (b) is identical to phase shift keying in
wireless communication, which is typical with a carrier. Example (a) can be considered as the real
part of example (b) though (a) can be encoded without the carrier as well. Logical information can
be encoded into multiple physical cubits by using majority logic coding, such as rate or population
8
coding, or by using temporal interval coding on a continuous non-discrete time scale.
Note that the present cubit formulation is mostly for computing, not for physics. We inherited
the standard bra and ket notation from quantum physics to describe cubits in a consistent manner
with qubits. If we dare to approximately relate cubit states to boson qubit ensemble states by taking
the classical limit of the occupation numbers for 0i and 1i, i.e., n0, n1 → ∞, and by neglecting
the off-diagonal terms, we get
0ii ∼ phn0i,
1ii ∼ phn1i.
(10)
(11)
Therefore, cubits may be regarded as population coded qubits in ensemble average. This relation
is to be revisited later for both bosons and fermions.
3 Operations for cubits
QIC can have various gate primitives for linear and nonlinear operations. Linear operations are
natural outcomes of wave superpositions, and the corresponding primitives are coupler with full
cubits for reversible operations, and adder or combiner (and coupler with half cubits) for irre-
versible operations, respectively. Splitter is the reverse operation of combiner. The fixed multiplier
with the constant W for unnormalized cubits has to be either lossy (W < 1) or active (W > 1).
Nonlinear operations such as wave multiplication are also possible. These operations for classi-
cal waves may look familiar to those with wireless communication backgrounds since they are
typically used for splitting and combining, and for up- and down-conversion of wireless signals.
9
Integrate and fire operation makes discrete binary decisions of either fire 1ii or not fire 0ii for
normalized half qubit out of non-discrete continuous states aii.
Figure 3 compares linear and nonlinear operations for single- and multiple-input gates with
bits, cubits, and qubits in classical, wave-based, and quantum computing. The mathematical space
corresponding to the number system of choice for each basic unit is represented by T . For example,
T = SU(2) for qubits. For classical computing, NOT and XOR gates perform linear operations
over Cartesian product states T ⊕ T while AND and OR gates perform nonlinear operations. Fred-
kin or Toffoli gates are listed as classical reversible gates 28. The most general form of a nonlinear
operation can be represented as a memory, for which any output can be defined at the expense of
2N resources for N address inputs. QC gates perform linear operations in SU(2) or in its tensor
product space, T ⊗ T. The measurements nonlinearly reduce qubits in the tensor product states to
ordinary bits in the Cartesian product states, though they are still projective linear operations in the
original tensor product states, but nonlinearity can arise if multiple qubit measurement results are
reduced in Cartesian product states. Measurements in QC can be related to integrated and fire in
QIC as a decision process of superposed states to one of the orthogonal states.
4 Log-scale encoding to bridge QC and QIC
Let us consider a decision problem with each decision represented by a qubit: pi = h1pii2 and
pi = 1 − pi = h0pii2. In other words,
Inii = pii = ¯pi 0i + p 1i .
(12)
10
When the state consists of a tensor product state of independent (i.e., no entanglement) qubit states
as p0i ⊗ p1i ⊗ ... ⊗ pni, the probability P{d1...dn} of having a decision {d1...dn} with di = 0 or 1
is
P{d1...dn} =
n
Yj=1
h1pjidj h0pji1−dj 2.
When we take the log of both sides,
log P{d1...dn} = 2
n
Xj=1
log h1pjidj h0pji1−dj .
When wij identical qubits are involved in decision Pi, the equations become
and
Pi{d1...dn} =
n
Yj=1
h1pjidj h0pji1−dj 2wij .
log Pi{d1...dn} = 2
n
Xj=1
wij log h1pjidj h0pji1−dj .
Note that no cloning theorem does not allow the duplication for arbitrary qubit states.
(13)
(14)
(15)
(16)
The following log relationship can translate Tensor-product qubit to Cartesian-product cubit
as:
hh1Outiii ∼ log Pi{d1...dn},
hhdiIniii ∼ log h1piidi h0pii1−di .
(17)
(18)
In other words, the probability in a cubit are interpreted as a log scale encoding of the probability in
a qubit. Log-scale encoding is a standard technique in probability calculations, such as used in log
likelihood estimations. It simplifies probability calculations for a wide dynamic range inputs and
11
seems consistent with that fact that the biological brain can compute with a wide dynamic range of
sensory signals 29. The encoding with excitatory and inhibitory neurons for p and p in a dual-rail
manner is possible. The log scaling encoding can also provide a natural rectifying capability with
appropriate biasing bi as
hh1Outiii = 2
n
Xj=1
wij hhdjInjii + bi.
(19)
When the probabilities are normalized per unit time interval, they corresponds to the spike rate
which is the inverse of the time-to-spike interval 30. The neuron dynamics in this model is driven
by the digital number of replicated spike paths rather than the analog strength of each spike and its
synaptic weight. The log-scale encoding can significantly reduce the number of spikes for a given
signal dynamic range. When cubits are flying, Eq.19 can be expressed, by using Eq. 9, as
Outi(t)ii = Z Xj
Wij(t′) Inj(t − t′)ii dt′ + bi,
(20)
by substituting Wij(t) = 2wijδ(t − dij), where dij is the delay between node i and j and is related
to the path length Lij by dij = Lij/v.
Comparison of cubit and qubit operations for a binary neuron as an example is shown in
Fig. 4. Probabilities of cubits and qubits for a binary neuron can be inter-related by log encoding.
The present model describes inferencing. Eventually, weight update rules for learning has to be
worked out, ideally based on direct temporal correlations of spikes (i.e., cubits) rather than indirect
formulations based on rate coding 31, 32.
QIC can become a superior one for non-quantum applications, so should not always be con-
sidered as an inferior version of QC. Indeed, the approach is aiming at expanding the applica-
12
tion coverage of QC-like computing by appropriately loosened constraints of QC without much
affecting its advantages. This quantum-inspired approach can be interpreted as depth=1 noisy
intermediate-scale quantum computing (NISQ) 33 with ensembles of flying qubit inputs as spikes
in density matrices. QIC for neural operations interpreted as ensembles of parallel and concurrent
QC operations is shown in Fig. 4 (a). It can support nonbinary weights by the cubit inputs as
ensembles of concurrent and parallel qubit inputs:
IniihhIn ∼ ntral(¯p 0ih0 + p 1ih1).
(21)
A calculated result for ensembles as a function of ntrial is shown in Fig. 4 (b). Each trial consists
of ten parallel qubit measurements. The variance is expected to converge as ∼ 1/√ntrial for cubits
by concurrently executing entire trials in parallel. This concurrent and parallel trial is valid for
both bosons and fermions, and we can write as
n0 = ntrial ¯p,
n1 = ntrialp.
(22)
(23)
As summarized in Fig. 4 (c) QIC distinguish itself from classical computing in terms of paral-
lel execution and superposed states and from QC in terms of log-scale encoding and concurrent
execution. Since cubits are classical consisting of ensembles of orthogonal qubits, they can be
duplicated to constitute nonbinary weights without being constrained by the no cloning theorem.
Concurrent ensembles of try and measure executions can significantly accelerate and scale a wide
range of less entangled AI workloads without QC-unique constraints even under noises in data
and environment. Thus, the present QIC approach has a good potential to deliver much better
throughput and scalability for the main-stream AI workloads than strict QC counterpart.
13
When these primitives are combined, neuron models to describe an abstract level PHY can be
constructed, as exemplified in Fig. 4 (a). It consists of wave processes in continuous time model.
The output activation is triggered by input correlations, not input-output (scattering model). It can
accept both excitatory and inhibitory inputs. Weight matrix is expressed by Fredkin gate as a cou-
pler with weight adjustment, details of which are described in Appendix A. Fredkin gate can output
the leftover wave energy from the other port. This otherwise wasted wave energy would be useful
for weight update with learning. Simulation results is shown in Fig. 4 (b) using a neural channel
and circuit modeling technique discussed in another paper 34, 35. When the leakage parameter of
integrate and fire block is appropriately set, spike temporal correlations can be extracted by appro-
priately adjusting the leakage time constant of the membrane potential. When ∆d = di+1 − di =
200 ps (upper), spikes from multiple inputs are less temporally correlated, and thus output does
not fire. When ∆d = 100 ps (lower), spikes are more temporally correlated and output fires.
5 Discussion
We have shown that computing with cubits can be performed in parallel by using superposed
states, similar to qubits. Thus, here we address the differences between QC and QIC, starting
with the advantages of QIC for classical computing tasks. In short, cubits are better fit for ac-
celerating general classical algorithms. Using cubits can avoid the complications associated with
reversibility, cloning, and measurement. This is important given that many real-world algorithms
(e.g., those for AI tasks) are not reversible and require copying of data. in addition, there is less
overhead for data IO when using cubits since information in classical domain does not have to be
14
converted into information in quantum domain in Tensor-product states with a smaller number of
qubits. The dephasing time for classical waves is often much longer even at room temperature,
which is also the case for spike signals in the biological brain. The use of cubits thus offers the
potential for deeper circuits with less error correction overhead. In contrast, QIC is not suitable for
quantum information processing. No exponential tensor-product linear state spaces are natively
available, and nonlocal entanglement features are missing. Even though both cubits and qubits can
be reduced to the same classical bit operations, they behave differently in their own mathematical
spaces. This makes it fundamentally difficult to use cubits to emulate qubits for quantum informa-
tion processing. The following arguments in Appendix B may provide a better understanding of
this situation by illustrating the extent to which cubits can be used to perform algorithms involving
entanglement. More rigorous formulation of QIC may provide a new perspective on computational
complexity arguments 36.
Our approach is not as extreme as reversible computing 37 in terms of energy per computation
limit, since it allows for and asks for dissipatively losing non-essential information. Thus, it can
alleviate logical complications associated with stringent reversible computing constraints and more
suitable for inherently irreversible AI workloads. We exploit elastic waves rather than matters (i.e.,
not billiard balls) as information carrier for stable solid-state implementation. The use of orders of
magnitude slower waves than the speed of light is essential for energy efficiency and integration.
Our conjecture is that this is what the biological brain has already taken advantage of with spike
signals for energy efficiency.
15
Compared with existing neuromorphic computing approaches, our approach can better incor-
porate time as a resource for computation and communication, which is considered essential for
SNNs 22. Otherwise, though spike signaling can improve tolerance to noise and disturbance includ-
ing baseline wonder, simple rate coded spiking neural networks would result in exponentially less
efficient than binary coded analog neural networks in terms of coding efficiency in neural channels
35. Thus, though many prior art neuromorphic research results claim energy efficiency, in reality
the total power reduction would be limited when spiking IOs are inefficiently coded. We expect
that, in addition to coding efficiency, supporting temporal coding, such as interpulse-interval cod-
ing 38, natively in continuous time with desired stochasticity without being affected by undesired
temporal jitters due to signal collision and congestion can provide a new perspective. Importance
of temporal delay in biological brains has also addressed in neuroscience 39.
The proposed QIC should not be considered the same as standard analog signal processing.
The underlying physical processes are quite distinctive. The former is elastic and nondissipative
while the latter is often inelastic and dissipative.
In other words, the signal is represented by
nonequlibrated wave dynamics, rather than equilibrated potential. It can be considered as a minia-
turized version of computing with radio frequency (RF) passives, exploiting wave-based signal
processing features, such as superposition, naturally without signal collision nor unwanted leak-
age. Signals are physically and logically directive and nonblocking to each other thanks to the
wave nature. As was mentioned earlier, it is essential to reduce the wavelength of the spike signal
with respect to the transmission length by reducing the spike signal velocity. This aspect has not
been addressed much in conventional analog neuron models.
16
QIC can become a fair candidate for the microscopic computing model for the artificial brain,
in particular considering to realize SNNs with a good temporal degrees of freedom. It also provides
a new perspective on the microscopic computing model for the biological brain whether quantum
effects are mandatory to explain brain computing functionalities, in addition to quantum effects in
biochemistry perspectives. From the computing model point of view, entanglement aspects would
be a key to identify whether quantum effects are mandatory in order to explain brain functionalities.
Further arguments on entanglement with QIC is given in Appendix B.
6 Conclusion
We proposed quantum-inspired computing as energy- and computationally-efficient microscopic
computing model for AI workloads in analogy to quantum computing. Parallel computing advan-
tage with superposed states will be incorporated in a more general natural computing context with
less constraints. Log-scale encoding of information is shown to play a critical role in bridging the
computing models of classical and quantum waves. Our goal is to redefine AI computing model
with quantum-inspired approach and ultimately unify the computing model of AI and QC, hoping
that this path will bring us to better understand the microscopic computing model of the brain.
Appendix A: Implementation sketch
There are various candidates to implement cubits and their operations, such as spin waves, acoustic
waves, slow optics, and ion density waves. Here, we show implementation sketch for key primitive
building blocks using nanostructured electronics in Fig. 6. It may have a better path to be migrated
17
in existing semiconductor technologies, since no signal wave conversion is needed. The purpose
is to show that the proposed computing model has a reasonable path toward implementation, but
further study is needed to identify the most suitable way of realizing the device. In Fig. 6 (a),
weight matrix diagram is represented by Fredkin gate for cubits. It has three inputs:
for full cubits and
cii = ¯c0ii + c 1ii ,
pii = ¯p 0ii + p 1ii ,
cii = ¯q 0ii + q 1ii
cii = c 1ii ,
pii = p 1ii ,
cii = q 1ii
for half cubits. Note that cii signal may not be spiking. The outputs are defined as
for full cubits and
cii = ¯c0ii + c 1ii ,
¯c pii + c qii = (¯c¯p + c¯q) 0ii + (¯cp + cq) 1ii ,
¯c qii + c pii = (¯c¯q + c¯p) 0ii + (¯cq + cp) 1ii
cii = c 1ii ,
¯c pii + c qii = (¯cp + cq) 1ii ,
18
(24)
(25)
(26)
(27)
(28)
(29)
(30)
(31)
(32)
(33)
(34)
¯c qii + c pii = (¯cq + cp) 1ii
(35)
for half cubits. As indicated in these equations, both superposition (addition) and multiplications
are involved in the cubit Fredkin gate. The Fredkin gate is a universal classical gate with constant
ancilla inputs 28. Fredkin gate is more suitable than Toffoli gate for our purpose since the number
of conserved 0's and 1's between inputs and outputs can lead to more efficient passive gate imple-
mentations for better energy efficiency. Normalized half cubits with 0ii for no signal and 1ii
for a full signal can be assumed. With this assumption, superposed cubit states can be represented
by the scalar strength of 1ii. Because the Fredkin gate conserves the number of 0's and 1's,
the gate can be implemented passive wave couplers without active amplification. With an active
amplification option, the gate can generate signal gain for a larger fanout assuming unnormalized
cubits. When Fredkin gates are combined, the weighted sums of any number of inputs and outputs
can be flexibly calculated by using analog weighted-sum calculation. The data cubits pii and
qii for the Fredkin gate are repartitioned in accordance with instructions from the control cubit
cii. When the blocks of the gate structure are combined, they can represent any repartitioning
across a large number of data cubits as a unitary (flux conserving) transformation.
Figure 6 (b) depicts Fredkin gate with delay for cubits with nanostructured electron wave
guides and utilizes lateral tunneling transistor structure 40 -- 42, with energy diagram for lateral tun-
neling transistor in Fig. 6 (c). Here, single-particle electron quantum wave functions are treated
as Cartesian product classical waves in computing. The gate illustrated in Fig. 6 is replicated in
parallel 43 using the lateral tunneling transistor structure. The device consists of coupled quantum
structures, such as wells and wires, and an alignment gate. Without gate bias, the subband edges of
19
the coupled quantum wells and wires are not aligned and thus are isolated. Application of an align-
ment gate voltage adjusts the subband edge alignment. When the two subband edges are perfectly
aligned, we can make this work as 50:50 splitter/combiner by appropriately choosing the design
parameters 34. The device is driven by quantum physics in z direction 42 but by classical physics
in xy direction for ensemble averaging. The structure can be replicated for nonbinary weights as
expected in Fig. 4. The concurrent averaging can be done by using the on-off encoding illustrated
in Fig. 2 (a) such as by using 0i for "without electrons" and 1i for "with electrons." The spatially
parallel and temporarily concurrent measurement capability for the ensemble averaging is a unique
capability of cubits in comparison with the measurements for qubits, which have to occur serially
with no cloning. This feature should be useful in handling large fanout connections as is done by
the brain. By combining blocks with this structure, a Wij larger than 2 × 2 can be constructed.
Weight update can be performed by changing the gate voltages. The effect of leakage current on
the on-off ratio is a serious problem for the building blocks in classical computing but not in QIC
with classical waves.
Figure 6 (d) shows integrated and fire diagram with combiner and splitter. Integrate and fire
building block with combiner and splitter with quantum well and wires is shown in Fig.6 (e) with
energy diagram for the integrate and fire building block in Fig. 6 (f). Details of this block has been
presented elsewhere 34.
20
Appendix B: Remarks on entanglement
Finally, let us spend some time to discuss how much QIC can be closer to QC. Multiple cubit
states can be represented in tensor products by using multipliers. For example, the Walsh function
can be used to represent such tensor product states as those illustrated in Fig. 6. For simplicity,
it is assumed that the physical pulses have square shapes in an unlimited bandwidth environment.
Entangled states such as
1
√2
(0ii0ii + 1ii1ii)
(36)
can be generated as a linear combination of the tensor product states. Since this is a local entangle-
ment, it differs from the nonlocal entanglement in qubits. Tensor product states can be constructed
in spatial degrees of freedom as well. Since the brain occupies relatively small amount of the
space, this local entanglement feature can provide a similar feature with nonlocal quantum entan-
glement 8. Fig. 6 shows four states, αii = 0ii, βii = 1ii, γii = 1/√2(0ii + 1ii), and
δi = 1/√2(−0ii + 1ii), created by using the phase encoding in Fig. 2 (b), presumably with a
carrier. Given these four states,
D(αii ,γii) − D(αii ,δii) + D(βii ,γii) + D(βii ,δii) = 2√2,
(37)
when D(xii ,yii) = hhxyii as is commonly used in wireless coherent detection algorithms 44.
This argument may suggest that this relation is arising from wave nature, not necessarily quantum
nature.
Bibliography
21
1. Ricciardi, L. M. and Umezawa, H. Brain and physics of many-body problems. Kybernetik 4
44 -- 48 (1967).
2. Stuart, C. I. J. Takahashi, Y. Umezawa, H. On the stability and non-local properties of memory.
J. Theor. Biol. 71 605 -- 618 (1978).
3. Stuart, C. I. J. Takahashi, Y. Umezawa, H. Mixed-system brain dynamics: Neural memory as a
macroscopic ordered state. J. Found. Phys. 9 301 -- 327 (1979).
4. Hameroff, S. Penrose, R. Orchestrated Objective Reduction of Quantum Coherence in Brain
Microtubules: The "Orch OR2" Model for Consciousness. Toward a Science of Consciousness
- The First Tucson Discussions and Debates MIT Press, Cambridge, MA 507 -- 540 (1996).
5. Pribram K. H. Quantum holography: Is it relevant to brain function? Information Sciences 115
97 -- 102 (1999).
6. Alfinito, E. Vitiello, G. The dissipative quantum model of brain: How does memory localize in
correlated neuronal domains. Information Science 128 217 -- 229 (2000).
7. Rebentrost, P., Mohseni, M., and Lloyd, S. Quantum support vector machine for big data clas-
sification. Phys. Rev. Lett. 113 130503 (2014).
8. Biamonte, J. Quantum machine learning. Nature 549 195-202 (2017).
9. Havlicek V., et al. Supervised learning with quantum-enhanced feature spaces. Nature 567 209 --
212 (2019).
10. Grush, R. and Churchland, P. S. Gaps in Penrose's toilings. J. Conscious. Stud. 2 10 -- 29 (1995).
22
11. Tegmark, M. Why the brain probably not a quantum computer. Inform. Sci. 128 155 -- 179
(2000).
12. Litt, A. et al. Is the Brain a Quantum Computer? Cognitive Science 30 593 -- 603 (2006).
13. Katayama, Y. Yamane, T. Nakano, D. An Energy-Efficient Computing Approach by Filling
the Connectome Gap. Unconventional Computation & Natural Computation 229 -- 241 (2014).
14. Katayama, Y, et al. Wave-Based Neuromorphic Computing Framework for Brain-Like Energy
Efficiency and Integration. IEEE Trans. Nanotechnol. 16 762 -- 769 (2016).
15. Tang, E. A quantum-inspired classical algorithm for recommendation systems. ACM STOC
(2019).
16. IBM Research Blog, AI Year in Review: Highlights of Papers and Predictions from IBM
Research AI (2018) https://www.ibm.com/blogs/research/2018/12/ai-year-review/.
17. Preskill, J. Quantum computing and the entanglement frontier. 25th Solvay Conference (2012).
18. Leggett, A. J. Macroscopic Quantum Systems and the Quantum Theory of Measurement. Prog.
Theor. Phys. 69 80 -- 100 (1980).
19. Friedmann, J. R. et al. Quantum superposition of distinct macroscopic states. Nature 406 43 --
46 (2000).
20. Feynman, R. P. Simulating physics with computers Int. J. Theor. Phys. 21 467-488 (1982).
21. Bennett, C. H. Quantum information and computation. Nature 404 247-255 (2000).
23
22. Maass, W. Networks of Spiking Neurons: The Third Generation Neural Network Models
Neural Networks 10 1659 -- 1671 (1997).
23. Indiveri, G. and Liu, S.-C. Memory and information processing in neuromorphic systems Proc.
IEEE 103 1379 -- 1397 (2015).
24. Merolla, P. A. et al. A million spiking-neuron integrated circuit with a scalable communication
network and interface Science 345 668 -- 673 (2014).
25. Ambrogio, s. et al. Equivalent-accuracy accelerated neural-network training using analogue
memory Nature 558 60 -- 67 (2018).
26. Neeley, M. Emulation of a Quantum Spin with a Superconducting Phase Qubit. Science, 325
722 -- 725 (2009).
27. Kobayashi, H. Tang, D. T. Application of Partial-Response Channel Coding to Magnetic
Recording Systems. IBM Journal of Research and Development 14, 368 -- 375 (1970).
28. Fredkin, E. Toffoli, T. Conservative Logic. Int. J. Theor. Phys. 21 219 -- 253 (1982).
29. Dehaene, S. The neural basis of the Weber-Fechner law: a logarithmic mental number line.
TRENDS in Cognitive Science 7 145 -- 147 (2003).
30. Singh, C, Levy, W. B. A consensus layer V pyramidal neuron can sustain interpulse-interval
coding. PLOS ONE 12 (2017).
31. Hinton, G. E. How to do backpropagation in a brain. NIPS Deep Learning Workshop (2007).
24
32. Bengio, Y. et al. STDP-Compatible Approximation of Backpropagation in an Energy-Based
Model. Neural Computation 29, 555 -- 577 (2017).
33. Preskill, J. Quantum computing in the NISQ era and beyond. Quantum 2 79 (2018).
34. Katayama, Y. Wave-Based Spiking Neural Networks with Nano-Structured Electronics. IEEE
NANO (2019).
35. Katayama, Y. Channel Model for Spiking Neural Networks Inspired by Impulse Radio MIMO
Transmission. IEEE GLOBECOM to be presented (2019).
36. Berstein, E. and Vazirani, Z. Quantum complexity theory. SIAM J. Comput. 26 1411-1473
(1997).
37. Bennett C. H. The Thermodynamics of Computation -- A Review. Int. J. Theor. Phys. 21 905 --
940 (1982).
38. Singh. C and Levy, W. B. A consensus layer V pyramidal neuron can sustain interpulse-interval
coding. PLOS ONE https://doi.org/10.1371/journal.pone.0180839 (2017).
39. O'Keefe, J. et al., "Place cells, navigational accuracy, and the human hippocampus". Philos.
Trans. Royal Soc. B: Biol. Sci. 353 1333 -- 1340 (1998).
40. Katayama, Y. Tsui, D. C. Lumped circuit model of two-dimensional to two-dimensional tun-
neling transistors. Appl. Phys. Lett. 62 2563 -- 2565 (1993).
25
41. Simmons, J. A. et al. Unipolar complementary bistable memories using gate-controlled neg-
ative differential resistance in a 2D-2D quantum tunneling transistor. IEEE IEDM, 755 -- 758
(1997).
42. Katayama, Y. New Complementary Logic Circuits Using Coupled Open Quantum systems.
IEEE Trans. Nanotechnology 4 527 -- 532 (2005).
43. Katayama, Y. Physical and Circuit Modeling of Coupled Open Quantum Systems. IEEE
NANO -- 5th IEEE Conference on Nanotechnology (2005).
44. Tse, D. Viswanath, P. Fundamentals of Wireless Communication (Cambridge University Press,
2005).
Acknowledgements The author is grateful to colleagues inside and outside IBM Research for their valu-
able discussions.
Competing Interests The author declares no competing interests.
Correspondence Correspondence and requests for materials should be addressed to Yasunao Katayama (email:
[email protected]).
26
Figure 1 Approaches for physical computing models of the brain.
In the proposed
approach, quantum-inspired approach for AI is investigated, ultimately aiming at unifying
AI and QC computing models with classical and quantum waves.
Figure 2 Flying vs. standing cubits. Theoretically the distinction may be superficial since
it is just a matter of the coordinate selection. Slower v results in smaller spatial area vσ
for each spike wave packet with the temporal width of σ and thus better integration and
energy efficiency 14. The line v = c represents the light cone in the Einstein-Minkowski
space time. The original spatial width at standing is ignored for simplicity.
Figure 3 Physical cubit encoding examples: (a) on-off encoding, (b) phase encoding,
(c) temporal encoding, (d) dual-rail encoding. To be more consistent with biological spike
signaling, RZ rather than NRZ is assumed.
Figure 4 Comparison of linear and nonlinear operations for single- and multiple-input
primitives in classical, quantum-inspired, and quantum computing. T represents mathe-
matical space corresponding to the number system of choice for each basic unit.
Figure 5 Comparison of cubit and qubit operations for an artificial neuron with binary
weight matrices in Cartesian and Tensor product states, respectively.
Figure 6 (a) Neural operations in QIC interpreted as ensembles of parallel and concur-
rent QC operations; (b) Calculated result for ensembles as a function of ntrial; (c) QIC
27
distinguish itself from classical computing in terms of parallel execution and superposed
states and from QC in terms of log-scale encoding and concurrent execution.
Figure 7 (a) The diagram of an example artificial neuron based on the proposed com-
puting model; (b) Simulation results with ∆d = 200 ps and 100 ps for upper an lower
graphs, respectively. By appropriately adjusting the leakage time constant of the mem-
brane potential MB, input spike temporal correlations can be extracted.
Figure 8 Implementation sketch for key primitive building blocks: (a) Weight matrix di-
agram represented by Fredkin gate for cubits; (b) Fredkin gate with delay for cubits with
nanostructured electron wave guides and utilizes lateral tunneling transistor structure 40;
(c) Energy diagram for lateral tunneling transistor when the subband edges are misaligned
and decoupled (left) and aligned and coupled (right); (d) Integrated and fire diagram with
combiner and splitter; (e) Integrate and fire building block with combiner and splitter with
quantum well and wires; (f) Energy diagram for the integrate and fire building block.
Figure 9 Cubit states are encoded into classical waves using the Walsh function. Tensor
product states and local entanglement can be generated.
Figure 10 Coherent detection of γii and δii with in-phase αii and quadrature βii.
28
AI
Inferencing
W
W
W
Learning
Present
approach
QC
Reversible
U
U
Conventional
approaches
(quantum
machine
learning)
e
d
u
t
i
l
p
m
A
Flying
Slow v
Fast v
v = c
No signal
can reach
Space
~vs
On-off
(half cubit)
Complex
phase
Temporal
phase
Dual rail
Basic unit
Linear
reversible
Classical
bit
NOT(T→T)
Fredkin/Toffoli
(T⊕T⊕T→T⊕T⊕T)
Quantum-inspired
cubit
Delay (T→T)
d
Coupler (T⊕T→T⊕T)
Quantum
qubit
All gates
(T→T)
(TⓧT→TⓧT) …
U
Linear
nonreversible
XOR(T⊕T→T)
(with mod 2)
Adder/combiner/splitter
(T⊕T→T)
Fixed multiplier (T→T)
W
Nonlinear
AND/OR(T⊕T→T)
Multiplier (TⓧT→T)
Measurement
(TⓧT→T⊕T)
Memory(T⊕T→T)
00
01
10
11
0
1
1
0
Integrate & fire
(T →T)
Binary weight matrix
Quantum PNC gate
Positive Negative Don't
control
control
care
(a)
Depth = 1 (effectively)
(b)
t
u
p
n
i
t
n
e
r
r
u
c
n
o
C
a
l
l
i
c
n
A
i
)
s
a
b
r
o
(
(c)
Classical
computing
Quantum-
inspired
computing
(QIC)
Parallel
Superposition
Concurrent
Log encoding
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
l
e
b
m
e
s
n
E
1
2
(cid:3)0⟩ (cid:7) 1⟩(cid:8)
0 1 2 3 4 5 6 7 8 9 101112
n
trial
i
S
n
k
C
o
n
c
u
r
r
e
n
t
o
u
t
p
u
t
Quantum computing (QC)
Depth=1
NISQ
Noisy intermediate-scale
quantum computing
(NISQ)
Ideal quantum
computing
(a)
(b)
D d = 200 ps
W
1-W
d1
d2
d3
d4
Bias
D d
Not fire
D d = 100 ps
Fire
=
(d)
(e)
(a)
(b)
z
x
y
(c)
No tunneling
(isolated)
Tunneling
(coupled)
(f)
E
Integrate
Tunneling
(Fire)
Self reset
Subband
edge
z
0⟩⟩
1⟩⟩
0⟩⟩
1⟩⟩
0⟩⟩0⟩⟩
0⟩⟩1⟩⟩
1⟩⟩0⟩⟩
1⟩⟩1⟩⟩
0⟩⟩0⟩⟩ + 1⟩⟩1⟩⟩
1⟩⟩
(cid:6)⟩⟩
(cid:7)⟩⟩
(cid:8)⟩⟩
(cid:5)⟩⟩
0⟩⟩
|
1201.0339 | 2 | 1201 | 2012-03-26T08:28:01 | Synthetic reverberating activity patterns embedded in networks of cortical neurons | [
"q-bio.NC",
"nlin.CD"
] | Synthetic reverberating activity patterns are experimentally generated by stimulation of a subset of neurons embedded in a spontaneously active network of cortical cells in-vitro. The neurons are artificially connected by means of conditional stimulation matrix, forming a synthetic local circuit with a predefined programmable connectivity and time-delays. Possible uses of this experimental design are demonstrated, analyzing the sensitivity of these deterministic activity patterns to transmission delays and to the nature of ongoing network dynamics. | q-bio.NC | q-bio |
Synthetic reverberating activity patterns embedded in networks of cortical neurons
R. Vardi,1 A. Wallach,2 E. Kopelowitz,1, 3 M. Abeles,1 S. Marom,2 and I. Kanter1, 3
1Gonda Interdisciplinary Brain Research Center,
and the Goodman Faculty of Life Sciences, Bar-Ilan University, Ramat-Gan 52900, Israel.
2 Network Biology Research Laboratories, Technion - Israel Institute of Technology, Haifa 32000, Israel.
3Department of Physics, Bar-Ilan University, Ramat-Gan 52900, Israel.
Synthetic reverberating activity patterns are experimentally generated by stimulation of a subset
of neurons embedded in a spontaneously active network of cortical cells in-vitro. The neurons are
artificially connected by means of conditional stimulation matrix, forming a synthetic local circuit
with a predefined programmable connectivity and time-delays. Possible uses of this experimen-
tal design are demonstrated, analyzing the sensitivity of these deterministic activity patterns to
transmission delays and to the nature of ongoing network dynamics.
I.
INTRODUCTION
The assumption of isomorphism between behavior and temporal patterns of reverberating synchronized neural
electrical activity stands at the basis of many large scale theories of the brain [1, 2]. Actual measurement and controlled
experimental manipulation of such temporal activity patterns are associated with severe difficulties and most of what
is presently known comes from theoretical and numerical studies of simplified components, which are very remote
from the full dynamical richness of neural network entities [3 -- 8]. Here, we artificially generate temporal patterns
of reverberating synchronized neural electrical activity by conditioned sequential stimulation of a circuit of neurons
embedded within a large scale, spontaneously active network of cortical cells in-vitro [9 -- 11]. Cortical neurons were
obtained from newborn rats within 24 h after birth using mechanical and enzymatic procedures described in earlier
studies [9 -- 15]. The neurons were plated directly onto substrate-integrated multi-electrode arrays (60 Ti/Au/TiN
extracellular electrodes) and allowed to develop functionally and structurally mature networks over a time period of
2-3 weeks [see material and methods section].
II. EXPERIMENTAL SETUP AND THE GCD CLUSTERS
The basic experimental design consists of circuits with six different neurons that are artificially coupled to each
other (Fig. 1 left column). The coupling is realized in the following way: Neuron 1 is stimulated electrically, if
generating a spike within a neuronal response latency ±2 ms the next neuron in the artificial circuit is stimulated
after a fixed delay. If the next neuron fires within a neuronal response latency ±2 ms its connecting neurons are
stimulated, etc . . . Spikes that happened during the delay are ignored. Care is taken to use recording electrodes
that exhibit well-formed identified spikes that are selectively and reliably evoked by a spatially unique stimulation
source. Consider, for instance, the case of the directed circuit shown in the left column of Fig. 1a, where a stimulus
is applied in a site that evokes one spike recorded in electrode number 1. Conditioned to the detection of a spike in
electrode number 1, a stimulus is applied after a fixed delay to a site that evokes a spike detected in electrode number
2 and so on, following the connectivity of the circuit. In this trivial realization, it is expected that as long as each
stimulus evokes a spike that is detected in the corresponding recording electrode, a chain of one spike per stimulation
cycle will be formed (Fig. 1a, middle and right columns). The other rows (Fig. 1b-e) show the results obtained by
different six neuron circuits. In each of these realizations, evoked patterns of activity relax to a theoretically expected
periodicity that equals the greatest common divisor (GCD) of the circuit loops, where the number of neurons that
fire in synchrony is exactly 6/GCD . For instance, the realization with 3 and 6 size loops (Fig. 1d) is governed by
periodicity equals to GCD(3,6)=3, three activity groups of zero-lag synchrony (ZLS), in each of which two neurons
fire simultaneously. The different examples of Fig. 1 demonstrate the sensitivity of the GCD, a global determinant
of periodic steady-state mode of activity, to the circuit's connectivity matrix. The resulting patterns of synchronized
periodic activities might seem counterintuitive, since the circuit consists solely of unidirectional connections (Fig. 1a-
d) and zero-lag neurons do not share the same input as in the relay mechanism where two neurons are unidirectionally
mastered by a third one. But as shown elsewhere [16, 17], the shared information in such graphs is a consequence
of the information mixing mechanism where the activation of a given node is determined by the activity of nodes in
different time steps. The mixing occurs during the transient trajectory, and varies between zero and 25 time steps in
the cases demonstrated here (Fig. 1a-b, right column).
2
(a)
5
4
6
3
init.
1
2
(b)
6
5
4
3
init.
1
2
n
o
r
u
e
N
(c)
5
4
3
6
init.
1
2
(d)
5
4
3
6
init.
1
2
(e)
5
4
6
3
init.
1
2
6
4
2
6
4
2
6
4
2
6
4
2
6
4
2
e
z
s
i
p
u
o
r
G
6
4
2
0
6
4
2
0
6
4
2
0
6
4
2
0
6
4
2
0
5
10
10
10
20
5
10
3
6
3
6
Time / 2 sec
5
10
10
20
5
10
3
6
3
Time / 2 sec
6
FIG. 1: (Colour on-line) Transients and reverberating activity patterns in six neuron circuits, using a conditioned stimulation
protocol following the circuit connectivity (arrows in the left column). All programmable delays set equal to 2 s and at
time zero neuron 1 fired. The second column shows spike trains of the six recording electrodes. In this column the abscissa
shows the experimental time, but the spikes are plotted on an expanded time scale of 2.5 ms per trace, indicating that we
related to clear spikes with good signal to noise ratio and with reliably repeating shapes. Some of the neurons rarely fired
also in between external stimuli (not shown). The third column sums up the number of simultaneously active electrodes.
Neurons that fire together in the reverberating phase are represented by the same colour (first column) and after the transient
trajectory of activity (second column). (a) A directed loop with 6 activity groups of one neuron each, and with the lack of a
transient. (b) GCD(5,6)=1 group of reverberating activity and 25 time steps transient trajectory. (c) GCD(4,6)=2 groups of
reverberating activity and 9 time steps transient trajectory. (d) GCD(3,6)=3 groups of reverberating activity and 5 time steps
transient trajectory. (e) A single directed loop with one bidirectional coupling forming a loop of size 2, GCD(2,6)=2 groups of
reverberating activity as in (c), but with 5 time steps transient trajectory.
III. RESPONSE FAILURE
Under conditions where the network spontaneous activity is made sparse using synaptic blockers, the stability
of our synthetically evoked reverberating activity patterns is largely determined by the response reliability of the
stimulated circuit neurons. While the response of a single cortical neuron to repeated stimuli is inherently noisy, a
reliable 1:1 response over extended time scales is achievable in our setup for 1-5 Hz stimulation frequencies, whereas
higher frequencies may lead to intermittent responsiveness [15]. Such response fluctuations compromise the stability
of activity patterns in the simple case of a chain of one spike per stimulation cycle (Fig. 1a). But in cases where
the reverberating pattern relies on several neurons that fire simultaneously, the pattern is resilient to intermittent
responsiveness of one neuron (Fig. 2).
One more critical factor (beyond the individual neuron response fluctuations) that destabilizes the maintenance of
reverberating activity patterns is considered: Each stimulus evokes a network response that relaxes on a time scale
3
6
4
2
6
4
2
0
(d)
(e)
(f)
40
Time / 2 sec
20
10
0
80
85
Time / 0.1 sec
90
(a)
(b)
(c)
n
o
r
u
e
N
i
e
z
s
p
u
o
r
G
p
e
t
s
1
0
/
.
i
e
k
p
S
6
4
2
6
4
2
0
20
0
35
FIG. 2: The robustness of synthetic reverberating activity patterns to transmission delay. The time of each spike is marked as
a dot. Spikes that resulted due to electrical stimulation are marked by large dots and the "spontaneous" spikes by small dots.
(a) Spike trains of the circuit 1c with 2 periodical modes. A response failure of the neuron labelled 2 at time step 37 is followed
by 5 time steps recovery period. (b) The number of electrodes with simultaneous activity in 2a. (c) Activity of the sixty
electrodes in 2a, using non-overlapping windows of 0.2 s, indicating that the network activity relaxes shortly after stimulation
of the circuit. (d) Spike trains of circuit 1b with transmission delay of 0.1 s where the reverberating mode overcomes response
failures of several neurons. (e) The number of electrodes with simultaneous activity in 2d. (f) Activity of the sixty electrodes
in 2d, using non-overlapping windows of 0.01 s, overlaps the time scale of relaxation from the evoked network response.
determined by the network properties [13, 14]. When the time scale of stimulation overlaps the time scale of relaxation
from the evoked network response, the stability of synthetic reverberating activity patterns is heavily challenged due
to correlations among response failures of the six neurons that form the embedded circuit. In Fig. 2 we demonstrate
the sensitivity of synthetic reverberating activity patterns to the relations between the two time scales. Having said
that, we often detected a "self-repairing" process, as the synthetic reverberating pattern overcomes response failures
of several neurons through a longer transient of recovery (Fig. 2d-e). While the general trend seems lawful (the lower
the GCD the stable the synthetic activity pattern), we were not able to define a simple rule that relates stability
to transmission delay and network response features; the parameter space is too complex, and includes multiple
spatiotemporal correlations at different levels of organization. Nevertheless, correlations between the single neuron
level (Fig. 2e) and the population level (Fig. 2f) are clearly demonstrated.
IV. SPONTANEOUS NETWORK ACTIVITY
The ongoing spontaneous activity of mature large scale cortical networks in-vitro is composed of synchronous events,
decorated with asynchronous sporadic activity of individual neurons in between these synchronies [10, 11, 13, 14, 18,
19]. This spontaneous ongoing activity may be made more synchronous by partial blockade of inhibitory synapses
[13, 14], or more asynchronous by activation of cholinergic receptors [18]. In what follows we examine the sensitivity of
the smallest homogeneous directed circuit leading to ZLS [16], to the nature of ongoing (spontaneous) network activity
using pharmacological manipulations. The circuit consists of four electrodes with 3 and 4 directed loops, GCD(3,4)=1
(Fig. 3a), and a stimulation to the electrode labelled 1 is leading to eight time steps transient trajectory to ZLS (Fig.
3b). In the first part of the experiment, Carbachol (20 µM), a modulator of cholinergic receptors, was added to the
4
network. As expected [18], Carbachol suppresses the synchronous mode of the network. The stability of ZLS with
transmission delay of 0.5 s was estimated by observing the activity periods in 20 trials. Specifically, we estimated
the "survival duration" of reverberating patterns by measuring the time to first occurrence where no evoked spikes
were detected in all four electrodes. The average survival duration is 43.7 s in the presence of Carbachol (Fig. 3c).
Figure 3d (left) shows that the above stability is maintained albeit significant intermittent responsiveness. The other
extreme pharmacological manipulation used here is addition of Bicuculline (5 µM), leading to a suppression of total
network activity, amplification of correlated bursting activity interleaved with long (around 20 s on average) periods
of practically complete quiescence. The stability of ZLS with transmission delay of 0.5 s was estimated as explained
above and shown in Fig. 3c. In this case, the bursts may block the synthetic reverberating pattern completely, leading
to 6.95 s average survival duration. As a result, most trials terminate before reaching the fully blown reverberating
pattern of activity. An atypical trial, where a reverberating pattern is arrived at, is shown in Fig. 3d (right). We
believe that the median survival duration with Bicuculline (1.5 s; Fig. 3c) on the background of the long average
periods of complete quiescence (20 s), reflects the emergence of burst activity simultaneously with the stimulation
activity of the circuit every 0.5 s.
(a)
init.
1
2
(c)
15
Carbachol
Bicuculline
4
3
10
l
s
a
i
r
T
#
5
5
Time / 0.5 sec
10
10 20
40
80
120
Survival duration (sec)
(b)
n
o
r
u
e
N
4
3
2
1
(d)
60
e
d
o
r
t
c
e
E
l
30
0
20
60
100
0
20
Time / 0.5 sec
60
Time / 0.5 sec
i
e
z
s
p
u
o
r
G
4
3
2
1
0
FIG. 3: Robustness of synthetic reverberating pattern to network spontaneous and burst activities. (a) The smallest ho-
mogeneous directed circuit leading to ZLS, four electrodes with 3 and 4 directed loops. (b) The transient trajectory to ZLS
where a stimulation is given to neuron labelled 1. (c) Histogram of sessions of 20 trials with transmission delay of 0.5 s. The
average/median survival duration with Carbachol (light blue) is 43.7/31 s and with Bicuculline (dark blue) 6.93/1.5 s. (d)
Raster plot of the activity of the sixty electrodes (black) and the sum of the intermittent responsiveness of the four electrodes
(violet). A typical trial with Carbachol leading to significant intermittent responsiveness (left), and an atypical trial with
Bicuculline where a burst activity leads to a sudden death of the synthetic reverberating pattern (right).
V. COMPLEX EXTERNAL STIMULI
The interplay between the GCD circuit loops and the selected synthetic reverberating pattern initiated by stimu-
lation to one electrode only, calls for generalization to stimulation given simultaneously to several electrodes. Fig. 4a
demonstrates that a circuit consisting of 3 and 6 directed loops, in which the initial stimulation is given to electrodes
labelled 1, 3 and 5 and the transmission delay is 2 s, results in ZLS, in spite of the fact that the GCD(3,6)=3. This
demonstration indicates that the number of groups is equal to the GCD of the spatial circuit loops and the periodicity
5
of the initial stimulation [16], i.e. GCD(2,3,6)=1 (Fig. 4a). One implication of the above spatiotemporally generalized
GCD rule for the number of reverberating groups, is that spontaneous or network-induced fluctuations in the size or
identity of zero-lag groups might lead to a phenomenon of switching between synthetic reverberating patterns: unex-
pected evoked (or omitted) spikes may be thought of as equivalent to initial stimulation through several electrodes
simultaneously. Indeed, we observed such transitions between long-lasting patterns (Fig. 4b).
VI. TRANSIENT TRAJECTORY
Since neural phenomena often occur on very short time scales, there seems to be an inherent temporal gap between
the relatively long transient trajectory to steady state pattern, and the postulate that computing and functionality of
the network are based on reverberating patterns. Moreover, it is clear that the igniting stimuli cannot be inferred from
the reverberating pattern as exemplified by the six neuron circuit (Fig. 4c) where all 63 different stimulations lead
to ZLS. An alternative computational view is akin to order-based representation [15, 20 -- 22], such that an inference
takes place in the transient trajectory of activity to the reverberating pattern, where different igniting stimuli result
in different or only partially overlapped transient trajectories (Fig. 4d-e). Under this computational framework,
the GCD based steady state pattern stands mainly as a marker for the end of transient trajectory activity. The
partial overlap between transient trajectories is otherwise unavoidable (each one of the stimulations points directly
to the steady state pattern), however it enables a probabilistic inference which is accumulated dynamically along the
trajectory, e.g. identifying a number or a subset of possible stimulated neurons (Fig. 4c).
VII. CONCLUDING REMARKS
There are several constraints, imposed by the experimental setup, that limit our capacity to generalize the above
results to the brain. For instance, each node of a synthetic reverberating pattern in our experimental design is
represented by a single neuron whose activation leads to a very strong excitation of the next neuron after an extremely
long delay. This is in sharp contrast to the size of a functional cortical entity. Moreover, the present setup does not
allow for mimicking a circuit with many members in each node and fluctuating delays among pairs of connecting
neurons, and ignored spikes that happened during the delays.
In an extensive set of simulation studies at the population dynamics level we demonstrate that when the above
constraints are relaxed, the general picture that emerges from our experimental analysis remains valid, e.g. Fig. 5.
Every neural cell was simulated using the well known Hodgkin Huxley model [23] with similar selected parameters
as in [16]. Synaptic background noise was simulated by a balance of input from 800 excitatory neurons firing at
about 1-3Hz, and 200 inhibitory neurons firing at about 50-100 Hz. The free parameters were set in a way such
that a single cell with no synaptic or external input fired randomly at about 5-7 Hz and the cell activity was noisy
around the resting potential with a variance of about 5 mV. The connection between excitatory neurons belonging to
different nodes was elected with probability 0.2 and time delays among connecting neurons are taken from a uniform
distribution between [29,31]ms. Our numerical explorations indicate that when a single circuit node is a population
of neurons, the signal to noise ratio of the steady state is enhanced with the population sizes, and the circuit activity
becomes less sensitive to background fluctuations. Nevertheless, the suitability of our neuronal circuits to describe
irregular firing patterns of cerebral cortical area [24, 25] needs a further investigation and might requires averaging
out the regular firing by looking at long recording times.
VIII. MATERIALS AND METHODS
A. Cell preparation
Cortical neurons were obtained from newborn rats (Sprague- Dawley) within 24 h after birth using mechanical
and enzymatic procedures described in earlier studies [9 -- 15]. Rats were euthanized by CO2 treatment according to
protocols approved by the National Institutes of Health. The neurons were plated directly onto substrate-integrated
multi-electrode arrays and allowed to develop functionally and structurally mature networks over a time period of
2-3 weeks. The number of plated neurons in a typical network is of the order of 1,300,000, covering an area of about
380 mm2. The preparations were bathed in MEM supplemented with heat-inactivated horse serum (5%), glutamine
(0.5 mM), glucose (20 mM), and gentamicin (10 g/ml), and maintained in an atmosphere of 370C, 5% CO2, and 95%
air in an incubator as well as during the recording phases. The experiments described in Figures 1, 2 and 4 were
conducted in the standard growth medium, supplemented with 50 µM APV (amino-5-phosphonovaleric acid), 6.25
6
(a)
init.
5
4
3
init.
(c)
6
init.
1
2
2
4
Time / 2 sec
6
n
o
r
u
e
N
6
4
2
(b)
n
o
r
u
e
N
6
4
2
(d)
6
n
o
r
u
e
N
4
2
160
Time / 0.0417 sec
170
180
6
4
2
2
4
6
10
8
4
Time / 2 sec
2
i
(e)
e
z
s
p
u
o
r
G
6
4
2
0
6
8 10 12
10
20
Time / 2 sec
Steady state reverberating patterns reflect features of the circuit's igniting stimuli.
FIG. 4:
(a) 3 and 6 directed loops
where initial stimulation is given to electrodes labelled by 1, 3 and 5. The periodicity of the initial stimulation is 2 and the
spatiotemporally generalized GCD(2,3,6)=1 leading to ZLS. (b) Spike trains of the circuit 1d with 3 periodical modes and
with transmission delay of 1/24 s. The network noise leads to a transition between two long-lasting modes, from three modes
to one mode during [165,170] time steps. At the steady state mode synchronized neurons are represented by the same colour
and the black line stands for the sums up the number of simultaneously active electrodes. (c) An exhaustive enumeration of
transients to ZLS of all the 63 (26
− 1) possible simultaneous stimulations of Fig. 1b. (d) Experimental result of the transient
to ZLS with initial simultaneous stimulations to electrodes labelled (1,3,6) (green) and (3,5) (violet) with transmission delay
of 2 s. (e) Sums up the number of electrodes with simultaneous activity for initial stimulation to electrodes labelled by (1,3,6)
(dashed-green), (3,5) (violet) and (1) (dashed-red).
µM CNQX (6-cyano-7-nitroquinoxaline-2,3-dione), and 3.125 µM Bicuculline; this cocktail of synaptic blockers made
the spontaneous network activity sparse. Up-regulation and down-regulation of network spontaneous activity (Fig. 3)
was achieved by addition of 20 µM Carbachol and 5 µM Bicuculline, respectively. In every event of pharmacological
manipulation, at least half an hour was allowed for stabilization of the effect.
B. Measurements and stimulation
An array of 60 Ti/Au/TiN extracellular electrodes, 30 µm in diameter, and spaced either 500 or 200 µm from each
other (MultiChannelSystems, Reutlingen, Germany) were used. The insulation layer (silicon nitride) was pre-treated
with polyethyleneimine (Sigma, 0.01% in 0.1 M Borate buffer solution). A commercial amplifier (MEA-1060-inv-BC,
MCS, Reutlingen, Germany) with frequency limits of 150-3,000 Hz and a gain of ×1,024 was used. Mono-phasic
square voltage pulses (200 µs, 100-900 mV) were applied through extracellular electrodes. Data was digitized using
data acquisition board (PD2-MF-64-3M/12H, UEI, Walpole, MA, USA). Each channel was sampled at a frequency
of 16 K sample/s. Action potentials were detected on-line by threshold crossing. Data processing and conditioned
stimulation were performed by using Simulink (The Mathworks, Natick, MA, USA) based xPC target application
7
(a)
5
4
6
3
1
5
1
5
init.
(b)
6
init.
(c)
6
2
4
2
3
4
3
init.
1
(d)
5
4
2
3
6
init.
1
2
(e)
init.
6
5
1
4
2
3
FIG. 5: Raster diagram of the firing activity of the neurons in the population dynamics were the first row is identical to the
first row of Fig. 1. Each circuit node is represented by 30 neurons and the entire six node circuit consists of 180 neurons.
Simulations parameters are similar to the parameters used in [16].
[12].
C. Cell selection
Each circuit node was represented by a stimulation source (source electrode) and a target for the stimulation- the
recorded electrode (target electrode). The stimulation electrodes (source and target) were selected as the ones that
evoked well isolated and well formed spikes and reliable response. This examination was done with stimulus intensity
of 800 mV, after 30 repetitions at a frequency of 5 Hz. After the electrodes selection the experiment began.
D. Stimulation Control
A circuit is formed by 6 or 4 nodes with programmable identical delays between nodes. A given node response was
defined as a spike occurring within a time window of the observed neuronal response latency ±2 ms following the
electrical stimulus. The activity of all target electrodes for each circuit stimulation was collected and entailed stimuli
were delivered in accordance to the adjacency matrix.
Acknowledgments
8
The authors thank Vladimir and Elleonora Lyakhov for invaluable technical assistance and Reut Timor for simula-
tion assistant. The research leading to these results has received funding from the European Union Seventh Framework
Program FP7 under grant agreement 269459 and grant of the Ministry of Science and Technology of the State of
Israel and MATERA grant agreement 3-7878.
[1] Hebb D. O., The Organization of Behavior: A Neuropsychological Theory (Wiley) 1949.
[2] Abeles M., Corticonics: Neural circuits of the cerebral cortex (Cambridge Univ. Press) 1991.
[3] Plenz D., Thiagarajan T. C., Trends in Neuroscience 30, 101 (2007).
[4] Morison R. S., Dempsy E. W., Amer. J. Physiol. 138, 297 (1943).
[5] Eccles J. C., The Neurophysiological basis of mind (Oxford Univ. Press) 1952.
[6] Steriade M., McCormick D. A., Sejnowski T. J., Science 262, 670 (1993).
[7] Ikegaya Y. et al., Science 23, 559 (2004).
[8] Gray C. M., Konig P., Engel A. K., Singer W., Nature 338, 334 (1989).
[9] Zrenner C., Eytan D., Wallach A., Their P., Marom S., Front. Neurosci. 4, 173 (2010).
[10] Marom S., Shahaf G., Quarterly Reviews of Biophysics 35, 63 (2002).
[11] Morin F. O., Takamura Y., Tamiya E., J. Biosci. Bioeng. 100, 131 (2005).
[12] Gal A. et al., J. Neurosci. 30, 16332 (2010).
[13] Eytan D., Marom S., J. Neurosci. 26, 8465 (2006).
[14] Thivierge J. P., Cisek P., J. Neurosci. 28, 7968 (2008).
[15] Shahaf G. et al. PLoSComput Biol. 4, e1000228 (2008).
[16] Kanter I. et al., EPL 93, 66001 (2011).
[17] Nixon M. et al. ,Phys. Rev. Lett. 106, 223901(2011).
[18] Tateno T., Jimbo Y., Robinson H. P. C., Neuroscience 134, 425 (2005).
[19] Wagenaar D., Pine J., Potter S. M., BMC Neurosci. 7, 11 (2006).
[20] Rabinovich M., Huerta R., Laurent G., Science 321, 48 (2008).
[21] Johansson R. S., Birznieks I., Nature Neurosci. 7, 170 (2004).
[22] Lee A. K., Wilson M. A., Neuron 36, 1183 (2002).
[23] Hodgkin A. L., Huxley A. F., J Physiol 117, 500 (1952).
[24] Shinomoto S. et. al., PloS Comput. Biol. 5, e1000433 (2009).
[25] Shimokawa T., Shinomoto S., Neural. Comput. 21, 1931(2009).
|
1908.05736 | 1 | 1908 | 2019-08-14T08:42:22 | Disagreeing about Crocs and socks: Creating profoundly ambiguous color displays | [
"q-bio.NC"
] | There is an increasing interest in the systematic disagreement about profoundly ambiguous stimuli in the color domain. However, this research has been hobbled by the fact that we could not create such stimuli at will. Here, we describe a design principle that allows the creation of such stimuli and apply this principle to create one such stimulus set - the crocs and socks. Using this set, we probed the color perception of a large sample of observers, showing that these stimuli are indeed categorically ambiguous and that we can predict the percept from fabric priors resulting from experience. We also relate the perception of these crocs to other color-ambiguous stimuli - the dress and the sneaker and conclude that differential priors likely underlie polarized disagreement in cognition more generally. | q-bio.NC | q-bio | Disagreeing about Crocs and socks: Creating profoundly ambiguous color displays
Pascal Wallisch & Michael Karlovich
Department of Psychology and Center for Neural Science, New York University, 4 Washington
Place, New York, NY 10003, USA.
Corresponding author: Pascal Wallisch - [email protected] - @pascallisch
Abstract: There is an increasing interest in the systematic disagreement about profoundly
ambiguous stimuli in the color domain. However, this research has been hobbled by the fact
that we could not create such stimuli at will. Here, we describe a design principle that allows the
creation of such stimuli and apply this principle to create one such stimulus set - "the crocs and
socks". Using this set, we probed the color perception of a large sample of observers, showing
that these stimuli are indeed categorically ambiguous and that we can predict the percept from
fabric priors resulting from experience. We also relate the perception of these crocs to other
color-ambiguous stimuli - "the dress" and "the sneaker" and conclude that differential priors
likely underlie polarized disagreement in cognition more generally.
Keywords: Disagreement, Dress, Color, Perception, Ambiguity, Uncertainty, Priors
Acknowledgments: We would like to thank Cecilia Bleasdale for taking the original dress picture
and anyone who either helped us to get the data, in particular Cecilia Hua and anyone else who
donated their data, shared the link or both.
If you want to contribute your data to a followup project, you can do so here:
https://www.surveymonkey.com/r/crocPerception
Author contributions: PW & MK conceived and designed the study. MK & PW created the
stimuli. PW designed the survey. PW & MK recorded the data. MK analyzed the stimuli. PW
analyzed the data. PW & MK wrote the manuscript.
Competing financial interests and commercial relationships: The authors declare no competing
financial interests, nor any commercial relationships. In particular, we want to proactively affirm
that there was no coordination or communication of any kind with Crocs Inc. or any affiliate,
regarding any aspects of this study. We chose our stimulus materials solely on the basis of color
considerations. We also would like to emphasize that this study was entirely self-funded.
Introduction
Strong perceptual illusions - systematic mismatches between the conditions in the external
world and subjective experience - have been known for a long time. Notoriously, prolonged
exposure to a moving stimulus leads to the vivid subjective experience of motion in the opposite
direction once the stimulus objectively stopped moving, a phenomenon known as the motion
aftereffect (Purkinje, 1825). Similar phenomena exist in the color domain - for instance, the
perceived hue of a colored patch can be shifted dramatically from what it would appear in
isolation, if presented with a suitably patterned background (Shevell & Monnier, 2003) or under
carefully modulated lighting conditions (Lotto & Purves, 1999), suggesting that observers take
both background and illumination into account when judging the appearance of a colored patch.
Importantly, all of these illusions are unidirectional, as they seem to affect all observers similarly,
predictably shifting the percept in the same direction.
This state of affairs changed in February 2015, when a categorically ambiguous stimulus in the
color domain surfaced. Most observers perceived this stimulus - the photo of a wedding dress -
as either black and blue (the appearance of the physical dress under sunlight) or as white and
gold, which could not be attributed to problems with fundamental color vision or a misapplication
of color labels (Lafer-Sousa et al., 2015). The reason this stimulus - "the dress" - is considered
categorically (or profoundly) ambiguous is owed to the fact that different observers perceive
different kinds - not degrees - of hue, and these hues are on opposite sides of the color wheel.
Notably, this is not a bi-stable stimulus - only a few observers (on the order of 1%) - experience
a percept that switches back and forth, akin to genuinely bistable stimuli like the Necker cube
(Rubin, 1921), binocular rivalry (Levelt, 1968) or moving plaids (Hupe & Rubin, 2003), all with
characteristic and known perceptual dynamics. Instead, most observers seem to be "locked in"
to a particular interpretation of the image, consistent with the notion that accumulated
experience - in the form of an illumination prior - drives the perception of "the dress" stimulus
(Witzel et al., 2017). Indeed, the specific kind of life experience that could shape the formation
of differential lighting priors, such as differing degrees of exposure to sunlight predicts both the
perception of "the dress" stimulus as well as the readiness of observers to assume particular
kinds of illuminations of the "the dress" (Wallisch, 2017a). However, while being able to explain
the perception of a specific stimulus display in a coherent and meaningful way is compelling, it
is also possible that we just managed to shoehorn an "explanation" of a very specific stimulus,
without being able to generalize. It would be reassuring if the principles invoked in the
explanation of such stimuli allowed us to create novel ambiguous displays at will. Should this
succeed, we could have more confidence that it is really these principles governing the
perception of such displays. Alas - to date - all other categorically color ambiguous display that
have surfaced since "the dress", including the Adidas Jacket (Wallisch, 2016), the flip-flops
(Wallisch, 2017b) or the sneaker (Wallisch, 2017c) were not generated purposefully with these
principles in mind, but rather arose spontaneously on the internet, allowing for residual doubt
whether we truly understand this phenomenon to persist. Here, we propose and use principles
that allow for the creation of categorically ambiguous color displays and explore how differential
priors determine the subjective perceptual experience evoked by these displays in human
observers.
Method
Stimulus design
We designed our stimuli according to the following principle, guided by the notion that patterned
backgrounds and lighting conditions powerfully influence the perception of a color display.
Specifically, we first maximize uncertainty by using a monochromatic object that could - in
principle be almost any color. Here, we used two versions of "Classic crocs™", namely
"Ballerina pink" and "New mint", out of the 28 different colors on offer. We then put these crocs
on an unpatterned, black background to make the display devoid of visual cues which could
allow observers to calibrate their percept (Land & McCann, 1971). Next, we carefully titrated
complementary light to equate the light reflected off the croc to shades of grey.
A final step relies on another piece of garment - in this case the white socks - to mirror the
properties of the illumination source. Thus, we introduce the concepts of a colored target object
(TCO) - this is the percept that we will ask our objects to judge, corresponding in this case to the
color of the crocs. We also introduce the concept of an anchoring or cueing object, which
reflects the color of the illumination, hence illumination mirroring object (IMO), in this case the
socks, see figure 1.
Figure 1: A schematic of the stimulus design principles. Top row: colored target objects (crocs)
under typical lighting conditions. We define sunlight or bright indoor light (e.g. incandescent or
LED) as "typical" lighting conditions, as these are conditions observers can be expected to
experience in everyday living. Bottom row: The same crocs under altered lighting conditions,
carefully titrated to equalize the reflectance. In this view, the crocs can be conceived of as a
filter. Left panel: Pink crocs as the target object. Right panel: Mint crocs as the target object.
Note that this schematic is highly stylized, with arbitrary units of power (AU) on the y-axis. In
reality, the power of the illumination is continuous, not discrete. In addition, the power of "white"
light is actually not equal across wavelengths; wavelengths in the middle of the visible range are
perceived as much brighter by human observers, due to human spectral sensitivity (Wald,
1964). A: Pink crocs, illuminated by white light appear pink because the croc filters wavelengths
in the middle of the range. B: Mint crocs, illuminated by white light appear mint green because
the croc filters wavelengths in the short and long range. C: Pink crocs (the CTO), illuminated by
light with more power in the middle range, as reflected by the appearance of the IMO, appear as
some shade of grey. D: Mint crocs (the CTO), illuminated by light with more power in the short
and long range, as reflected by the appearance of the IMO appear as some shade of grey. Note
that in all cases, the reflectance of the white sock corresponds to the illumination, mirroring the
properties of the light.
This part of the display - the IMO - is also subtly color ambiguous. Whereas this kind of plain
tube sock is likely to be white - as they usually are, when experienced in everyday life - the
person could possibly be wearing colored socks, in which case the socks provide no information
about the lighting. In other words, we use an object that could be any color, but usually isn't -
like plain tube socks to serve this mirroring function to create a conflict between color
appearance and typically experienced color. We predict that some people - those who have a
strong prior that socks like these are white - will use this cue to disambiguate the display and
discount the illuminant, but others - those who don't perceive the socks as white - to take the
appearance of the socks at face value. This is in line with other research suggesting that "priors"
- expectations and assumptions about the properties of stimuli that are typically derived from
experience - play a central role in cognition (Körding & Wolpert, 2004; Stocker & Simoncelli,
2006; Griffiths et al., 2008). To summarize, we predict that these principles will yield a
categorically color-ambiguous display. We conceive of this design principle as "Substantial
Uncertainty combined with Ramified or Forked Priors and Assumptions yields Disagreement"
(SURFPAD). We used the following seven images of crocs and socks as stimuli in this
research, see figure 2.
Figure 2: The crocs and socks images we created. A: crocs and Socks under typical (a
combination of daylight coming through a window and a LED (Feit electric, 60 Watt equivalent,
5000K, 800 Lumen) lighting conditions. WP: "Ballerina pink" crocs. WM: "New mint" crocs. B: The
"New mint" crocs under colored, specifically pink/orange light, yielding two specific stimuli with
distinct properties, as reflected by the socks, M1 and M2. C: The "Ballerina pink" crocs under
colored, specifically green light, yielding three specific stimuli with distinct properties, as
reflected by the socks, namely P0, P1 and P2. The lighting conditions in B and C were created
via a dimmable SMD 5050 LED light strip under blackout conditions and on black LifeSpanⓇ
treadmill matting. Note that the granularity of the dimming settings of the light strip in
combination with the reflectance properties of the crocs (e.g. the "New mint" version was
reflecting more strongly in the short wavelengths than anticipated) did not always allow us to
completely "whiten" the light reflected off the crocs. In those cases, we applied minimal digital
filtering with Adobe® Photoshop® CC 2019 software to close the distance. However, we strongly
believe that the use of a more sophisticated lighting system that allows for a sufficiently fine
titration of the different light channels would obviate the need for any digital post-production.
We first created P0 as a proof of concept. We then created two sets of pairs of stimuli using both
pink and mint crocs - and the same socks - under different lighting conditions, P1/P2 and
M1/M2, respectively.
We combined these croc displays we created with two others - "the dress" and "the sneaker",
both recognized as categorically ambiguous color displays, to use as stimuli in our study. Figure
3 shows the entire set, in combination with gamuts of the individual pixels in hsv space.
Figure 3: The complete set of stimulus types used in this study. In all cases, we show - for
stylistic reasons - a single croc and sock in the top of the croc displays to symbolize the stimulus
type. In the study, we used the complete images as depicted in figure 2. The bottom row for
each stimulus type depicts all pixels from both crocs and socks (but without black background)
as clouds in 2d projections of hsv space, in which individual pixels are represented as colored
dots. The axes of hsv space are oriented as depicted in the inset (to the right of M2 in B). These
pixel clouds were created using MATLAB's "colorcloud" function. In general, the parts in the dot
cloud that correspond to the part of the image (croc or sock) are aligned. For example, in P1 the
pixels from the sock are to the right and clearly separated from those of the croc. However, this
separation is not always as clear, e.g. in M2, where the point clouds blend together, but in
general the pixels corresponding to the sock are to the left of those corresponding to the croc.
A: "Ballerina pink" (Wp) and "New mint" (Wm) crocs under typical light, as in figure 2A. B: Wm
crocs under colored light, as in figure 2B. C: Wp crocs under colored light, as in figure 2C. D:
"Sneaker" and "Dress" displays and their corresponding colored pixel clouds. The "Dress" image
is courtesy Cecilia Bleasdale.
Participants
We logged data from participants online between July 5th and August 5th 2019. These were
recruited via social media, specifically Twitter and Facebook. The prompts on Twitter were: "If
you have 5 minutes, want to help us out to get to the bottom of why some people see "the
crocs" differently and want to see some pretty striking images check this out:" and "What colors
do you see, when looking at these #crocs? Last call, if you still want to donate your data. Should
only take a couple of minutes.", with a similar prompt on Facebook: "If you have 5 minutes, want
to help us out to get to the bottom of why some people see "the crocs" differently and are
curious to see some striking ambiguous images, donate your data by taking this brief survey."
These invitations yielded data from 5,944 participants. Of these, 108 participants indicated that
they were not taking the survey seriously and 80 participants reported a color vision impairment.
Thus, we retained data from 5,762 (~97%) participants for further analysis. Of these, 938
(16.3%) identified as male, 3,764 (65.3%) as female and the remainder did not state a gender.
The reported median age in our sample was 32 years. All study procedures followed the
principles of the Declaration of Helsinki, but were - according to the "final rule" (Revised
Common Rule) - exempt from review by the New York University Institutional Review Board
(UCAIHS).
Survey
We deployed a survey with 18 questions on surveymonkey.com. Nine of these questions
pertained to the perception of the stimulus displays. Of these, seven were the crocs and socks
displays shown in figure 2. One was "the dress" stimulus and one was a display of "the
sneaker", shown in figure 3D. Response options to the croc stimuli were "Pink/salmon",
"Green/mint", "Grey/White", "Orange/Beige" and "Other, please specify", ordered randomly for
each question. The response options for "the dress" and "the sneaker" stimuli were the standard
"white/gold", "black/blue", "blue/gold", "white/black", "switching" and "other" as well as
"pink/white", "grey/green", "pink/green", "grey/white", "switching" and "other", respectively.
Importantly, as observers could use the other images to calibrate their perception, each image
was shown in isolation, on a separate page, with participants unable to go back to a previous
false positives stemming
page. Participants were prompted with the question: "What color are the crocs in this image?"
for croc images, and: "What is the color of the dress in this image?" as well as: "What color is
the shoe in this image?" for "the dress" and sneaker images, respectively.
The other questions covered whether observers saw the socks as white, how familiar they are
with socks and crocs, what they believed about the source of the illumination and how confident
they are in the veracity of their beliefs in general. We also asked demographic questions about
age and gender as well as whether observers had a color vision deficit and whether they took
the survey seriously.
Data analysis
All data were analyzed with MATLAB (Mathworks, Inc., Sherborn, MA). Specifically, we use
Monte Carlo methods with 100,000 repetitions to bootstrap confidence intervals (Efron &
Tibshirani, 1994). To guard against potential
from multiple
comparisons, we adopted a conservative confidence interval of 99% for all analyses. Because
our data stems from voluntary online survey participation, it is not uncommon that participants
did not respond to some questions, e.g. on age or gender. We have no reason to assume that
these omissions are systematic, so we adopt a column-wise approach to handling missing data.
We report dependent variables in terms of proportions of observers who report a particular
percept. As our stimuli are often ambiguous, we report the proportion of our sample with a
percept that corresponds to the "Color Appearance Under Typical IlluminatiON" (CAUTION).
This avoids the awkwardness of reporting the proportion of participants who report a "veridical"
percept, as this concept is ill-defined in the color domain. Veridical pertains to correspondence
with reality, but - as has often been noted - while this makes sense for quantities of space and
time, the qualitative experience of color is largely generated by the brain and experienced by the
mind (Jackson, 1982). Photons have wavelengths, but not colors, and the wavelength of
photons is only loosely related to the color experienced by the human or animal observer. Here,
"typical illumination" means illumination conditions that observers are likely to encounter in
everyday living, such as daylight - when outdoors - or lighting sources designed to mimic
daylight when indoors.
Results
Are images of these crocs perceptually ambiguous under normal ("white") lighting conditions?
The first question that usually arises when discussing profoundly color ambiguous stimuli is
whether the disagreement indicates that one of the observers suffers from a color vision deficit.
In addition, we have no control over the particular viewing conditions such as viewing distance,
screen type, screen calibration, color temperature and the like, which one would be able to
control when doing such experiments under laboratory conditions. We cannot even be entirely
sure that participants know what a "croc" is or that an image was delivered each and every time,
for every participant, given that server uptimes are not perfect. In addition, one could reasonably
be skeptical that observers are reliably reporting their percepts in the first place, as we are not
paying them nor know their identity nor have any other interactions of any kind.
Thus, it is important to establish that participants can reliably discern and report the color of the
croc intended by the manufacturer (pink vs. mint) and that they agree with each other in spite of
all of these considerations and possibilities for discrepant reporting of percepts.
that
the color of
these crocs
the extremes of
We report the results of asking participants about the crocs under typical lighting conditions (WP
and WM) in figure 4.
Remarkably, despite all kinds of theoretical possibilities for misunderstandings and for things to
go wrong - whether due to technical issues or
participant characteristics of any kind or even
interactions between
them, about 98% of all
participants who provided a response have no
problems discerning the manufacturer intended color
of
the crocs as pink/salmon and green/ mint
respectively, from an image served over the internet,
and agreeing with each other about the category of
the color.
The exact p-value for either outcome obtained with
resampling methods is 0, assuming that people used
the available color categories at random. At no point
did
the null distribution come
anywhere close to the empirical result. Thus, we
conclude
is not
ambiguous to our participants under typical lighting
conditions.
Are images of these crocs perceptually ambiguous
under altered lighting conditions?
As we saw in figure 4, just putting the crocs on an
unpatterned black background does not render them
perceptually ambiguous with respect to color, so the
question remains whether they do become ambiguous
under suitably altered lighting conditions - along the
lines we laid out in the Method section.
The results are presented in figure 5. Briefly, we
managed to create five categorically color ambiguous
stimuli with different perceptual balancing points.
First, we created P0 as a proof of concept. We then
created two pairs of stimuli from each colored croc. All
5 images are categorically ambiguous with respect to
color - the modal percept for P0 and P1 is grey (43%
and 63%, respectively), for P2 green (40%). The M
percepts are dominated by beige or orange percepts
at 69% (M1) and 61% (M2), respectively. Overall, P2
was the most evenly balanced stimulus - therefore, P2
can be genuinely considered as having four distinct interpretations that are shared by more than
10% of the sample. Also, we note that the available color labels exhausted the space of
possibilities well - less than 1% of responses fell into "other" responses.
Figure 4: Perception of croc color
under typical lighting conditions. A:
Percentage of sample reporting
categorical percept on the x-axis in
response to Wp. x-axis: Color labels.
For reasons of economy, we
abbreviate "pink or salmon" as
"Pink", "green or mint" as "Green",
"grey or white" as "Grey" and "beige
or orange" as "Beige", here and in all
other figures. B: As in "A", but for WM.
The insets to the left of the y-axes
depict the stimulus that evoked the
response.
The chance of observing this pattern of results - using Chi-squared tests with the proportions
expected from the croc color perception under normal lighting perceptions is low, with a p-value
of below 1e-300 in each of the five cases. Overall, M1, M2 and P1 are conceptually most similar
to "the sneaker" (modal percept of grey in about 63% of observers) and "the dress" (modal
percept of white/gold in about 60% of observers). Therefore, these stimuli can be considered as
bi-stable with respect to the population (not individual observers) and P0 and P2 can be
considered tri- and quadra-stable with respect to the population, respectively, in contrast to the
unambiguous Wp and Wm.
Figure 5: Perception of color ambiguous stimuli created with altered lighting conditions. Rows,
from top to bottom: P0, P1, P2, M1, M2, in the order in which we created them. Y-axis:
Percentage of sample reporting this percept. X-axis: Categorical color percept. At the beginning
of each right, we depict the stimulus that evoked the observer response.
Are people responding randomly, because participants realize that illumination is not normal?
Whereas these results are striking at face value, it is conceivable that lighting conditions are so
obviously unsuitable that participants cannot reasonably be expected to report their actual
percept. Instead, they might be guessing as to what the color of the croc might have been. Such
a response could be akin to a situation under low light conditions, when vision is dominated by
rod-responses, as rods are more light-sensitive than cones, rendering vision effectively
monochromatic (Stabell & Stabell, 2002). However, we don't think this is plausible, for several
reasons. First, participants were not shy to point out problems if they arose, and had the option
to skip questions. Yet, both the "other" category is almost un-utilized, and most people did
respond to the questions. Second, close to 50% of participants did not suspect anything to be
amiss with the lighting in the images - guessing that they were either taken under sunlight or
indoor lighting, i.e. typical lighting conditions. Finally, and most compellingly, the responses
between stimuli are closely related - if someone perceived one image in a certain way, they
were far more likely to perceive the others similarly, see figure 6.
Figure 6: Croc display perceptions as a function of how P2 was perceived. A: In the left panel,
we depict the absolute probability of perceiving a croc display in a certain way, as a function of
how P2 was perceived (represented by stylized colored crocs on the x-axis). Rows, from top to
bottom: P0, P1, M1, M2. B: As in A, but the center panel depicts differences from the baserates
depicted in figure 5. C: As in B, but the right panel shows the same information as ratios.
We used P2 perception as a base, because this was the most ambiguous of our stimulus
displays. Note that the conditional probability of seeing a particular color is boosted in all 16
cases, sometimes considerably. The average boost was 21 percentage *points*, relative to the
percentages shown in figure 5. Sometimes, the result of this boost is dramatic, i.e. the
conditional probability of seeing M2 as beige if P2 was seen as beige is 90% (up from 70%).
The conditional probability of seeing P1 as pink if P2 was seen as pink is 86% (up from 18%).
On average, the ratio is boosted almost 2.5 fold. To summarize, there is considerable
consistency within observers and between percepts. This is unlikely to happen accidentally,
such as when observers are unsure of which option to pick, particularly as the response choices
are randomized and distributed over various pages - importantly - as we noted in the Method
section, participants cannot go back and revisit previous stimuli to compare.
Note that the pink percept seems to "gain the most" from having seen the reference stimulus as
pink, but this could be simply because this percept had the most to gain to begin with (lowest
absolute prevalence in the sample). Also note that this pattern of results - 16 positive
differences or 16 ratios larger than 1 - is in itself significant, as we would expect such a result
less than 1 in 65,000 cases by chance alone, translating to a p-value of 1.53e-5. As these are
conditional probabilities, and the individuals involved different people, these are independent
events, so this calculation is valid. Another way to interpret figure 6 is that the P2 percept
represents a strong attractor - in whichever way observers perceive P2 - and for whatever
reason - increases the probability that other croc displays are seen in the same way. In other
words, a propensity to see these stimuli in a certain way cuts across all of these stimuli and
might constitute a characteristic of the individual observer.
What determines the perception of the crocs displays?
After establishing that participants can agree on the color of the stimuli under typical lighting
conditions, but that this consensus breaks down under conditions of SURFPAD lighting, and
that this seems to be an individual characteristic, questions about this individual characteristic
come into focus. What - specifically - predisposes an individual observer to perceive these croc
displays in a certain way - and different from other observers? In the case of "the dress", a key
factor turned out to be illumination priors - assumptions about the illumination derived from
experience (Witzel et al., 2017; Wallisch, 2017a). We asked observers several questions with
this in mind and explore the relative role of fabric and illumination priors here.
Given the previous result - perception of all 5 crocs is tightly inter-related and given that we
asked the question about sock color right after asking about P2 - which might have drawn
attention to the color of the socks for subsequent stimuli, we focus our discussion of these
matters on the percepts involving P2, see figure 7.
Figure 7: Disagreement about the perception of P2 as a function of responses to relevant
survey questions. Each row represents the responses to a self-report question as the
independent variable. The dependent variable in all cases is the reported perception of P2.
Black vertical bars denote the 99% confidence interval.
The primary driver of P2 perception seems to be how the socks were seen. If the socks were
seen as white, the probability that the crocs were perceived as pink is over 12x that of
participants who did not perceive the socks as white. Dovetailing nicely with this effect is the
observation that the opposite is true for green and grey perception, if the socks were not seen
as white, but in the other direction. Observers were more than 2x as likely to perceive the crocs
as grey and over 4x as likely to perceive them as green if they did not perceive the socks as
white, compared to when they did not. The fact that the confidence intervals are much larger for
the "unsure" group is largely due to this being by far the smallest group. A secondary driver of
disagreement seems to be lighting conditions. Consistent with our interpretation, if participants
did perceive the lighting as abnormal, they were slightly - but significantly - more likely to
perceive P2 as pink and less likely to perceive P2 as grey or green, perhaps because they
questioned the prima facie appearance of the crocs.
Finally, answers to arbitrary questions do not systematically distinguish participants responding
in systematic ways. For instance, we did not expect self-reported croc experience to have any
perceptual effects, mostly due to the fact that even being familiar with crocs does not allow to
disambiguate which of the 28 classic crocs is on display. In sum, unspecific experience without
a clear mechanism does nothing to reduce the uncertainty of the croc color, whereas
assumptions about lighting or socks do.
What determines whether socks were seen as white?
Along the lines of Wallisch (2017a), we suspect that experience is a primary predictor of
assumptions about lighting or sock color, and lifestyle choices a primary predictor of experience.
We did not ask a question about lifestyle choices regarding light, as - unlike in "the dress" case,
where chronotypes have an obvious impact on sunlight exposure, everything else being equal,
there is no clear and reasonably common predictor of exposure to colored LED lights. However,
one can ask about direct experience with wearing plain tube socks, so we did ask that question.
Figure 8 summarizes the results with regard to the perceptual sequelae of answering this
question.
Figure 8: Perceptual consequences of wearing white tube socks. Left panel: Proportion seeing
socks as white as a function of tube socks being worn. Right panel: Same, but seeing P2 as
pink as the dependent variable. Black vertical bars represent the 99% confidence interval.
The effects in figure 8 are obviously less strong than what would be needed to comfortably
claim that we understand this aspect of the phenomenon well, but then again it is both
remarkable and interesting to see these systematic and significant effects in the first place,
given how much random variability is likely involved in these self-reported responses. In
addition, there are certainly other ways to be exposed to tube socks than wearing them, i.e. one
might see them in sports, on TV, in ads, or in stores, but none of these are easy to probe in self-
reports. Put differently, we would not expect strong effects on the basis of this simple self-report
question, as wearing the socks is clearly not a very close proxy to actual overall exposure.
Interestingly, the perceptual effects on P2 trade off clearly, i.e. all of the increased pink
perception comes from a decrease in perceiving P2 as green, with the proportions of grey and
beige entirely unaffected (not pictured in figure 8), just as one would expect if the socks are
reflecting the light, allowing observers to discount it.
How broad is the effect of these priors on perception more generally?
Now that we established the power of these priors on the perception of P2 with respect to color,
one can wonder how far these priors extend. Does P2 perception predict perception of "the
sneaker" and "the dress"? An answer in the affirmative could hint at common underlying
mechanisms and observer characteristica. Does a strong sock prior also predict perception of
"the sneaker" or "the dress"? For answers to these questions, see figure 9.
Figure 9: Priors and non-croc perception. A: Descriptive statistics for sneaker perception. Y-
axis: Proportion of the sample reporting seeing Pink/White (P/W), Grey/Green (G/G),
Pink/Green (P/G), Grey/White (G/W), Switching (S) or Other. B: Descriptive statistics for dress
perception. Y-Axis: Proportion of the sample reporting seeing White/Gold (W/G), Blue/Black
(B/B), Blue/Gold (B/G) and Black/White (B/W) or Switching (S). C: Probability of seeing the
sneaker as pink (pink bars) or grey (grey bars) as a function of seeing P2 as pink, green, grey or
beige (as indicated by the croc pictograms on the x-axis). D: Probability of seeing the dress as
White/Gold (Yellow bars) or Black/Blue (Blue bars) as a function of seeing P2 as pink, green,
grey or beige (as indicated by the croc pictograms on the x-axis). E: Probability of seeing the
sneaker as pink (pink bars) or grey (grey bars) as a function of sock perception. F: Probability of
seeing the dress as White/Gold (Yellow bars) or Black/Blue (Blue bars) as a function of sock
perception. Black vertical bars represent the 99% confidence interval.
The first thing to note is that "the sneaker" is most perceptually similar to P1 - grey perception
strongly dominates, with the remainder mostly taken up by pink, as could be anticipated from
the point clouds in figure 3. Second, we broadly replicate the proportions described in the
perception of "the dress" as reported in Lafer Sousa et al. (2015), or Wallisch (2017a), even four
years later. Third, P2 perception strongly determines sneaker perception. Almost 90% of those
who saw P2 as pink saw the sneaker as pink, whereas just over 10% of those who saw P2 as
green did. Almost 90% of those who saw P2 as green saw the sneaker as grey, whereas just
over 10% of those who saw P2 as pink did. The numbers for seeing P2 as grey is almost the
same as seeing P2 as green, whereas seeing P2 as beige falls somewhere in between grey
and pink. In contrast, P2 perception had little to no influence on the perception of "the dress",
suggesting that crocs and sneaker perception share a common mechanism - perhaps fabric
priors, specifically white sock priors - whereas the perception of "the dress" is primarily driven by
a different mechanism - perhaps lighting priors, specifically sunlight priors. Fourth, the effects of
assuming the socks are white reach across percepts, predicting not only P2 - and indeed the
perception of the other croc displays - but also the sneaker. This could be due to a number of
reasons. For instance, this could be a manifestation of a tendency to apply fabric priors across
the board. Or this pattern of results could simply reflect the fact that people who think socks are
white are the same people who think that shoelaces - and maybe other objects like that - are
white. It could also reflect the workings of another, not yet well understood mechanism or
personal characteristic, of which sock perception is simply a marker of. Notably, these socks are
often worn in athletic contexts - it would not be too far of a stretch to suggest that they might be
worn with - sneakers - that will tend to happen to have white shoelaces. Finally, consistent with
what we already observed here, how the sock is perceived seems to have very little impact on
how the dress is perceived, again suggesting that the perception of "the dress" is not governed
by fabric priors, but primarily by lighting priors. In addition, the lack of an impact of sock
perception on dress perception suggests that there isn't simply some overall bias in color
perception that affects all of perception affecting the color perception of socks, crocs, sneakers
and dresses alike. The complete lack of an impact on dress perception rules out this possibility -
priors and how they manifest in perception seem to be fairly specific.
What is the cognitive penetrability of these effects?
Most perceptual effects are notoriously immune to awareness and cognitive interventions
(Firestone & Scholl, 2016). Telling people that the physical "dress" is blue and black does not
change one's percept, should it be white and gold. However, there is reason to believe that the
effect driven croc perception might be different. For instance, if one were to take one's percept
at face value, a pink croc under typical lighting conditions, as in Wp would likely be perceived as
grey or green under altered lighting conditions, as these are the colors of the pixels that make
up the croc (see figure 3). However, if one were to recognize that these lighting conditions are
irregular, one might scrutinize them and give more weight to one's priors, importantly one's
white sock prior, if such a prior exists, allowing an observer to color calibrate the image and
perhaps recover the color under typical light. To account for this possibility, we asked one
question along these lines, namely whether observers consider it possible that their beliefs
could be wrong. This question is somewhat difficult to interpret, as a lot of factors might go into
answering it, including self-awareness, meta-cognition, modesty, as well as unequal efforts
among people to make sure that their beliefs are accurate. In hindsight, we should have asked
something more closely related to perception, i.e. beliefs in naive realism, but we do think it is
fair to assume that this question can be taken as a proxy for confidence in the veracity of one's
own cognition, given everything else is equal. Of course, everything else won't be equal on the
level of the individual, which is why a large sample size is critical. See figure 10 for a summary
of these effects.
Figure 10: Cognitive penetrability of percepts. In each panel, we report the proportion of
participants who report a color perception that corresponds to the color appearance under
typical illumination conditions (CAUTION, see Method) as a function of whether participants
allowed for the possibility that some of their beliefs might be wrong, for a given stimulus display.
Black vertical bars represent the 99% confidence interval.
The effects illustrated in this figure are consistent and perhaps even ironic. The more confident
observers are in their beliefs being correct, the less likely they are to recover CAUTION, which
is not implied by the pixel clouds in figure 3. In other words, these observers trust the
appearance of the shoes, without second-guessing this appearance, perhaps due to confidence
in the general veracity of one's cognition. This extends to CAUTION perceptions of socks and
sneakers, which are significantly modulated by the answer to this question. As we noted above -
the perception of "the dress" is not influenced by meta-cognition. We have confidence that the
confidence intervals in P0 would shrink considerably with more participants - we had by far the
least respondents to this question. Whether this is ironic or not - echoing other meta-cognitive
effects (Dunning, 2011) depends on one's conception of what color the crocs "really" are -
CAUTION or the color of the individual pixels as seen in the point clouds of figure 3, or if that
question even makes sense, as the concept of veridicality is not well defined for color in the first
place.
What is the effect of demographics?
Finally, we wondered about the effects of demographics on the perception of these stimuli. Prior
research on "the dress" (e.g. Lafer-Sousa et al., 2015; Wallisch, 2017a) suggest that gender
effects are miniscule, but age effects could be considerable, although a relative dearth of old
observers is a concern. See figure 11 for an overview of these demographic effects.
Figure 11: Demographic effects. A: Stimuli where CAUTION decreases with age. Circles: Dress.
Right-pointed triangles: Sneaker. Left-pointed-triangles: P1. Down-pointed triangles: P2. Red
dotted lines represent SEM. B: Stimuli where there is no clear age-effect. Squares: M1.
Diamonds: M2. Up-pointed triangles: P0. Red dotted lines represent SEM. C: Effects of gender
on CAUTION. Each pair of bars corresponds to the stimulus displays labeled on top of it. Black
vertical bars represent the 99% confidence interval.
To summarize, most pink croc stimuli, as well as the dress and sneaker exhibit strong age-
effects, with a gradual decrease in CAUTION as a function of age. This contrasts with a sharp
decline of the White/Gold modal percept with retirement age for the dress (Wallisch, 2017a),
which we replicated here and a sharp rise in Blue/Gold percepts (not pictured). As noted in
Wallisch (2017a), it is hard to meaningfully interpret these effects, as age is a carrier variable of
so many other effects, including physiological effects and cohort effects. We also observe a lack
of an age related decline in CAUTION for the mint croc stimuli, which could possibly be due to a
floor effect - there is little CAUTION for these stimuli to begin with. We attribute the lack of an
age effect for P0 to a lack of participants, as in figure 10 - the error bars are simply too wide to
interpret this trajectory meaningfully. Finally, as already discussed, gender effects are miniscule,
if they exist at all. There seems to be marginal gender effect on the perception of P1 and its
related stimulus, the sneaker, but if they exist, these effects are small and hard to interpret. This
addresses one of the concerns related to this study, namely that the majority of our sample
identified as female. This might be owed to the fact that we framed the recruitment for our study
in terms of donating data, and women have been shown to be more proactive in charitable
giving in general (Piper & Schnepf, 2008; Mesch et al., 2011). As figure 11 illustrates, whereas
this might have biased the gender-composition of the sample, it is unlikely that this imbalance
affected the perceptual effects reported in this study.
Discussion
We managed to create visual displays such that observers categorically disagree about the
colors of the objects depicted, even though the same objects evoke near complete consensus
under typical lighting conditions. We did so by systematically avoiding, minimizing and reducing
cues that observers typically use to determine the color appearance of an object, such as
patterned backgrounds (Jenness & Shevell, 1995), color memory (Witzel et al., 2011) or familiar
lighting (Maloney, 1999), while at the same time introducing a part of the object that mirrors the
properties of the lighting, but for which some - but not all - participants have a prior that they use
to anchor their color percept, a process which we term SURFPAD. We were also able to show
that the perception of the different color displays was systematically linked within a given
observer, extending even to other color-ambiguous stimuli, such as the "sneaker". We take this
to mean that a disposition to see these stimuli in a certain way amounts to a personality
characteristic. This personality characteristic is - in this case - best characterized as a "fabric
prior" for white socks, which strongly determines the perception of the displays we created. This
fabric prior - in turn - seems to have been determined by experience with this type of sock, akin
to the illumination priors that determine the perception of the "dress" due to differential exposure
to sunlight (Wallisch, 2017a).
We believe that these effects are consistent with the operation of known cognitive mechanism
involved in color vision, most importantly color constancy (McCann et al., 1976). In this special
case, color constancy seems to be derived from a kind of color memory (Jin & Shevell, 1996;
Hansen et al., 2006), in a subset of our observers. Remarkably, this mechanism allows these
observers to recover the color that corresponds to CAUTION, even though there are virtually no
individual pixels in the object that suggest CAUTION, a striking empirical instance of the whole
being different from the sum of its parts, as suggested theoretically by Koffka (1922). Put
differently, instead of alerting the organism to the profound uncertainty inherent in these
displays, the brain seems to err on the side of arriving at actionable conclusions by operating
inadvertent mechanisms akin to autocorrect - educated guesses to recover the stimulus
situation by making assumptions about the color of lighting (in the case of the dress) or fabric
(here). It is increasingly well recognized that perceptual mechanisms in the brain utilize
information beyond the present stimulus in the form of priors, assumptions, expectations and
predictions, in order to cope with and act in an environment that is inherently more uncertain
than one would prefer (Press & Yon, 2019). This line of reasoning also suggests that while we
understand the color appearance of objects under typical lighting conditions relatively well - to
the point of being able to reasonably describe color appearance in terms of a linear system
(Shevell, 2003), the same might be true under carefully altered lighting conditions, where
differential priors play a much larger role in determining the subjective color experience.
There are several limitations inherent in this study that potentially threaten the generality of
these conclusions. Most of these pertain to our specific implementation. First, in order to rapidly
obtain data from many observers, we logged the data online, instead of in a lab. This will
necessarily introduce considerable uncertainty as to the stimulus situation that the participants
actually encountered, such as screen size or color calibration, which would not be a concern
under controlled laboratory conditions. Similarly, we would have wanted to take more complete
control of the lighting conditions when creating these stimuli, involving both a high-end LED
lighting system and a photometer. Then again, that we readily found these effects despite the
noise introduced by presenting the stimuli online speaks to both their strength and robustness,
akin to all the other color-ambiguous stimuli (the "dress", "adidas jacket", "flip-flops", etc.) that
surfaced and thrived "in the wild", without relying on carefully controlled laboratory conditions. In
addition, our observers were able to agree on the color of these objects under typical lighting
conditions, so it is unlikely that this ambiguity is introduced as an artifact of different screen
settings. Similarly, whereas we strongly suggest future studies in this domain to more carefully
control the lighting conditions when creating the ambiguous displays, our efforts still have merit
as a proof of concept. What seems to matter critically is the relative position of the color clouds
from the two objects - the colored target object (CTO), the croc in this case and the cueing or
anchoring illumination-mirroring object (IMO), in this case the sock - in hsv space. Displays
seem to evoke particularly strong disagreement between observers if there is a separation
between the two clouds and if they form parallel columns in this space, whether this
configuration is achieved by a careful control of lighting conditions or digital filtering. The
creation of such stimuli in hsv space will also allow us to address the issue of beige perception.
Are participants perceiving these stimuli as beige because they tried to color-correct their
percept, but failed (ending up at beige instead of pink), or are the actual pixels - as suggested
by the point clouds in figure 3, particularly for M1 and M2 far enough in the beige range to make
a beige croc color plausible a priori? Cleaner stimuli - that are clearly separated and mostly
avoid the beige range should cut down on beige perception, increasing the perception of pink
(for the Ps) and grey (for M).
Another concern is that we illustrated the SURFPAD principle only with one type of object -
crocs and socks, while claiming that it is a general design principle. This is a clear case of a
concern that can readily be addressed in future studies, using any number of colored stimuli,
such as colored candles as the CTO and the wick as the IMO or colored tissue boxes as the
CTO and white tissues as the IMO. With some imagination, this principle can likely be extended
to other visual domains, such as distance or motion, and perhaps even to auditory stimuli.
Based on our experiences of applying SURFPAD to crocs and socks, we predict that a suitable
application of these principles would yield stimulus displays with similar levels of disagreement
between observers.
However, the biggest concern regarding this kind of work is that we overlooked some low level
explanation for the phenomena we observed, like so many others in perception (Firestone &
Scholl, 2016). It would not be the first time that a high profile and high level explanation is better
explained by other factors, which is usually the case when the experimenter inadvertently fails
to notice or rule out covariates that correlate and drive the dependent variable even better. A
prominent example of this would be Mischel who thought that willpower determines the ability to
resist the temptation of marshmallows (Mischel et al., 1989), a phenomenon better explained by
trust (Watts et al., 2018). This can happen even in experimental work, when the independent
variables created by the experimenter mask a confound. For instance, it has been proposed that
observers perceiving a rotating line to lead a simultaneously flashed line suggests the operation
of motion extrapolation mechanisms to guide dynamic action (Nijhawan, 1994). However, this
phenomenon can be more readily explained in terms of a low-level feature - the flashed line has
lower effective contrast than the rotating one, and lower contrast is associated with longer
neural delays (Raiguel et al., 1999). This suggests that the rotating line is seen as leading
because its signal arrives in the brain before that of the flashed line, so it should be possible to
negate this advantage - and the lead - by increasing the relative contrast of the flashed line,
which is what can be shown empirically (Purushothaman et al., 1998).
While we concede this theoretical possibility, we do not think it likely in this case. First, the
SURFPAD principle leverages a combination of known mechanisms and principles of visual
processing, without relying on exotic postulates or remote theoretical possibilities. Second,
given how rare such color ambiguous displays are - with the advent of smartphones, there are
now well over a trillion images taken per year (Richter, 2017), and only a handful of these are
known to be color-ambiguous, whereas all the displays we created according to this principle
are color ambiguous, creating such displays inadvertently is about as plausible as suggesting to
go to the moon without a solid understanding of gravity or to build a nuclear reactor without a
reasonable understanding of the basic physics governing nuclear reactions involved.
Finally, suggesting that there are more powerful low level features at work that we somehow
overlooked strains credulity, given the sheer size of the effects at play - the white sock prior
alone yields a 12-fold range in the pink percept of P2, which in turn determines close to 90% of
the sneaker percept. One would be hard-pressed to envision more powerful predictors of these
percepts.
Of course, it is true that - in principle - a final determination on this issue won't be possible
without future work. If we claim that socks are critical to bring about disagreement of the croc
color, it stands to reason that we should create and ask observers about crocs altogether
without socks. As this might look somewhat strange, a fair comparison would be to use the
same exact crocs, but making sure that the socks do not reflect the light. Using perfectly white
socks should not allow observers to use the socks to color calibrate the rest of the display.
Similarly, it might be of interest to ask observers directly about the color of shoelaces to further
elucidate the link between sneaker perception and croc perception, and the role of fabric priors
and their specificity as they apply to perception.
Our results do suggest that disagreements due to differential illumination priors (as in the dress)
and disagreements due to differential fabric priors (as in the crocs) are similar, but separate
phenomena. We believe that this is due to the fact that the "dress" phenomenon is situated on
the daylight axis between yellow and blue, whereas our crocs displays leverage the opposite
axis between pink and green, which is likely why there is so little cross-display information. In
other words, we can't predict how someone saw the dress based on how they saw the crocs or
vice versa, because these displays operate on different axes through color space. We propose
to create a display involving both fabric priors and daylight priors involving grey, yellow and blue
crocs as well as yellow and blue lights, using SURFPAD principles.
A bigger point of this study - beyond crocs, socks and color - is that this principle pertains to the
nature of disagreement and polarization. Given some of the feedback we received, there is no
question that some people are annoyed by these kinds of ambiguous displays, as being
confronted with these displays threatens their self-perception as being accurate observers of
and effective operators in the external world. Of course, this kind of self-perception is validated
all the time - under typical conditions, people's interpretation of the situation does correspond to
reality, and the interpretation of others. What makes the operation of these cognitive
"autocomplete" or "autocorrect" mechanisms so pernicious is that they do not flag to the
observer when they had to be leveraged to bring about the perceptual interpretation. They are
simply presented to the observer as "the percept", regardless of how it came about, perhaps in
an attempt to facilitate and encourage action.
Note that we would expect these effects to be least worrisome in perception, for several
reasons. First, the brain devotes considerable resources to the processing of perceptual
information (Wallisch & Movshon, 2008). Second, there are - under natural conditions - multiple
redundancies both within and across modalities (Landy & Kojima, 2001; Hillis et al., 2002;
Körding et al., 2007). Third, there was likely immense evolutionary pressures to get the right
answer. If the brain is indeed epistemically immodest - biased towards deriving actionable
conclusions despite profound uncertainty - as suggested by our results, it seems important that
these conclusions are not catastrophic to the integrity of the organism. So whereas these
mechanisms are particularly tractable in the perceptual domain, we suspect that they play a
larger role in cognition beyond perception, particularly as we now live in environments and
systems that are complex enough that there are almost always considerable degrees of
freedom for observers to come up with their own interpretation (Wallisch & Whritner, 2017). Put
differently, we are now - cognitively - living in a SURFPAD world. We are increasingly
encountering conditions of tremendous and polarized differences in the interpretations of current
events. This is somewhat surprising, as it is happening in spite of all the information that is now
available about the state of the world, which could drive consensus. But it is important to note
that given the complexity of the world, we are almost always in a regimen of high SU (due to the
complexity of reality and the social systems we inhabit itself). So if there is fundamental
disagreement (D), it is likely not due to disagreements about the evidence - shrouded in SU -
but rather the cognitively invisible RFPA. Where do these ramified or forked priors and
assumptions come from?
It is no secret that media and social media have effectively created increasingly divergent
differential priors, leading to starkly differing interpretations of events, regardless of what
actually happens. Alarmingly, both media and social media seem to have created cognitive
versions of the socks, anchors that appear politically correct to a particular echo chamber
(Brady et al., 2019). To explore reductions in the difference in interpretations, future work on
SURFPAD could explore the malleability of these differential priors - does priming with suitable
stimuli just before the study update the prior? Is it possible to override the prior by introducing
other display elements that could restore consensus, e.g. a colored background that also
reflects the illumination?
Thus, in terms of fostering conflict resolution, we advise neither to highlight the disagreement
itself - which typically garners all the attention, nor the SU, which is inevitable in a complex
world, but rather to focus squarely on the roots of disagreement - the RFPA. Differential priors,
created in a strong and polarized fashion and under conditions of profound uncertainty will make
dramatically differing interpretations of the same events and situations inevitable.
It is appropriate to address this issue with the urgency that it necessitates if one is to avoid the
adverse outcomes associated with mounting and polarized disagreement.
References
Brady, W. J., Crockett, M., & Van Bavel, J. J. (2019). The MAD Model of Moral Contagion: The
role of motivation, attention and design in the spread of moralized content online.
Dunning, D. (2011). The Dunning -- Kruger effect: On being ignorant of one's own
ignorance. In Advances in experimental social psychology. Academic Press.
Efron, B., & Tibshirani, R. J. (1994). An introduction to the bootstrap. CRC press.
Firestone, C., & Scholl, B. J. (2016). Cognition does not affect perception: Evaluating the
evidence for "top-down" effects. Behavioral and Brain Sciences, 39.
Griffiths, T. L., & Tenenbaum, J. B. (2006). Optimal predictions in everyday cognition.
Psychological Science, 17(9), 767-773.
Hansen, T., Olkkonen, M., Walter, S., & Gegenfurtner, K. R. (2006). Memory modulates color
appearance. Nature Neuroscience, 9(11), 1367.
Hillis, J. M., Ernst, M. O., Banks, M. S., & Landy, M. S. (2002). Combining sensory information:
mandatory fusion within, but not between, senses. Science, 298(5598), 1627-1630.
Hupé, J. M., & Rubin, N. (2003). The dynamics of bi-stable alternation in ambiguous motion
displays: a fresh look at plaids. Vision Research, 43(5), 531-548.
Jackson, F. (1982). Epiphenomenal Qualia. The Philosophical Quarterly, 32(127)
Koffka, K. (1922). Perception: an introduction to the Gestalt-Theorie. Psychological Bulletin,
19(10), 531.
Jenness, J. W., & Shevell, S. K. (1995). Color appearance with sparse chromatic context. Vision
Research, 35(6), 797-805.
Jin, E. W., & Shevell, S. K. (1996). Color memory and color constancy. JOSA A, 13(10), 1981-
1991.
Körding, K. P., & Wolpert, D. M. (2004). Bayesian integration in sensorimotor learning. Nature,
427(6971), 244.
Körding, K. P., Beierholm, U., Ma, W. J., Quartz, S., Tenenbaum, J. B., & Shams, L. (2007).
Causal inference in multisensory perception. PLoS one, 2(9), e943.
Lafer-Sousa, R., Hermann, K. L., & Conway, B. R. (2015). Striking individual differences in color
perception uncovered by "the dress" photograph. Current Biology, 25 (13), R545 -- R546.
Land, E. H., & McCann, J. J. (1971). Lightness and retinex theory. Josa, 61(1), 1-11.
Landy, M. S., & Kojima, H. (2001). Ideal cue combination for localizing texture-defined edges.
JOSA A, 18(9), 2307-2320.
Levelt, W. J. (1965). On binocular rivalry (Doctoral dissertation, Van Gorcum Assen).
Lotto, R. B., & Purves, D. (1999). The effects of color on brightness. Nature Neuroscience,
2(11), 1010.
Maloney, L. T. (1999). Physics-based approaches to modeling surface color perception. Color
Vision: From genes to perception, 387-416.
McCann, J. J., McKee, S. P., & Taylor, T. H. (1976). Quantitative studies in retinex theory a
comparison between theoretical predictions and observer responses to the "color mondrian"
experiments. Vision Research, 16(5), 445-IN3.
Mesch, D. J., Brown, M. S., Moore, Z. I., & Hayat, A. D. (2011). Gender differences in charitable
giving. International Journal of Nonprofit and Voluntary Sector Marketing, 16(4), 342-355.
Mischel, W., Shoda, Y., & Rodriguez, M. I. (1989). Delay of gratification in children. Science,
244(4907), 933-938.
Nijhawan, R. (1994). Motion extrapolation in catching. Nature.
Piper, G., & Schnepf, S. V. (2008). Gender differences in charitable giving in Great Britain.
Voluntas: International Journal of Voluntary and Nonprofit Organizations, 19(2), 103-124.
Purkinje, J. E. (1825). Beobachtungen und Versuche zur Physiologie der Sinne. Neue Beitrage
zur Kenntniss des Sehens in subjektiver Hinsicht. Berlin, Germany: Reimer.
Press, C., & Yon, D. (2019). Perceptual Prediction: Rapidly Making Sense of a Noisy World.
Current Biology, 29(15), R751-R753.
Purushothaman, G., Patel, S. S., Bedell, H. E., & Ogmen, H. (1998). Moving ahead through
differential visual latency. Nature, 396(6710), 424.
Monnier, P., & Shevell, S. K. (2003). Large shifts in color appearance from patterned chromatic
backgrounds. Nature Neuroscience, 6(8), 801.
Raiguel, S. E., Xiao, D. K., Marcar, V. L., & Orban, G. A. (1999). Response latency of macaque
area MT/V5 neurons and its relationship to stimulus parameters. Journal of Neurophysiology,
82(4), 1944-1956.
Richter, F. (2017). https://www.statista.com/chart/10913/number-of-photos-taken-worldwide/
Rubin, E. (1921). Visuell wahrgenommene Figuren: Studien in psychologischer Analyse.
Gyldendalske Boghandel.
Shevell, S. K. (Ed.). (2003). The science of color. Elsevier.
Stabell, B., & Stabell, U. (2002). Effects of rod activity on color perception with light adaptation.
JOSA A, 19(7), 1249-1258.
Stocker, A. A., & Simoncelli, E. P. (2006). Noise characteristics and prior expectations in human
visual speed perception. Nature Neuroscience, 9(4), 578.
Wald, G. (1964). The receptors of human color vision. Science, 145(3636), 1007-1016.
Wallisch, P., & Movshon, J. A. (2008). Structure and function come unglued in the visual cortex.
Neuron, 60(2), 195-197.
Wallisch, P. (2016): https://slate.com/technology/2016/03/the-science-of-the-black-and-blue-
dress-one-year-later.html
Wallisch, P. (2017a). Illumination assumptions account for individual differences in the
perceptual interpretation of a profoundly ambiguous stimulus in the color domain:"The dress".
Journal of Vision, 17(4), 5-5.
Wallisch, P. (2017b): https://slate.com/technology/2017/04/heres-why-people-saw-the-dress-
differently.html
Wallisch, P. (2017c) https://slate.com/technology/2017/10/the-shoe-is-the-new-dress.html
Wallisch, P., & Whritner, J. A. (2017). Strikingly low agreement in the appraisal of motion
pictures. Projections, 11(1), 102-120.
Watts, T. W., Duncan, G. J., & Quan, H. (2018). Revisiting the marshmallow test: A conceptual
replication investigating links between early delay of gratification and later outcomes.
Psychological Science, 29(7), 1159-1177.
Witzel, C., Valkova, H., Hansen, T., & Gegenfurtner, K. R. (2011). Object knowledge modulates
colour appearance. i-Perception, 2(1), 13-49.
Witzel, C., Racey, C., & O'Regan, J. K. (2017). The most reasonable explanation of "the dress":
Implicit assumptions about illumination. Journal of Vision, 17(2), 1-1.
|
1902.05568 | 1 | 1902 | 2019-02-14T19:12:32 | Tractography and machine learning: Current state and open challenges | [
"q-bio.NC",
"cs.LG"
] | Supervised machine learning (ML) algorithms have recently been proposed as an alternative to traditional tractography methods in order to address some of their weaknesses. They can be path-based and local-model-free, and easily incorporate anatomical priors to make contextual and non-local decisions that should help the tracking process. ML-based techniques have thus shown promising reconstructions of larger spatial extent of existing white matter bundles, promising reconstructions of less false positives, and promising robustness to known position and shape biases of current tractography techniques. But as of today, none of these ML-based methods have shown conclusive performances or have been adopted as a de facto solution to tractography. One reason for this might be the lack of well-defined and extensive frameworks to train, evaluate, and compare these methods.
In this paper, we describe several datasets and evaluation tools that contain useful features for ML algorithms, along with the various methods proposed in the recent years. We then discuss the strategies that are used to evaluate and compare those methods, as well as their shortcomings. Finally, we describe the particular needs of ML tractography methods and discuss tangible solutions for future works. | q-bio.NC | q-bio |
DanielJörgens2
MaximeDescoteaux*1
Pierre-Marc
ORIGINAL ARTICLE
Tractographyandmachinelearning: Currentstate
andopenchallenges
PhilippePoulin1
Jodoin*1
1DepartmentofComputerScience,
UniversitédeSherbrooke,Sherbrooke,
Québec,Canada
2DepartmentofBiomedicalEngineering
andHealthSystems,KTHRoyalInstituteof
Technology,Stockholm,Sweden
*Co-lastauthor
Correspondence
PhilippePoulin,DepartmentofComputer
Science,UniversitédeSherbrooke,
Sherbrooke,Québec,Canada
Email: [email protected]
Fundinginformation
FRQNT;UniversitédeSherbrooke
InstitutionalChairinNeuroinformatics;
NSERCDiscoverygrantfromPr
DescoteauxandJodoin
Supervisedmachinelearning(ML)algorithmshaverecently
beenproposedasanalternativetotraditionaltractography
methodsinordertoaddresssomeoftheirweaknesses. They
canbepath-basedandlocal-model-free,andeasilyincorpo-
rateanatomicalpriorstomakecontextualandnon-localde-
cisionsthatshouldhelpthetrackingprocess. ML-basedtech-
niqueshavethusshownpromisingreconstructionsoflarger
spatialextentofexistingwhitematterbundles,promising
reconstructionsoflessfalsepositives,andpromisingrobust-
nesstoknownpositionandshapebiasesofcurrenttractog-
raphytechniques. Butasoftoday,noneoftheseML-based
methodshaveshownconclusiveperformancesorhavebeen
adopted as a de facto solution to tractography. One rea-
sonforthismightbethelackofwell-definedandextensive
frameworkstotrain,evaluate,andcomparethesemethods.
In this paper, we describe several datasets and evalu-
ationtoolsthatcontainusefulfeaturesforMLalgorithms,
alongwiththevariousmethodsproposedintherecentyears.
Wethendiscussthestrategiesthatareusedtoevaluateand
comparethosemethods,aswellastheirshortcomings. Fi-
nally,wedescribetheparticularneedsofMLtractography
methodsanddiscusstangiblesolutionsforfutureworks.
KEYWORDS
DiffusionMRI,tractography,machinelearning,benchmark
1
INTRODUCTION
2
1
Inthefieldofdiffusionmagneticresonanceimaging(dMRI),tractographyreferstotheprocessofinferringstreamline
structuresthatarelocallyalignedwiththeunderlyingwhitematter(WM)dMRImeasurements[1]. Asimpleapproach
toobtainsuchstreamlinesisaniterativeprocessinwhich,startingfromaseedpoint,anestimateofthelocaltissue
orientation is determined and followed for a certain step length before repeating the orientation estimation at the
newposition. Thetrackingproceduremaybedeterministic[2,3](ateachpoint,thealgorithmfollowsthestrongest
orientation)orprobabilistic[4,5,6](ateachpoint,thealgorithmsamplesadirectioncloselyalignedwiththestrongest
orientation). Trackingmayalsobeglobalassomemethodsrecoverstreamlinesallatonce[7,8,9]. Inbetweenthelocal
andglobalmethodsisthecategoryofshortest-pathmethods,includingfrontevolution,simulateddiffusion,geodesic,
andgraph-basedapproaches[1]. Ultimately,thecollectionofalltrajectoriescreatedinthatwayiscalledatractogram.
In traditional methods, the estimate of the local tissue fiber orientation is usually inferred from an explicit and
local model which fits the (local) diffusion data. These local models include diffusion tensor models [3, 10], multi-
tensormodels[11],andothermethodsthataimatreconstructingthefiberorientationdistributionfunction(fODF)
likeconstrainedsphericaldeconvolution(CSD)[12,13], tonameafew. However, thechoiceofthebestmodelisby
itselfdifficult[14,15],asitdependsonvariousfactorssuchasdataacquisitionprotocolortargetedWMregions,and
thereforehasadirectinfluenceonthequalityofanobtainedtractogram[16]. Moreover,traditionalmethodsbased
onlocalorientationalonearepronetomakecommonmistakes,suchasmissingthefullspatialextentofbundlesand
producingagreatamountoffalsepositiveconnections[15].
Another important factor for the performance of a tractography method are the actual rules that regard the
progressionofasinglestepaswellassimpleglobalpropertiesofanindividualstreamline. Traditionalmethodsmay
defineseveralengineered,or"manually-defined",high-levelruleswiththeaimofimprovingtheanatomicalplausibility
oftherecoveredtractogram. Instancesoftheseareconstraintsonstreamlinelength(i.e. filteringstreamlinesthatare
too long or too short), streamline shape (e.g. filtering streamlines with sharp turns), or progression rules that make
streamlines"bounceoff"theWMborderwhentheyareabouttoleavetheWMmaskwithacertainangle[17,18]. In
thesamewayasmodelingnoiseandartifacts,anddefiningtherightlocalmodel,alsothedesignofthesehigh-levelrules
hasadirectimpactontheperformanceofatractographymethod[16,15].
PHILIPPEPOULIN ET AL.
Toaddresstheseinherentdifficulties,recentproposalssuggestthatmachinelearning(ML)algorithms,supervised
or unsupervised, may be used to implicitly learn a local, global or contextual fiber orientation model as well as the
trackingprocedure. Approachesrangingfromtheapplicationofself-organizingmaps(SOM)[19,20],randomforests
(RF)[18,21],MultilayerPerceptrons(MLP)[22,23],GatedRecurrentUnits(GRU)[24,25,26],aswellasConvolutional
NeuralNetworks(CNN)[27]andAutoencoders[28],havebeenemployedatthecoreoftractographytodrivestreamline
progression. Apart from the differences in their underlying architecture, these ML methods differ substantially in
aspectsoftheexactproblemformulation,e.g. definitionoftheinputdatatothemodel,modelingthepredictionsasa
regression[24,28]orclassificationproblem[21,22],oreventhegeneraltractographyapproach,i.e. whole-brain[24,25]
orbundle-specific[27,28,29,26]. Thefactalonethattheseapproachesdifferinseveralaspects,makesitdifficultto
drawconclusionsonthevalueofeachoftheindividualmodelingchoices.
Furthermore, whiletheabovementionedapproachesconstitutethemainideasforapplyingMLdirectlytothe
process of tractography, machine learning and especially deep learning (DL) methods have been applied in related
fields. StackedU-NetswereproposedtosegmentthevolumeofindividualwhitematterbundlesfromimagesoffODF
peaks[30]. Itwasalsosuggestedtopredictfiberorientationsfromrawdiffusiondatabasedonconvolutionalneural
networks(CNN)[31]. Severalideasforstreamlineclusteringorstreamlinesegmentationhavebeenproposed,including
a CNN based on landmark distances [32], a long short-term memory (LSTM)-based siamese network for rotation
PHILIPPEPOULIN ET AL.
3
invariantstreamlinesegmentation[33],andaCNNapproachforstreamlineclusteringbasedonthesequenceoftheir
coordinates [34, 35]. Even though the mentioned works are closely related to tractography and contribute to the
commongoalofimprovedanalysisofthewhitematteranatomyofthehumanbrain,werestrictourfocusexclusivelyon
thedirectapplicationofML(andespeciallyDL)fortractography,withtheexplicitgoalofproducingstreamlinesand
addressingtheweaknessesoftraditionalmethods. Forthatreason,werefertheinterestedreadertotherespective
referencesformoredetails.
Animportantfactorforeffectivelyadvancingthisfieldofresearchisacommonandappropriatemethodologyfor
trainingandevaluatingtheperformanceofdifferentapproaches,whichiscurrentlylacking. Overtheyears,multiple
challenges have been proposed to assess the performance of conventional tractography methods, and a clear and
exhaustivereviewisprovidedbySchillingetal.[14]. However,wearguethatthedesignofthesechallengesistypically
inappropriateforMLmethods. Infact,the2015ISMRMTractographyChallenge[15](alongwiththeTractometerevalua-
tiontool[16])hasbeenadoptedasthetoolofchoiceforbenchmarkingnewMLtractographypipelines[24,18,23,25].
Unfortunately,severalinherentflawsarisingspecificallyinthecontextofMLmakeitdifficulttoperformafaircom-
parisonbetweentheresultsobtainedfromdifferentMLpipelines. Inparticular,diffusiondatapreprocessingisleftto
participants(dissimilarinputs),trackingseedsandatrackingmaskarenotalwaysgiven(varyingtestenvironment),
thetestdiffusionvolumeissometimesusedfortraining(datacontamination),trainingstreamlinesarenotprovided
(disparatetrainingdata), and testing on a single synthetic subject means that any computed estimator of a model's
performance is unreliable (small sample size). Against the background of a prospectively increasing number of ML-
basedapproachestacklingtheproblemoftractography,acarefullydesignedevaluationframeworkthatappropriately
addressesthespecificrequirementsofMLmethodshasthepotentialtosupportandfacilitateresearchinthisfieldin
theupcomingyears.
Inthispaper,wefollowathreefoldstrategy. First,weintroducethecurrentlyavailabledatasetsandevaluation
toolsalongwithusefulfeaturesandweaknessesregardingmachinelearning. Then,weprovideacomprehensivereview
ofexistingML-basedtractographyapproachesandderiveasetofkeyconceptsdistinguishingthemfromeachother.
Subsequently,weidentifyanddiscussthestrategiesforevaluationoftractographypipelinesandidentifyissuesand
limitations arising when applied to ML-based tractography methods. We finally describe important features for an
appropriateevaluationframeworkthecommunityoughttoadoptinthenearfuturetobetterpromotedata-driven
streamlinetractographyandpointoutthepotentialadvantagesforresearchindata-drivenstreamlinetractography.
2 ANNOTATED DATASETS AND EVALUATION TOOLS
Overtheyears,manydiffusionMRIdatasetswereproducedandannotated,eitheraspartofachallengeorresearch
papers. Inthissection,weoverviewseveraldatasetsthathavebeenusedtotrainand/orvalidatesupervisedlearning
algorithmsfortractography. Specifically,weselecteddatasetsthatofferbothdiffusiondataandstreamlines. Selected
datasetsalsoneededtohaveeitherclearlydefinedevaluationmetrics,ortobelargeenough(morethan50subjects)to
beconsideredasstandalonetrainingsets. Weincludedatasetsthatareeitherpubliclyavailableorsimplymentionedin
aresearchpaperwithoutapublicrelease.
Weexcludeddatasetsorchallengesfocusedonnon-humananatomy(e.g. ratormacaque),wherethegroundtruthis
hardertodefineandresultsmightbehardertogeneralizetohumananatomy(fordata-drivenalgorithms),likethe2018
VOTEM Challenge [36] (my.vanderbilt.edu/votem/). Moreover, we left out datasets focused only on pathological
caseslikethe2015DTIChallenge[37],becauseweconsiderittooearlyfordata-driventractographyalgorithms,atleast
untilmoreconclusiveresultsonhealthysubjects. Wealsoexcludedtractographyatlaseswhentrackingwasdoneona
4
PHILIPPEPOULIN ET AL.
singlediffusionvolume,usuallyaveragedovermultiplesubjects(e.g. HCP842[38]),becauseresultstendtobeoverly
smoothandunsuitedforMLmethods. However,weincludearecentcasewhentrackingwasdoneforeachsubject: the
100-subjectsWMatlasofZhangetal.[39].
Whilealltheselecteddatasetsareusefulinonewayoranotherfordata-drivenmethods,theydifferinmultiple
ways,whicharedetailedinthefollowingsubsectionsandsummarizedinTable1. Thelistedpropertiesarethefollowing:
• Name: Thedatasetnameandreference
Year: Theyearofpublicationofthedatasetorpaperusingthedataset
•
Public: Isthedataset(diffusiondataandstreamlines)publiclyavailable?
•
Real: Isthediffusiondataarealacquisitionorisitsimulated?
•
• Human: Doesthediffusiondatarepresentthehumanbrainanatomy?
Subjects: Thenumberofsubjectsoracquisitions
•
Bundles: Thenumberofbundlesortracks(ifstreamlinesareavailable)
•
• GT:Isagroundtruthknown? Forrealacquisitions,streamlinesvalidatedbyahumanexpert(e.g. neuroanatomist)
areconsideredasGTdespitethefactthattheseannotationsaresubjecttointer-raterandintra-ratervariations.
• Metrics: Well-definedevaluationmetricsareavailablewiththisdataset.
Split: Isthedatasetsplitintoatrainingandtestingsetthatfutureworkscanrelyon?
•
Notethatthenotionof"groundtruth"referstoanindisputablebiologically-validatedlabelassignedtoanobserved
variable. Inmedicalimaging,suchgroundtruthmaybeobtainedwithabiopsy[40],throughoutcarefulcomplementary
analysis[41]orbyhavingseveralexpertsagreeingonagivendiagnostic[42]. Unfortunately,suchrestrictivedefinition
ofagroundtruthisunreachablemostofthetime,especiallyforwhitemattertracksobtainedfromtractography,where
noexpertcantrulyassesstheexistence(ornon-existence)ofagivenstreamlineinahumanbrainfromMRIimages
only. Infact,onlysynthetically-generatedstreamlinesorman-madephantomscanbeconsideredasreal"groundtruth".
Despitethat,forthepurposeofthispaper,wealsousetheterm"groundtruth"foranydatathathasbeenmanually
validatedbyahumanexpert,typicallyaneuro-anatomist. Inthemedicalimagingfield,thisannotateddatawouldbe
calledagoldstandard,whileintheartificialintelligencecommunity,itmightbecalledweaklyannotateddata. Although
suchannotationsdonotmeetthefundamentaldefinitionofagroundtruth,itisnonethelesswidelyacceptedbythe
medicalimagingAIcommunity[43].
2.1 TheFiberCupdatasetandtheTractometertool
OriginalFiberCupTractographyContest(2009)
Fillardetal.proposedtheFiberCupTractographyContest[44,51]inconjunctionwiththe2009MICCAIconference. The
goalwastoquantitativelycomparetractographymethodsandalgorithmsusingaclearandreproduciblemethodology.
TheybuiltarealisticdiffusionMR7-bundlephantomwithvaryingconfigurations(crossing,kissing,splitting,bending).
Theorganizersacquireddiffusionimageswithb-valuesof2000,4000,and6000s/mm2,andusedisotropicresolutions
of3mmand6mm,resultingin6differentdiffusiondatasets. Contestantswereprovidedalldatasets(butnottheground
truth)andwerefreetoapplyanypreprocessingtheywantedonthediffusionimages. Evaluationwasdonebychoosing
16 specific voxels, or seed points, in which a unique fiber bundle is expected. Participants were expected to submit
a single fiber bundle for each of those seed voxels. Quantitative evaluation was done by comparing the 16 pairs of
candidateandgroundtruthfibersusingasymmetricRootMeanSquareError(sRMSE).
WhiletheFiberCupTractographyContestmakesagoodtestcaseforsimpleconfigurations,itdoesnotrepresenta
truehumananatomyanddoesnotimposeachoiceofb-valueandpreprocessing,whichcaninducesignificantdifferences
PHILIPPEPOULIN ET AL.
TABLE 1 Annotateddatasets.
Year
2009
2012
2011
2012
2013
2015
2018
2018
2018
2018
2018
Name
Fibercup[44]
SimulatedFibercup[45]
Tracula[46]
HARDI2012[47]
HARDI2013[48]
ISMRM2015[15,16]
HAMLET[29]
PyT(BIL&GIN)[49]
BST(BIL&GIN)[50]
TractSeg(HCP)[30]
Zhangetal.(HCP)[39]
Real Human
(cid:88)(cid:3)
Public
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
Subjects
1
1
67
2
2
1
83
410
39
105
100
GT Metrics
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
5
Split
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
58+198 (cid:88)(cid:3)
(cid:88)(cid:3)
Bundles
7
7
18
7
20
25
12
2
5
72
indata-drivenmethods. Also,itdoesnotprovideanytrainingstreamlines,andisthususefulonlyasavalidationtool
for ML-based methods. Furthermore, the fact that it contains only one subject makes it hard to evaluate the true
generalization capability of an ML method trained and tested on that dataset. However, it is the only dataset that
provides seed points in order to have a uniform test environment, which is of utmost importance when comparing
ML-basedalgorithms. Intheend,itisunclearifforML-basedmethodstherewouldbeanycorrelationbetweenagood
performanceontheFiberCupcontestandgoodperformanceonhumananatomy.
Tractometerevaluationtool(2013)
In2013,Côtéetal.developedtheTractometerevaluationtool,tobeusedalongsidetheoriginalFiberCupdata,with
theaimofprovidingquantitativemeasuresthatbetterreflectbrainconnectivitystudies. UsingaRegionofInterest
(ROI)-based filtering method, a complete tractogram can be evaluated on global connectivity metrics, such as the
numberofvalidandinvalidbundles. Furthermore,theyproposetwoseedingmasks: acompletemask(mimickingabrain
WMmask),andaROImask(mimickingGM-WMinterfaces). Thetractometerwasdesignedtoaddressthefactthat
"metricsaretoolocalandvulnerabletotheseedsgiven,and,asaresult,donotcapturetheglobalconnectivitybehavior
ofthefibertrackingalgorithm"[16].
SimulatedFiberCup(2014)
In2014,Neheretal.proposedasimulatedversionoftheFiberCup,allowingnewtrackingalgorithmstobetestedusing
multipleacquisitionparameters[52]. ThesimulateddatacanbeusedalongsidetheTractometertooldesignedforthe
originalphantom.
Wilkinsetal.alsodevelopedasyntheticversionoftheFiberCupdataset,butdidnotpubliclyreleasethedata[45].
Unfortunately,withregardstoMLmethods,thesimulatedFiberCupdatasetsuffersfromthesameshortcomingsasthe
originalFiberCupdatasetasitcontainsonlyonenon-humansubjectwhosedataisnotsplitaprioriintoatrainingand
testingset.
PHILIPPEPOULIN ET AL.
6
2.2 Tracula(2011)
Yendikietal.[46]publishedtheTraculamethodforautomatedprobabilisticreconstructionof18majorWMpathways.
Itusespriorinformationontheanatomyofbundlesfromasetoftrainingsubjects. Thetrainingsetwasbuiltfrom34
schizophrenia patients and 33 healthy controls, using a 1.5T Siemens scanner as part of a multi-site MIND Clinical
ImagingConsortium[53]. Thediffusionimagesinclude60gradientdirectionsacquiredwithab-valueof700s/mm2,
alongwith10b=0images,withanisotropicresolutionof2mm. Whole-braindeterministictrackingwasperformed,
followedbyexpertmanuallabelingusingROIsfor18majorWMbundles. Thedatasetalsoincludesameasureofthe
inter-raterandintra-ratervariabilityfortheleftandrightuncinate.
Toourknowledge,thisistheearliestapparitionofalarge-scalehumandatasetwithexpertannotationofstreamlines.
Itisalsotheonlydatasetthatincludesameasureofinter-raterandintra-ratervariability,whichisadesirablefeaturefor
MLmethods(alsodiscussedlaterinSection4.4). Unfortunately,thecompletesetofdiffusionimagesandstreamlines
hasbeenincorporatedintothemethodandisnotpublic.
2.3 HARDIReconstructionChallenges
HARDIReconstructionChallenge(2012)
Daduccietal.organizedthe2012HARDIReconstructionChallenge[48]attheISBI2012conference. Thegoalofthe
challengewastoquantitativelyassessthequalityofintra-voxelreconstructionsbymeasuringthepredictednumber
offiberpopulationsandtheangularaccuracyofthepredictedorientations. Atrainingsetwasreleasedpriortothe
challenge,andatestsetwasusedtoscorethealgorithms. Assuch,the2012HARDIdatasetcontainsdiffusionimages
butnostreamlines.
Participantscouldrequestacustomacquisition(onlyonce)bysendingalistofsamplingcoordinatesinq-space,
andtheorganizerswouldthenproduceasimulatedsignalforthegivenparameters. A 16 × 16 × 5volumewasthen
produced,containingsevendifferentbundlesattemptingtorecreaterealistic3-Dconfigurations. Themetricsproposed
by the authors are ill-posed for ML-based methods because of the limited context available and the focus on local
performances. LiketheFiberCup,itwouldonlybeusefulasavalidationtoolgiventhelackoftrainingstreamlines,a
limitednumberofbundles(onlyseven)andalimitednumberofnon-humansubjects(onlytwo).
HARDIReconstructionchallenge(2013)
The2013HARDIReconstructionChallenge[47]wasorganizedoneyearlaterattheISBI2013conference. ForML-based
methods,threeimprovementsarerelevantcomparedtothe2012challenge: amorerealisticsimulationofthediffusion
signal, a new evaluation system based on connectivity analyses and a larger set of 20 bundles. Indeed, data-driven
methodstrytolearnanimplicitrepresentationwithoutimposingamodelonthesignal,whichmeansthatthesignal
usedfortrainingandtestingshouldbeascloseaspossibletothatinclinicalpractice. Furthermore,themainbenefitof
data-drivenmethodsistheabilitytousecontextinordertomakegoodpredictionsinamultitudeofconfigurations,
which means they have the potential to particularly improve connectivity analyses. Therefore, it would be a better
validationtoolforML-basedmethodsthanthe2012HARDIReconstructionChallenge. Nonetheless,thedatasetsuffers
fromaninherentlimitationasitcontainsonlytwonon-humansubjects.
PHILIPPEPOULIN ET AL.
7
ISMRMTractographyChallenge(2015)
FIGURE 1 2015ISMRMTractographyChallengedatagenerationprocess(Takenfrom www.tractometer.org)
2.4
This dataset has been designed for a tractography challenge organized in conjunction with the 2015 ISMRM con-
ference[15]. Duringthechallenge,participantswereaskedtoreconstructstreamlinesfromasynthetichuman-like
diffusion-weighted MR dataset which was simulated with the aim of replicating a realistic, clinical-like acquisition,
includingnoiseandartifacts. Theavailabledataconsistsofadiffusiondatasetwith32b=1000s/mm2 imagesandone
b=0image,with2mmisotropicresolution,aswellasaT1-likeimagewith1mmisotropicresolution. Sincealldatawas
generated from an expert segmentation of 25 bundles, in theory, a perfect tracking algorithm should only produce
exactlythesespecificbundles. Unfortunately,asfortheHARDIandFiberCupdatasets,the2015ISMRMTractography
Challengecontainsdatafromalimitednumberofsubjects(onlyone)andlacksaclearseparationbetweentrainingand
testingdata. Nonetheless, incombinationwiththeTractometer tool[16], thisdatasethasoftenbeenusedtoassess
ML-basedtractographymethods. Figure1showsthedatagenerationprocessforthechallenge.
Once a tractogram has been generated using the challenge diffusion data, the tractometer tool uses a "bundle
recognitionalgorithm"[54]toclusterthestreamlinesintobundles. Thegeneratedbundlesarethencomparedtothe
groundtruth,producinggroupsof"validbundles"and"invalidbundles",dependingonwhichregionsofthebrainthe
streamlinesconnect. Streamlinesthatdonotcorrespondtoagroundtruthbundleareclassifiedas"Noconnections"
streamlines. ThemetricscomputedbythemodifiedTractometerfortheTractographyChallengeareasfollows:
• Validbundles(VB):Thenumberofcorrectlyreconstructedgroundtruthbundles.
InvalidBundles(IB):Thenumberofreconstructedbundlesthatdonotmatchanygroundtruthbundles.
•
• ValidConnections(VC):Theratioofstreamlinesinvalidbundlesoverthetotalnumberofproducedstreamlines.
InvalidConnections(IC):Theratioofstreamlinesininvalidbundlesoverthetotalnumberofproducedstreamlines.
•
• NoConnections(NC):Theratioofstreamlinesthatareeithertooshortordonotconnecttworegionsofthecortex
overthetotalnumberofproducedstreamlines.
•
•
8
PHILIPPEPOULIN ET AL.
BundleOverlap(OL):Theratioofgroundtruthvoxelstraversedbyatleastonestreamlineoverthetotalnumberof
groundtruthvoxels.
BundleOverreach(OR):Theratioofvoxelstraversedbyatleastonestreamlinethatdonotbelongtoaground
truthvoxeloverthetotalnumberofgroundtruthvoxels.
F1-score(F1): Theharmonicmeanofrecall(OL)andprecision(1-OR).
Thedefinitionofstreamline-orientedmetrics(VB,IB,VC,IC,NC)andvolume-orientedmetrics(OL,OR,F1)means
thatthereisnosinglenumberthatcanfullyassesstheperformanceofanalgorithm. Forexample,deterministicmethods
oftenscorehigheronstreamline-orientedmetricscomparedtoprobabilisticmethods. Assuch,athoroughreviewofall
scoresmustbeperformedinordertoproperlycomparealgorithms,andinmanycases,thechoiceofanalgorithmover
anothermaydependonaspecificuse-case(e.g. bundlereconstructionvs. connectivityanalysis).
•
2.5 HAMLET(2018)
Tovalidatetheirmethod,Reisertetal.[29]usedadatasetof83humansubjectsfromtwoindependentcohorts. Thefirst
cohortcomprises55healthyvolunteers,allscannedbyaSiemens3TTIMPRISMAMRIscanner. Thesecondcohorthas
28volunteersscannedwithaSiemensTIMTRIO.Thefirstcohortwasusedfortrainingwhilethesecondonewasused
fortesting. Subjectsinthesecondcohortwerescannedtwicefortest-retestexperiments,someuniquecharacteristic
tothatdataset. Thereferencestreamlineswereobtainedbyfirsttrackingthewholebrainwithglobaltractography,
andthenbysegmentingthestreamlinesfor12bundleswithaselectionalgorithminMNIspace. Unfortunately,the
recoveredstreamlineshavenotbeenmanuallyvalidatedbyanexpert.
2.6 DatasetsbasedontheBIL&GINdatabase
Bundle-SpecificTractography(2018)
Rheaultetal.proposedabundle-specifictrackingmethodbasedonanatomicalpriorsthatimprovestrackinginthe
centrumsemiovalecrossingregions[50]. Usingmultipletractographyalgorithms, theytrackedandsegmentedfive
bundles(ArcuateFasciculus-AFleft/right,CorpusCallosum-CC,PyramidalTracts-PyTleft/right)in39subjectsfrom
theBIL&GINdatabase[55]. Tocomparealgorithms,theyusedanautomaticbundlesegmentationmethodbasedon
clearanatomicaldefinitions. Inaddition,theydefinedseveralperformancemetrics,suchasbundlevolume,ratioofvalid
streamlines,andefficiency. However,thetractogramsandautomaticbundlesegmentationprocedurewereneithermade
publicnorvalidatedbyanexpert. Suchadataset,alongwiththeevaluationprocedure,couldbeextremelyusefulto
assessifdata-drivenmethodscanreliablylearnthestructureofaspecificbundleandreconstructitinunseensubjects.
Apopulation-basedatlasofthehumanpyramidaltract(2018)
Chenotetal.createdastreamlinedatasetoftheleftandrightPyTbasedonapopulationof410subjects[49],alsofrom
theBIL&GINdatabase[55]. Todoso,theycombinedmanualROIsalongthebundles'pathwayandthebundle-specific
tractographyalgorithmofRheaultetal.[50]. Thequalityofthesegmentationsandthehighnumberofsubjectswould
makethisanoteworthytrainingdatasetfordata-drivenmethods. UnfortunatelyforMLmethods,onlytwobundles
wereexamined. Furthermore,whiletheprobabilitymapsoftheatlashavebeenrenderedpublic,thetractogramsare
stillunavailable.
9
PHILIPPEPOULIN ET AL.
2.7 DatasetsbasedontheHCPdatabase
TractSeg(2018)
Wasserthaletal.proposedadata-drivenmethodforfastWMtractsegmentationwithouttractography[30]. Indoing
so,theybuiltanimpressivedatasetof72manually-validatedbundlesfor105subjectsfromtheHumanConnectome
Project(HCP)diffusiondatabase[56,57]. Tractogramswereobtainedviaafour-stepsemi-automaticapproach:
1. Tractography(Multi-ShellMulti-TissueCSD[5])
2.
3. Tractrefinement(ManualROIs[59]+QuickBundles[60])
4. Manualqualitycontrolandcleanup
Initialtractextraction(TractQuerier[58])
To the best of our knowledge, this is the largest public database to include both diffusion data and reference
streamlines. Nofurtherpreprocessingofthediffusiondataisneededbecauseofthestandardprocedureof[57]. The
authorsdefinedvolume-orientedmetricssuchasDicescore[61],butdidnotofferanystreamline-orientedmetrics
astheirmethodpredictsavolumesegmentation. Thehighnumberofsubjectsandbundlesmakesthisaremarkable
trainingset.
Inasubsequentpaper,thesameauthorsre-usedasubsetof20bundlesoftheTractSegdatasettotrainandvalidate
their TOM ML algorithm [27]. However, as for original 72-bundle dataset, the TOM dataset does not come with a
predefinedsetoftrainingandtestingdataandnoformalevaluationprotocolthatuserscouldrelyonhasbeenproposed.
Zhangetal. (2018)
Zhang et al. [39] built a WM fiber atlas using 100 HCP subjects. They first generated streamlines for all subjects
usingatwo-tensorunscentedKalmanfiltermethod[62],andsampled10,000streamlinesfromeachsubjectaftera
tractographyregistrationstep. Then,usingahierarchicalclusteringmethod,theauthorsgeneratedaninitialWMfiber
atlas of 800 clusters. Finally, an expert neuroanatomist reviewed the annotations in order to accept or reject each
cluster, and provided the correct annotations when the initial annotation was rejected. The final, proposed atlas is
comprisedof58bundles(eachcomposedofmultipleclusters),alongwith"198shortandmediumrangesuperficialfiber
clustersorganizedinto16categoriesaccordingtothebrainlobestheyconnect"[39].
Whiletheatlasispublic,thesampledstreamlinesfromthe100subjectsareallmergedintothesingletemplate. In
orderforMLmethodstobenefitfromthisdataset,thestreamlineswouldneedtobeseparatedbackintothespaceof
theparticularoriginalsubjects. Forthisreason,wedonotconsiderthisdatasettobe"public",inthecontextofmachine
learning.
3 MACHINE LEARNING METHODS FOR TRACTOGRAPHY
Forthisreview,weregardallsupervisedmachinelearningmethodspublishedinpeer-reviewedjournals,conferencesor
onarXiv(arxiv.org)andbiorXiv(biorxiv.org). Weaddedtherequirementthatmethodsneededtobespecifically
designedfortractography,i.e. withthepurposeofpredictingacontextualstreamlinedirection(andnotreconstructing
alocal,non-conditionalfODForclusteringstreamlines). Thiscriterionincludeswhole-brainaswellasbundle-specific
tractographymethods. AsummaryofthemainpropertiesforallreviewedmethodsisprovidedinTable2.
10
TABLE 2 Mainpropertiesofdata-drivenmethodsfortractography.
PHILIPPEPOULIN ET AL.
Prediction
Classification
Regression
Regression
Classification
Regression
Regression
Regression
Regression
Implicit
stop
(cid:88)(cid:3)
(cid:88)(cid:3)
(cid:88)(cid:3)
dMRI
input
ResampledDWI
SH
SH
SH
SH
ResampledDWI
RawDWI
fODFpeaks
Spatial
context
50samples
1x1x1voxel
1x1x1voxel
1x1x1voxel
1x1x1voxel
3x3x3voxels
EntireWM
EntireWM
Method
Neheretal.[18]
Poulinetal.[24]
Poulinetal.[26]
Benouetal.[25]
Jörgensetal.[22]
Wegmayretal.[23]
Wasserthaletal.[27]
Reisertetal.[29]
Model
RF
GRU
GRU
GRU
MLP
MLP
CNN
CNN-like
Temporal
context
1lastdirection
2lastdirections
4lastdirections
Full
Full
Full
N/A
N/A
RF: Random Forest; MLP: Multilayer perceptron; GRU: Gated recurrent unit; CNN: Convolutional neural network; SH: Spherical har-
monics coefficients; fODF: fiber Orientation Distribution Function; Implicit stop: indicates if a method learns its tracking stopping
criterionorifitreliesonausualexplicitcriterion.
RandomForestclassifier
Tothebestofourknowledge,Neheretal.werethefirsttoproposeamachinelearningalgorithmfor(deterministic)
tractography[21]. TheyemployaRFclassifiertolearnamappingfromrawdiffusionmeasurementstoadirectional
proposalforstreamlinecontinuation. Aftercollectingseveralofsuchproposalsinalocalneighborhoodofthecurrent
streamlineposition(radius: 25%ofthesmallestsidelengthofavoxel),theseareaggregatedinavotingschemetofinally
arriveatasingledirectioninwhichtogrowthestreamline.
To define reference streamlines for their experiments, the authors employ several tractography pipelines and
traintheirclassifieroneachoftheresultingtractograms. Theydeterminethebesttrainedmodelbyevaluatingthe
performance of each on a replication of the FiberCup phantom (based on the Tractometer metrics of [16]). Finally,
comparingtheperformanceofthelattertoallotherreferencepipelines,theyreportasuperiorperformanceoftheir
trackingmodeloverallotherapproaches. Whiletractogramswerescoredonasimulatedphantom(i.e. norealanatomy),
extendedexperimentspresentedinasubsequentpaper[18]confirmthesuperiorityoftheirapproachonthe2015
ISMRMTractographyChallengedataset(simulateddataofahumananatomy).
GatedRecurrentUnit(GRU)Tracking
Hypothesizing"thattherearehigh-orderdependenciesbetween"thelocalorientationatapointofastreamlineand
theorientationsatallotherpointsonthesamestreamline,Poulinetal.proposedarecurrentneuralnetwork(RNN)
based on a GRU [24] to learn the tracking process. Their method implements an implicit model mapping diffusion
measurementstolocalstreamlineorientationswhichnotonlydependsonmeasurementsinalocalcontext,butonall
datapreviouslyseenalongtheextentofaparticularstreamline. Asopposedto[18,21],theRNNmodelisimplemented
asaregressionapproach. Intheirexperiments,theauthorsshowthatarecurrentmodel(whentrainedonreference
streamlines obtained using deterministic CSD-based tractography [6]) was able to outperform most of the original
submissionsinthe2015ISMRMTractographyChallengewithrespecttotheTractometerscores(discussedinsection2.4).
11
PHILIPPEPOULIN ET AL.
DeepTracker
Inasubsequentpaper,Poulinetal.[26]againsuggestedusingaGRU,butinabundle-specificfashion. Whilethemodel
architectureisverysimilar,itwastrainedonadatasetof37realsubjects,eachwithacuratedsetofstreamlinesfor
bundles. Aftertrainingasinglemodelforeachoftheselectedbundles,theauthorsshowedpromisingresultscompared
toexistingmethods,perhapsindicativethatthedifficulttaskoflearningtotrackstreamlinesnecessitatesmoredata
thanpreviouslythought.
DeepTract
Morerecently,BenouandRiklin-Raviv[25]proposedaGRU-basedrecurrentneuralnetworksimilartothatofPoulin
etal.[24]. Intheirmethod,theydirectlyusetheresampleddiffusionsignalasinputtothemodel(likeNeheretal.[18]),
in order to estimate a discrete, streamline-specific fODF representation which they refer to as "conditional fODF"
(CfODF).Insteadofpredictinga3Dorientationvectorusingaregressionapproach,theauthorsimplementtheirmodel
asaclassifierenablingthemtointerprettheprobabilitiesobtainedfordiscretesampleddirections(i.e. theclasses)as
thementionedCfODF.ThisfODF-basedformulationfurtherallowsforaninherentlydefinedcriterionforstreamline
termination based on the entropy of the CfODF. The proposed model can be employed for both deterministic and
probabilistictractography.
Like Poulin et al. [24], the authors trained and tested their method on the 2015 ISMRM Tractography Challenge
dataset. Theyreportresultsaftertrainingtheirmethodonthedatasetgroundtruthaswellasonstreamlinesobtained
withtheMITKdiffusiontool[63].
Multi-LayerPerceptronPoint-WisePrediction
Jörgensetal.[22]proposeamulti-layerperceptron(MLP)topredictthenextstepofastreamline. Like[18,21,24],
theirmethodtakesasinputthediffusionsignalandthusavoidsexplicitdMRImodel-fitting. Theauthorsimplemented
differentconfigurationsoftheirproposedMLPsuchasthreedifferentinputscenarios(point-wiseinputvsregion-wise
inputwithandwithoutconsideringpreviousorientations),differentapproachestoaggregatetheoutput(maximum
likelihood,mathematicalexpectationofthecategoricalpredictionandregression)aswellasthevotingschemeproposed
byNeheretal.[21]. Resultsrevealthatthebestconfigurationsarethosehavingtheprevioustwodirectionsincludedin
theinputofthenetworkthusshowingthattemporalcontextisakeycomponentfordata-driventractography. Also,the
regressionandclassificationapproachesledtosimilarresultsandtheuseofregion-wiseinformationdidnotprovide
anysubstantialimprovementovertheuseofpoint-wiseinformation.
LikePoulinetal.[24]andBenouandRiklin-Raviv[25],theauthorstrainedandtestedtheirmethodonthe2015
ISMRMTractographyChallengedataset(butdidnotusetheTractometertool). Unfortunately,theydidnotestimatethe
trackingcapabilitiesoftheirmethodastheyonlymeasuredpoint-wiseangularerrorswhenpredictingthenextstepofa
streamline.
Multi-LayerPerceptronRegressionTracking
Asimilarapproachsuggestedby[23]employsaMLPtopredictthenextdirectionofastreamlinethroughregression.
At each point, the input of the model is given by all diffusion measurements in a cubic neighborhood, along with a
certainnumberofpreviousstepsforthecurrentstreamline. Inthatway,theauthorsprovidetheMLmodeldirectlywith
diffusioninformationinalocalneighborhood(spatialcontext)aswellasanotionof"history"ofthecurrentstreamline
(temporalcontext). Definingtheirreferencestreamlinesastractogramsobtainedwithastandardtractographymethod
frominvivodatasets,theytraintheirmodelonthreesubjectsfromtheHCPdatabase. Experimentalvalidationson
the 2015 ISMRM Tractography Challenge dataset reveal that their model outperforms some ML methods [18, 24] in
12
PHILIPPEPOULIN ET AL.
mostTractometermetrics. However,asdemonstratedbylowoverlapscores,theauthorsacknowledgethattheirmodel
produces"ratherconfinedbundleswithlittlespread",especiallyincontrastto[18,24]. Whilethestrengthofthismodel
istoexplicitlyprovideinformationfromalocalneighborhood,likeforJörgensetal.[22],thenotionofcontextalongthe
streamlineislimitedandneedstobedefinedbeforetraining. Sincetheidealtemporalcontext(intermsofstreamline
length, orsteps)isstillunknown, thiscouldpotentiallyprohibitthemodelfromtakingadvantageofallinformation
relevanttostreamlinecontinuation.
Tractorientationmappingusinganencoder-decoderCNN
Wasserthaletal.[27]proposedadata-driven,bundle-specifictrackingmethod. AsopposedtotheotherMLmethods
reportedinthispaper,theauthorsdonottrytodirectlyreconstructstreamlinesperse. Instead,theirproposedTract
OrientationMapping(TOM)methodpredictsbundle-specificfODFpeaksthatarethenusedbyadeterministictracking
method. First,CSDisusedtoextractthreeprincipaldirectionsinallWMvoxels. Then,aU-NetCNN[64]istrainedto
mapthesefODFpeakstobundle-specificpeaks,i.e. peaksthatareonlyrelevantforthestreamlinesofagivenbundle.
TheirCNNtakesasinput9channels(thethreefODFpeaks)andoutputs60channels,i.e. a3Dbundle-specificfODF
vectorforeachofthe20bundlestheyarelookingtorecover. Whiletherecoveredbundle-specificpeakscanbeusedin
differentways,theauthorsshowthatusingthemdirectlyasinputtoadeterministicMITKdiffusiontractographygives
someofthebestresults. Theapproachwastrainedandtestedon105HCPsubjects,eachwithreferencestreamlines
producedbyasemi-automaticdissectionof20largeWMbundles(whichtheyrecentlyrenderedpublic[30]).
HAMLET
Inasimilarlineofthought,intheirHAMLETproject(HierarchicalHarmonicFiltersforLearningTractsfromDiffusionMRI)
Reisertetal.[29]maprawsphericalharmonicsoforder2toasphericaltensorfield. Inthatsense,likeWasserthaletal.
[27],theirMLmethoddoesnotoutputstreamlinesbutinsteadvoxel-wisebundle-specifictensorsthatcansubsequently
beusedasinputtoaclassicaltractographymethod. Themagnitudeoftheproduced 3 × 3tensorindicatesthepresence
ofaspecificbundlewhereasthetensororientationpredictsthelocalstreamlinedirection. Theirmethodimplementsa
multi-resolutionCNNwithrotationcovariantconvolutionoperations. Theytrainedandtestedtheirmethodontwo
in-housedatasetscomprisingatotalof83humansubjects. The12bundlesandtheirassociatedreferencestreamlines
havebeenobtainedwithglobaltractographyandautomaticbundleselectionmethod. Unfortunately,thereferencedata
wasnotmanuallyvalidatedbyahumanexpert,andtheydidnotperformanycomparisonsagainstothertractography
methods.
4 RESULTS & DISCUSSION
4.1 Resultsonthe2015ISMRMTractographyChallenge
The2015ISMRMTractographychallengeistheonlydatasetthathasbeenusedtoassessperformanceofseveraldata-
driventractographymethodsandisthus,asoftoday,theonlyavailablecommongroundonwhichtocomparemethods.
Itwasusedbyfourdifferentpapersnamely,theRandom-ForestofNeheretal.[18],theGRUofPoulinetal.[24]and
BenouandRiklin-Raviv[25],andtheMLPofWegmayretal.[23]. Experimentalresultsreportedbytheauthorshave
beentranscribedinTable3,andcomparedwithoriginalsubmissionsinFigure2. Notethatthemetricsmarkedasnot
available(N/A)arethosetheauthorsdidnotreportintheiroriginalpaper.
Ascanbeseen,resultsvaryalotandthereisnocleartrendshowingwhichmethodperformsbest,especiallygiven
the nature of the evaluation metrics. As mentioned in section 2.4, methods can be evaluated using bothstreamline-
PHILIPPEPOULIN ET AL.
13
TABLE 3 Tractometerresults. TheBundlesandConnections(%)metricsarestreamline-orientedmetricswhereasthe
Avg. bundle(%)metricsarevolume-orientedmetrics
Model
Bundles
Random-Forest[18]
GRU[24]
MLP[23]
GRU(DeepTract)[25]
Valid
23
23
23
23
Invalid
94
130
57
51
Valid
52
42
72
41
Connections(%)
Invalid
N/A
46
N/A
33
Noconnection
N/A
13
N/A
23
Avg. bundle(%)
Overreach
37
35
28
17
F1-score
61
65
26
44
Overlap
59
64
16
34
(a)
(b)
(c)
FIGURE 2 2015ISMRMTractographyChallengeoriginalsubmissions(1-20)andnewresults(21-24)
orientedmetricsandvolume-orientedmetrics,whicharenotalwayscorrelated. Forexample,amethodmayhavealarge
numberofvalidconnectionsbutalowoverlap(liketheMLPofWegmayretal.)whichmeansthatalthoughthemodel
wasabletorecovermostvalidbundles,thegeneratedstreamlinesdonotproperlycoverthespatialextentofthose
bundles. Also,amethodcanbemoreconservativeandscorebestintermsofinvalidconnectionsandoverreachlikethe
GRUofBenouandRiklin-Raviv,butatthesametimehavealowratioofvalidconnectionsandapoorbundleoverlap.
Ontheotherhand,theRandom-ForestofNeheretal.doesnotscorebestinanycategory,butiscompetitiveaccording
toallmetrics(itslargeF1-scoreunderlinesthatitisamorebalancedmethodcomparedtoMLPandDeepTract). Ontop
ofthat,allmethodsweretrainedandevaluateddifferently,soanycomparisonbasedonthereportedresultsshouldbe
donewithextremecare.
14
TABLE 4 Differencesindata
Random-Forest[18]
dwidenoise+dwipreproc
Method
GRU[24]
MLP[23]
DeepTract[25]
Preprocessing
None
dwipreproc
N/A
PHILIPPEPOULIN ET AL.
Trainingsubjects
5HCPsubjects
Challengesubject
3HCPsubjects
Challengesubject
Referencestreamlines
CSD(Deterministic)
CSD(Deterministic)
iFOD(Probabilistic)
Q-Ball(Probabilistic)
WMmask
Notneeded
GroundTruth
N/A
Notneeded
4.2 The2015ISMRMTractographyChallengeasanevaluationtoolforMLalgorithms
As mentioned before, the 2015 ISMRM Tractography Challenge has been adopted as the de facto evaluation tool to
compareMLtractographymethods. However,thestrengthsandweaknessesofthattoolshouldbethoroughlyreviewed
to understand and trust any technique reporting results with it. In this section, we present what we consider to be
importantissueswiththewayinwhichthistoolhasbeenusedtoassesstheperformanceofdata-drivenmethods. In
particular,wedetailthediscrepanciesbetweenthefourML-basedmethods,differencesthatmayexplainsomeofthe
resultsinTable3andpotentiallyundermineanyconclusionthatonecoulddrawfromit. Letusmentionthatsomeof
theseissueswiththe2015ISMRMdatasetaretypicalforthefieldoftractographyasawhole.
Table4presentsasummaryofthedifferencesinhowthetoolisused. Notethatthenotavailable(N/A)markisused
foranyinformationtheauthorsdidnotmentionintheiroriginalpaper.
Dissimilarinputs
ThefourMLmethodsuseadifferentpreprocessingpipeline. Amongtheproposedalgorithms,twoappliedMRtrix's
dwidenoiseordwipreproc(www.mrtrix.org),anotheronedenoisedusing[65]andcorrectedforeddycurrentsandhead
motion,andanotheronedidnotapplyanypreprocessingatall. Moreover,someusedthediffusionsignaldirectlyas
input,whileothersresampledittoaspecificnumberofgradientdirections. Insomecases,sphericalharmonicswere
fittedtothesignalandtheSHcoefficientswerefedasinputtothemodel. Finally,thenon-recurrentmodelsarealso
givenavariablenumberofpreviousstreamlinedirectionsasinput.
The output of each of these pipelines contain various degrees of information. For example, fODF peaks are in
theoryalreadyalignedwiththemajorWMpathways,andinformationmaybelostdependingonthespecificmodelused
torecoverthepeaksfromthediffusionsignal. Ontheotherhand,usingtherawdiffusionsignalmightcontainmore
informationbutismoredifficulttounderstandandprocess,andthusadata-drivenmodelmightrequiremorecapacity
tousesuchaninput. Withoutathoroughinvestigationoftheinformationcontainedineachoutput,anyvariationsinthe
Tractometerresultscouldbeattributedtothevariationsinpreprocessing. Sincewecurrentlydonothaveanyindication
ofwhatisusefulfordata-drivenalgorithms,itisimpossibletocompareMLmethodsiftheydonotusethesameinput
data.
Varyingtestenvironment
Sincenowhitemattermaskisprovided,itmustbecomputedbyeachparticipantincaseitisneededfortracking. Outof
thefourMLmethodsthatwereevaluatedonthechallenge,twoneededWMmasks;oneusedthegroundtruthmask,
andtheotherdidnotmentionhowthemaskwascomputed. Furthermore,sincenotrackingseedsaresuppliedwiththe
dataeither,theirarrangemententirelydependsontheWMmask(andthenumberofseedspervoxel,whichisalsonot
given).
Giventhenatureofstreamlinetractography,smallvariationsofthetrackingmaskorthetrackingseedscouldhave
PHILIPPEPOULIN ET AL.
15
asubstantialimpactontheresultingstreamlinesandbythatalsoontheobtainedevaluationmetrics. Also,eventhough
computingastoppingcriterionwithinthealgorithmisaworthyimprovement,itisadifferenttaskthantracking,and
shouldbeevaluatedseparately. Consequently,allmethodsshouldbeprovidedthesametrackingmaskandseedsto
reduceasmuchaspossiblethenumberoffreevariablesduringevaluation.
Datacontamination
The use of ML methods requires special care when dealing with available data. Since machine learning models are
obtainedbyderivingimplicitrulesdirectlyfromgivendata(i.e.trainingdata),testingthetruegeneralizationcapabilities
oftheserulesmustbedoneusingadifferentandunseensetofdata(i.e. testdata).
Two methods suffer from data contamination, orleakage [66]: the GRU in [24] and the MLP in [23]. Here, data
contaminationreferstotheusageofthesamediffusiondatafortrainingandtesting. Thismeansthatthetruegeneral-
izationcapabilitiesofthetestedmethodonnew,thusunseensubjectsarestillunknown,sincethemodelhasalready
seenthespecificdiffusionpatternsthatareneededinorderto"explore"attesttime,andthereforehasbeengivenan
"unfair"advantage.
Disparatetrainingdata
Allmethodsuseddifferentreferencestreamlinesandsubjectsfortraining. Asmentionedearlier,someemployedthe
testdiffusiondatadirectly,whileothersreliedonavaryingnumberofsubjectsfromtheHCPdatabase. Twomethods
useddeterministicCSDtracking[6]togeneratereferencestreamlines,oneusedQBItracking[67](probabilistic)and
thelastoneusediFODtracking[5](alsoprobabilistic). Inordertoprovideauniformbasisforcomparison,thesame
comprehensivestreamlinetrainingsetshouldbeavailabletoeveryalgorithm.
Simulationasasubstituteforhumanacquisition
While thediffusion signal ofthe 2015 ISMRM datasetis typical ofthat of ahuman brain, itis nonetheless obtained
throughsimulation. Assuch,resultsonthatdatasetshouldnotbeseenasameasureoffutureperformanceonreal
humansubjects,atleastnotwithoutfurtherempiricalevaluation. Furthermore,atthegivenresolutionandusingthis
particularconfigurationof25bundles,falsepositivestreamlinesthatwouldotherwisebeplausiblegiventheunderlying
anatomyofarealscanmightbeimpossibletoavoid. Indeed,someauthorstriedtrainingtheirmodelsusingtheground
truthbundles,andstillproducedover50invalidbundlesinbothcases[18,25].
Smallsamplesize
The2015ISMRMTractographyChallengedatasethasonlyonesubject,whichmakesithardtoassessthefutureperfor-
manceofadata-drivenalgorithm[68]. Inordertocomputeunbiasedestimatesoffutureperformance,arichertestset
withmoresubjectsisneeded. Also,givenmoresubjects,bootstrappingmethods[69](i.e. samplingwithreplacement)
couldhelptobuildmoreaccurateestimators.
4.3 Otherresults
Some authors report local performance measures, such as the mean angular error [22]. However, local metrics do
nottakeintoaccountcompoundingerrors,whichcanhaveamajoreffectonglobalstructure. Consequently,global
evaluationmetricsshouldbepreferred.
Tractographypapersoftenreportavisualevaluationonunseen,invivosubjects,asaqualitativeevaluation. For
example,Figures3and4comparesomeoftheproposeddata-drivenapproacheswithstandardtractographymethodson
16
PHILIPPEPOULIN ET AL.
FIGURE 3 ComparisonbetweentheRFofNeheretal.(toprow),andclassicaldeterministicCSDstreamline
tractography(bottomrow). ResultsobtainedonHCPsubject992774. (Takenfrom[18]withauthorizationfromthe
authors)
FIGURE 4 Comparisonofvarioustrackingmethods: A:Deterministic,B:DeterministicBundle-Specific
(DET-BST)[50],C:ProbabilisticparticlefilterBST(PROB-PF-BST)[17],D:DeepTracker[26]. Resultsobtainedona
BIL&GINsubject. (Takenfrom[26]withauthorizationfromtheauthors)
whitematterbundleswithknownanatomy. However,inabsenceofagroundtruthortheexpertiseofaneuroanatomist,
itishardtodrawdefinitiveconclusionsonthequalityofsuchresults. Inaddition,Reisertetal.[29]presentedcorrelation
plotstoassessreproducibility,butonlyofferedqualitativecomparisonswiththereferencestreamlineswithoutany
quantitativeresults. Togaintrustinthesedata-drivenmethods,amorerigorousapproachisneeded.
Finally,mostMLmethodsofferareductionincomputationtimecomparedtotraditionalmethods. Thisisanon-
negligiblebenefit,shouldthesemethodsbeadoptedinpractice.
4.4 Proposedguidelinesforadata-driventractographyevaluationframework
ConsideringtheMLtractographyevaluationissuespreviouslyunderlined,wediscussinthissectionthefundamental
elements of a better framework we believe the community should adopt in the upcoming years. We start with the
essentialcharacteristicssuchaframeworkshouldhave,followedwithusefulfeatures.
17
PHILIPPEPOULIN ET AL.
Essentialcharacteristics
Firstandforemost,anidealdata-driventractographyevaluationframeworkshouldcomewithapublicandfree-to-use
datasetthatanyonecouldeasilyrelyon. Thedatasetshouldincludeimagesofrealhumanacquisitionsalongwitha
carefulexpertselectionofgroundtruthstreamlines. Itisimportanttoavoidanybiastowardsaspecifictractography
algorithm. Inordertoachievethis,thestreamlinescouldbefirstgeneratedbyalargenumberofdifferent(andideally
orthogonal) deterministic, probabilistic and global algorithms and then segmented by expert annotators according
tostrictanatomicaldefinitionsforagivennumberofbundles. Whilesuchmanualannotationwouldbetedious,time
consumingandevenerrorprone,weconsiderthisanindispensablesteptowardsbuildingarealisticandusefuldataset
forML-baseddevelopment. Theneedforsuchagoldstandardthatquantifieshumanvariabilityiswell-knowninother
fields,suchasautomaticimagesegmentation,cellcountingorinmachinelearning[70,71,72,73]. Despitethefactthat
simulatedbrainimagescomewithapixel-accuratesetofgroundtruthstreamlinesthatcanbegeneratedinamatterof
seconds,bydefinitionsyntheticdiffusionsignalsareover-simplisticpicturesofrealdataand,assuch,cannotprovide
anyguaranteeofsubsequentperformancefordata-drivenmethodsonrealdata.
AlthoughthereisnoconsensusregardingthemostdesirablefeaturesaMLtractographyalgorithmshouldhaveand
howitshouldbeevaluated,byitsverynature,anyMLevaluationframeworkshouldaimatmeasuringhowanalgorithm
canfaithfullyreproduceataskitwastrainedfor. Assuch,areasonabledatasetshouldincludeasufficientlylargenumber
of well-separated training and testing images. Thus, statistics resulting from such a dataset would not suffer from
contaminationandthereportedmetricswouldbereliableandunbiasedestimatesofthetruegeneralizationpowerof
aMLalgorithm. Inaddition,toensurethattheobserveddifferencesbetweenmultiplealgorithmsareresultingfrom
theintrinsicpropertiesofthemodelandnotcausedbysomefeatureoftheevaluationframework,thenumberoffree
variablesshouldbereducedtoaminimum. Consequently,thetrackingmasksandseedsshouldbeprovidedtogether
withclearlypreprocesseddiffusiondata,sothattheproposedmethodscanbeevaluatedinequalconditions. There
shouldbemultiple"classes"ofinputdata,dependingonwhetheranalgorithmsupportsDWIsamples,SHcoefficientsor
fODFpeaks. Furthermore,theinitialdiffusionsignalshouldhavethesamestatisticalpropertiesforthetrainingandthe
testingset. Finally,theacquiredimagesshouldideallybeacquiredatdifferentMRIscannerswithdifferentacquisition
protocolsinordertoavoidoverfittingissues.
Evaluationmetricsshouldalsobeboundtothepurposeoftractographyalgorithms. Consideringthattractography
ismostlyusedforbundlereconstruction,tractometrystudiesandconnectivityanalyses,anidealevaluationframework
shouldincludetwosetsofmetrics: 1)metricsmeasuringhowaMLmethodcanfaithfullyreproduceasetofpredefined
bundlesitwastrainedtorecover(tractometry),and2)metricsmeasuringhowitcanconnectmatchingregionsofthe
brain,i.e. producevalidconnections(connectivity). Furthermore,sincemanyapplicationsusetractographyalgorithms
toproducealargenumberofstreamlines(withmanyfalsepositives),whicharethenfilteredoutbyapost-processing
algorithmsuchasRecoBundles[54],theframeworkshouldreportresultsbeforeandafterpost-processing. Thiswould
underlinethetruerecallpowerofadata-drivenalgorithm,whichisafundamentalcharacteristicoftract-basedand
connectivity-basedapplications[15].
Lastly,thesizeofanidealdatasetisofprimaryimportance. Whileasmall-sizeddatasetcouldbepronetooverfitting,
itwouldbecostlytocreateaverylargedatasetandalsodifficulttoensureacoherentmanualannotation. Onerule
ofthumbthatcanbeusedtoidentifythe"correct"sizeofadatasetisthroughtheinspectionofthelearningcurveof
severalMLmodels[74]. Thesecurvesshowthemodelperformanceasafunctionofthetrainingsamplesize. Typically,
theperformanceofseveralmodelssaturatesforasufficientdatasetsize. Althoughimperfect,thisprocedureisagood
heuristicforestimatingthesizeofthedataset.
PHILIPPEPOULIN ET AL.
18
Otherusefulfeatures
Despiteanythoroughmanualannotationprotocol,manuallyannotatedbundlescanbesubjecttonon-negligibleinter-
raterandintra-ratervariability. Assuch,ausefulcharacteristicofaMLtractographydatasetwouldbeameasureof
those variations. This would be obtained by having several experts annotating the dataset, and at least one expert
annotatingittwiceormoretimes. Suchmeasureswouldprovideaminimalboundbeyondwhichadata-drivenalgorithm
couldbeconsidered"asgoodasanexpert". Anotherveryusefultoolwouldbeanopenlyaccessibleonlineevaluation
system. Given such a system, people could upload their test results in order to compare them with the test ground
truth. Inthatway,anautomaticrankingproceduresimilartothatofKagglecouldbeusedtosortvariousMLalgorithms
basedontheirachievedscores. Whilenorankingmethodisperfect,itwouldnonethelessprovideacommonevaluation
frameworkthatpeoplecouldrelyon.
AnidealdatasetwouldalsocoverthewholefieldofdiffusionMRIacquisitionprotocols,fromHCP-likeresearch
acquisitionstoclinicalacquisitions. Itwouldincludesingleb-valueaswellasmultipleb-valuesdata,alongwithmore
sophisticatedacquisitionprotocolssuchasb-tensorencoding. Itwouldalsoneedlowresolutionimagestogetherwith
high-resolutionimages. Sincedataharmonizationisalsoaproblemfordata-drivenalgorithms,acquisitionfromseveral
sitesareneededfortest-reteststudies. Annotatedpathologicalcaseswouldcompletethedatasetbyallowingcareful
preliminarystudiesonhowML-basedmethodscanbereliedoninunhealthypatients.
Finally,sincetractographyisusedmoreandmoreinpre-clinicalapplications,asubsetofmanuallyannotatedrodent
ormacaquebrainswouldbeofgreatinteresttotrainandtestfutureMLalgorithms(likethe2018VOTEMChallenge[36],
forexample).
This is, of course, the ultimate wish list. But, in the era of open data and open science, it needs to be done by
the community, for the community. We can already see this work in progress with more and more accessible and
reproducibledatabeingpublishedeveryyear.
5 CONCLUSION
Inthispaper,weprovidedanexhaustivereviewofthecurrentstateoftheartofmachinelearningmethodsinthefield
oftractography. Wedescribedtheexistingdatasetsthatcomprisebothdiffusiondataandreferencestreamlines,which
couldgenerallybeusefulfornewtrackingmethodsbasedonML.Inparticular, wethoroughlyexaminedthewidely
usedevaluationtoolfordata-driventrackingmethods,the2015ISMRMTractographyChallenge,anddetailedflawsand
shortcomingwhenusedtoassessdata-drivenalgorithms. Basedonourfindings,wesuggestedgoodpracticesthatwe
believewouldfosterthedevelopmentofanewevaluationframeworkforML-basedtractographymethodswiththe
potentialtoeffectivelyadvancethisfieldofresearch.
Thereisnodoubtthatmachinelearningtractographywillhaveanimportantroletoplayinthefuturetosolvesome
oftheopenproblemsoftractography. Atthemoment,however,allexistingmethodsshowtheoreticalpotentialand
inlimitedtestcases. Methodshaveyettomakesoliddemonstrationsoftheirperformanceandefficiencyinpractice.
ThereisstillnoML-basedtractographytoolthatisascalableandusableonanygivendiffusionMRIdataset. Thisistrue
forhealthydatasetsbutevenmoresoforpathologicalbrains. Hence,itisfairtosaythatML-basedtractographyisstill
atitsinfancyandisnotreadyfor"prime-time",butisnonethelessaveryfertilefieldofresearchtomakemeaningful
contributionstothefieldofconnectivitymapping.
PHILIPPEPOULIN ET AL.
REFERENCES
[1] Jeurissen B, Descoteaux M, Mori S, Leemans A. Diffusion MRI fiber tractography of the brain. NMR in Biomedicine
19
2017;p.e3785.
[2] Yeh FC, Verstynen TD, Wang Y, Fernández-Miranda JC, Tseng WYI. Deterministic diffusion fiber tracking improved by
quantitativeanisotropy. PloSone2013;8(11):e80713.
[3] BasserPJ,PajevicS,PierpaoliC,DudaJ,AldroubiA. InvivofibertractographyusingDT-MRIdata. Magneticresonance
inmedicine2000;44(4):625 -- 632.
[4] Behrens TE, Berg HJ, Jbabdi S, Rushworth MF, Woolrich MW. Probabilistic diffusion tractography with multiple fibre
orientations: Whatcanwegain? Neuroimage2007;34(1):144 -- 155.
[5] Tournier JD, Calamante F, Connelly A. Improved probabilistic streamlines tractography by 2nd order integration over
fibreorientationdistributions. In: Proceedingsoftheinternationalsocietyformagneticresonanceinmedicine, vol.18;
2010.p.1670.
[6] TournierJD,CalamanteF,ConnellyA. MRtrix: diffusiontractographyincrossingfiberregions. InternationalJournalof
ImagingSystemsandTechnology2012;22(1):53 -- 66.
[7] Reisert M, Mader I, Anastasopoulos C, Weigel M, Schnell S, Kiselev V. Global fiber reconstruction becomes practical.
Neuroimage2011;54(2):955 -- 962.
[8] ManginJF,FillardP,CointepasY,LeBihanD,FrouinV,PouponC.Towardglobaltractography.Neuroimage2013;80:290 --
296.
[9] Jbabdi S, Woolrich MW, Andersson JL, Behrens T. A Bayesian framework for global tractography. Neuroimage
2007;37(1):116 -- 129.
[10] Pierpaoli C, Jezzard P, Basser PJ, Barnett A, Di Chiro G. Diffusion tensor MR imaging of the human brain. Radiology
1996;201(3):637 -- 648.
[11] Caan MW, Khedoe HG, Poot DH, Arjan J, Olabarriaga SD, Grimbergen KA, et al. Estimation of diffusion properties in
crossingfiberbundles. IEEEtransactionsonmedicalimaging2010;29(8):1504 -- 1515.
[12] Tournier JD, Calamante F, Gadian DG, Connelly A. Direct estimation of the fiber orientation density function from
diffusion-weightedMRIdatausingsphericaldeconvolution. NeuroImage2004;23(3):1176 -- 1185.
[13] Descoteaux M, Deriche R, Knösche TR, Anwander A. Deterministic and probabilistic tractography based on complex
fibreorientationdistributions. IEEEtransactionsonmedicalimaging2009feb;28(2):269 -- 86.
[14] Schilling KG, Daducci A, Maier-Hein K, Poupon C, Houde JC, Nath V, et al. Challenges in diffusion MRI tractography --
Lessonslearnedfrominternationalbenchmarkcompetitions. Magneticresonanceimaging2018;.
[15] Maier-Hein KH, Neher PF, Houde JC, Côté MA, Garyfallidis E, Zhong J, et al. The challenge of mapping the human con-
nectomebasedondiffusiontractography. Naturecommunications2017;8(1):1349.
[16] Côté MA, Girard G, Boré A, Garyfallidis E, Houde JC, Descoteaux M. Tractometer: towards validation of tractography
pipelines. Medicalimageanalysis2013;17(7):844 -- 857.
[17] GirardG,WhittingstallK,DericheR,DescoteauxM. Towardsquantitativeconnectivityanalysis: reducingtractography
biases. Neuroimage2014;98:266 -- 278.
[18] NeherPF,CôtéMA,HoudeJC,DescoteauxM,Maier-HeinKH. Fibertractographyusingmachinelearning. NeuroImage
2017sep;158:417 -- 429.
20
PHILIPPEPOULIN ET AL.
[19] DuruDG,OzkanM. SOMBasedDiffusionTensorMRAnalysis. In: ImageandSignalProcessingandAnalysis,2007.ISPA
2007.5thInternationalSymposiumonIEEE;2007.p.403 -- 406.
[20] DuruDG,OzkanM. Self-organizingmapsforbraintractographyinMRI. In: NeuralEngineering(NER),20136thInterna-
tionalIEEE/EMBSConferenceonIEEE;2013.p.1509 -- 1512.
[21] Neher PF, Götz M, Norajitra T, Weber C, Maier-Hein KH. A Machine Learning Based Approach to Fiber Tractography
UsingClassifierVoting. Springer,Cham;2015.p.45 -- 52.
[22] Jörgens D, Smedby Ö, Moreno R. Learning a Single Step of Streamline Tractography Based on Neural Networks.
ComputationalDiffusionMRISpringer;2018.p.103 -- 116.
In:
[23] WegmayrV,GiuliariG,HoldenerS,BuhmannJ. Data-drivenfibertractographywithneuralnetworks. In: 2018IEEE15th
InternationalSymposiumonBiomedicalImaging(ISBI2018)IEEE;2018.p.1030 -- 1033.
[24] Poulin P, Côté MA, Houde JC, Petit L, Neher PF, Maier-Hein KH, et al. Learn to Track: Deep Learning for Tractography.
Springer,Cham;2017.p.540 -- 547.
[25] BenouI,Riklin-RavivT.DeepTract:AProbabilisticDeepLearningFrameworkforWhiteMatterFiberTractography.arXiv
preprintarXiv:1812051292018;.
[26] PoulinP,RheaultF,St-OngeE,JodoinPM,DescoteauxM. Bundle-WiseDeepTracker: Learningtotrackbundle-specific
streamline paths. In: Proceedings of the International Society for Magnetic Resonance in Medicine ISMRM-ESMRMB;
2018..
[27] Wasserthal J, Neher PF, Maier-Hein KH. Tract orientation mapping for bundle-specific tractography. In: International
ConferenceonMedicalImageComputingandComputer-AssistedInterventionSpringer;2018.p.36 -- 44.
[28] Lucena OASd, Deep Learning for Brain Analysis in MR Imaging. São Paulo, Brazil: [sn]; 2018. http://repositorio.
unicamp.br/jspui/handle/REPOSIP/332646.
[29] Reisert M, Coenen VA, Kaller C, Egger K, Skibbe H. HAMLET: Hierarchical Harmonic Filters for Learning Tracts from
DiffusionMRI. arXivpreprintarXiv:1807010682018;.
[30] Wasserthal J, Neher P, Maier-Hein KH. TractSeg-Fast and accurate white matter tract segmentation. NeuroImage
2018;183:239 -- 253.
[31] KoppersS,MerhofD. DirectEstimationofFiberOrientationsUsingDeepLearninginDiffusionImaging. Springer,Cham;
2016.p.53 -- 60.
[32] NgattaiLamPD,BelhommeG,FerrallJ,PattersonB,StynerM,PrietoJC. TRAFIC:FiberTractClassificationUsingDeep
Learning. ProceedingsofSPIE -- theInternationalSocietyforOpticalEngineering2018feb;10574.
[33] PatilSM,NigamA,BhavsarA,ChattopadhyayC. SiameseLSTMbasedFiberStructuralSimilarityNetwork(FS2Net)for
RotationInvariantBrainTractographySegmentation. undefined2017;.
[34] GuptaV,ThomopoulosSI,RashidFM,ThompsonPM. FiberNET:AnEnsembleDeepLearningFrameworkforClustering
WhiteMatterFibers. Springer,Cham;2017.p.548 -- 555.
[35] GuptaV,ThomopoulosSI,CorbinCK,RashidF,ThompsonPM. FIBERNET2.0: Anautomaticneuralnetworkbasedtool
forclusteringwhitematterfibersinthebrain. In: 2018IEEE15thInternationalSymposiumonBiomedicalImaging(ISBI
2018)IEEE;2018.p.708 -- 711.
[36] Thomas C, Frank QY, Irfanoglu MO, Modi P, Saleem KS, Leopold DA, et al. Anatomical accuracy of brain connec-
tions derived from diffusion MRI tractography is inherently limited. Proceedings of the National Academy of Sciences
2014;111(46):16574 -- 16579.
PHILIPPEPOULIN ET AL.
21
[37] PujolS,WellsW,PierpaoliC,BrunC,GeeJ,ChengG,etal. TheDTIchallenge: towardstandardizedevaluationofdiffu-
siontensorimagingtractographyforneurosurgery. JournalofNeuroimaging2015;25(6):875 -- 882.
[38] Yeh FC, Panesar S, Fernandes D, Meola A, Yoshino M, Fernandez-Miranda JC, et al. Population-averaged atlas of the
macroscalehumanstructuralconnectomeanditsnetworktopology. NeuroImage2018;178:57 -- 68.
[39] ZhangF,WuY,NortonI,RigoloL,RathiY,MakrisN,etal. Ananatomicallycuratedfiberclusteringwhitematteratlasfor
consistentwhitemattertractparcellationacrossthelifespan. NeuroImage2018;.
[40] Thon A, Teichgräber U, Tennstedt-Schenk C, Hadjidemetriou S, Winzler S, Malich A, et al. Computer aided detection in
prostatecancerdiagnostics: Apromisingalternativetobiopsy? Aretrospectivestudyfrom104lesionswithhistological
groundtruth. PLOSONE201710;12(10):1 -- 21.
[41] Clinic C, Alzheimer's Disease: Overview of Diagnostic Tests; 2014.
clevelandclinic.org/health/diagnostics/9176-alzheimers-disease-overview-of-diagnostic-tests/.
[Online; accessed 3-January-2019]. https://my.
[42] BernardO,BoschJG,HeydeB,AlessandriniM,BarbosaD,Camarasu-PopS,etal.Standardizedevaluationsystemforleft
ventricular segmentation algorithms in 3D echocardiography. IEEE transactions on medical imaging 2016;35(4):967 --
977.
[43] MenzeBH,JakabA,BauerS,Kalpathy-CramerJ,FarahaniK,KirbyJ,etal. Themultimodalbraintumorimagesegmenta-
tionbenchmark(BRATS). IEEEtransactionsonmedicalimaging2015;34(10):1993.
[44] Fillard P, Descoteaux M, Goh A, Gouttard S, Jeurissen B, Malcolm J, et al. Quantitative evaluation of 10 tractography
algorithmsonarealisticdiffusionMRphantom. Neuroimage2011;56(1):220 -- 234.
[45] WilkinsB,LeeN,SinghM. DevelopmentandevaluationofasimulatedFiberCupphantom. In: InternationalSymposium
onMagneticResonanceinMedicine(ISMRM'12);2012.p.1938.
[46] Yendiki A, Panneck P, Srinivasan P, Stevens A, Zöllei L, Augustinack J, et al. Automated probabilistic reconstruction of
white-matter pathways in health and disease using an atlas of the underlying anatomy. Frontiers in neuroinformatics
2011;5:23.
[47] Daducci A, Canales-Rodríguez EJ, Descoteaux M, Garyfallidis E, Gur Y, Lin YC, et al. Quantitative comparison of
IEEE transactions on medical imaging
reconstruction methods for intra-voxel fiber recovery from diffusion MRI.
2014;33(EPFL-ARTICLE-183667):384 -- 399.
[48] DaducciA, Caruyer E, Descoteaux M,Houde J, Thiran J. HARDI reconstructionchallenge 2013. In: Proceedingsof the
IEEEInternationalSymposiumonBiomedicalImaging(ISBI),SanFrancisco,CA;2013..
[49] ChenotQ,Tzourio-MazoyerN,RheaultF,DescoteauxM,CrivelloF,ZagoL,etal. Apopulation-basedatlasofthehuman
pyramidaltractin410healthyparticipants. BrainStructureandFunction2018;p.1 -- 14.
[50] Rheault F, St-Onge E, Sidhu J, Maier-Hein K, Tzourio-Mazoyer N, Petit L, et al. Bundle-specific tractography with incor-
poratedanatomicalandorientationalpriors. NeuroImage2018;.
[51] Poupon C, Laribiere L, Tournier G, Bernard J, Fournier D, Fillard P, et al. A diffusion hardware phantom looking like a
coronal brain slice. In: Proceedings of the International Society for Magnetic Resonance in Medicine, vol. 18; 2010. p.
581.
[52] Neher PF, Laun FB, Stieltjes B, Maier-Hein KH. Fiberfox: facilitating the creation of realistic white matter software
phantoms. Magneticresonanceinmedicine2014;72(5):1460 -- 1470.
[53] White T, Magnotta VA, Bockholt HJ, Williams S, Wallace S, Ehrlich S, et al. Global white matter abnormalities in
schizophrenia: amultisitediffusiontensorimagingstudy. Schizophreniabulletin2009;37(1):222 -- 232.
22
PHILIPPEPOULIN ET AL.
[54] GaryfallidisE,CôtéMA,RheaultF,SidhuJ,HauJ,PetitL,etal. Recognitionofwhitematterbundlesusinglocalandglobal
streamline-basedregistrationandclustering. NeuroImage2018;170:283 -- 295.
[55] MazoyerB,MelletE,PercheyG,ZagoL,CrivelloF,JobardG,etal. BIL&GIN:aneuroimaging,cognitive,behavioral,and
geneticdatabaseforthestudyofhumanbrainlateralization. Neuroimage2016;124:1225 -- 1231.
[56] VanEssenDC,SmithSM,BarchDM,BehrensTE,YacoubE,UgurbilK,etal. TheWU-Minnhumanconnectomeproject:
anoverview. Neuroimage2013;80:62 -- 79.
[57] Glasser MF, Sotiropoulos SN, Wilson JA, Coalson TS, Fischl B, Andersson JL, et al. The minimal preprocessing pipelines
fortheHumanConnectomeProject. Neuroimage2013;80:105 -- 124.
[58] Wassermann D, Makris N, Rathi Y, Shenton M, Kikinis R, Kubicki M, et al. The white matter query language: a novel
approachfordescribinghumanwhitematteranatomy. BrainStructureandFunction2016;221(9):4705 -- 4721.
[59] Stieltjes B, Brunner RM, Fritzsche K, Laun F. Diffusion tensor imaging: introduction and atlas. Springer Science & Busi-
nessMedia;2013.
[60] Garyfallidis E, Brett M, Correia MM, Williams GB, Nimmo-Smith I. Quickbundles, a method for tractography simplifica-
tion. Frontiersinneuroscience2012;6:175.
[61] TahaAA,HanburyA. Metricsforevaluating3Dmedicalimagesegmentation: analysis,selection,andtool. BMCmedical
imaging2015;15(1):29.
[62] Reddy CP, Rathi Y. Joint multi-fiber NODDI parameter estimation and tractography using the unscented information
filter. Frontiersinneuroscience2016;10:166.
[63] MITK, MITK Diffusion Imaging; 2018.
DiffusionImaging.
[Online; accessed 3-January-2019].
http://www.mitk.org/wiki/
[64] Ronneberger O, Fischer P, Brox T. U-Net: Convolutional Networks for Biomedical Image Segmentation.
ImageComputingandComputer-AssistedIntervention(MICCAI),vol.9351ofLNCS;2015.p.234 -- 241.
In: Medical
[65] ManjónJV,CoupéP,ConchaL,BuadesA,CollinsDL,RoblesM. Diffusionweightedimagedenoisingusingovercomplete
localPCA. PloSone2013;8(9):e73021.
[66] Kaufman S, Rosset S, Perlich C, Stitelman O. Leakage in data mining: Formulation, detection, and avoidance. ACM
TransactionsonKnowledgeDiscoveryfromData(TKDD)2012;6(4):15.
[67] AganjI,LengletC,SapiroG. ODFreconstructioninq-ballimagingwithsolidangleconsideration. In: BiomedicalImaging:
FromNanotoMacro,2009.ISBI'09.IEEEInternationalSymposiumonIEEE;2009.p.1398 -- 1401.
[68] RaudysSJ,JainAK. Smallsamplesizeeffectsinstatisticalpatternrecognition: Recommendationsforpractitioners. IEEE
TransactionsonPatternAnalysis&MachineIntelligence1991;(3):252 -- 264.
[69] EfronB,TibshiraniRJ. Anintroductiontothebootstrap. CRCpress;1994.
[70] Kleesiek J, Petersen J, Döring M, Maier-Hein K, Köthe U, Wick W, et al. Virtual raters for reproducible and objective
assessmentsinradiology. Scientificreports2016;6:25007.
[71] Entis JJ, Doerga P, Barrett LF, Dickerson BC. A reliable protocol for the manual segmentation of the human amygdala
anditssubregionsusingultra-highresolutionMRI. Neuroimage2012;60(2):1226 -- 1235.
[72] BoccardiM,BocchettaM,ApostolovaLG,BarnesJ,BartzokisG,CorbettaG,etal. DelphidefinitionoftheEADC-ADNI
HarmonizedProtocolforhippocampalsegmentationonmagneticresonance. Alzheimer's&Dementia2015;11(2):126 --
138.
PHILIPPEPOULIN ET AL.
23
[73] PiccininiF,TeseiA,PaganelliG,ZoliW,BevilacquaA. Improvingreliabilityoflive/deadcellcountingthroughautomated
imagemosaicing. Computermethodsandprogramsinbiomedicine2014;117(3):448 -- 463.
[74] BeleitesC,NeugebauerU,BocklitzT,KrafftC,PoppJ. Samplesizeplanningforclassificationmodels. AnalyticaChimica
Acta2013;760:25 -- 33.
|
1604.01943 | 1 | 1604 | 2016-04-07T10:02:02 | Brief wide-field photostimuli evoke and modulate oscillatory reverberating activity in cortical networks | [
"q-bio.NC"
] | Cell assemblies manipulation by optogenetics is pivotal to advance neuroscience and neuroengineering. In in vivo applications, photostimulation often broadly addresses a population of cells simultaneously, leading to feed-forward and to reverberating responses in recurrent microcircuits. The former arise from direct activation of targets downstream, and are straightforward to interpret. The latter are consequence of feedback connectivity and may reflect a variety of time-scales and complex dynamical properties. We investigated wide-field photostimulation in cortical networks in vitro, employing substrate-integrated microelectrode arrays and long-term cultured neuronal networks. We characterized the effect of brief light pulses, while restricting the expression of channelrhodopsin to principal neurons. We evoked robust reverberating responses, oscillating in the physiological gamma frequency range, and found that such a frequency could be reliably manipulated varying the light pulse duration, not its intensity. By pharmacology, mathematical modelling, and intracellular recordings, we conclude that gamma oscillations likely emerge as in vivo from the excitatory-inhibitory interplay and that, unexpectedly, the light stimuli transiently facilitate excitatory synaptic transmission. Of relevance for in vitro models of (dys)functional cortical microcircuitry and in vivo manipulations of cell assemblies, we give for the first time evidence of network-level consequences of the alteration of synaptic physiology by optogenetics | q-bio.NC | q-bio | Brief wide-field photostimuli evoke and modulate oscillatory
reverberating activity in cortical networks
R Pulizzi1, G Musumeci1, C Van Den Haute2-3, S Van De Vijver1, V Baekelandt2, M Giugliano1, 4-5
1 Theoretical Neurobiology & Neuroengineering, University of Antwerp, Antwerp, Belgium;
2 Laboratory of Neurobiology and Gene Therapy, Katholieke Universiteit Leuven, Leuven, Belgium;
3 Leuven Viral Vector Core, Katholieke Universiteit Leuven, Leuven, Belgium
4 Department of Computer Science, University of Sheffield, S1 4DP Sheffield, UK
5 Laboratory of Neural Microcircuitry, Brain Mind Institute, EPFL, CH-1015 Lausanne, Switzerland
Author's contribution: conceived and designed the experiments: RP, GM, and MG. Performed the experiments: RP
simulations: MG. Contributed
and GM. Analysed
reagents/materials/analysis tools: CVDH, VB, and MG. Wrote the paper: RP and MG. The authors declare no
competing financial interests.
the data: SVDV and RP. Performed computer
We are grateful to Mr. D. Van Dyck and M. Wijnants for excellent technical assistance, to Drs. J. Couto, D. Linaro,
R. Schneggenburger, U. Egert, and S. Marom, for helpful discussions, to Drs. M. Mattia, E. Vasilaki, M. Deger, and two
anonymous Reviewers for comments on an earlier version of the manuscript, and to Dr. P. Hegemann for the gift of the
CHR2 L132C/T159C-mCherry. Financial support from the 7th Framework Programme of the European Commission
(FP7-PEOPLE-ITN "NAMASEN" grant n. 264872, FP7-PEOPLE-IAPP "NEUROACT" grant n. 286403, FP7-ICT-
FET project "ENLIGHTENMENT", grant n. 284801), the Interuniversity Attraction Poles Program (IUAP) of the
Belgian Science Policy Office, and the University of Antwerp is kindly acknowledged.
Correspondence should be addressed to Dr. Michele Giugliano, Theoretical Neurobiology and Neuroengineering,
University of Antwerp, Campus Drie Eiken, Universiteitsplein 1, 2610 Wilrijk, Belgium. E-mail:
[email protected]
Authors' preprint
Photo-activated network responses in vitro
Abstract
Cell assemblies manipulation by optogenetics is pivotal to advance neuroscience and neuroengineering.
In in vivo applications, photostimulation often broadly addresses a population of cells simultaneously,
leading to feed-forward and to reverberating responses in recurrent microcircuits. The former arise from
direct activation of targets downstream, and are straightforward to interpret. The latter are consequence of
feedback connectivity and may reflect a variety of time-scales and complex dynamical properties.
We investigated wide-field photostimulation in cortical networks in vitro, employing substrate-integrated
microelectrode arrays and long-term cultured neuronal networks. We characterized the effect of brief light
pulses, while restricting the expression of channelrhodopsin to principal neurons. We evoked robust
reverberating responses, oscillating in the physiological gamma frequency range, and found that such a
frequency could be reliably manipulated varying the light pulse duration, not its intensity. By pharmacology,
mathematical modelling, and intracellular recordings, we conclude that gamma oscillations likely emerge as
in vivo from the excitatory-inhibitory interplay and that, unexpectedly, the light stimuli transiently facilitate
excitatory synaptic transmission. Of relevance for in vitro models of (dys)functional cortical microcircuitry
and in vivo manipulations of cell assemblies, we give for the first time evidence of network-level
consequences of the alteration of synaptic physiology by optogenetics.
Keywords
Optogenetics; substrate-integrated microelectrode arrays; in vitro cortical networks; gamma rhythms;
oscillations; reverberating activity;
Introduction
Since first expressing channelrhodopsin in neurons 1, optogenetics 2 raised immense interest as one of the
most promising techniques in neuroscience 3-8. Optogenetics enables the manipulation of the activity of
genetically-identified neurons with unprecedented temporal resolution 9-14. Therefore, both in fundamental
research and neuroprosthetics, the idea of regulating neuronal firing by light (e.g. in a closed loop) quickly
emerged as a pivotal causative approach to advance our correlative understanding of brain (dys)functions.
Nevertheless, the limits and most useful features of the stimulation (e.g. intensity, time course, frequency,
temporal pattern, etc.), to be exploited in a control system, have not been exhaustively explored, but see 15,16.
Despite the low spatial resolution in most in vivo applications, but see 17,18, as light broadly addresses the
feed-forward projections expressing an opsin, the target neuronal populations can be efficiently controlled
5,19,20. In this case, optogenetic control of downstream neuronal firing is straightforward and irreplaceably
elegant and simple. However, when recurrent collaterals expressing an opsin dominate, e.g. as in the cortex,
it is expected that light stimuli elicit additional reverberating responses, due to local recurrent pathways.
These responses arise as second-order effects from the recruitment or suppression of neuronal activity and
may reflect unanticipated time-scales, and complex dynamical microcircuit properties. These add complexity
to the interpretation of optogenetic stimulation and to its use for closed loop control, particularly in vivo
where they might be hard to isolate and study in depth. A complementary approach for dissecting
reverberating activity may come from using of reduced in vitro experimental preparations 21-23, offering
substantial advantages over in silico simulations. In fact, while addressing the impact of recurrent
connectivity greatly benefits from detailed and simplified neuronal models 24,25, introducing opsins's
biophysics into large-scale network simulations has been infrequent so far, but see 26,27,28. In addition, most of
Hodgkin-Huxley-like models fail to capture the extended dynamical properties of neurons and synapses 29,30.
Here we explore in vitro the collective response evoked by light in recurrent microcircuits, employing
mature cultured neurons, dissociated from the rat neocortex, and let developing ex vivo on microelectrode
arrays (MEAs) for several weeks. As neurons reorganized into functional recurrent networks 31,32, MEAs
allowed non-invasive and chronic access to their coordinated activity 33-36. We selectively targeted the
expression of a variant of channelrhodopsin in principal neurons and delivered brief wide-field light stimuli,
while recording spiking activity from the MEA microelectrodes.
We observed that evoked network responses outlasted the stimuli, as for focal extracellular electrical
stimulation 37,38, and also contained transient oscillations in the mean firing rate with frequency in the gamma
range 75. Unexpectedly, the duration of the light pulses reproducibly altered the temporal course of the
reverberating response and modulated the frequency of its oscillations from ~50 to ~150 cycles/s. A simple
mathematical model suggested a possible synaptic mechanism underlying our observations, which was later
confirmed by performing intracellular recordings and monitoring synaptic release events immediately before
2/23
Authors' preprint
Photo-activated network responses in vitro
and after the light stimuli. All in all, these results enhance our understanding on how reverberating activity
can be manipulated, offers an experimental model to study gamma-range rhythms in a dish, and provide for
the first time, evidence of unexpected network dynamical consequences of the alteration of synaptic
physiology by optogenetics.
Results
We cultured large-scale networks of rat primary cortical neurons, over 57 arrays of microelectrodes
(MEAs) integrated in transparent glass substrates. Over four weeks in vitro, neurons developed functional
synaptic connections and reached a mature pattern of spontaneous activity 34. By detecting action potentials
recorded at each microelectrode, we investigated the collective neuronal responses upon brief wide-field
light stimulation (Fig. 1A). In 39 MEAs, we transduced cells by AAV to express a variant of
channelrhodopsin (CHR2 LC-TC) fused to a red-fluorescent protein (mCherry), targeting selectively
CaMKIIα-positive cells (i.e. putative glutamatergic neurons) 53 (Fig. 1D). This mutant opsin responds to blue
light with stronger currents, reduced proton permeation, and enhanced calcium (Ca2+) selectivity, compared
to its wild type 40. Immunocytochemistry confirmed our desired high level of opsin expression, transduction
efficiency (i.e. ~80% of all neurons), and low apoptosis (not shown). Routine MEA recordings showed no
differences with control cultures in terms of network development and cell survival, inferred by analyzing
spontaneous bursting activity (Fig. 1B-C) 22,43 and by quantifying the number of active microelectrodes (i.e.
those detecting >1 spike in 50s; Fig. 1C).
Figure 1: Opsin expression in cultured neurons does not alter the spontaneous network electrical
activity. Transparent glass microelectrode arrays (MEAs), used as substrates for long-term primary cortical
cultures, were employed to record light-evoked (A) and spontaneous (B) neuronal electrical activity, 28 days
after plating. Brief light pulses were delivered wide field by a computer-controlled LED, focused on the
entire inner area of the MEA (A), and found to elicit neuronal spiking activity analyzed from extracellular
raw voltage signals. Transduction of ChR2 and mCherry by AAV vectors, with a CaMKIIα promotor, altered
neither the spontaneous network bursting (B), nor induced significant changes in neuronal firing rate, burst
rate and duration, or in the number of active microelectrodes (C, quantified over 11 control and 46
transduced MEAs). NeuN immunostaining (green) and fluorescent mCherry (red) imaging highlighted (D)
the ChR2 expression in putative excitatory neurons, with spatially unrestricted presence over the cell surface.
Authors' preprint
3/23
Photo-activated network responses in vitro
Light-evoked network responses. We employed a computer-controlled blue LED to deliver brief wide-
field pulses of light over the inner bottom area of each MEA. Simultaneously we recorded the neuronal spike
responses, detected extracellularly at each of the 59 microelectrodes of the MEA (Fig. 1A, inset). The
duration of light pulses was systematically varied in the set {0.1, 0.5, 1, 2, 5, 10, 20} ms, while the light
intensity was kept constant to its full power value (i.e. 2 mW/mm2 at the sample). This allowed us to vary
reliably the amplitude of light-activated CHR2 currents, as validated by patch-clamp experiments (see
Supplemental Figure S1). Each pulse was repeated 60 times at a very low rate (0.2Hz), thus reducing the
likelihood of short- and long-term plastic changes in neuronal responses 55.
Regardless of the pulse duration, the evoked network response outlasted the stimulus and exhibited an
early and a late component, as apparent in the shape of the peri-stimulus histogram (PSTH) of the spike
times (Fig. 2A). Control cultures displayed as expected no evoked response (not shown). Similar to
electrically evoked responses 37, the two components were related to the direct neuronal activation and the
reverberating interactions mediated by recurrent connectivity. When the intensity and not the duration of the
light pulse were changed, considerably less stable and reproducible late responses were observed (not
shown), correlating with weaker early responses.
The direct component, occurring in the first few milliseconds after the stimulus onset, originated from
the intense sudden depolarization caused by the CHR2 activation. It had low temporal jitter across
microelectrodes and repetitions, and was followed by a transient decrease of activity. The reverberating
response arose several tens of milliseconds after (i.e. 30-50ms), as a renewed progressive build up and then
exhaustion of neuronal firing. It lasted up to several hundreds of milliseconds and reflected in part the slow
deactivation kinetics of CHR2 LC-TC, and in part the effect of recurrent network interactions. As a
consequence of the high reliability of the first evoked 1-2 spikes/electrode, observed intracellularly in
individual neurons (not shown), the early peak in the PSTH was rather sharp and highly reproducible, with
its amplitude and latency not significantly affected by the total pharmacological blockade of synaptic
transmission. Amplitude and latency of the early peak in the PSTH did not significantly correlate with the
light pulse duration (Fig. 2B). Instead, the peak amplitude and latency of the reverberating response were
affected by the total pharmacological blockade of synaptic transmission (not shown) and significantly
correlated with the light pulse duration (Fig. 2C-D; r = 0.42 and r = -0.36, respectively, p < 0.0001): the
longer the pulse, the weaker and more delayed the reverberating response. Differences in peak amplitudes
and latencies, upon varying pulse duration, were highly significant (p<0.01) across the entire dataset (n = 39
MEAs), but only for the reverberating responses, not for the early responses.
Examining the evoked responses in more details, employing higher temporal resolution (i.e. 1ms bins),
we also found signatures of global oscillatory activity in the PSTHs, with striking similarity to physiological
gamma-range oscillations. Figure 3A illustrates, for a representative MEA, the spikes evoked by the stimulus
as a raster plot of their times of occurrence and its corresponding PSTH, across 59 microelectrodes and 60
repetitions. The estimate of the time-varying power spectrum of the PSTH (Fig. 3B) clearly revealed the
appearance of a fading oscillation with frequencies in the gamma range (i.e. ~[40; 100] cycles/sec), as
conventionally measured 75ms after the stimulus onset, and qualified as dominant in terms of signal-to-noise
ratio (see the Methods). Across all MEAs tested, oscillations completely faded away ~200ms after the onset
of the stimulus, and generally slowed down at a rate of ~0.3cycles/sec per msec over time. Most importantly,
the frequency of the oscillations significantly correlated with the pulse duration (r = 0.55, p<0.0001): the
longer the pulse, the faster the dominant oscillation (Fig. 3C-D).
In addition, an initially low variability of the spike count, estimated by the Fano factor at each
microelectrode across repetitions 56, immediately raised to higher values (i.e., from 0.4 to 1) within 20-50ms
from the stimulus onset (not shown). Moreover, the occurrence time of the spikes recorded at individual
microelectrodes showed no entrainment with the oscillation cycles of the PSTH over time, but rather
irregular discharges with spike rates generally below 50 spikes/sec (not shown). Overall, these results
suggest that excitatory-inhibitory feedback, rather than a beating-wave arising from spike synchronization
drift over time, is the prevailing mechanism of the observed oscillations.
Authors' preprint
4/23
Photo-activated network responses in vitro
Figure 2: Light-evoked network responses. Brief light pulses evoked network-wide spiking
responses, only in cultures transduced by AAV. Responses were quantified over 60 repetitions by computing
the peri-stimulus histogram of spike times (PSTH) (A), detected over 5 ms bins across all microelectrodes
(Fig. 1A). In all experiments, the time course of the PSTH was stereotyped and consisted of an early and late
phase, similarly to what was described for extracellular electrical stimulation. While intensity was fixed, the
duration of light pulses was changed systematically and it was unexpectedly found to significantly modulate
the late (C-D) but not the early phase of the PSTH (B). Longer pulses significantly (p<0.001) delayed (C)
and weakened (D) the late response peak, without affecting the early phase (circles in A). Evaluating the
non-parametric Kendall's rank coefficient, a significant correlation (p<0.0001) was only found between pulse
duration and late-peak latency (0.47) or between pulse duration and late-peak amplitude (-0.36). Panels B-D
summarize the results over 39 MEAs, while the insets show averages across several MEAs, over five distinct
sets of sister cultures, remarking the consistency of the findings across different biological preparations.
Similarity of evoked and spontaneous network events. Inspired by a previous report 68, we analysed in
more detail the time course of the population mean firing rate during spontaneous network-wide events (i.e.
network bursts - Fig. 1B) 34,43, in the absence of any light stimulation. While for evoked responses the spike
times were related to the stimulus onset and their histograms averaged accordingly across repetitions (Fig.
2A), for spontaneous events no temporal reference frame exists by definition. We thus rigidly shifted all
spikes, detected simultaneously across the microelectrodes during each spontaneous burst, by a common
interval adapted to maximise the firing rate time course similarity across bursts 68 (see the Methods).
Unexpected temporal features then emerged in the average firing rate, which could not be revealed by
routine analysis methods based on firing rate peaks 43 or threshold-crossing alignment 43 of spontaneous
bursts. We found that some but not all MEAs showed prominent signatures of spontaneous transient
oscillations in the time course of the firing rate within a network burst, occurring in the same physiological
gamma frequency of evoked rhythms (Fig. 4), similarly to the light evoked responses.
GABAA receptors are necessary for the evoked global gamma oscillations. In order to investigate
the synaptic mechanisms of the evoked oscillations, we pharmacologically abolished fast inhibitory synaptic
transmission. As inhibition is necessary for gamma-range rhythmogenesis in vivo 75, we tested the hypothesis
that excitation-inhibition interplay sustains light-evoked oscillations in vitro. When a competitive selective
blocker of GABAA postsynaptic receptors was bath-applied, not only the spontaneous network activity
significantly increased in terms of firing and burst rate (Fig. 5A) as described 34,57,58. In addition, while light-
Authors' preprint
5/23
Photo-activated network responses in vitro
evoked responses still consisted of an early and a late component, no oscillatory activity occurred as no
single peak in the power spectrum of the PSTHs qualified as a dominant frequency (n = 17 MEAs) (Fig. 5B-
C). Hence, the pharmacological manipulation of network activity corroborates the interplay between
excitatory and inhibitory neurons, necessary for the light-evoked network oscillations.
Figure 3: Oscillatory light-evoked responses are modulated by the duration of light pulses. Typical
light-evoked responses are plotted for three light pulse durations (i.e., 1, 5, and 20 ms) as (A) spike-time
rasters (black dots) over all electrodes and trials, and as spike-time PSTH (red traces), computed over 1 ms
bins and peak normalized. The presence of transient oscillations was quantified by estimating the time-
varying power spectrum of each PSTH (B, left panels) and further slicing it 75 ms after the stimulus onset
(B, right panels, white dashed line; see the Methods). A clear dominant frequency (see Material and
Methods) was reliably observed in 37 out of 39 MEAs (for 1,2, and 5 ms long stimuli) or in 30 out of 39
MEAs (for 0.1, 0.5, 10, and 20 ms long stimuli). The dominant frequency showed a progressive drift towards
slower oscillations over time (B), and it significantly varied for increasing values of the pulse durations
6/23
Authors' preprint
Photo-activated network responses in vitro
(p<0.001; C-D). Evaluating the non-parametric Kendall's rank coefficient, a significant (p<0.0001)
correlation was found between dominant frequency and pulse duration (0.55).
Figure 4: Oscillatory firing activity during spontaneous network bursting. By aligning for maximal
similarity the spike trains, recorded by MEA microelectrodes for each spontaneous network burst, and
averaging as in 68, some (A-B) but not all MEAs (C) showed prominent signatures of oscillations in the same
physiological gamma frequency of evoked rhythms. The power spectra (insets) further quantify that, when
occurring, oscillations were dominated by physiological gamma range frequency.
Authors' preprint
7/23
Photo-activated network responses in vitro
Figure 5: Inhibition is necessary for light-evoked oscillations. The pharmacological blockade of GABAA
receptors disinhibited the network and altered quantitatively but not qualitatively the spontaneous activity,
significantly (p<0.05) increasing the burst rate and decreasing the burst duration (A). However, light-evoked
responses were qualitatively affected by network disinhibition as they did not contain oscillatory components
(n = 17 MEAs), as revealed by the power spectrum analysis (see Fig. 3) and the lack of a dominant
oscillation frequency (B-C).
Oscillations in silico by the excitatory-inhibitory feedback loop. In order to further investigate the
consequences of the excitatory-inhibitory feedback loop, we considered a classic firing-rate mathematical
model 50. The in vitro cortical networks considered here are composed by both excitatory and inhibitory
neurons, showing a prominent pattern of recurrent synaptic connectivity 34. We therefore defined a model
composed of an excitatory population and an inhibitory population with recurrent excitation and feedback
inhibition (Fig. 6A), and we analysed and numerically simulated its response to an external constant input.
As excitatory neurons express ChR2 in the experiments, only the excitatory population was activated by the
external input. After linearization of eqs. 4-6, the approximate conditions for stable oscillatory activity could
be formulated in terms of average synaptic efficacies. Assuming identical strength of excitatory synapses for
simplicity, i.e. 𝐽"= 𝐽%%= 𝐽&%, recurrent excitation should not compromise the stability of the system, i.e.
𝐽"<2𝛼%*+, but must be strong enough to recruit the inhibitory population, i.e. 𝐽"> 𝛼%𝛼&𝐽%& *+. This range
of values for 𝐽" exists if the feedback inhibition is sufficiently strong, i.e. 𝐽%&>𝛼&*+ (Fig. 6C). Under these
𝑓=(2𝜋)*+(2𝜏)*3 4𝛼%𝛼&𝐽%&𝐽"− 𝛼%𝐽" 3
for 𝐽%&=0 the term below the square root would be negative, and supports the interpretation of the
Importantly, any (unmodelled) mechanism that should increase the efficacy of excitatory synapses 𝐽", e.g.
oscillations (Fig. 6B). Similarly, any (unmodelled) mechanism that should decrease 𝐽", e.g. as a recovery to
hypotheses, oscillations occur at a frequency f, given as the imaginary part of the conjugate complex
eigenvalues of the connectivity matrix A (eq. 8), irrespective of the amplitude and duration of the external
constant input 47:
as in short-term synaptic facilitation, would be sufficient to increase f and thus speed up the global
experimental results obtained under blockade of GABAA.
its resting value following an intense facilitation, would account for the progressive slowdown of the
oscillations through time.
(1)
Equation 6 implies that feedback inhibition is necessary in our simplified in silico scenario (Fig. 6C), as
Authors' preprint
8/23
Photo-activated network responses in vitro
Figure 6: A firing rate model links evoked oscillations to excitatory-inhibitory interactions. By a firing-
rate mathematical model (A) we qualitatively reproduced oscillatory transient activity in silico (B). Recruited
by the external stimulation of excitatory neurons, the negative feedback of the inhibitory loop generated a
well-known alternating pull able to decrease neuronal firing. The red and green traces represent qualitatively
the mean firing rate of inhibitory and excitatory neurons, respectively. Altering the efficacy of synaptic
connections in a series of numerical simulations (B), resulted in different frequencies of the emerging
network rhythm: increasing the strengths of excitatory connections (i.e. Jie and Jee) increased the frequency of
the oscillations. As in the experiments (Fig. 5), inhibition was necessary for the oscillations, as they
completely vanished upon removal of the inhibitory feedback (C, Jei = 0). However, the model could account
per se neither for the dependency of the oscillation frequency on the duration of the light pulse, nor for its
drift over time. As a small increase in the average excitatory synaptic efficacy is sufficient in silico to speed
up oscillations, the duration of the light pulse might indirectly affect synaptic efficacy in the experiments, as
in short-term facilitation.
Short-term facilitation of synaptic release probability at excitatory synapses. The insights obtained
with the model suggest that light, and its duration, might proportionally affect the excitatory synaptic
efficacy. We therefore hypothesised that the link between light duration and oscillation frequency might
involve the influx and transient accumulation of Ca2+ in the excitatory synaptic boutons, as in the residual
Ca2+ hypothesis for short-term synaptic facilitation 59,60. This hypothetical link, occurring by the activation of
CHR2 LC-TC and voltage-gated calcium-channels, would lead to a (transient) increase of the action
potential-dependent excitatory synaptic efficacy (i.e. analogous to 𝐽" in eq. 1) and it would compete with the
endogenous Ca2+ recovery processes. That the synaptic release probability should have been reversibly
affected by the light stimulus, and depend linearly on its duration was a prediction of our model.
Intending to test our prediction, we performed whole-cell intracellular recordings from the soma of
neurons not expressing CHR2 LC-TC (n = 5 cells), by selecting mCherry-negative cells through
videomicroscopy. Then, we blocked action potential (AP) initiation by bath application of a selective blocker
9/23
Authors' preprint
Photo-activated network responses in vitro
of voltage-gated sodium channels (i.e. tetrodotoxin, TTX). Recording in voltage-clamp, we quantified
number and instantaneous rate of AP-independent excitatory postsynaptic events 61,62. Such events are the
result of neurotransmitter release, occurring in connected presynaptic boutons spontaneously or reflecting a
transient increase in the free Ca2+ concentration in the boutons 63,64. We found that immediately after the light
stimulus, the presynaptic release probability abruptly and reversibly increased (Fig. 7C), and it was linearly
correlated at its peak with the duration of the light pulse (Pearson's r = 0.7, p<0.0001). This observation is
compatible with the residual Ca2+ accumulation in the synaptic terminal. In addition, the presynaptic release
probability reversed significantly over time, exponentially decaying with a time constant of ~200ms (Fig.
7D), similarly to the recovery from short-term facilitation.
Figure 7: Brief light pulses transiently increase the rate of AP-independent synaptic release events.
Intracellular whole-cell voltage-clamp recordings were performed while blocking sodium-currents by TTX,
suppressing initiation and propagation of action potentials (APs). The time course of intrinsic and synaptic
currents, elicited by brief light pulses (blue arrow and dotted line), reveal fast on-kinetics and slow off-
kinetics of ChR2-LCTC (A), which strongly prevails on synaptic currents in mCherry-positive neurons.
Recording from mCherry-negative neurons, only synaptic currents are observed (B). The light pulses
induced a transient increase in the rate of spontaneous synaptic release event significantly correlated with the
duration of the pulse (0.7, p<0.0001) (C). The time course of the release events, quantified in 100 ms bins
(D), decayed exponentially with a time constant of ~200 ms, roughly independent on the pulse duration.
Authors' preprint
10/23
Photo-activated network responses in vitro
Discussion
Our results show that dissociated neuronal cultures are an interesting model for the study of gamma-
range oscillations and are useful to investigate some non-trivial consequences of wide-field stimulation in
recurrently connected neurons. Through this experimental model, we report for the first time how to robustly
evoke in vitro an oscillatory reverberating network response, and reproducibly modulate its frequency, within
a range of physiological importance. Many earlier studies reported on spontaneous and evoked oscillations in
vitro 65-68, but none described a fine experimental control of its features. This might prove very relevant for in
vitro models of (dys)functional cortical microcircuitry 69,70 and for in vivo optogenetic manipulations of
cortical cell assemblies.
We interpreted the emergence of gamma-range damped oscillations from the known interplay between
excitation and inhibition, supported by pharmacological experiments and mathematical modeling. An
alternative interpretation is offered from computational studies: the firing rate of a population of weakly-
coupled neurons show damped oscillations when all cells respond simultaneously and regularly to the same
strong stimulus onset 71-73. Roughly the opposite of a beating-wave, the damped oscillations observed in our
experiments might be the consequence of jitter and progressive random shifts of otherwise initially
synchronous and regular spike trains. The onset of the light pulse might transiently reset and de facto
synchronize the firing of the majority of the cells. Neurons would then start to fire regularly and roughly
synchronously, while progressively going out of phase due to distinct membrane properties, heterogeneous
CHR2 expression, synaptic interactions, etc.
In such a scenario however, single-neuron firing rates would roughly match the global rhythm of PSTH,
with an action potential per cycle. Then, the blockade of feedback synaptic inhibition would not necessarily
disrupt the oscillations. For these reasons, and for the physiological range of rhythms observed, the theory of
sparsely synchronized oscillations proposed in excitatory-inhibitory network architectures to explain the
emergence of gamma oscillations in vivo 74-76, might better describe our in vitro data. In fact, the
pharmacological blockade of GABAA receptors completely abolished the oscillations in the PSTH, the spike
count across trials showed high variability, and the temporal pattern of firing detected at single
microelectrodes appeared to be not completely entrained with the oscillation cycles.
We note however that the blockade of GABAA receptors also removes an inhibitory tone in the
excitatory neurons, steepening their frequency-current curves and depolarising their resting potential. It is
then possible that these two effects might alter the conditions for oscillations emergence, in a scenario where
excitation alone generates the rhythm. Future experiments, mimicking the inhibitory tone in disinhibited
networks by Optogenetics (e.g. expressing CHR2 and Archaerhodopsin
in excitatory neurons),
complemented by accurate biophysical spiking neural model, will be needed to explore the validity of this
alternative hypothesis.
In summary, the activity we observed in vitro captures many of the feature of in vivo gamma-range
rhythms, such as its short-life, the requirement for an intact excitatory-inhibitory loop, and concurrence with
irregular single-neuron firing 75.
To further strengthen this possible conclusion, it would be intriguing in the future to measure the
dynamical response 77 of both mCherry-positive/negative cells in vitro, as well as of short-term synaptic
plasticity 74-76. Combining these quantitative descriptions together, as described by Wang (2010), one could
predict whether or not gamma-range global rhythm can be sustained by recurrent connectivity at the
observed firing rates. This might serve as an experimental validation of the theory of sparsely synchronized
oscillations 74-76, ultimately linking intrinsic and synaptic properties to emerging oscillations in the same in
vitro biological preparation.
An important finding of our experiments is the proposed causal relationship between duration of light
pulses and excitatory synaptic efficacy. While the probability of synaptic release was already shown to be
altered upon optogenetic stimulation 78, here we describe a network-level correlate of this effect for the first
time. All our experimental results are consistent with our hypothesis of light-induced short-term facilitation
of excitatory synaptic efficacy. This requires the reasonable assumption that the transient and reversible
increase in Ca2+, observed in terms of an enhanced quantal synaptic release, affects the action potential-
dependent synaptic release probability, as in the residual Ca2+ hypothesis for explaining activity-dependent
short-term facilitation 59,60. The impact of CHR2 on Ca2+ transients has already documented postsynaptically
in dendrites and spines 78,79, although linked to voltage-gated calcium-channels activation. Here we advanced
the hypothesis that direct calcium influx through CHR2 at presynaptic boutons is responsible for the
observed phenomenon.
Authors' preprint
11/23
Photo-activated network responses in vitro
Ultimately, a full spiking neuron model, including a modified version of the Tsodyks-Markram model of
short-term synaptic plasticity 80 would be relevant to test the CHR2 LC-TC contribution on synaptic
variables. This model would extend the very basic predictions offered by our analysis of the mean-field
model (Fig. 6; eq. 1). Nonetheless, the rate model we examined was simple enough to test the excitation-
inhibition feedback loop hypothesis for the oscillations, as well as to explicitly suggest a candidate
observable for the additional intracellular experiments of Fig. 7A-B.
An alternative explanation of the relation between the light pulse duration and oscillation frequency
might involve the transient alteration of the excitability of (excitatory) neurons expressing CHR2. In fact, as
CHR2 LC-TC takes time to deactivate, the total effective membrane ionic conductance in those neurons is
larger immediately after the stimulation than before it. This effect would however not account for the
increase in the frequency with an increase in the duration of the light pulses. An increase in the effective
membrane conductance is known to decrease the slope of single-cell input-output transfer function 81, thus
leading to opposite consequences than those observed, for the frequency f of oscillation (see 𝛼% in eq. 1). In
addition, while our hypothesis of presynaptic Ca2+ accumulation explains why progressively longer pulses
result in faster oscillations, a stimulus-dependent change in the total membrane conductance per se would not
explain any temporal integration of the stimulus feature. Its value would be dependent only on the time since
the onset of CHR2 activation, irrespective of the pulse duration.
We also remark that the observed slowdown of the oscillation instantaneous frequency (Fig. 3B) is
consistent with our hypothesis of a transient accumulation of Ca2+, and its expected depletion during a
subsequent recovery phase, as well as with the short-term synaptic depression. As mCherry was expressed
over the entire neuronal morphology (Fig. 1D), the spatially unrestricted expression of CHR2, presumably
even at excitatory synaptic boutons, supports the Ca2+-accumulation hypothesis. Thus, as the brief light pulse
instantaneously primed a synapse, subsequent (endogenous) presynaptic action potentials might find easier to
release from the same synapse, as in short-term (spiking) activity-dependent facilitation 59,60.
Our intracellular recordings performed under TTX (Fig. 7B-C) however represent only indirect
experimental evidence of the presynaptic facilitation of synaptic efficacy. Intracellular recordings, performed
simultaneously from a pair of synaptically connected neurons, would be necessary to quantify whether action
potential-dependent synaptic release appears potentiated immediately after delivering light stimulation.
While wide-field photostimulation might not be entirely appropriate in that context, a similar experiment
might conclusively prove that CHR2 affects recurrent synaptic transmission by altering synaptic efficacy. In
addition, with recently developed opsins e.g. Chronos, 82 with altered Ca2+ permeability or faster closing
kinetics, we predict that the dependency of the oscillation frequency on the light stimulation duration should
change dramatically.
We finally note that the validity of our linear analysis of the mathematical model does not extend to very
low firing rate regimes, due to the non-linear response properties of neurons near their firing threshold. In
such a regime, oscillations take a more complicated form and their frequency is also determined by the
balance between excitation and inhibition.
The finding of a signature of gamma rhythms also during spontaneous bursting, in some but not all
MEAs (Fig. 4), is also of great significance to interpret the light activated responses. Despite based on very
short (i.e. 5 mins) recordings for each MEA, originally meant as a viability control, this finding generalizes
to mature large-scale networks what was first reported in developing small neuronal circuits 68: an "innate"
mode of operation of recurrent neuronal circuits. In addition, our observation indirectly confirms a broad
heterogeneity in the excitation-inhibition ratio in cortical networks, reflected in the variability of the
spontaneous rhythms frequency from ~50 to ~100 cycles/s (Fig. 4) (or the lack of it) 88. In addition,
transitions were sometimes observed (e.g. Fig. 4A, lower panel) between non-oscillating and oscillating
modes of intra-burst activity, consistent with the broad range of time-scales for single-neuron and network
excitability 88. Finally, our results further support the dynamical character of such an excitation-inhibition
ratio 88, as it can be readily manipulated by internal or external inputs (i.e. light pulses).
Overall, in line with many previous studies 4,8,14,83, our results stress the necessity to take into account
the possible impact on recurrent connectivity when designing optogenetic experimental protocols. Moreover,
we emphasize the importance of further developments in the field, particularly confining the opsin
expression in specific regions of a neuron 84, based on exact experimental needs. In fact, the light-induced
short-term facilitation of excitatory synaptic efficacy might prove to be a bug for some applications and not a
feature. Ultimately, besides further investigating and clarifying the impact of optogenetic manipulations in
Authors' preprint
12/23
Photo-activated network responses in vitro
large-scale recurrent cortical networks, particularly those affecting synaptic release probability, our work
provides a means to manipulate the network responses, for the benefit of future principled design of closed-
loop control paradigms.
Large-scale cultured networks are acknowledged as an experimental preparation, where structural,
dynamical, and plastic properties of in vivo cortical networks may be effectively investigated 22, 32, 34, 38, 43, 89-
95. Due to their inherent 2-dimensional geometry, lower effective cell packing density, and random
arrangement of cells' morphology in the plane, they cannot capture faithfully the features of electrical local
field potentials (LFPs) as recorded in 3-dimensional layered structures as brain slices and in vivo 96. While
our (evoked or spontaneous) in vitro gamma oscillations occur prominently in the network-wide probability
of firing (Fig. 3A), intact brain rhythms are often reflected also in the spatially-integrated activity of LFPs.
As we examined the low frequency components of single-trial MEA raw recordings, a weak signature of the
global firing rhythm could be indeed observed (Supplemental Figure S2) although with orders of magnitude
worse signal-to-noise ratio and more difficult interpretation 97 than the spike PSTHs.
and
In conclusion, the emergence of oscillations at physiological gamma frequency and their modulation by
the selective activation of excitatory neurons makes large-scale cultured networks, MEAs, and optogenetics
relevant for studying and manipulating network phenomena under highly-controlled conditions (see also
85,86,87), as well as for validating theoretical models of population dynamics with simplified hypotheses on
unstructured connection topology.
Materials and Methods
Fw–AAAGCTAGCACTTGTGGACTAAGTTTGTTCACATCCC–
Viral vectors production. To replace the cytomegalovirus (CMV) promoter in the Adeno-Associated
Viral vector (AAV) transfer plasmid 77, a 364 bp fragment of the mouse α-subunit of Ca2+/Calmodulin-
dependent protein kinase II promoter (CaMKIIα) was PCR-amplified from plasmid 20944 (Addgene, UK),
using
Rev–
AAAAGCGCTGATATCGCTGCCCCCAGAACTAGGGGCCACTCG– as primers. This PCR fragment was
ligated in the AAV transfer plasmid, using NheI and Eco47 III, after the CMV promoter was cut out. eGFP
was then replaced by CHR2 L132C/T159C-mCherry. The opsin coding sequence was cloned by PCR
amplification and the final transfer plasmid was sequence-verified prior to viral vectors production. AAV2/7
viral vectors were produced at the Leuven Viral Vector Core, as described earlier 77. Briefly, HEK 293T cells
(ATCC, Manassas, VA, USA) were seeded in the Hyperflask at 1.2x108 cells per viral vector production, in
Dulbecco's Modified Eagle's Medium (DMEM) with 2% Fetal Calf Serum. The following day the medium
was replaced by serum-free Optimem and cells were transfected with the pAAV transfer plasmid, the rep/cap
plasmid, and the pAd.DELTA.F6 plasmid in a 1:1:1 ratio. The supernatant was harvested five days after
transient transfection, concentrated using tangential flow filtration, and purified using an iodixanol step
gradient. Aliquots were stored at -80°C and quantified, using real-time PCR, as Genome Copies per ml
(GC/ml): titers varied between 9.8x1011 GC/ml and 1.6x1012 GC/ml.
Neuronal culturing and transduction. Primary cultures of mammalian neurons, dissociated from the
postnatal rat neocortex, were prepared as in 30, in accordance with international and institutional guidelines
on animal welfare. All experimental protocols were specifically approved by the Ethical Committee of the
Department of Biomedical Sciences of the University of Antwerpen (permission n. 2011_87), and licensed
by the Belgian Animal, Plant and Food Directorate-General of the Federal Department of Public Health,
Safety of the Food Chain and the Environment (license n. LA1100469). Briefly, cerebral cortices (excluding
the hippocampus) were removed from the brains of Wistar pups (P0), flushed with ice-cold phosphate
buffered saline (PBS) containing 20 mM glucose, and roughly minced with a blade. Hemi-cortices were
enzymatically digested by adding 0.05 mg/ml trypsin to the solution and by gentle agitation, for 15 min at
37 °C (GFL 1083, Gesellschaft für Labortechnik mbH, Burgwedel, Germany). After stopping the digestion
by adding 2 ml of heat-inactivated Normal Horse Serum (NHS), the tissue fragments were left to sediment at
room temperature and then mechanically triturated by a 10 ml Falcon pipette, filtered through a nylon
monofilament cell strainer (40µm, #352340, BD Falcon, Franklin Lakes, NJ, USA), and centrifuged at room-
temperature (220 g for 5 min; GS-6, Beckmann-Coulter, Brea, CA, USA). The resulting cell pellet was
resuspended in culture medium, composed of 4ml Minimum Essential Medium (MEM), 5% NHS, 50 µg/ml
gentamycin, 0.1 mM L-glutamine, and 20 mM sucrose, and diluted to reach a final surface plating density of
6,500 cell/mm2 on glass coverslips (15mm, VD1-0015-Y2MA, Laborimpex, Forest/Vorst, Belgium) and
Authors' preprint
13/23
Photo-activated network responses in vitro
commercial substrate-integrated arrays of microelectrodes (MEAs; Multi Channel Systems GmbH,
Reutlingen, Germany).
MEAs with a regular 8x8 arrangement of 60 Titanium Nitrate (TiN) microelectrodes, each with 30µm
diameter and 200µm spacing (60MEA200/30iR-ITO-gr, Multi Channel Systems), were employed. Prior to
cell seeding, MEAs and coverslips were coated overnight with polyethyleneimine (0.1% wt/vol in millQ
water at room temperature), and then extensively washed with milliQ water and let air-drying. Seeded MEAs
and coverslips were maintained in an incubator for up to 6 weeks at 37 °C in 5% CO2 and 100% R.H. (5215,
Shellab, Cornelius, OR, USA), with MEAs sealed by fluorinated Teflon membranes (Ala-MEA-Mem, Ala
Science, Farmingdale, NY, USA) and coverslips stored in Petri dishes.
The culture transduction was performed in vitro five days (DIV5) after seeding, by partly replacing the
medium of each MEA or coverslip with a 1:50 dilution of viral particles AAV-CaMKIIα-CHR2(LC-TC)-
mCherry in fresh medium. At DIV8, fresh and pre-warmed medium was added to reach 1 ml final volume in
MEAs and coverslips. From DIV10 on, half of their culture medium volume was replaced every 2 days with
fresh, pre-warmed medium. All reagents were obtained from Sigma-Aldrich (St. Louis, MO, USA) or Life
Technologies (Gent, Belgium).
Immunocytochemistry. Cell identity, density, and opsin expression selectivity and efficiency were
evaluated after DIV28 by immunocytochemistry, by imaging mCherry-positive cells and neuronal nuclei
(NeuN) stained by an Anti-NeuN monoclonal antibody (clone A60, #MAB377, Millipore, Billerica, MA,
USA) (Fig. 1D). Control and transduced cultures on coverslips were washed three times with PBS and fixed
by 4% paraformaldehyde dissolved in PBS (10 min at 37 °C). After washing twice with ice-cold PBS, cells
membranes were permeabilised (0.25% Triton X-100 in PBS-Tween for 10 min) and pre-incubated with 10%
goat serum in PBS-Tween for 30min. Primary Anti-NeuN (1:150) incubation was performed overnight at
4 °C. After repeated PBS washing, the incubation with 1:150 Alexa-488 goat anti-mouse (A11001,
Molecular Probes, Eugene, OR, USA) was performed for 1h. After further PBS washing, the coverslips were
mounted on glass slides by a medium containing 4′,6-diamidino-2-phenylindole (DAPI), which stained all
cell nuclei (Vectashield H-1500, Vector Laboratories, Burlingame, CA, USA).
Several microphotographs of NeuN immunolabelling, DAPI staining, and mCherry red-fluorescence were
acquired under epifluorescent microscopy (Olympus DP71, Tokyo, Japan) and processed for semiautomatic
cell-counting (ImageJ, NIH, Bethesda, MD, USA), examining four non-overlapping random fields (320µm ×
240µm) over at least three distinct coverslips, seeded during two different cell dissociation sessions.
MEAs extracellular electrophysiological recordings. Before each experiment, MEAs were transferred
from the incubator to an electronic amplifier (MEA-1060-Up-BC, Multi Channel Systems) with 1–3000 Hz
bandwidth and amplification factor of 1200, placed inside an electronic-friendly incubator (37 °C, 5% CO2,
~20 R.H.). MEAs were left in the incubator to accommodate for 5min before starting the recordings, and
were used to detect non-invasively the spontaneous and (light) evoked neuronal electrical activity. An
MCCard A/D board and the MCRack software (Multi Channel Systems) were used to sample (25
kHz/channel), digitize (16 bits), and store on disk the raw voltage traces individually detected by each of the
60 microelectrodes for subsequent analysis. Recordings took place after 28 DIVs, as continuous as well as
triggered acquisitions of raw traces. Assessing the health of each MEA, five minutes of spontaneous activity
was routinely recorded, before delivering the photo-stimulations (Fig. 1B-C). Light-evoked activity was then
recorded across the entire MEA, by monitoring 200ms proceeding and 1000ms following the photo-stimulus
onset.
Intracellular electrophysiological recordings. Glass coverslips were employed for patch-clamp
experiments in sister cultures of at least 20DIV, using an Axon Multiclamp 700B amplifier (Molecular
Devices, USA) controlled by the LCG software 78. Patch pipettes were pulled on a horizontal puller (P97,
Sutter, Novato, USA) from filamented borosilicate glass capillaries (World Precision Instrument Inc. (WPI),
USA), and had a resistance of 4–5 MΩ, when filled with an intracellular solution containing (in mM): 115 K-
gluconate, 20 KCl, 10 HEPES, 4 Mg-ATP, 0.3 Na2 -GTP, 10 Na2 –phosphocreatine (pH adjusted to 7.4 with
NaOH). An extracellular solution, containing (in mM): NaCl 145, KCl 4, MgCl2 1, CaCl2 2, HEPES 5, Na-
pyruvate 2, glucose 5 (pH adjusted to 7.4 with NaOH), was constantly perfused at a rate of 1 ml/min and at a
temperature of 37±1 °C for each experiment.
Authors' preprint
14/23
Photo-activated network responses in vitro
Miniature post-synaptic currents (mPSCs) were recorded under voltage-clamp at a hyperpolarized holding
potential of -70mV, under the whole-cell configuration from the soma of mCherry-negative neurons, after
low-pass filtering at 6 kHz and sampling at 30 kHz the acquired traces. Photo-activated responses were also
recorded monitoring the membrane currents, while patching mCherry-positive neurons. Repeated stimulation
trials, acquiring 2 s preceding and 2 s following each photo-stimulation were performed, as in MEA
experiments. Cells with resting potential above −55 mV, or with a series resistance larger than 25 MΩ, were
not included for analysis and discarded. In all the experiments, the Active Electrode Compensation 41 was
employed by default through LCG.
Pharmacology. In some MEA experiments, Gabazine (GBZ, SR95531) was bath-applied prior to the
electrophysiological recordings at the final concentrations of 30 µM, as a selective competitive blocker of
GABAA receptors (Fig. 5). In all intracellular experiments (Fig. 7A-B), 1 µM tetrodotoxin (TTX) was added
to the extracellular solution to selectively block sodium channels and thus suppress the generation of action
potentials in all neurons. All drugs were obtained from Abcam (Cambridge, England, UK).
Data Analysis and Statistics. For MEA recordings, extracellular voltage waveforms of light-evoked
neuronal responses were analyzed on-the-fly by custom scripts written in MATLAB (The MathWorks, Natik,
MA, USA), while the package QSpike Tools 80 was employed for analyzing spontaneous activity and
extracting standard observables such as spike times, spike rate, burst rate, and burst duration 22,43. Raw
voltage traces recorded at each microelectrode (Fig. 1A) were zero-phase band-pass filtered (400-3000Hz)
and then processed by an adaptive peak-detection algorithm, based on voltage-threshold crossing and
including an elementary sorting of positive and negative amplitude peaks of the corresponding spike
waveforms 44-46. The times of occurrence of extracellular spikes, for each recording channel and each
repetition, were stored on disk for subsequent analysis together with their corresponding spoke waveform.
Spike waveforms with negative peak voltage amplitudes were sorted and included in the analysis.
Specifically, their time stamps were used to estimate network-wide evoked instantaneous spike response
probability, computed as a Peri-Stimulus spike Time Histogram (PSTH) over 5 ms bins (Figs. 2A) and whose
features were subsequently extracted (Fig. 2B-D).
Spontaneous network bursts were also studied, analyzing the corresponding firing probability over time
by computing for each burst a spike time histogram (STH) over 5 ms bins. The overall firing rate profile
during a burst, averaged over all the bursts recorded from a MEA, was however not obtained arbitrarily
aligning the STHs recorded to the time of their peak 43. Instead, STHs were first arbitrarily shifted in time
and then aligned for averaging, by a series of time intervals that maximized their mutual similarity, through
maximizing their correlation coefficients 68.
Frequency-domain analysis was also performed, estimating the spectrogram 47 of each PSTH (estimated
over 1 ms bins, Fig. 3A) corresponding to a different photo-stimulus duration (Fig. 3B). In details, the Fast
Fourier Transform with a sliding window (50 ms width, 50% overlap) was employed to extract the time-
varying spectrum of frequencies contained in the PSTH 48, averaging over the 60 trials and normalizing
across frequencies by the PSTH spectrum of the 200 ms preceding the photo-stimulus onset. In addition, the
power spectrum computed 75 ms after the stimulus onset (Fig. 3B-C) was conventionally employed to
determine existence and location of the dominant oscillation frequency, defined arbitrarily as the frequency
corresponding to the highest peak of the spectrum, which necessarily exceeded the median value of the
spectrum over all frequencies by three times (Fig. 3C).
In intracellular recordings, current traces were low-pass filtered at 1.5 kHz and analysed using custom-
written MATLAB scripts to count the number or the instantaneous rate of mPSCs.
Data is presented as mean ± standard error of the mean (SEM), except for Figures 1C and 5A, where box
plots show the median (horizontal bar) and the mean values (circle), the 25th and 75th percentiles (box's
edges), and the most extreme data points (whiskers). Distributions were compared using the Kolmogorov-
Smirnov test and only differences with at least p<0.05 were considered significant. Correlations were
assessed by computing the non-parametric Kendall's rank correlation coefficient and its significance level 86.
Wide-field photo-stimulation. A blue light-emitting diode (LED) (Rebel, Quadica Development,
Canada), powered by an externally dimmable DC driver (LuxDrive, Randolph, USA), was employed to
deliver wide-field photo-stimulations. The LED was further equipped with a parabolic reflector lens, to
Authors' preprint
15/23
Photo-activated network responses in vitro
collimate the light beam and distribute the emitted power on the bottom of MEAs and coverslips uniformly
(Fig. 1A).
In MEA experiments, photo-stimuli were delivered at full power (i.e. 2 mW/mm2, measured at the sample
by a calibrated photodiode; 818-ST2-UV, Newport Spectra-Physics, Netherlands) by a voltage-controlled
stimulus generator (STG1002, Multi Channel Systems) connected to the DC driver. The generator was
programmed over USB to define timing and waveform of each stimulus. 60 identical square pulses were
consecutively delivered at a 0.2 Hz repetition rate, and the specific effect of the pulse duration [0.1, 0.5, 1, 2,
5, 10, 20] ms was investigated. Thus, stimulation sessions differed only in the pulse durations, whose value
was chosen in a shuffle order across successive sessions.
In the intracellular experiments, an analog output of the acquisition D/A board was connected to the LED
DC driver and the LCG software employed to deliver the photo-stimuli as in MEA experiments, repeated 20
times per condition at a 0.2 Hz.
In our experiments, the pulse duration T was varied while keeping its intensity Lmax to full power to
improve proportionality between stimulus and net inward charge, as tested by voltage-clamp experiments
(Supplemental Figure S1). In fact, the equation describing the fraction x of open CHR2 channels, from a
minimal 2-state (non-inactivating) kinetic biophysical model,
𝑥𝑡 = −𝑏∙𝑥𝑡 +𝑎∙𝐿𝑡 ∙ 1−𝑥(𝑡) (2)
∙ 9∙:BCDE3∙;
;∙9∙:BCDE; ≈𝑇 (3)
𝑥
1−𝑥 9 :
;
9∙:BCDE;∙ 𝑇+ 1−𝑒H∙9∙:BCDE;
9∙:BCD
predicts a net charge proportional to T, only for a large intensities Lmax.
(4)
I𝑡 = 𝛼&ℎ&𝑡 −𝜃& E,
frequency-current response function. In the case of a network topology including recurrent excitation and
reciprocal excitation and inhibition (Fig. 6A), the following coupled (non-linear) differential equations fully
describe the dynamics of the system:
Firing-rate model. A Wilson and Cowan rate-based mathematical model was defined and numerically
simulated 50,51, to review the consequences of recurrent and feed-back interactions between an excitatory (e)
and an inhibitory (i) neuronal population. Each population was assumed to be composed of identical and
indistinguishable neurons and thus collectively described by a single state variable (i.e. he(t) and hi(t)). This
variable approximated the total average incoming synaptic inputs to a generic unit of each population. The
time-varying output mean firing rate of a population (i.e. E and I) was then assumed, for simplicity, to
depend only on its total synaptic input, as in a threshold-linear frequency-current curve (eq. 4):
𝐸𝑡 = 𝛼%ℎ%𝑡 −𝜃% E
where [ ]E indicates the positive part of its argument, θ a minimal activation threshold, and α the slope of the
𝜏%ℎ%𝑡 = −ℎ%𝑡 +𝐽%%𝐸(𝑡)−𝐽%&𝐼𝑡 +𝐸"
𝜏&ℎ%&𝑡 = −ℎ&𝑡 +𝐽&%𝐸𝑡 −𝐽&&𝐼𝑡 ,
where 𝐽%%,𝐽%&,𝐽&%,𝐽&& indicate the average synaptic efficacies of recurrent excitation, reciprocal excitation and
reciprocal inhibition (i.e. 𝐽&&=0, see Figure 6A). 𝐸" represents the time-varying external input current,
induced by CHR2 activation, and 𝜏% and 𝜏& are the time constants of decay of the synaptic currents in the
CHR2 LC-TC, 𝐸"(𝑡) was approximated as a brief square pulse followed by a much longer decaying
exponential function. By a change of variables, 𝑥=ℎ%−𝜃% and 𝑦=ℎ&−𝜃&, and assuming that both he and
hi at a given moment are larger 𝜃% and 𝜃&, respectively, eqs. 5-6 can be rewritten as a linear system of
𝑥𝑦 =𝐴 𝑥𝑦 +𝑏,
𝐴= (𝛼%𝐽%%−1)/𝜏%
𝛼%𝐽&%/𝜏&
The values of the parameters in A, and in particular of the mean synaptic couplings 𝐽%%,𝐽%&,𝐽&%,𝐽&&,
inhibition, and mutual inhibition, respectively. For simplicity of the subsequent analysis, we assumed no
lack of presynaptic activity, for each population respectively. Taking into account the rather slow kinetics of
𝑏= (𝐸"−𝜃%)/𝜏%
−𝜃&/𝜏&
differential equations, using linear algebra notation:
determine the eigenvalues of the matrix A and thus the dynamics of the output network response (e.g. x(t)) to
the external stimulus (i.e. E0(t)), being exponential or oscillatory over time 52. The parameters employed for
16/23
Authors' preprint
−𝛼&𝐽%&/𝜏%
−(𝛼&𝐽&&−1)/𝜏&
where
(5)
(6)
(7)
(8)
Photo-activated network responses in vitro
the simulations of Figure 6 were: 𝐽&&=0,𝐽%&∈{0,12},𝐽&%=𝐽%%∈{0.32,0.61,0.9,1.2}, 𝛼&=𝛼%=1, 𝜃%=
𝜃&=0.1,𝜏%=𝜏&=0.1,𝐸"=20∙ 1+ 𝑢𝑡−1 −𝑢(𝑡−1.5) +𝑢(𝑡−1.5)∙𝑒*(\*+.])/".] , where 𝑢𝑡 is
equal to 0 for t<0 and equal to 1 otherwise.
Authors' preprint
17/23
Photo-activated network responses in vitro
References
1 Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G. & Deisseroth, K. Millisecond-timescale, genetically targeted
optical control of neural activity. Nature neuroscience 8, 1263-1268, doi:10.1038/nn1525 (2005).
2 Deisseroth, K. et al. Next-generation optical technologies for illuminating genetically targeted brain circuits. The
for Neuroscience 26, 10380-10386,
the Society
journal of
Journal of neuroscience
doi:10.1523/JNEUROSCI.3863-06.2006 (2006).
the official
:
3 Williams, S. C. & Deisseroth, K. Optogenetics. Proceedings of the National Academy of Sciences of the United
States of America 110, 16287, doi:10.1073/pnas.1317033110 (2013).
4 Fenno, L., Yizhar, O. & Deisseroth, K. The development and application of optogenetics. Annual review of
neuroscience 34, 389-412, doi:10.1146/annurev-neuro-061010-113817 (2011).
5 Yizhar, O., Fenno, L. E., Davidson, T. J., Mogri, M. & Deisseroth, K. Optogenetics in neural systems. Neuron 71, 9-
6 Boyden, E. S. A history of optogenetics: the development of tools for controlling brain circuits with light. F1000
34, doi:10.1016/j.neuron.2011.06.004 (2011).
Biol Rep 3, 11, doi:10.3410/B3-11 (2011).
7 Boyden, E. S. Optogenetics: using light to control the brain. Cerebrum 2011, 16 (2011).
8 Hausser, M. Optogenetics: the age of light. Nat Methods 11, 1012-1014, doi:10.1038/nmeth.3111 (2014).
9 Liu, X. et al. Optogenetic stimulation of a hippocampal engram activates fear memory recall. Nature 484, 381-385,
doi:10.1038/nature11028 (2012).
10 Bernstein, J. G. & Boyden, E. S. Optogenetic tools for analyzing the neural circuits of behavior. Trends in cognitive
sciences 15, 592-600, doi:10.1016/j.tics.2011.10.003 (2011).
11 Dranias, M. R., Ju, H., Rajaram, E. & VanDongen, A. M. Short-term memory in networks of dissociated cortical
neurons. The Journal of neuroscience : the official journal of the Society for Neuroscience 33, 1940-1953,
doi:10.1523/JNEUROSCI.2718-12.2013 (2013).
12 El Hady, A. et al. Optogenetic stimulation effectively enhances intrinsically generated network synchrony. Frontiers
in neural circuits 7, 167, doi:10.3389/fncir.2013.00167 (2013).
13 Inagaki, H. K. et al. Optogenetic control of Drosophila using a red-shifted channelrhodopsin reveals experience-
dependent influences on courtship. Nat Methods 11, 325-332, doi:10.1038/nmeth.2765 (2014).
14 Roux, L., Stark, E., Sjulson, L. & Buzsaki, G. In vivo optogenetic identification and manipulation of GABAergic
interneuron subtypes. Current opinion in neurobiology 26, 88-95, doi:10.1016/j.conb.2013.12.013 (2014).
15 Tchumatchenko, T., Newman, J. P., Fong, M. F. & Potter, S. M. Delivery of continuously-varying stimuli using
channelrhodopsin-2. Frontiers in neural circuits 7, 184, doi:10.3389/fncir.2013.00184 (2013).
16 Malyshev, A., Goz, R., LoTurco, J. J. & Volgushev, M. Advantages and limitations of the use of optogenetic
approach in studying fast-scale spike encoding. PloS one 10, e0122286, doi:10.1371/journal.pone.0122286 (2015).
17 Vaziri, A. & Emiliani, V. Reshaping the optical dimension in optogenetics. Current opinion in neurobiology 22, 128-
137, doi:10.1016/j.conb.2011.11.011 (2012).
18 Schoenenberger, P., Grunditz, A., Rose, T. & Oertner, T. G. Optimizing the spatial resolution of Channelrhodopsin-2
activation. Brain Cell Biol 36, 119-127, doi:10.1007/s11068-008-9025-8 (2008).
19 Petreanu, L., Huber, D., Sobczyk, A. & Svoboda, K. Channelrhodopsin-2-assisted circuit mapping of long-range
callosal projections. Nature neuroscience 10, 663-668, doi:10.1038/nn1891 (2007).
20 Gradinaru, V., Mogri, M., Thompson, K. R., Henderson, J. M. & Deisseroth, K. Optical deconstruction of
parkinsonian neural circuitry. Science 324, 354-359, doi:10.1126/science.1167093 (2009).
21 Lau, P. M. & Bi, G. Q. Synaptic mechanisms of persistent reverberatory activity in neuronal networks. Proceedings
the United States of America 102, 10333-10338,
the National Academy of Sciences of
of
doi:10.1073/pnas.0500717102 (2005).
22 Eytan, D. & Marom, S. Dynamics and effective topology underlying synchronization in networks of cortical
neurons. The Journal of neuroscience : the official journal of the Society for Neuroscience 26, 8465-8476,
doi:10.1523/JNEUROSCI.1627-06.2006 (2006).
23 Tateno, T., Jimbo, Y. & Robinson, H. P. Spatio-temporal cholinergic modulation in cultured networks of rat cortical
neurons: spontaneous activity. Neuroscience 134, 425-437, doi:10.1016/j.neuroscience.2005.04.049 (2005).
24 Rieubland, S., Roth, A. & Hausser, M. Structured connectivity in cerebellar inhibitory networks. Neuron 81, 913-
929, doi:10.1016/j.neuron.2013.12.029 (2014).
25 Barak, O., Tsodyks, M. & Romo, R. Neuronal population coding of parametric working memory. The Journal of
neuroscience : the official journal of the Society for Neuroscience 30, 9424-9430, doi:10.1523/JNEUROSCI.1875-
10.2010 (2010).
26 Grossman, N. et al. The spatial pattern of light determines the kinetics and modulates backpropagation of
optogenetic action potentials. Journal of computational neuroscience 34, 477-488, doi:10.1007/s10827-012-0431-7
(2012).
Authors' preprint
18/23
Photo-activated network responses in vitro
27 Witt, A. et al. Controlling the oscillation phase through precisely timed closed-loop optogenetic stimulation: a
computational study. Frontiers in neural circuits 7, doi:10.3389/fncir.2013.00049 (2013).
28 Williams, J. C. et al. Computational optogenetics: empirically-derived voltage- and
channelrhodopsin-2 model. PLoS computational biology 9, e1003220, doi:10.1371/journal.pcbi.1003220 (2013).
29 Gal, A. et al. Dynamics of Excitability over Extended Timescales in Cultured Cortical Neurons. Journal of
light-sensitive
Neuroscience 30, 16332-16342, doi:10.1523/jneurosci.4859-10.2010 (2010).
30 Reinartz, S., Biro, I., Gal, A., Giugliano, M. & Marom, S. Synaptic dynamics contribute to long-term single neuron
response fluctuations. Frontiers in neural circuits 8, doi:10.3389/fncir.2014.00071 (2014).
31 Marom, S. & Eytan, D. Learning in ex-vivo developing networks of cortical neurons. Progress in brain research
147, 189-199, doi:10.1016/S0079-6123(04)47014-9 (2005).
32 Corner, M. A., van Pelt, J., Wolters, P. S., Baker, R. E. & Nuytinck, R. H. Physiological effects of sustained blockade
of excitatory synaptic transmission on spontaneously active developing neuronal networks--an inquiry into the
reciprocal linkage between intrinsic biorhythms and neuroplasticity in early ontogeny. Neurosci Biobehav Rev 26,
127-185 (2002).
33 Potter, S. M. & DeMarse, T. B. A new approach to neural cell culture for long-term studies. Journal of neuroscience
methods 110, 17-24 (2001).
34 Marom, S. & Shahaf, G. Development, learning and memory in large random networks of cortical neurons: lessons
beyond anatomy. Quarterly reviews of biophysics 35, 63-87 (2002).
35 Pine, J. Recording action potentials from cultured neurons with extracellular microcircuit electrodes. Journal of
neuroscience methods 2, 19-31 (1980).
36 Gross, G. W., Williams, A. N. & Lucas, J. H. Recording of spontaneous activity with photoetched microelectrode
surfaces from mouse spinal neurons in culture. Journal of neuroscience methods 5, 13-22 (1982).
37 Wagenaar, D. A., Pine, J. & Potter, S. M. Effective parameters for stimulation of dissociated cultures using multi-
electrode arrays. Journal of neuroscience methods 138, 27-37, doi:10.1016/j.jneumeth.2004.03.005 (2004).
38 Jimbo, Y., Kawana, A., Parodi, P. & Torre, V. The dynamics of a neuronal culture of dissociated cortical neurons of
neonatal rats. Biological cybernetics 83, 1-20 (2000).
39 Van der Perren, A. et al. Efficient and stable transduction of dopaminergic neurons in rat substantia nigra by rAAV
2/1, 2/2, 2/5, 2/6.2, 2/7, 2/8 and 2/9. Gene therapy 18, 517-527, doi:10.1038/gt.2010.179 (2011).
40 Linaro, D., Couto, J. & Giugliano, M. Command-line cellular electrophysiology for conventional and real-time
closed-loop experiments. Journal of neuroscience methods 230, 5-19, doi:10.1016/j.jneumeth.2014.04.003 (2014).
41 Brette, R. et al. High-resolution intracellular recordings using a real-time computational model of the electrode.
Neuron 59, 379-391, doi:10.1016/j.neuron.2008.06.021 (2008).
42 Mahmud, M., Pulizzi, R., Vasilaki, E. & Giugliano, M. QSpike tools: a generic framework for parallel batch
preprocessing of extracellular neuronal signals recorded by substrate microelectrode arrays. Frontiers in
neuroinformatics 8, 26, doi:10.3389/fninf.2014.00026 (2014).
43 van Pelt, J., Wolters, P. S., Corner, M. A., Rutten, W. L. & Ramakers, G. J. Long-term characterization of firing
dynamics of spontaneous bursts in cultured neural networks. IEEE transactions on bio-medical engineering 51,
2051-2062, doi:10.1109/TBME.2004.827936 (2004).
44 Quiroga, R. What is the real shape of extracellular spikes? Journal of neuroscience methods 177, 194-198,
doi:10.1016/j.jneumeth.2008.09.033 (2009).
45 Quiroga, R. Q. Spike sorting. Current biology : CB 22, R45-46, doi:10.1016/j.cub.2011.11.005 (2012).
46 Quiroga, R. Q., Nadasdy, Z. & Ben-Shaul, Y. Unsupervised spike detection and sorting with wavelets and
superparamagnetic clustering. Neural computation 16, 1661-1687, doi:10.1162/089976604774201631 (2004).
47 Oppenheim, A. V., Schafer, R. W. & Buck, J. R. Discrete-time signal processing (2nd ed.). (Prentice-Hall, Inc.,
1999).
48 Mattia, M., Ferraina, S. & Del Giudice, P. in NeuroImage Vol. 52 812-823 (2010).
49 Press, W. H., Teukolsky, S. A., Vetterling, W. T. & Flannery, B. P. Numerical Recipes 3rd Edition: The Art of
Scientific Computing. (Cambridge University Press, 2007).
50 Wilson, H. R. & Cowan, J. D. Excitatory and inhibitory interactions in localized populations of model neurons.
Biophysical journal 12, 1-24, doi:10.1016/S0006-3495(72)86068-5 (1972).
51 Dayan, P. & Abbott, L. F. Theoretical Neuroscience: Computational and Mathematical Modeling of Neural Systems.
(The MIT Press, 2005).
52 Brogan, W. L. Modern control theory (3rd ed.). (Prentice-Hall, Inc., 1991).
53 Liu, X. B. & Murray, K. D. Neuronal excitability and calcium/calmodulin-dependent protein kinase type II: location,
location, location. Epilepsia 53 Suppl 1, 45-52, doi:10.1111/j.1528-1167.2012.03474.x (2012).
54 Prigge, M. et al. Color-tuned channelrhodopsins for multiwavelength optogenetics. The Journal of biological
chemistry 287, 31804-31812, doi:10.1074/jbc.M112.391185 (2012).
55 Chiappalone, M., Massobrio, P. & Martinoia, S. Network plasticity in cortical assemblies. The European journal of
neuroscience 28, 221-237, doi:10.1111/j.1460-9568.2008.06259.x (2008).
56 Koch, C. & Segev, I. Methods in Neuronal Modeling: From Ions to Networks. (MIT Press, 1998).
Authors' preprint
19/23
Photo-activated network responses in vitro
57 Giugliano, M., Darbon, P., Arsiero, M., Luscher, H. R. & Streit, J. Single-neuron discharge properties and network
activity in dissociated cultures of neocortex. Journal of neurophysiology 92, 977-996, doi:10.1152/jn.00067.2004
(2004).
58 Gullo, F. et al. Orchestration of "presto" and "largo" synchrony in up-down activity of cortical networks. Frontiers
in neural circuits 4, 11, doi:10.3389/fncir.2010.00011 (2010).
59 Markram, H., Lubke, J., Frotscher, M. & Sakmann, B. Regulation of synaptic efficacy by coincidence of
postsynaptic APs and EPSPs. Science 275, 213-215 (1997).
60 Zucker, R. S. & Regehr, W. G. Short-term synaptic plasticity. Annual review of physiology 64, 355-405,
doi:10.1146/annurev.physiol.64.092501.114547 (2002).
61 Turrigiano, G. G., Leslie, K. R., Desai, N. S., Rutherford, L. C. & Nelson, S. B. Activity-dependent scaling of
quantal amplitude in neocortical neurons. Nature 391, 892-896, doi:10.1038/36103 (1998).
62 Prange, O. & Murphy, T. H. Correlation of miniature synaptic activity and evoked release probability in cultures of
cortical neurons. The Journal of neuroscience : the official journal of the Society for Neuroscience 19, 6427-6438
(1999).
63 Lou, X., Scheuss, V. & Schneggenburger, R. Allosteric modulation of the presynaptic Ca2+ sensor for vesicle
fusion. Nature 435, 497-501, doi:10.1038/nature03568 (2005).
64 Kochubey, O., Lou, X. & Schneggenburger, R. Regulation of transmitter release by Ca(2+) and synaptotagmin:
insights from a large CNS synapse. Trends in neurosciences 34, 237-246, doi:10.1016/j.tins.2011.02.006 (2011).
65 Plenz, D. & Kitai, S. T. Generation of high-frequency oscillations in local circuits of rat somatosensory cortex
cultures. Journal of neurophysiology 76, 4180-4184 (1996).
66 Cunningham, M. O. et al. A role for fast rhythmic bursting neurons in cortical gamma oscillations in vitro.
Proceedings of the National Academy of Sciences of the United States of America 101, 7152-7157,
doi:10.1073/pnas.0402060101 (2004).
67 Gireesh, E. D. & Plenz, D. Neuronal avalanches organize as nested theta- and beta/gamma-oscillations during
development of cortical layer 2/3. Proceedings of the National Academy of Sciences of the United States of America
105, 7576-7581, doi:10.1073/pnas.0800537105 (2008).
68 Shein Idelson, M., Ben-Jacob, E. & Hanein, Y. Innate synchronous oscillations in freely-organized small neuronal
circuits. PloS one 5, e14443, doi:10.1371/journal.pone.0014443 (2010).
69 Thut, G., Miniussi, C. & Gross, J. The functional importance of rhythmic activity in the brain. Current biology : CB
22, R658-663, doi:10.1016/j.cub.2012.06.061 (2012).
70 Basar, E. & Guntekin, B. A review of brain oscillations in cognitive disorders and the role of neurotransmitters.
Brain research 1235, 172-193, doi:10.1016/j.brainres.2008.06.103 (2008).
71 Mattia, M. & Del Giudice, P. Population dynamics of interacting spiking neurons. Physical review. E, Statistical,
nonlinear, and soft matter physics 66, 051917 (2002).
72 Knight, B. W., Omurtag, A. & Sirovich, L. The approach of a neuron population firing rate to a new equilibrium: an
exact theoretical result. Neural computation 12, 1045-1055 (2000).
73 Mainen, Z. F. & Sejnowski, T. J. Reliability of spike timing in neocortical neurons. Science 268, 1503-1506 (1995).
74 Brunel, N., Wang, X.J. What Determines the Frequency of Fast Network Oscillations With Irregular Neural
Discharges? I. Synaptic Dynamics and Excitation-Inhibition Balance. Journal of neurophysiology, 415-430,
doi:10.1152/jn.01095.2002 (2003).
75 Buzsaki, G. & Wang, X. J. Mechanisms of gamma oscillations. Annual review of neuroscience 35, 203-225,
76 Wang, X. J. Neurophysiological and computational principles of cortical rhythms in cognition. Physiol Rev 90,
doi:10.1146/annurev-neuro-062111-150444 (2012).
1195-1268, doi:10.1152/physrev.00035.2008 (2010).
77 Köndgen, H. et al. The dynamical response properties of neocortical neurons to temporally modulated noisy inputs
in vitro. Cerebral cortex 18, 2086-2097, doi:10.1093/cercor/bhm235 (2008).
78 Schoenenberger, P., Scharer, Y. P. & Oertner, T. G. Channelrhodopsin as a tool to investigate synaptic transmission
and plasticity. Experimental physiology 96, 34-39, doi:10.1113/expphysiol.2009.051219 (2011).
79 Zhang, Y. P. & Oertner, T. G. Optical induction of synaptic plasticity using a light-sensitive channel. Nat Methods 4,
139-141, doi:10.1038/nmeth988 (2007).
80 Tsodyks, M., Pawelzik, K. & Markram, H. Neural networks with dynamic synapses. Neural computation 10, 821-
81 Chance, F. S., Abbott, L. F. & Reyes, A. D. Gain modulation from background synaptic input. Neuron 35, 773-782
835 (1998).
(2002).
82 Klapoetke, N. C. et al. Independent optical excitation of distinct neural populations. Nat Methods 11, 338-346,
83 Hegemann, P. & Nagel, G. From channelrhodopsins
to optogenetics. EMBO Mol Med 5, 173-176,
84 Grubb, M. S. & Burrone, J. Channelrhodopsin-2 localised to the axon initial segment. PloS one 5, e13761,
doi:10.1038/nmeth.2836 (2014).
doi:10.1002/emmm.201202387 (2013).
doi:10.1371/journal.pone.0013761 (2010).
Authors' preprint
20/23
Photo-activated network responses in vitro
85 Wagenaar, D. A., Madhavan, R., Pine, J. & Potter, S. M. Controlling bursting in cortical cultures with closed-loop
multi-electrode stimulation. The Journal of neuroscience : the official journal of the Society for Neuroscience 25,
680-688, doi:10.1523/JNEUROSCI.4209-04.2005 (2005).
86 Fong, M. F., Newman, J. P., Potter, S. M. & Wenner, P. Upward synaptic scaling is dependent on neurotransmission
rather than spiking. Nat Commun 6, 6339, doi:10.1038/ncomms7339 (2015).
87 Newman, J. P. et al. Optogenetic feedback control of neural activity. Elife 4, doi:10.7554/eLife.07192 (2015).
88 Haroush, N., Marom, S. Slow dynamics in features of synchronized neural network responses. Frontiers in
computational neuroscience 9:40, doi:10.3389/fncom.2015.00040 (2015).
89 Droge, M. H., Gross, G. W., Hightower, M. H., & Czisny, L. E. Multielectrode analysis of coordinated, multisite,
rhythmic bursting in cultured cns monolayer networks. The Journal of neuroscience : the official journal of the
Society for Neuroscience 6, 1583–1592 (1986).
90 Jimbo, Y., Robinson, H. P., & Kawana, A. Simultaneous measurement of intracellular calcium and electrical activity
from patterned neural networks in culture. IEEE Transactions in Biomedical Engineering 40, 804–810. doi:
10.1109/10.238465 (1993).
91 Maeda, E., Robinson, H. P., & Kawana, A. The mechanisms of generation and propagation of synchronized bursting
in developing networks of cortical neurons. The Journal of neuroscience : the official journal of the Society for
Neuroscience 15, 6834–6845 (1995).
92 Shahaf, G., & Marom, S. Learning in networks of cortical neurons. The Journal of neuroscience : the official
journal of the Society for Neuroscience 21, 8782–8788 (2001).
93 Wagenaar, D. A., Pine, J., & Potter, S. M. An extremely rich repertoire of bursting patterns during the development
of cortical cultures. BMC Neuroscience 7:11. doi: 10.1186/1471-2202-7-11 (2006).
94 Ham, M. I., Bettencourt, L. M., McDaniel, F. D., & Gross, G. W. Spontaneous coordinated activity in cultured
networks: analysis of multiple ignition sites, primary circuits, and burst phase delay distributions. Journal of
Computational Neuroscience 24, 346–357. doi: 10.1007/s10827-007-0059-1 (2008).
95 Stegenga, J., le Feber, J., & Rutten, W. L. C. Changes within bursts during learning in dissociated neural networks.
Conf. Proc. IEEE Eng. Med. Biol. Soc. 2008, 4968–4971. doi: 10.1109/IEMBS.2008.4650329 (2008).
96 Einevoll, G. T., Kayser, C., Logothetis, N. K., & Panzeri, S. Modelling and analysis of local field potentials for
studying the function of cortical circuits. Nature Reviews Neuroscience 14, 770–785, doi:10.1038/nrn3599 (2013).
97 Waldert, S., Lemon, R. N., & Kraskov, A. Influence of spiking activity on cortical local field potentials. Journal of
Physiology 591.21: 5291–5303 (2013).
Supplementary Information
Authors' preprint
21/23
Photo-activated network responses in vitro
Figure S1. ChR2-LCTC current dynamics. Voltage-clamp experiments were performed under TTX in
neurons expressing ChR2-LCTC, in order to explore the inward current dynamics. Varying the stimulus
duration modulates the peak current amplitude in four representative neurons (A-D).
Authors' preprint
22/23
Photo-activated network responses in vitro
Figure S2. Light-evoked oscillatory activity is also reflected in single-trial raw voltage traces.
Reminiscent of network rhythm signatures in local field potentials in vivo or in brain slices, the impact of
brief (1msec) photoactivation was apparent also in single trials raw extracellular voltage recordings, low-
pass filtered below 110 cycles/sec. For each MEA microelectrode, filtered voltage traces recorded ~200ms
preceding (A) or following (B) the stimulus onset are displayed: the apparent transient increase in signal
variance, recorded at most microelectrodes, correspond to reverberating network-wide spiking activity (Figs.
2A, 3). Estimating the power spectrum of a sample low-pass filtered raw trace (C-D) reveals light-induced
oscillations in the same range as the spike PSTH (Fig. 3C), although with a much worst signal-to-noise ratio
(Fig. 3C).
Authors' preprint
23/23
|
1908.10101 | 2 | 1908 | 2019-08-28T09:44:21 | What is the dynamical regime of cerebral cortex? | [
"q-bio.NC"
] | Many studies have shown that the excitation and inhibition received by cortical neurons remain roughly balanced across many conditions. A key question for understanding the dynamical regime of cortex is the nature of this balancing. Theorists have shown that network dynamics can yield systematic cancellation of most of a neuron's excitatory input by inhibition. We review a wide range of evidence pointing to this cancellation occurring in a regime in which the balance is loose, meaning that the net input remaining after cancellation of excitation and inhibition is comparable in size to the factors that cancel, rather than tight, meaning that the net input is very small relative to the cancelling factors. This choice of regime has important implications for cortical functional responses, as we describe: loose balance, but not tight balance, can yield many nonlinear population behaviors seen in sensory cortical neurons, allow the presence of correlated variability, and yield decrease of that variability with increasing external stimulus drive as observed across multiple cortical areas. | q-bio.NC | q-bio |
What is the dynamical regime of cerebral cortex?
Yashar Ahmadian1,2 and Kenneth D. Miller3
1Institute of Neuroscience, Departments of Biology and Mathematics, University of Oregon, OR, USA
2As of March 2020: Computational and Biological Learning Lab, Department of Engineering, University of
Cambridge, Cambridge, UK
3Center for Theoretical Neuroscience, Swartz Program in Theoretical Neuroscience, Kavli Institute for Brain
Science, and Dept. of Neuroscience, College of Physicians and Surgeons and Morton B. Zuckerman Mind
Brain Behavior Institute, Columbia University, NY, USA
August 29, 2019
Summary
Many studies have shown that the excitation and inhibition received by cortical neurons remain roughly
balanced across many conditions. A key question for understanding the dynamical regime of cortex is the
nature of this balancing. Theorists have shown that network dynamics can yield systematic cancellation
of most of a neuron's excitatory input by inhibition. We review a wide range of evidence pointing to this
cancellation occurring in a regime in which the balance is loose, meaning that the net input remaining
a(cid:129)er cancellation of excitation and inhibition is comparable in size to the factors that cancel, rather than
tight, meaning that the net input is very small relative to the cancelling factors. This choice of regime has
important implications for cortical functional responses, as we describe: loose balance, but not tight bal-
ance, can yield many nonlinear population behaviors seen in sensory cortical neurons, allow the presence
of correlated variability, and yield decrease of that variability with increasing external stimulus drive as
observed across multiple cortical areas.
In what regime does cerebral cortex operate? This is a
fundamental question for understanding cerebral corti-
cal function. The concept of a "regime" can be defined in
various ways. Here we will focus on a definition in terms
of the balance of excitation and inhibition: how strong
are the excitation and inhibition that cortical cells receive,
and how tight is the balance between them? As we will
see, the answers to these questions have important im-
plications for the dynamical function of cortex.
We first consider several more fundamental distinctions
in cortical regime. First, neurons may fire in a regular
or irregular fashion, where regular firing refers to emit-
ting spikes in a more clock-like manner, while irregular
firing refers to emi(cid:130)ing spikes in a more random man-
ner, like a Poisson process. Cortex appears to be in an ir-
regular regime (Shadlen and Newsome, 1998; So(cid:129)ky and
Koch, 1993), though some areas are less irregular than
others (Maimon and Assad, 2009). Second, neurons may
fire in a synchronous regime, meaning with strong cor-
relations between the firing of di(cid:128)erent neurons, or an
asynchronous regime, meaning with weak (or no) corre-
lations. Cortical firing, particularly in the awake state,
generally appears to be in an asynchronous regime (Co-
hen and Kohn, 2011; Doiron et al., 2016; Ecker et al., 2014,
2010), although some conditions may show more syn-
chronous firing (DeWeese and Zador, 2006; Poulet and
Petersen, 2008; Stevens and Zador, 1998; Tan et al., 2014).
Thus, we will take cortex to be in an asynchronous irreg-
ular regime. Brunel (2000) first defined conditions on net-
works of excitatory and inhibitory neurons that led them
to operate in the asynchronous irregular regime.
1
A third distinction is whether a network goes to a stable
fixed rate of firing for a fixed input (with noisy fluctua-
tions about that fixed rate given noisy inputs), or whether
it shows more complex behaviors, such as movement be-
tween multiple fixed points, oscillations, or chaotic wan-
dering of firing rates. We will focus on the case of a
single fixed point, which seems likely to reasonably ap-
proximate at least awake sensory cortex (see discussion
in Miller, 2016). The fixed point is taken to be stable,
meaning that the network dynamics cause firing rates
to return to the fixed point levels a(cid:129)er small perturba-
tions. Finally, for a given fixed point, recurrent excitation
may be strong enough to destabilize the fixed point in
the absence of feedback inhibition; that is, if inhibitory
firing were held frozen at its fixed point level, a small per-
turbation of excitatory firing rates would cause them to
either grow very large or collapse to zero. In that case,
the fixed point is stabilized by feedback inhibition, and
the network is known as an inhibition-stabilized network
(ISN). Alternatively, the recurrent excitation may be weak
enough to remain stable even without feedback inhibi-
tion. A number of studies have found strong evidence
that at least primary visual and auditory cortices are ISNs
both at spontaneous (Sanzeni et al., 2019) and stimulus-
driven (Adesnik, 2017; Kato et al., 2017; Ozeki et al., 2009)
levels of activity.
Note that for some of the distinctions we describe be-
tween regimes, there is a sharp transition from one
regime to the other, while for others the transition is
gradual. We use the word "regime" in either case to de-
scribe qualitatively di(cid:128)erent network behaviors.
The assumption that cortex is in an irregularly-firing
regime (as well as its operation as an ISN) strongly points
to the need for some kind of balance between excita-
tion and inhibition. Stochasticity of cellular and synap-
tic mechanisms (Mainen and Sejnowski, 1995; O'Donnell
and van Rossum, 2014; Schneidman et al., 1998) and in-
put correlations (DeWeese and Zador, 2006; Stevens and
Zador, 1998) may contribute to irregular firing. How-
ever, a number of authors have argued that, assuming
inputs are un- or weakly-correlated, then irregular fir-
ing will arise if the mean input to cortical cells is sub-
or peri-threshold, so that firing is induced by fluctua-
tions from the mean rather than by the mean itself (Amit
and Brunel, 1997; Troyer and Miller, 1997; Tsodyks and
Sejnowski, 1995; van Vreeswijk and Sompolinsky, 1996).
This is referred to as the fluctuation-driven regime, as
opposed to the mean-driven regime in which the mean
input is strongly suprathreshold and spiking is largely
driven by integration of this mean input. The fluctuation-
driven regime yields random, Poisson-process-like firing,
because fluctuations are equally likely to occur at any
time, whereas the mean-driven regime yields regular fir-
ing.
Given the strength of inputs to cortex (to be discussed
below), the mean excitation received by a strongly-
responding cell is likely to be su(cid:128)icient to drive the cell
near or above threshold. Therefore, for the mean input to
be subthreshold, the mean inhibition is likely to cancel a
significant portion of the mean excitation; that is, the ex-
citation and inhibition received by a cortical cell should
be at least roughly balanced (Tsodyks and Sejnowski,
1995; van Vreeswijk and Sompolinsky, 1996). Consis-
tent with the idea that inhibition balances excitation,
many experimental investigations have suggested that
cortical or hippocampal excitation and inhibition remain
balanced or inhibition-dominated across varying activ-
ity levels (Anderson et al., 2000; Atallah and Scanziani,
2009; Barral and Reyes, 2016; Dehghani et al., 2016; Galar-
reta and Hestrin, 1998; Graupner and Reyes, 2013; Haider
et al., 2006, 2013; Higley and Contreras, 2006; Marino
et al., 2005; Okun and Lampl, 2008; Shu et al., 2003; Wehr
and Zador, 2003; Wu et al., 2008, 2006; Yizhar et al., 2011;
Zhou et al., 2014).
Excitation and inhibition may be balanced in at least two
ways. First, inhibitory and excitatory synaptic weights
may be co-tuned, so that cells that receive more (or less)
excitatory weight receive correspondingly more (or less)
inhibitory weight (Bhatia et al., 2019; Xue et al., 2014).
This does not ensure balancing of excitation and inhibi-
tion received across varying pa(cid:130)erns of activity. Second,
given su(cid:128)iciently strong feedback inhibitory weights, the
network dynamics may ensure that the mean inhibition
and mean excitation received by neurons remain bal-
anced across pa(cid:130)erns of activity, without requiring tun-
ing of synaptic weights. Here, we will focus on this sec-
ond, dynamic form of balancing.
As we will discuss, theorists have described mecha-
nisms by which inhibition and excitation dynamically re-
main balanced, keeping neurons in the fluctuation-driven
regime, without any need for fine tuning of parameters
such as synaptic weights. This dynamical balance can be
a "tight balance", which we define to mean that the exci-
tation and inhibition that cancel are much larger than the
residual input that remains a(cid:129)er cancellation, or a "loose
2
balance", meaning that the canceling inputs are compa-
rable in size to the remaining residual input (terms to de-
scribe balanced networks are not yet standardized; see
Appendix 1 for comparison of our usage to other nomen-
clatures). The question of whether the balance is tight or
loose has important implications for the behavior of the
network. Here we will review these issues and argue that
the evidence is most consistent with a loosely balanced
regime.
A Theoretical Problem: How to Achieve Input
Mean and Fluctuations That Are Both Compa-
rable in Size to Threshold?
How do cortical neurons stay in the irregularly fir-
ing regime? There are two requirements to be in the
fluctuation-driven regime, which yields irregular firing:
(1) The mean input the neurons receive must be sub- or
peri-threshold; (2) Input fluctuations must be su(cid:128)iciently
large to bring neuronal voltages to spiking threshold suf-
ficiently o(cid:129)en to create reasonable firing rates. We will
measure the voltage e(cid:128)ects of a neuron's inputs in units
of the voltage distance from the neuron's rest to thresh-
old; this distance, typically around 20 mV for a corti-
cal cell (e.g., Constantinople and Bruno, 2013, Fig. 3K), is
equal to 1 in these units. Thus, a necessary condition for
being in the irregularly firing regime is that the voltage
mean driven by the mean input (henceforth abbreviated
to "the mean input") should have order of magnitude 1,
which we write as O(1), or smaller. The second require-
ment above then dictates that the voltage fluctuations
driven by the input fluctuations from the mean (hence-
forth abbreviated to "input fluctuations") should also be
O(1). In particular, this means that the ratio of the mean
input to the input fluctuations should be O(1). (Note, we
use the O() notation simply to indicate order of magni-
tude, and not in its more technical, asymptotic sense of
the scaling with some parameter as that parameter goes
to zero or infinity.)
Several authors have considered the requirements for
these conditions to be true (Renart et al., 2010; Tsodyks
and Sejnowski, 1995; van Vreeswijk and Sompolinsky,
1996, 1998). Following these authors, we assume the net-
work is composed of excitatory (E) and inhibitory (I) neu-
ron populations, which receive excitatory inputs from an
external (X) population. The la(cid:130)er could represent any
cortical or subcortical neurons outside the local cortical
network, for example, the thalamic input to an area of pri-
mary sensory cortex. As a simplified toy model of the as-
sumption that the network is in the asynchronous irreg-
ular regime, we assume that the cortical cells fire as Pois-
son processes without any correlations between them, as
do the cells in the external population.
Suppose that a neuron receives KE excitatory inputs. Sup-
pose these inputs produce EPSPs that have an exponen-
tial time course, with mean amplitude JE and time con-
stant τE, and have mean rate rE. Then the mean depolar-
ization produced by these excitatory inputs is JEKErEτE.
Defining nE = rEτE to be the mean number of spikes of an
input in time τE, we find that the mean excitatory input
to the neuron is
µE = JEKEnE
(1)
Let σE denote the standard deviation (SD) of fluctuations
in this input. Assuming the spike counts of pre-synaptic
neurons are uncorrelated, their spike count variances just
add. Because they are firing as Poisson processes, the
variance in a neuron's spike count in time τE is equal to
its mean spike count nE. Thus, the variance of input from
one pre-synaptic neuron is J2
EnE, and so the variance in
the total input is KEJ2
EnE and
σE = JE
(2)
Therefore the ratio of the mean to the SD of the neuron's
excitatory input is
(cid:112)KEnE
=(cid:112)KEnE,
µE
σE
(3)
(4)
(5)
independent of JE. Similar reasoning about the neuron's
inhibitory or external input leads to all the same expres-
sions, except with E subscripts replaced with I or X sub-
scripts to represent quantities describing the inhibitory
or external input the cell receives.
Again assuming that the di(cid:128)erent populations are uncor-
related so that their contributed variances add, the total
or net input the neuron receives has mean, µ, and stan-
dard deviation, σ, given by:
(cid:113)
µ = µE + µX − µI
X + σ2
σ =
I
σ2
E + σ2
We imagine that KE and KI are the same order of magni-
tude, O(K) for some number K, and similarly nE and nI
are O(n) for some number n. We also assume µX and σX
are the same order of magnitude as µEorµI and σEorσI,
Kn is O(1), both µ and
respectively, or smaller. Then, if
σ can simultaneously be made O(1) with suitable choice
of the J's (generically, i.e. barring special cases in which
√
3
√
the elements of µ precisely cancel). This means that
the irregularly-firing regime is self-consistent: having as-
sumed that neurons are in the irregularly-firing regime,
we arrive at expressions for the mean and SD of their in-
put that indeed can keep the network in this regime.
Kn (cid:29) 1
If K is very large, however -- large enough that
-- then the ratio of the mean to the SD of each type of
input, and hence of the net input, is much greater than
1. Van Vreeswijk and Sompolinsky (1996,1998) consid-
ered the case of such very large K and showed how the
network could remain in the AI regime. They proposed
choosing the J's proportional to 1√
, so that the standard
K
deviations σE and σI are O(1) (Eq. 2), but then by Eq. 3 the
√
means µE and µI are large, O(
K). Then, for the neurons
to be in the asynchronous irregular regime, inhibitory in-
put, −µI, must cancel or "balance" a su(cid:128)icient portion of
the excitatory input, µE + µX, so that the mean input, µ,
is O(1).
If a neuron's mean excitatory and inhibitory inputs very
precisely cancel each other, so that the mean net input µ
is much smaller than either of these factors alone, we say
there is tight balance. If the net input is more comparable
in size to the factors that are cancelling, we call this loose
balance. The two cases may be distinguished by the size
of a dimensionless balance index β:
The Tightly Balanced Solution
K
Van Vreeswijk and Sompolinsky (1996,1998) showed that,
for very large K, and hence very large Kn, and all J's
∝ 1√
, the network dynamics would produce a tightly
balanced solution provided only that some mild (inequal-
ity) conditions on the weights are satisfied, without any
requirements for fine tuning. This is known more gen-
erally as the "balanced network" solution. To under-
stand this solution, we define the mean number of in-
puts, PSP amplitude, and time constant from population
B (B ∈ {E, I, X}) to a neuron in population A (A ∈ {E, I})
to be KAB, JAB and τAB respectively. We define the mean
e(cid:128)ective weight from population B to a neuron in popu-
lation A as WAB = JABKABτAB. Le(cid:130)ing rB be the average
firing rate of population B, then WABrB = JABKABnB, the
mean input from population B to population A. We as-
sume that WABrB = O(
K) for all A, B. The requirements
for balance are then that the mean net input to both ex-
citatory and inhibitory cells, uE and uI respectively, are
O(1), where (from Eqs. 1,4),
√
uE = WEErE − WEIrI + WEXrX
uI = WIErE − WIIrI + WIXrX
(7)
(8)
If we define the external inputs to the network IE =
WEXrX, II = WIXrX, then these equations can be wri(cid:130)en
as the vector equation
β =
µ
µE + µX
(6)
where u ≡
(cid:18) uE
uI
(cid:19)
, r ≡
(cid:19)
u = Wr + I
(cid:18) rE
(cid:18) IE
(cid:18) WEE −WEI
(cid:19)
rI
WIE −WII
, I ≡
II
.
(cid:19)
(9)
, and W is
K
Note that, using the above analysis,
in the limit of
large K considered by Van Vreeswijk and Sompolinsky
(1996,1998), β ∼ 1√
. Tight balance means that the bal-
ance index is very small, β (cid:28) 1; in loose balance, the
index is not so small, say 0.1 < β < 1, very roughly. As
we will see, whether the network is in tight or loose bal-
ance has important implications for the network's behav-
ior and computational ability. (Note that, in general, the
degree of balance can be di(cid:128)erent in di(cid:128)erent neurons in
the same network. In the above discussion we assumed
that di(cid:128)erent neurons of the same E/I type are statisti-
cally equivalent, e.g., in terms of the number and activity
of presynaptic inputs; this is the case in the randomly
connected network of Van Vreeswijk and Sompolinsky
(1996,1998). In that case β would not vary systematically
between neurons of the same type.)
4
the weight matrix W =
The balanced network solution arises by noting that the
le(cid:129) side of Eq. 9 is very small (O(1)) relative to the indi-
√
vidual terms on the right (O(
K)). So we first find an
approximate solution r0 to Eq. 9 in which the small le(cid:129)
side is replaced by 0 to yield the equation for perfect bal-
ance, i.e. all inputs perfectly cancelling: Wr0 + I = 0, or
r0 = −W−1I, where W−1 is the matrix inverse of W.
Note that r0 is O(1), because the elements of W and
I are all the same order of magnitude, so their ratio is
generically O(1). We can then write r as an expansion
+ …, where r0, r1, … are
in powers of
all O(1), to obtain a consistent solution: u = Wr1√
+ …
where the first term on the right is O(1), as desired, and
the remaining terms are very small (O(cid:16) 1√
(cid:17) or smaller).
, r = r0 + r1√
K
1√
K
K
K
The authors showed that, with some mild general con-
ditions on the weights W and inputs I, this tightly bal-
anced solution would be the unique stable solution of
the network dynamics. That is, for a given fixed input
I, the network's excitatory/inhibitory dynamics will lead
it to flow to this balanced solution for the mean rates:
r = −W−1I + O(cid:16) 1√
(cid:17).
K
We immediately see two points about the tightly bal-
√
anced (
Kn (cid:29) 1) solution:
1. Mean population responses are linear in the inputs.
−W−1I is a linear function of the input I. Tight bal-
ance implies that nonlinear corrections to r ≈ r0 =
−W−1I are very small relative to this linear term,
except for very small external inputs, so mean re-
sponse r is for practical purposes a linear function
of the input.
2. External input must be large relative to the net input
and to the distance from rest to threshold. The exter-
nal input I must have the same order of magnitude
as the recurrent input Wr0, so that balance can oc-
cur, Wr0 = −I, with rates that are O(1). If I were
smaller, the firing rates r ≈ r0 would correspond-
ingly be unrealistically small.
In the above treatment we focused on population-
averaged responses, rE and rI. We emphasize that the
balancing only applies to the mean input across neurons
of each type, and leaves una(cid:128)ected input components
with mean of zero across a given type; while the mean
input is very large in the tightly balanced regime, zero-
mean input components can be O(1) and yet elicit O(1)
responses in individual neurons (see e.g. (Hansel and van
Vreeswijk, 2012; Pehlevan and Sompolinsky, 2014; Sadeh
and Ro(cid:130)er, 2015)). Furthermore, even in the tightly bal-
anced regime, individual neurons can exhibit nonlinear-
ities in their responses, but these are washed out at the
level of population-averaged responses. We also note that
synaptic nonlinearities, e.g. synaptic depression, which
were neglected here, can allow a tightly balanced state
with nonlinear population-averaged responses (Mongillo
et al., 2012).
A Loosely Balanced Regime
As we have seen, if Kn is O(1), the mean and the fluc-
tuations of the input that neurons receive can both be
5
O(1) without requiring any balancing. Given that there is
both excitatory and inhibitory input, there will always be
some input cancellation or "balancing" -- some portion of
the input excitation will be cancelled by input inhibition,
leaving some smaller net input. When Kn is O(1), all of
these quantities -- the excitatory input, the inhibitory in-
put, and the net input a(cid:129)er cancellation -- will generically
be O(1), and thus balancing is "loose" -- the factors that
cancel and the net input a(cid:129)er cancellation are of compa-
rable size, and the balance index β is not small.
However, the fact that there is some inhibition that can-
cels some excitation does not by itself imply interesting
consequences for network behavior. We will use the term
"loosely balanced solution" to refer more specifically to
a solution having two features: (1) the dynamics yields
a systematic cancellation of excitation by inhibition like
that in the tightly balanced solution. In particular, in the
loosely balanced networks on which we will focus, a sig-
nature of this systematic cancellation is that the net in-
put a neuron receives grows sublinearly as a function of
its external input (we will make this more precise below);
(2) this cancellation is "loose", as just described. As we
will discuss, such a loosely balanced solution produces
various specific nonlinear network behaviors that are ob-
served in cortex.
Ahmadian et al. (2013) showed that such a loosely bal-
anced solution would naturally arise from E/I dynam-
ics for recurrent weights and external inputs that are
not large, provided that the neuronal input/output func-
tion, determining firing rate vs. input level, is supralinear
(having ever-increasing slope) over the neuron's dynamic
range. They modeled this supralinear input/output func-
tion as a power law with power greater than 1 (Fig. 1).
Such a power-law input-output function is theoretically
expected for a spiking neuron when firing is induced by
input fluctuations rather than the input mean (Hansel
and van Vreeswijk, 2002; Miller and Troyer, 2002), and is
observed in intracellular recordings over the full dynamic
range of neurons in primary visual cortex (V1) (Priebe
and Ferster, 2008). Of course, a neuron's input/output
function must ultimately saturate but, at least in V1, the
neurons do not reach the saturating portion of their in-
put/output function under normal operation. For the
loosely balanced solution to arise, some general condi-
tions on the weight matrix, similar to those for the tightly
balanced network solution but less restrictive, must also
be satisfied.
postsynaptic neuron's gain, hence the e(cid:128)ective synaptic
strengths increase with increasing gains.
The increasing e(cid:128)ective synaptic strengths lead to two
regimes of network operation.
For very weak exter-
nal drive and thus weak network activation, all e(cid:128)ective
synaptic strengths are very weak, for both externally-
driven and network-driven synapses. External drive to
a neuron is delivered monosynaptically, via the weak
externally-driven synapses.
In contrast, assuming that
the network is inactive in the absence of external in-
put, network drive involves a chain of two or more weak
synapses: the weak externally driven synapses activate
cortical cells, which then drive the weak network-driven
synapses. From the same principle that x2 (cid:28) x when
x (cid:28) 1, the network drive is therefore much weaker than
the external drive. Thus, the input to neurons is dom-
inated by the external input, with only relatively small
contributions from recurrent network input. In sum, in
this weakly-activated regime, the neurons are weakly cou-
pled, largely responding directly to their external input
with li(cid:130)le modification by the local network.
With increasing external (stimulus) drive and thus in-
creasing network activation, the gains and thus the ef-
fective synaptic strengths grow. This causes the relative
contribution of network drive to grow until the network
drive is the dominant input. At some point, the e(cid:128)ective
E → E connections become strong enough that the net-
work would be prone to instability -- a small upward fluc-
tuation of excitatory activities would recruit su(cid:128)icient re-
current excitation to drive excitatory rates still higher,
which if unchecked would lead to runaway, epileptic ac-
tivity (and to ever-growing e(cid:128)ective synaptic strengths
and thus ever-more-powerful instability). However, if
feedback inhibition is strong and fast enough, the inhi-
bition will stabilize the network, that is, it becomes an
ISN. This stabilization is achieved by a loose balancing of
excitation and inhibition, as we will explain in more de-
tail below. Thus, in this more strongly-activated regime,
the neurons are strongly coupled and are loosely balanced.
Note that the input driving spontaneous activity may al-
ready be strong enough to obscure the weakly coupled
regime, as suggested by the finding that V1 under spon-
taneous activity is already an ISN (Sanzeni et al., 2019).
As in the tightly balanced network, the network's exci-
tatory/inhibitory dynamics lead it to flow to this loosely
balanced solution, without any need for fine tuning of pa-
rameters. Because this mechanism involves stabilization,
by inhibition, of the instability induced by the supralinear
Figure 1. The supralinear (power-law) neuronal
transfer function. The transfer function of neurons in
cat V1 is non-saturating in the natural dynamic range
of their inputs and outputs, and is well fit by a supralin-
ear rectified power-law with exponents empirically found
to be in the range 2-5. Such a curve exhibits increasing
input-output gain (i.e. slope, indicated by red lines) with
growing inputs, or equivalently with increasing output
firing rates. Gray points indicate a studied neuron's aver-
age membrane potential and firing rate in 30ms bins; blue
points are averages for di(cid:128)erent voltage bins; and black
line is fit of power law, r = [V − θ]p
+, where r is firing rate,
V is voltage, [x]+ = x, x > 0, =0 otherwise; θ is a fi(cid:130)ed
threshold; and p, the fi(cid:130)ed exponent, here is 2.79. Figure
modified from (Priebe et al., 2004).
In the presence of a supralinear input/output function,
the loosely balanced solution arises as follows. Whereas
previously we considered the e(cid:128)ects of increasing K when
√
recurrent and external inputs were all O(
K), now we
consider the more biological case of increasing external
input (i.e., stimulus) strength while recurrent weights are
at some fixed level. The supralinear input/output func-
tion means that a neuron's gain -- its change in output
for a given change in input -- is continuously increas-
ing with its activation level. This in turn means that ef-
fective synaptic strengths are increasing with increasing
network activation. The e(cid:128)ective synaptic strength mea-
sures the change in the postsynaptic cell's firing rate per
change in presynaptic firing rate. This is the product of
the actual synaptic strength -- the change in postsynaptic
input induced by a change in presynaptic firing -- and the
6
Figure 2. Loose vs. tight balance. We illustrate the external (blue), recurrent or within-network (green) and net
(orange, equal to external plus recurrent) inputs to a typical excitatory cell in a Stabilized Supralinear Network. At all
biological ranges of external input (stimulus) strength, the balance is loose, as exhibited by the le(cid:129) set of three arrows
(representing the external, recurrent and net inputs): the net input is comparable in size to the other two. Nevertheless
the balance systematically tightens with increasing external input (right set of arrows), as the net input grows only
sublinearly with growing external input strength. At very high (possibly non-biological) levels of external input, the
balance can become very tight, with the net input much smaller in magnitude than the external and recurrent inputs.
input/output function of individual neurons along with
E → E connections, it has been termed the Stabilized
Supralinear Network (SSN) (Ahmadian et al., 2013; Rubin
et al., 2015).
To describe the mathematics of this mechanism, we again
consider an excitatory and an inhibitory population along
with external input. We define the vectors r, u and I and
the matrix W as before. Then the power-law input/output
function means that the network's steady state firing rate
rSS for a steady input I satisfies
rSS = k(u)p
+ = k(WrSS + I)p
+
(10)
7
where (v)+ is the vector v with negative elements set to
zero, (v)p
+ means raising each element of (v)+ to the power
p, p is a number greater than 1 (typically, 2 to 5, Priebe
and Ferster, 2008), and k is a constant with units Hz
(mV )p
(and the units of W, r, and I are mV
, Hz, and mV respec-
Hz
tively). It is convenient to absorb k into e(cid:128)ective weights
and inputs by writing(cid:102)W = k1/pW,(cid:101)I = k1/pI, so the equa-
If we let ψ = (cid:107)(cid:102)W(cid:107) represent a norm of(cid:102)W (think of it
neuron), and similarly let c = (cid:107)(cid:101)I(cid:107) represent a typical in-
as a typical total E or I recurrent weight received by a
(11)
tion becomes
rSS = ((cid:102)WrSS +(cid:101)I)p
+
1020304050External input (mV)402002040Inputs to E cell (mV)externalrecurrentnet put strength, then it turns out the network regime is con-
trolled by the dimensionless parameter
α = ψcp−1.
(12)
α1/p
(cid:16) r1
+ …(cid:17)
As we increase the strength of external drive and thus
of network activation, c and thus α is increasing. When
α (cid:28) 1, the network is in the weakly coupled regime; for
α (cid:29) 1, the network is in the strongly coupled regime; and
the transition between regimes generically occurs when
α is O(1) (Ahmadian et al., 2013).
The loosely-balanced solution then turns out to be of the
form
r = −W−1I + c
ψ
(13)
where r1 is dimensionless and O(1), and the higher-order
(see
terms (indicated by …) involve higher powers of
Appendix 2). Equation 13 is precisely the same equa-
in the case
tion as for the tightly balanced solution,
In the
that the input/output function is a power law.
√
tightly balanced network, ψ and c are both O(
K), so
α is O((K)p/2), i.e. very large, and the
in the sec-
(cid:17), as expected. The loosely-
ond term becomes O(cid:16) 1√
balanced solution arises, however, when α is O(1). In par-
ticular, in the biological case of fixed weights but increas-
ing stimulus drive, and given the supralinear neuronal in-
put/output functions, the same E/I dynamics that lead to
the tightly balanced solution when inputs are very large
will already yield a loosely balanced solution when inputs
areO(1). The conditions for this loosely balanced solution
to arise are further discussed in Appendix 2.
1
α1/p
1
α1/p
K
(cid:18)(cid:16) c
(cid:17)1/p(cid:19)
(cid:17) = O
The fact that the solution is loosely balanced can be seen
by computing the balance index, β (Eq. 6). The network
excitatory drive is O(ψr), the external drive is O(c), and
because the first term on the right side of Eq. 13 can-
cels the external input, the net input a(cid:129)er balancing is
W (which is O(ψ)) times the 2nd term on the right side
. Since 1/p < 1,
the net input thus grows sublinearly with growing exter-
nal input strength, c, as illustrated in Fig. 2. Moreover, it
of Eq. 13, or O(cid:16) c
follows that the balance index (Eq. 6) is O(cid:16) c/α1/p
is O(cid:16) 1
anced solution this is very small, O(cid:16) 1√
(cid:17) which
(cid:17) (assuming the order of magnitude of the re-
(cid:17), but the loosely
current input, ψr, is the same as or smaller than that of
the external input strength, c). Again, for the tightly bal-
c+ψr
α1/p
α1/p
ψ
balanced solution arises when this is O(1).
K
8
In more complex models (involving many neurons with
structured connectivity and stimulus selectivity), loosely-
balanced solutions still arise when α is O(1). That is, even
in such cases the full nonlinear steady state equations,
Eq. (10), can yield biologically plausible solutions, and
when that happens the net inputs to activated neurons
grow sublinearly with growing external input strength,
and balance indices are O(1). The case of structured net-
works with stimulus selectivity is further discussed in Ap-
pendix 2.
We can now see that the loosely balanced regime di(cid:128)ers
from the tightly balanced in the two points summarized
previously:
1. In the loosely balanced regime, mean population re-
sponses are nonlinear in the inputs. This is because,
when balance is loose, the second term in Eq. 13,
which is not linear in the input, cannot be ne-
glected relative to the first, linear term. In particu-
lar, the nonlinear population behaviors observed in
the loosely balanced regime with a supralinear in-
put/output function closely match the specific non-
linear behaviors observed in sensory cortex (Rubin
et al., 2015), as we will discuss below.
2. In the loosely balanced regime, external input can be
comparable to the net input and to the distance from
rest to threshold.
What Regime Do Experimental Measurements
Suggest?
As we have seen above, the same model can give a loosely
balanced solution (Eq. 13) when α is O(1) (e.g., when c
and ψ are both O(1)), but give a tightly balanced solu-
tion when α is large (e.g., when c and ψ are both O(
K)).
Which of these regimes is supported by experimental
measurements?
√
Measurements of Biological Parameters
√
Kn? We saw in Eq. 3 that the ratio of
How large is
the mean to the standard deviation, µY/σY, of the input
of type Y (Y ∈ {E, I, X}) received by a neuron is equal
to √
KYnY, where KY is the number of inputs of type Y
a given neuron receives and nY is the average number of
spikes one of these inputs fires in a PSP decay time τy
KE = 200 KE = 1000 KE = 5000
1.4
4.5
14.1
3.2
10.0
31.6
rE = 0.1 Hz
rE = 1 Hz
rE = 10 Hz
0.6
2.0
6.3
Table 1. Values of √
KEnE for varying KE and rE, for τ = 20ms.
KEnE.
(cid:113)(cid:104)J2
(nY = rYτy, where rY is the average firing rate of one of
these inputs). Here we estimate √
Note that √
KYnY is actually an upper bound for the ratio
µY/σY for a given input type, because we have neglected
a number of factors that would increase fluctuations for
a given mean. These include (i) correlations among in-
puts which, even if weak, can significantly boost the input
SD, σY, without altering input mean; (ii) variance in the
weights, JY, which would increase the estimate of σY by a
Y(cid:105) /(cid:104)JY(cid:105)2; and (iii) network e(cid:128)ects that can am-
factor
plify input variances by creating firing rate fluctuations,
although this amplification may be small for strong stim-
uli (Hennequin et al., 2018). Furthermore, the ratio µ/σ of
total input is expected to be smaller than the ratio µY/σY
for any single type. This is because σ2 involves the sum
of three variances (Eq. 5), while µ involves a di(cid:128)erence of
one mean from the sum of two others (Eq. 4), represent-
ing the e(cid:128)ect of loose balancing.
Given these considerations, we are primarily concerned
with estimating the overall magnitude of √
KEnE rather
than detailed values.
If this magnitude is very much
larger than the observed µ/σ ratio in vivo, then tight bal-
ancing may be needed to explain the in vivo ratio. To esti-
mate the in vivo µ/σ ratio, we note that, in anesthetized
cat V1, σ varies from 1 to 7 mV and µ ranges from 0 to 15
mV (20 mV in one case) for a strong stimulus (Finn et al.,
2007; Sadagopan and Ferster, 2012). While these authors
did not give the paired µ and σ values for individual cells,
it seems reasonable to guess from these values that the
value of µ/σ for the total input to these cells is generally
in the range 0−15. Finn et al. (2007) also reported that, at
the peak of a simple cell's voltage modulation to a high-
contrast dri(cid:129)ing grating, the ratio σ/µ had an average
value of about 0.15 (here, we are taking µ to be the mean
voltage at the peak). This suggests that the average value
of µ/σ at peak activation is around 1/0.15 = 6.7.
How large is √
KEnE? In a study of input to excitatory
cells in layer 4 of rat S1 (Schoonover et al., 2014), the
EPSP decay time τE was around 20 ms. From 1800 to
4000 non-thalamic-recipient spines were found on stud-
KEnE ranges from 3.3√
ied cells which, with an estimated average of 3.4 synapses
per connection between layer 4 cells (Feldmeyer et al.,
1999), corresponds to a KE -- the number of other corti-
cal cells providing input to one cell -- of 530 to 1200. If
rE is expressed in Hz, then √
rE to
4.9√
rE. Thus, even if average input firing rates were 10
Hz, which would be very high for rodent S1 (Barth and
Poulet, 2012) (note that the average is over all inputs, not
just those that are well driven in a given situation, and
so is likely far below the rate of a well-driven neuron),
these ratios would be 10.4−15.5. For more realistic rates
of 0.1−1Hz (Barth and Poulet, 2012), these ratios would
be 1.0−4.9. All of these are comparable in magnitude to
observed in vivo levels of µ/σ.
More generally, estimates across species and cortical ar-
eas of the number of spines on excitatory cells, and thus
of the number of excitatory synapses they receive, range
from 700 to 16,000, with numbers increasing from pri-
mary sensory to higher sensory to frontal cortices (Ama-
trudo et al., 2012; Elston, 2003; Elston and Fujita, 2014; El-
ston and Manger, 2014). Estimates of the mean number of
synapses per connection between excitatory cells range
from 3.4 to 5.5 across di(cid:128)erent areas and layers studied
(Fares and Stepanyants, 2009; Feldmeyer et al., 1999, 2002;
Markram et al., 1997). These numbers yield a KE of 130
to 4700. In Table 1, we show the value of √
KEnE for KE
ranging from 200 to 5000 (rounded upward to bias results
most in favor of a need for tight balancing) and for rates
rE of 0.1 to 10 Hz. The results are all comparable to the
µ/σ's observed in vivo, except for the most extreme case
considered (KE = 5000, rE = 10Hz), and even that case
is only o(cid:128) by a factor of 2. Thus, the numbers strongly
argue that tight balancing is not needed for the ratio of
voltage mean to variance to have values as observed in
vivo.
External input is comparable in strength to net in-
put. Several studies have silenced cortical firing while
recording intracellularly to determine the strength of ex-
ternal input, with cortex silenced, relative to the net in-
put with cortex intact. These find the external input to
9
be comparable to the net input, consistent with the loose
balance scenario, rather than much larger as the tight bal-
ance scenario requires.
Ferster et al. (1996) cooled V1 and surrounding V2 in anes-
thetized cats to block spiking of almost all cortical cells,
both excitatory and inhibitory, leaving axonal transmis-
sion (e.g. of thalamic inputs) intact, though with weak-
ened release. By measuring the size of EPSPs evoked by
electrical stimulation of the thalamic lateral geniculate
nucleus (LGN) in thalamic-recipient cells in layer 4 of V1,
they could estimate the degree of voltage a(cid:130)enuation of
EPSPs induced by cooling. Correcting for this a(cid:130)enua-
tion, they estimated that the first harmonic voltage re-
sponse to an optimal dri(cid:129)ing luminance grating stimulus
of a layer 4 V1 cell was, on average, about 1/3 as great
with cortex cooled as with cortex intact, suggesting that
the external input to cortex is smaller than the net input
with cortex intact. Chung and Ferster (1998) and Finn
et al. (2007) assayed the same question by using cortical
shock to silence the local cortex for about 150 ms, during
which time the voltage response to an optimal flashed
grating was measured. They found that on average the
transient voltage response in layer 4 cells with cortex si-
lenced was about 1/2 the size of that with cortex intact
(Chung and Ferster, 1998), and more generally ranged
from 0% to 100% of the intact cortical response (Finn et al.,
2007). This again suggests that the external input to cor-
tex is smaller than the net input, i.e. the external input is
O(1), consistent with loose but not tight balance.
Total excitatory or inhibitory conductance is com-
parable to threshold. The above results suggest that
depolarization due to thalamus alone is less than that in-
duced by the combination of thalamic and cortical ex-
citation plus cortical inhibition, i.e. a(cid:129)er cortical "bal-
ancing" has occurred. One can also ask what propor-
tion of the total excitation is provided by thalamus. This
has been addressed in voltage-clamp recordings in anes-
thetized mice by silencing cortex through light activa-
tion of parvalbumin-expressing inhibitory cells express-
ing channelrhodopsin.
In layer 4 cells of V1 (Li et al.,
2013b; Lien and Scanziani, 2013) and primary auditory
cortex (A1) (Li et al., 2013a), mean stimulus-evoked ex-
citatory conductance with cortex silenced was estimated
to be 30-40% of that with cortex intact.
This tells us that the external and cortical contributions
to excitation are comparable. How large are they com-
pared to the excitation needed to depolarize the cell from
rest to threshold, which is typically a distance of about
20 mV (Constantinople and Bruno, 2013)? With cortical
spiking intact, these authors (Li et al., 2013a; Lien and
Scanziani, 2013) found mean stimulus-evoked peak exci-
tatory currents ranging from 60 to 150pA for various stim-
uli. Even assuming a membrane resistance of 200 MΩ,
which seems on the high end for in vivo recordings (Li
et al. (2013a) reported input resistances of 150−200 MΩ),
these would induce depolarizations of 12 to 30 mV; that
is, the total excitatory current is comparable to threshold,
i.e. it is O(1).
A similar result can be found from decomposition of exci-
tatory and inhibitory conductances from current-clamp
recordings at varying hyperpolarizing current levels. In
cat V1 cells for an optimal visual stimulus, one finds that
peak excitatory and inhibitory stimulus-induced conduc-
tances, gE and gI respectively, are typically < 10nS and
almost always < 20nS, on top of stimulus-independent
conductances (gL, for leak conductance) around 10nS
(Anderson et al., 2000; Ozeki et al., 2009). A study
of response to whisker stimulation in rat barrel cortex
found excitatory and inhibitory conductances of ≤ 5ns
(Lankarany et al., 2016). The depolarization that the
stimulus-induced excitatory conductance would induce
by itself is
VE, where VE is the driving potential of ex-
citatory conductance, about 50 mV at spike threshold of
around −50 mV (e.g. Wilent and Contreras, 2005). Using
the cat V1 numbers, this means that the depolarization
driven by excitatory conductance is typically < 25 mV
and almost always < 33 mV. Hyperpolarization driven
by the inhibitory conductance alone would be 0.4 to 0.6
times these values, given inhibitory driving force of −20
to −30 mV at spike threshold. These values are all quite
comparable to the distance from rest to threshold, ∼ 20
mV, that is, they are O(1).
How large is the expected mean excitatory input?
We have seen that the expected mean depolarization in-
duced by recurrent excitation is JEKEnE, where JE is the
mean EPSP amplitude. Based on the measurements of
Lien and Scanziani (2013) and Li et al. (2013b), discussed
above, total excitation may be about 1.5 times greater
than recurrent excitation. JE can be di(cid:128)icult to estimate,
because some of the KE anatomical synapses may be very
weak and not sampled in physiology, and because synap-
tic failures, depression, or facilitation can alter average
EPSP size relative to measured EPSP sizes. Furthermore,
measurements are variable, for example JE for layer 4 to
layer 4 connections in rodent barrel cortex has been es-
gE
gE +gL
10
√
√
√
√
Kn > 10 − 14 (compare values of
timated to be 1.6 mV in vitro (Feldmeyer et al., 1999) vs.
0.66 mV in vivo (Schoonover et al., 2014). If we assume
typical values for JE of 0.5 − 1 mV, then 1.5JEKEnE would
Kn > 7 − 10 and exceed 150 mV
exceed 75 mV for
for
Kn in Table 1).
We can very roughly guess that neural responses may be-
come be(cid:130)er described by tight rather than loose balance
somewhere in this range of mean excitatory input (and
corresponding
Kn). While the measurements of exci-
tatory currents and conductances described above argue
that such a range is not reached in primary sensory cor-
tex, it could conceivably be reached (Table 1) in areas with
higher KE, e.g. frontal cortex.
Nonlinear Behaviors
Sensory cortical neuronal responses display a variety of
nonlinear behaviors that, as we'll describe, are expected
from the SSN loosely balanced regime but not from the
tightly balanced regime. Many of these nonlinearities
are o(cid:129)en summarized as "normalization" (Carandini and
Heeger, 2012), meaning that responses can be fit by a phe-
nomenological model of an unnormalized response that
is divided by (normalized by) some function of all of the
unnormalized responses of all the neurons within some
region. To describe these nonlinear behaviors, we must
first define the classical receptive field (CRF): the local-
ized region of sensory space in which appropriate stimuli
can drive a neuron's response.
One nonlinear property is sublinear response summation:
across many cortical areas, the response to two stimuli si-
multaneously presented in the CRF is less than the sum of
the responses to the individual stimuli, and is o(cid:129)en closer
to the average than the sum of the individual responses
(reviewed in Carandini and Heeger, 2012; Reynolds and
Chelazzi, 2004). An additional nonlinearity is that the
form of the summation changes with the strength of the
stimulus: summation becomes linear for weaker stimuli
(Heuer and Bri(cid:130)en, 2002; Ohshiro et al., 2013). It is of-
ten di(cid:128)icult to determine if such nonlinear behaviors are
computed in the recorded area or involve changes in the
inputs to that area. However, some recent experiments
studied summation of response to an optogenetic and a
visual stimulus, a case in which the inputs driven by each
stimulus should not alter those driven by the other. Sub-
linear summation of responses to these stimuli was found
(Nassi et al., 2015; Wang et al., 2019, but see Histed, 2018),
which became linear for weak stimuli (Wang et al., 2019).
11
Another set of nonlinearities involve interaction of a CRF
stimulus and a "surround" stimulus, which is located out-
side the CRF. Across many cortical areas, surround stim-
uli can suppress response to a CRF stimulus ("surround
suppression"; reviewed in Angelucci et al., 2017; Rubin
et al., 2015), but this e(cid:128)ect varies with stimulus strength.
When the center stimulus is weak, a surround stimulus
can facilitate rather than suppress response (Ichida et al.,
2007; Polat et al., 1998; Sato et al., 2014; Schwabe et al.,
2010; Sengpiel et al., 1997). Similarly, the summation field
size -- the size of a stimulus that elicits strongest response,
before further expansion yields surround suppression -- is
largest for weak stimuli and shrinks with increasing stim-
ulus strength (Anderson et al., 2001; Cavanaugh et al.,
2002; Nienborg et al., 2013; Sceniak et al., 1999; Shushruth
et al., 2009; Song and Li, 2008; Tsui and Pack, 2011). The
summation field size in feature space -- the optimal range
of simultaneously presented motion directions in mon-
key area MT -- similarly shrinks with increasing stimulus
strength (Liu et al., 2018).
Additional nonlinearities include a decrease, with in-
creasing stimulus strength, in the ratio of excitation to
inhibition received by neurons (Adesnik, 2017) and in the
wavelength of a characteristic spatial oscillation of activ-
ity (Rubin et al., 2015).
All of these nonlinear cortical response properties, and
more, follow naturally (Ahmadian et al., 2013; Rubin et al.,
2015) from the two regimes of the loosely balanced sce-
nario with a supralinear input/output function, along
with simple assumptions on connectivity (e.g. that con-
nections decrease in strength and/or probability with
spatial distance, e.g. Markov et al., 2011, or with di(cid:128)er-
ence in preferred features, e.g. Cossell et al., 2015; Ko et al.,
2011). In contrast, as described previously, the tightly bal-
anced scenario causes population-averaged responses to
be linear responses to the input (individual neurons, but
not the population average, may have nonlinear behav-
iors), and thus appears inconsistent with these nonlin-
ear cortical behaviors, which in most cases are consis-
tent enough across neurons that they should character-
ize the mean population response. Synaptic nonlineari-
ties can give nonlinear population-averaged behavior in
the tightly balanced regime (Mongillo et al., 2012), but
it has not been claimed or demonstrated that this could
produce specific nonlinearities like those seen in cortical
responses.
Correlations and Variability
Across many cortical systems, the correlated component
of neuronal variability is decreased when a stimulus is
given, with variability decrease seen both in neurons that
respond to the stimulus and those that don't respond
(Churchland et al., 2010). This is also naturally explained
by the loosely balanced SSN network (Hennequin et al.,
2018).
In the strongly coupled regime of the loosely
balanced SSN network, increasing stimulus strength in-
creases the strength with which correlated pa(cid:130)erns of ac-
tivity inhibit themselves, thus damping their responses
to input fluctuations. The tightly balanced state repre-
sents the end state of this process -- a fully asynchronous
regime in which correlations are completely suppressed
(Renart et al., 2010; van Vreeswijk and Sompolinsky, 1998;
with dense connectivity, the mean correlation is propor-
tional to 1/K, and the standard deviation of the distribu-
tion of correlations is proportional to 1/
K Renart et al.
(2010); recall that K is meant to be a very large number
to achieve the tightly balanced state). Thus, the tightly
balanced state appears incompatible with the observed
decrease in correlated variability induced by a stimulus,
because the state has essentially no correlated variabil-
ity. However, it should be noted that variants of the
tightly balanced network involving structured connectiv-
ity can yield finite correlated variability among prefer-
entially connected neurons while maintaining tight bal-
ance, although average correlation over all neuron pairs
can still go to zero with increasing
K (Litwin-Kumar
et al., 2012; Rosenbaum et al., 2017).
√
√
Discussion
We have seen that many independent lines of evidence
are all consistent with cortex being in a loosely balanced
regime, and are inconsistent with tight balance. We de-
fine balance to mean that the dynamics yields a system-
atic cancellation of excitation by inhibition; a signature
of this for the loosely balanced scenario that we con-
sider is that the net input a neuron receives a(cid:129)er can-
cellation grows sublinearly as a function of its external
input. Loose balance means that the net input a(cid:129)er can-
cellation is comparable in size to the factors that cancel,
whereas tight balance means that the net input is very
small relative to the cancelling factors. In both cases, the
net input a(cid:129)er cancellation is comparable in size to the
distance from rest to threshold so that neuronal firing can
12
be in the fluctuation-driven regime that produces irregu-
lar firing like that observed in cortex.
One line of evidence for loose balance involves a vari-
ety of measurements on the numbers and/or strengths of
the inputs cells receive, including spine counts, strengths
of external and total input, and strengths of excitatory
and of inhibitory input. These measurements show that
the expected ratio of mean to standard deviation of the
network input before any tight balancing is already con-
sistent with the ratios observed for a cell's net input as
judged by its voltage response. That is, tight cancella-
tion is not needed to achieve the ratios observed. These
measurements further show that external input and net-
work input are comparable in size to the net input re-
maining a(cid:129)er cancellation, and that they and the total
excitatory and total inhibitory input are all comparable to
the distance from rest to threshold, consistent with loose
but not tight balancing. Other lines of evidence include
a variety of nonlinear population response properties of
sensory cortical neurons, as well as the presence of corre-
lated variability in neural responses and its decrease upon
presentation of a stimulus, all of which emerge naturally
from loose balance with a supralinear input/output func-
tion, but appear largely incompatible with tight balance.
It should be emphasized that the number of excitatory
synapses received by an excitatory cell, KE, increases
from primary sensory to higher sensory to frontal cor-
tex (e.g. Elston, 2003). Higher numbers are expected to
push in the direction of tighter balance. The expected ra-
tio of input mean to standard deviation and the expected
size of the mean input both can become high enough to
potentially yield tight balance for the highest KE's, par-
ticularly if higher average firing rates rE are assumed.
Our other arguments depend largely, but not entirely, on
measurements from sensory cortex. The measurements
of net input and external input are all from primary sen-
sory cortex. The studied nonlinear response properties
are primarily from both lower and higher visual cortices
(reviewed in Rubin et al., 2015). Suppression of correlated
variability by a stimulus, however, has been observed in
frontal and parietal as well as sensory cortex (Churchland
et al., 2010). In sum, while the evidence strongly favors
loose balance in sensory cortex, the evidence as to the
regime of frontal cortex is weaker.
The seminal discovery of the tightly balanced network
(van Vreeswijk and Sompolinsky, 1996, 1998) solved a key
problem in theoretical neuroscience: how can neurons re-
main in the fluctuation-driven regime, so that they have
irregular firing with reasonable firing rates, without re-
quiring fine tuning of parameters? The answer was that
when external and network inputs were very large, the
network's dynamics could robustly tightly balance the
excitation and inhibition that neurons receive, leaving a
net input a(cid:129)er cancellation that is negligibly small rela-
tive to the factors that cancel. This allows both the mean
and standard deviation of the net input to be comparable
to the distance from rest to threshold despite the very
large size assumed for the factors that cancel, yielding
the fluctuation-driven regime. This achievement along
with the model's mathematical tractability have made it
a popular model for the theoretical study of neural cir-
cuits. However, for all of the reasons stated above, this
tightly balanced regime does not seem to match observa-
tions of at least sensory cortical anatomy and physiology.
The loosely balanced solution shows that, when neuronal
input/output functions are supralinear, the same dynam-
ical balancing can arise from network dynamics without
fine tuning, but in a regime in which external and net-
work inputs are not large. Instead, the balancing arises
when these inputs, and the net input remaining a(cid:129)er can-
cellation of excitatory and inhibitory input, are all com-
parable in size to one another and to the distance from
rest to threshold. Furthermore, for weak inputs this same
scenario produces a weakly-coupled, feedforward-driven
regime which explains the observation that summation
changes from sublinear, or suppressive, for stronger stim-
uli to linear, or facilitative, for weak inputs.
√
The tightly balanced network demonstrated that a net-
work could self-consistently generate its own variability.
As we showed in the section "How large is
Kn?", the
loosely balanced regime can also generate realistic levels
of variability. However, biologically there is no need for
the network to generate all of its own variability, as all
inputs to cortex are noisy (and there are other sources of
noise such as stochasticity of cellular and synaptic mech-
anisms (Mainen and Sejnowski, 1995; O'Donnell and van
Rossum, 2014; Schneidman et al., 1998) and input corre-
lations (DeWeese and Zador, 2006; Stevens and Zador,
1998)).
In at least one case (Sadagopan and Ferster,
2012), the noise derived from the cortical area's input was
shown to be large enough to potentially fully account for
the noise seen in the cortical neurons. In the SSN net-
work, the network will amplify input noise in the weakly
coupled regime, and then decrease noise for increasingly
strong inputs in the strongly coupled, loosely balanced
regime; the result is that, for higher input strengths, noise
can be reduced to the level driven by the inputs (e.g., see
Fig. 2D of Hennequin et al., 2018), consistent with the ob-
servations of Sadagopan and Ferster (2012).
In conclusion, we believe that at least sensory, and per-
haps all of, cortex operates in a regime in which the in-
hibition and excitation neurons receive are loosely bal-
anced. This along with the supralinear input/output
function of individual neurons and simple assumptions
on connectivity explains a large set of cortical response
properties. A key outstanding question is the computa-
tional function or functions of this loosely balanced state
and the response properties it creates (e.g., see Echeveste
et al., 2019; G. Barello and Y. Ahmadian, in preparation).
Appendix 1: Nomenclature for balanced solu-
tions
There is no standard nomenclature for describing bal-
anced solutions. Here we have used loose vs. tight bal-
ance to describe, given systematic cancellation, whether
the remainder a(cid:129)er cancellation is comparable to, or
much smaller than, the factors that cancel.
Deneve and Machens (2016) used loose balance to mean
that fast fluctuations of excitation and inhibition are un-
correlated, although they are balanced in the mean, as
in the sparsely-connected network of van Vreeswijk and
Sompolinsky (1998); and used tight balance to mean that
fast fluctuations of excitation and inhibition are tightly
correlated with a small temporal o(cid:128)set, as in the densely-
connected network of Renart et al. (2010) and in the
spiking networks of Deneve, Machens and colleagues in
which recurrent connectivity has been optimized for e(cid:128)i-
cient coding (Barre(cid:130) et al., 2013; Boerlin et al., 2013; Bour-
doukan et al., 2012). All of these networks are tightly bal-
anced under our definition.
Hennequin et al. (2017) also defined balance to be tight
if it occurs on fast timescales, and loose otherwise, but
they implied that this is equivalent to our definition, that
tight balance means the remainder is small compared to
the factors that cancel, and loose balance means the re-
mainder is comparable to the factors that cancel. The im-
plied equivalence rests on the fact that tight balance un-
√
der our definition produces very large (i.e., O(
K)) nega-
tive eigenvalues (in linearization about the fixed point)
which means very fast dynamics, approaching instan-
taneous population response as K and hence negative
13
eigenvalues go to infinity. We point out, however, that
loose balance under our definition can produce negative
eigenvalues large enough to produce quite fast dynamics,
with e(cid:128)ective time constants on the order of single neu-
rons' membrane time-constant, or even as small as a few
milliseconds, depending on parameters.
and inhibition are balanced on fast time scales, which
depends on whether there are substantial shared input
fluctuations across neurons. "Finite" vs. "instantaneous"
time scales of balancing can distinguish whether relax-
ation rates -- the rates of balancing shared fluctuations --
are moderately-sized vs. very large.
An additional source of confusion is that there are two
forms of fast fluctuations, with di(cid:128)erent behaviors. Fast
fluctuations can be shared (correlated) across neurons,
or they can be independent. The large negative eigen-
values in tightly balanced networks (in our definition) af-
fect shared fluctuations corresponding to eigen-modes in
which the activities of excitatory and inhibitory neurons
fluctuate coherently. Thus, shared fluctuations are bal-
anced on fast time scales. By contrast, spatial activity
pa(cid:130)erns in which neurons fluctuate independently are
largely una(cid:128)ected by those eigenvalues, and need not be
balanced.
Fluctuations due to changes in population mean rates of
the external input are shared, and so this form of fluctu-
ation is followed on fast time scales in all balanced net-
works (at finite rates in loosely balanced networks, and
approaching instantaneous following in tightly balanced
networks). Fluctuations also arise from network and ex-
ternal neuronal spiking noise.
In networks with sparse
connectivity (van Vreeswijk and Sompolinsky, 1998), this
yields independent fluctuations in di(cid:128)erent neurons, and
thus independent fluctuations of excitation and inhibi-
tion on fast time scales (though their means are bal-
anced).
In networks with dense connectivity (Renart
et al., 2010), these spiking fluctuations become shared
fluctuations due to common inputs arising from the
dense connectivity, and so in these networks excitation
and inhibition are balanced on fast time scales. To sum-
marize, all balanced networks will balance shared fluctu-
ations, such as those due to changing external input rates,
on fast time scales; but excitation and inhibition can
nonetheless be unbalanced on fast time scales in sparse
networks, due to independent fluctuations induced by
spiking noise.
To conclude, we would argue for a future standardized
terminology for dynamically-induced balancing of exci-
tation and inhibition, in which "loose" vs. "tight" balance
refer to our definition as to whether the remainder af-
ter cancellation is comparable to, or much smaller than,
the factors that cancel. We suggest the use of "temporal"
vs. "mean" balance to refer to whether or not excitation
Appendix 2: When do balanced solutions arise?
We consider a rate model in which the neuron's in-
put/output function is described by some function f (x),
which is zero for x ≤ 0 and monotonically increasing for
x ≥ 0. Then the network's steady-state firing rate rSS for
a steady input I is
rSS = f (WrSS + I)
(14)
where f acts element by element on its argument, that is,
f (u) is a vector whose ith element is f (µi) (the f 's might
di(cid:128)er for di(cid:128)erent neurons, which we neglect for simplic-
ity). As before, we let ψ = (cid:107)W(cid:107) and c = (cid:107)I(cid:107). We define
the dimensionless and O(1) matrix J = W/ψ and vec-
tor g = I/c, so that J and g represent the relative synaptic
strengths and relative input strengths, respectively, while
their overall magnitudes and dimensions are in ψ and c.
Then, as in Ahmadian et al. (2013), we can define the di-
mensionless variable y = ψ
c r, and Eq. 14 can be rewri(cid:130)en
ySS = ψ
f (c(JySS + g))
(15)
c
Note that this equation ensures that ySS ≥ 0. Note also
+ ((x)+ = x, x ≥ 0; = 0, otherwise),
that, when f (x) = (x)p
then this equation can be rewri(cid:130)en ySS = α (JySS + g)p
+
where α = ψcp−1. This is the origin of the dimensionless
constant α mentioned in the main text.
If we define f −1(0) = 0, then because f is monotoni-
cally increasing for non-negative arguments, it is invert-
ible over that range, i.e. f −1(x) is defined for x ≥ 0. We
can then rewrite Eq. 15 as
ySS
= (JySS + g)+
(16)
(cid:19)
f −1(cid:18) c
ψ
1
c
If we assume
(JySS + g)i ≥ 0 for all i,
(17)
that is, if (JySS)i ≥ −(g)i for all i, then we can replace
the right side of Eq. 16 with JySS + g (without the ()+).
This condition, Eq. 17, is a condition on the solution ySS,
which we must check is self-consistently met for any so-
lution we derive under this assumption. Note also that,
14
from Eq. 15, the condition (JySS + g)i > 0 is met if and
only if (ySS)i > 0, so if we find a solution ySS that has
all positive elements, it will automatically satisfy Eq. 17.
Given this assumption, a bit of further manipulation then
yields
J−1f −1(cid:18) c
(cid:19)
ySS = −J−1g +
1
c
ySS
ψ
(18)
SS ≡ −J−1g, is the balancing term, which
The first term, y0
cancels the external input g, i.e. (Jy0
SS + g) = 0. If the sec-
ond term becomes small relative to the first in some limit,
then the tightly balanced solution, ySS ≈ −J−1g or equiv-
alently rSS ≈ −W−1I, exists in that limit, while a loosely
balanced solution (balance index O(1)) arises when the
2nd term is comparable in size to the first. (More careful
analysis is needed to ensure that this solution is stable,
and that there are not also other solutions.) Note that
Eq. 18 gives an equation of the form Eq. 13 when we (1)
Take f −1(x) = (x)1/p
and (2) Multiply both sides of Eq. 18
by c/ψ to convert ySS to rSS.
SS ≡ −J−1g are > 0, a self-
Assuming all the elements of y0
consistent solution in which the second term in Eq. 18
becomes small can be found in at least three cases:
+
• If c and ψ are scaled by the same factor, which
becomes arbitrarily large, then there is a self-
consistent solution in which ySS is converging to
−J−1g. Then the f −1 factor is not changing (except
for the small changes due to the changes in ySS as
it converges), but it is multiplied by the factor 1
,
c
which becomes arbitrarily small; thus the second
term becomes arbitrarily small, regardless of the
particular structure of f . This is the case studied
for the tightly balanced solution, where both c and
ψ are taken proportional to
K with K very large.
(Note that the mean field equations derived in (van
Vreeswijk and Sompolinsky, 1998) di(cid:128)er from the
generic steady-state rate equations, Eq. (14), in that
they also involve the self-consistently calculated
input fluctuation strengths, σA; the scaling argu-
ment given here nevertheless holds in that case
too.)
√
• Suppose c is scaled for fixed ψ, which is the bio-
logical case in which synaptic strengths are fixed
and the strength of the external input is varied
from small to large. Then if f −1(x) grows more
slowly than linearly in increasing x, then the 1
c
factor shrinks faster than the f −1 term grows, so
15
again there is a self-consistent solution in which
ySS is converging to −J−1g and the second term
becomes arbitrarily small with increasing c. This is
the case studied for the loosely balanced solution
in the SSN, in which f (x) grows supralinearly with
x and therefore f −1(x) grows sublinearly with x.
• We again suppose c is scaled for fixed ψ, but now
imagine that f −1(x) grows faster than linearly in in-
creasing x, i.e. f (x) is sublinear (for example, f (x) =
(x)p
+ for 0 < p < 1). Then there is a self-consistent
solution in which y → −J−1g as c → 0, with the
second term in Eq. 18 going to zero as c → 0. This
case is the reverse of the SSN: the strongly coupled,
balanced regime arises for c → 0, while the weakly
coupled, feedforward-driven regime arises for large
c.
In sum, if the elements of −J−1g are positive, then a self-
consistent tightly-balanced solution arises for any f if c
and ψ are scaled together by an increasing factor; for
supralinear f if c is scaled by an increasing factor; or for
sublinear f if c is scaled by a decreasing factor. In all of
these cases, for moderate sizes of the scaled parameter(s)
(e.g., for the SSN, for α = O(1)) such that the second term
of Eq. 18 is comparable in size to the first, a loosely bal-
anced solution should arise. Note that, since rSS = c
ψ ySS,
then from Eq. 18 the net input a(cid:129)er balancing should
grow with increasing external input c as f −1(c); this is
sublinear in c for the SSN case of supralinear f .
If one considers a two-population model -- a population of
E cells and a population of I cells, with each population's
average rate represented by a single variable -- then con-
ditions on J and g can be defined such that the elements
of −J−1g are positive and the balanced fixed point is sta-
ble and is the only fixed point (Ahmadian et al., 2013;
Kraynyukova and Tchumatchenko, 2018; van Vreeswijk
and Sompolinsky, 1998). On the other hand, when the
E element of −J−1g is negative, Eq. 18 cannot serve as a
basis for an asymptotic expansion with the leading term
−J−1g, and the tightly balanced state does not exist.
(Given (−J−1g)E < 0, if the tightly balanced state ex-
isted -- meaning that the second term of Eq. 18 becomes
much smaller than the first while Eq. 18 is applicable --
then ySS must have crossed zero to become negative, but
once that has happened we could no longer proceed past
Eq. 16 and Eq. 18 would no longer be applicable, which is
a contradiction; hence the tightly balanced state cannot
exist.) However, the loosely balanced state can still arise
c J−1f −1(cid:16) c
ψ ySS(t − 1)(cid:17), given appropri-
in this case in a broad parameter regime of J and g, and
can be found as the fixed point of the iterative equation,
ySS(t) = −J−1g + 1
ate initial conditions ySS(0); see Ahmadian et al. (2013). In
this case, with increasing c, rE grows, but then saturates
and starts decreasing, and eventually is pushed down to
0. However, if we assume that the maximal external input
(i.e. the maximal c; for example, the maximal firing rate of
the thalamic input to a primary sensory cortical area) can
only drive rE to saturation or slightly beyond, this repre-
sents a viable model of cortical systems (Ahmadian et al.,
2013; Hennequin et al., 2018; Rubin et al., 2015).
A two-population model accurately describes the behav-
ior of an unstructured model with many E and I neu-
rons, i.e. with random connectivity and with neurons in
each population receiving comparable stimulus inputs. In
some cases this model also can form a good approxima-
tion to the behavior of a multi-neuron circuit with struc-
tured connectivity and stimulus selectivity (Ahmadian
et al., 2013). More generally, though, in such a structured
circuit with localized connectivity, for larger/stronger lo-
calized stimuli, some set of neurons (e.g., neurons not
selective for the stimulus) may eventually receive a net
inhibition and become silent, meaning that the condi-
tion of Eq. 17 is not met and Eq. 18 does not apply.
(However, if the connectivity is translation-invariant --
the same at any position in the model -- and the ex-
ternal input extends more narrowly than the network
connections, then a balanced fixed point can still be at-
tained, Rosenbaum and Doiron, 2014.) Nonetheless, we
find in simulations (Ahmadian et al., 2013; Hennequin
et al., 2018; Rubin et al., 2015) that for reasonable stimu-
lus input strengths, SSN behavior is reasonably described
by the two-population model, in that (1) there is a transi-
tion with increasing input strength from a weakly cou-
pled, feedforward-driven regime to a strongly-coupled,
loosely balanced regime in which the input to excited
neurons grows sublinearly as a function of the external in-
put strength; and (2) if we define the W and g of the two-
population model as describing the net input received by
a cell in the larger, structured model -- e.g., WEE repre-
sents the mean summed synaptic strength from excita-
tory cells to a single excitatory cell, and gE represents
the mean external input received by stimulus-selective
excitatory cells -- then reasonably good insight into the
operating regime of the larger model can be obtained
from the analysis of the two-population model presented
here and, in much more detail, in Ahmadian et al. (2013);
Kraynyukova and Tchumatchenko (2018).
We believe the same overall analysis of a transition from
a weakly coupled regime to a strongly coupled, loosely
balanced regime will apply to multi-population models
incorporating multiple subtypes of inhibitory cells (e.g.
Garcia Del Molino et al., 2017; Kuchibhotla et al., 2017;
Litwin-Kumar et al., 2016), but more detailed aspects
of the analysis of the two-population model (Ahmadian
et al., 2013; Kraynyukova and Tchumatchenko, 2018) need
to be generalized to that case.
Acknowledgements
We thank Larry Abbo(cid:130), Mario DiPoppa and Agostina
Palmigiano for many helpful
comments on the
manuscript, David Hansel, Gianluigi Mongillo and Al-
fonso Renart for many useful discussions, and Randy
Bruno for help with references. YA is supported by start-
up funds from the University of Oregon. KDM is sup-
ported by NSF DBI-1707398, NIH U01NS108683, NIH
R01EY029999, NIH U19NS107613, Simons Foundation
award 543017, and the Gatsby Charitable Foundation.
References
Adesnik, H. (2017).
Synaptic Mechanisms of Feature
Coding in the Visual Cortex of Awake Mice. Neuron,
95:1147 -- 1159.
Ahmadian, Y., Rubin, D. B., and Miller, K. D. (2013). Anal-
ysis of the stabilized supralinear network. Neural Com-
putation, 25:1994 -- 2037.
Amatrudo, J. M., Weaver, C. M., Crimins, J. L., Hof, P. R.,
Rosene, D. L., and Luebke, J. I. (2012).
Influence of
highly distinctive structural properties on the excitabil-
ity of pyramidal neurons in monkey visual and pre-
frontal cortices. J. Neurosci., 32(40):13644 -- 13660.
Amit, D. and Brunel, N. (1997). Dynamics of a recurrent
network of spiking neurons before and following learn-
ing. Network: Comput. Neural Syst., 8:373 -- 404.
Anderson, J. S., Carandini, M., and Ferster, D. (2000). Ori-
entation tuning of input conductance, excitation, and
inhibition in cat primary visual cortex. J. Neurophysiol.,
84:909 -- 926.
16
Anderson, J. S., Lampl, I., Gillespie, D. C., and Ferster, D.
(2001). Membrane potential and conductance changes
underlying length tuning of cells in cat primary visual
cortex. J. Neurosci., 21:2104 -- 2112.
Chung, S. and Ferster, D. (1998). Strength and orienta-
tion tuning of the thalamic input to simple cells re-
vealed by electrically evoked cortical suppression. Neu-
ron, 20:1177 -- 89.
Angelucci, A., Bijanzadeh, M., Nurminen, L., Federer, F.,
Merlin, S., and Bresslo(cid:128), P. (2017). Circuits and mech-
anisms for surround modulation in visual cortex. Ann.
Rev. Neurosci., 40:425 -- 451.
Atallah, B. V. and Scanziani, M. (2009).
Instantaneous
modulation of gamma oscillation frequency by balanc-
ing excitation with inhibition. Neuron, 62:566 -- 577.
Barre(cid:130), D., Deneve, S., and Machens, C. (2013).
Barral, J. and Reyes, A. (2016). Synaptic scaling rule pre-
serves excitatory-inhibitory balance and salient neu-
ronal network dynamics. Nat. Neurosci., 19:1690 -- 1696.
Fir-
ing rate predictions in optimal balanced networks. In
Pereira, F., Burges, C., Bo(cid:130)ou, L., and Weinberger, K.,
editors, Advances in Neural Information Processing Sys-
tems, pages 1538 -- 1546. MIT Press.
Barth, A. L. and Poulet, J. F. (2012). Experimental evi-
dence for sparse firing in the neocortex. Trends Neu-
rosci., 35:345 -- 355.
Bhatia, A., Moza, S., and Bhalla, U. S. (2019). Precise
excitation-inhibition balance controls gain and timing
in the hippocampus. Elife, 8.
Boerlin, M., Machens, C. K., and Deneve, S. (2013). Predic-
tive coding of dynamical variables in balanced spiking
networks. PLoS Comput. Biol., 9(11):e1003258.
Bourdoukan, R., Barre(cid:130), D., Deneve, S., and Machens, C.
(2012). Learning optimal spike- based representations.
In Pereira, F., Burges, C., Bo(cid:130)ou, L., and Weinberger,
K., editors, Advances in Neural Information Processing
Systems, pages 2285 -- 2293. MIT Press.
Brunel, N. (2000). Dynamics of sparsely connected net-
works of excitatory and inhibitory spiking neurons. J
Comput Neurosci, 8:183 -- 208.
Carandini, M. and Heeger, D. J. (2012). Normalization as
a canonical neural computation. Nat. Rev. Neurosci.,
13:51 -- 62.
Churchland, M. M., Yu, B. M., Cunningham, J. P., Sugrue,
L. P., Cohen, M. R., Corrado, G. S., Newsome, W. T.,
Clark, A. M., Hosseini, P., Sco(cid:130), B. B., Bradley, D. C.,
Smith, M. A., Kohn, A., Movshon, J. A., Armstrong,
K. M., Moore, T., Chang, S. W., Snyder, L. H., Lisberger,
S. G., Priebe, N. J., Finn, I. M., Ferster, D., Ryu, S. I., San-
thanam, G., Sahani, M., and Shenoy, K. V. (2010). Stim-
ulus onset quenches neural variability: a widespread
cortical phenomenon. Nat. Neurosci., 13:369 -- 378.
Cohen, M. R. and Kohn, A. (2011). Measuring and inter-
preting neuronal correlations. Nat. Neurosci., 14:811 --
819.
Constantinople, C. M. and Bruno, R. M. (2013). Deep cor-
tical layers are activated directly by thalamus. Science,
340:1591 -- 1594.
Cossell, L., Iacaruso, M. F., Muir, D. R., Houlton, R., Sader,
E. N., Ko, H., Hofer, S. B., and Mrsic-Flogel, T. D.
(2015). Functional organization of excitatory synaptic
strength in primary visual cortex. Nature, 518:399 -- 403.
Dehghani, N., Peyrache, A., Telenczuk, B., Le Van (cid:131)yen,
M., Halgren, E., Cash, S. S., Hatsopoulos, N. G., and
Destexhe, A. (2016). Dynamic Balance of Excitation
and Inhibition in Human and Monkey Neocortex. Sci
Rep, 6:23176.
Deneve, S. and Machens, C. K. (2016). E(cid:128)icient codes and
balanced networks. Nat. Neurosci., 19:375 -- 382.
DeWeese, M. and Zador, A. (2006). Non-Gaussian mem-
brane potential dynamics imply sparse, synchronous
activity in auditory cortex. J. Neurosci., 26:12206 -- 12218.
Doiron, B., Litwin-Kumar, A., Rosenbaum, R., Ocker,
G. K., and Josic, K. (2016). The mechanics of state-
dependent neural correlations. Nat. Neurosci., 19:383 --
393.
Cavanaugh, J. R., Bair, W., and Movshon, J. A. (2002). Na-
ture and interaction of signals from the receptive field
center and surround in macaque V1 neurons. J. Neuro-
physiol., 88:2530 -- 2546.
Echeveste, R., Aitchison, L., Hennequin, G., and Lengyel,
M. (2019). Cortical-like dynamics in recurrent circuits
optimized for sampling-based probabilistic inference.
biorxiv, doi: h(cid:130)p://dx.doi.org/10.1101/696088.
17
Ecker, A. S., Berens, P., Co(cid:130)on, R. J., Subramaniyan,
M., Denfield, G. H., Cadwell, C. R., Smirnakis, S. M.,
Bethge, M., and Tolias, A. S. (2014). State dependence
of noise correlations in macaque primary visual cortex.
Neuron, 82:235 -- 248.
Ecker, A. S., Berens, P., Keliris, G. A., Bethge, M., Lo-
gothetis, N. K., and Tolias, A. S. (2010). Decorre-
lated neuronal firing in cortical microcircuits. Science,
327(5965):584 -- 587.
Elston, G. N. (2003). Cortex, cognition and the cell:
new insights into the pyramidal neuron and prefrontal
function. Cereb. Cortex, 13:1124 -- 1138.
Elston, G. N. and Fujita, I. (2014). Pyramidal cell develop-
ment: postnatal spinogenesis, dendritic growth, axon
growth, and electrophysiology. Front Neuroanat, 8:78.
Elston, G. N. and Manger, P. (2014). Pyramidal cells in V1
of African rodents are bigger, more branched and more
spiny than those in primates. Front Neuroanat, 8:4.
Fares, T. and Stepanyants, A. (2009). Cooperative synapse
formation in the neocortex. Proc. Natl. Acad. Sci. U.S.A.,
106(38):16463 -- 16468.
Feldmeyer, D., Egger, V., Lubke, J., and Sakmann, B.
(1999). Reliable synaptic connections between pairs
of excitatory layer 4 neurones within a single "barrel"
of developing rat somatosensory cortex.
J. Physiol.,
521:169 -- 190.
Feldmeyer, D., Lubke, J., Silver, R. A., and Sakmann, B.
(2002). Synaptic connections between layer 4 spiny
neurone-layer 2/3 pyramidal cell pairs in juvenile rat
barrel cortex: physiology and anatomy of interlaminar
signalling within a cortical column. J Physiol, 538:803 --
822.
Garcia Del Molino, L. C., Yang, G. R., Mejias, J. F., and
Wang, X. J. (2017). Paradoxical response reversal of
top-down modulation in cortical circuits with three in-
terneuron types. Elife, 6.
Graupner, M. and Reyes, A. D. (2013). Synaptic input
correlations leading to membrane potential decorre-
lation of spontaneous activity in cortex.
J. Neurosci.,
33:15075 -- 15085.
Haider, B., Duque, A., Hasenstaub, A. R., and McCormick,
D. A. (2006). Neocortical network activity in vivo is gen-
erated through a dynamic balance of excitation and in-
hibition. J. Neurosci., 26:4535 -- 4545.
Haider, B., Hausser, M., and Carandini, M. (2013). Inhibi-
tion dominates sensory responses in the awake cortex.
Nature, 493:97 -- 100.
Hansel, D. and van Vreeswijk, C. (2002). How noise con-
tributes to contrast invariance of orientation tuning in
cat visual cortex. J. Neurosci., 22:5118 -- 5128.
Hansel, D. and van Vreeswijk, C. (2012). The mechanism
of orientation selectivity in primary visual cortex with-
out a functional map. J. Neurosci., 32:4049 -- 4064.
Hennequin, G., Agnes, E. J., and Vogels, T. P. (2017). In-
hibitory Plasticity: Balance, Control, and Codepen-
dence. Annu. Rev. Neurosci., 40:557 -- 579.
Hennequin, G., Ahmadian, Y., Rubin, D. B., Lengyel,
M., and Miller, K. D. (2018). The Dynamical Regime
of Sensory Cortex: Stable Dynamics around a Sin-
gle Stimulus-Tuned A(cid:130)ractor Account for Pa(cid:130)erns of
Noise Variability. Neuron, 98(4):846 -- 860.
Heuer, H. W. and Bri(cid:130)en, K. H. (2002). Contrast depen-
dence of response normalization in area MT of the rhe-
sus macaque. J. Neurophysiol., 88:3398 -- 3408.
Ferster, D., Chung, S., and Wheat, H. (1996). Orientation
selectivity of thalamic input to simple cells of cat visual
cortex. Nature, 380:249 -- 252.
Higley, M. and Contreras, D. (2006). Balanced excita-
tion and inhibition determine spike timing during fre-
quency adaptation. J. Neurosci., 26:448 -- 457.
Finn, I. M., Priebe, N. J., and Ferster, D. (2007). The emer-
gence of contrast-invariant orientation tuning in sim-
ple cells of cat visual cortex. Neuron, 54:137 -- 152.
Histed, M. H. (2018). Feedforward Inhibition Allows Input
Summation to Vary in Recurrent Cortical Networks.
eNeuro, 5(1).
Galarreta, M. and Hestrin, S.
Frequency-
dependent synaptic depression and the balance of ex-
citation and inhibition in the neocortex. Nat. Neurosci.,
1:587 -- 594.
(1998).
Ichida, J. M., Schwabe, L., Bresslo(cid:128), P. C., and Angelucci,
A. (2007). Response facilitation from the "suppressive"
receptive field surround of macaque V1 neurons.
J.
Neurophysiol., 98:2168 -- 2181.
18
Kato, H. K., Asinof, S. K., and Isaacson, J. S. (2017).
Network-Level Control of Frequency Tuning in Audi-
tory Cortex. Neuron, 95(2):412 -- 423.
Maimon, G. and Assad, J. A. (2009). Beyond Poisson: in-
creased spike-time regularity across primate parietal
cortex. Neuron, 62:426 -- 440.
Ko, H., Hofer, S. B., Pichler, B., Buchanan, K. A., Sjostrom,
P. J., and Mrsic-Flogel, T. D. (2011). Functional speci-
ficity of local synaptic connections in neocortical net-
works. Nature, 473:87 -- 91.
Kraynyukova, N. and Tchumatchenko, T. (2018). Stabi-
lized supralinear network can give rise to bistable, os-
cillatory, and persistent activity. Proc. Natl. Acad. Sci.
U.S.A., 115(13):3464 -- 3469.
Kuchibhotla, K. V., Gill, J. V., Lindsay, G. W., Papadoy-
annis, E. S., Field, R. E., Sten, T. A., Miller, K. D., and
Froemke, R. C. (2017). Parallel processing by cortical
inhibition enables context-dependent behavior. Nat.
Neurosci., 20:62 -- 71.
Lankarany, M., Heiss, J. E., Lampl, I., and Toyoizumi, T.
(2016). Simultaneous Bayesian Estimation of Excita-
tory and Inhibitory Synaptic Conductances by Exploit-
ing Multiple Recorded Trials. Front Comput Neurosci,
10:110.
Li, L. Y., Li, Y. T., Zhou, M., Tao, H. W., and Zhang, L. I.
(2013a).
Intracortical multiplication of thalamocorti-
cal signals in mouse auditory cortex. Nat. Neurosci.,
16:1179 -- 1181.
Li, Y. T., Ibrahim, L. A., Liu, B. H., Zhang, L. I., and Tao,
H. W. (2013b). Linear transformation of thalamocor-
tical input by intracortical excitation. Nat. Neurosci.,
16:1324 -- 1330.
Lien, A. D. and Scanziani, M. (2013). Tuned thalamic exci-
tation is amplified by visual cortical circuits. Nat. Neu-
rosci., 16:1315 -- 1323.
Litwin-Kumar, A., Chacron, M. J., and Doiron, B. (2012).
The spatial structure of stimuli shapes the timescale of
correlations in population spiking activity. PLoS Com-
put. Biol., 8:e1002667.
Litwin-Kumar, A., Rosenbaum, R., and Doiron, B. (2016).
Inhibitory stabilization and visual coding in cortical
circuits with multiple interneuron subtypes. J. Neuro-
physiol., 115:1399 -- 1409.
Liu, L. D., Miller, K. D., and Pack, C. C. (2018). A Uni-
fying Motif for Spatial and Directional Surround Sup-
pression. J. Neurosci., 38(4):989 -- 999.
Mainen, Z. F. and Sejnowski, T. J. (1995). Reliability of
spike timing in neocortical neurons. Science, 268:1503 --
1506.
Marino, J., Schummers, J., Lyon, D. C., Schwabe, L., Beck,
O., Wiesing, P., Obermayer, K., and Sur, M. (2005). In-
variant computations in local cortical networks with
balanced excitation and inhibition. Nature Neurosci.,
8:194 -- 201.
Markov, N. T., Misery, P., Falchier, A., Lamy, C., Vezoli,
J., (cid:131)ilodran, R., Gariel, M. A., Giroud, P., Ercsey-
Ravasz, M., Pilaz, L. J., Huissoud, C., Barone, P., De-
hay, C., Toroczkai, Z., Van Essen, D. C., Kennedy, H.,
and Knoblauch, K. (2011). Weight consistency speci-
fies regularities of macaque cortical networks. Cereb.
Cortex, 21(6):1254 -- 1272.
Markram, H., Lubke, J., Frotscher, M., Roth, A., and Sak-
mann, B. (1997). Physiology and anatomy of synaptic
connections between thick tu(cid:129)ed pyramidal neurones
in the developing rat neocortex. J. Physiol. (Lond.), 500
( Pt 2):409 -- 440.
Miller, K. D. (2016). Canonical computations of cerebral
cortex. Curr. Opin. Neurobiol., 37:75 -- 84.
Miller, K. D. and Troyer, T. W. (2002). Neural noise can
explain expansive, power-law nonlinearities in neural
response functions. J. Neurophysiol., 87:653 -- 659.
Mongillo, G., Hansel, D., and van Vreeswijk, C. (2012).
Bistability and spatiotemporal irregularity in neuronal
networks with nonlinear synaptic transmission. Phys.
Rev. Le(cid:130)., 108:158101.
Nassi, J. J., Avery, M. C., Cetin, A. H., Roe, A. W., and
Reynolds, J. H. (2015). Optogenetic Activation of Nor-
malization in Alert Macaque Visual Cortex. Neuron,
86:1504 -- 1517.
Nienborg, H., Hasenstaub, A., Nauhaus, I., Taniguchi,
H., Huang, Z. J., and Callaway, E. M. (2013). Con-
trast dependence and di(cid:128)erential contributions from
somatostatin- and parvalbumin-expressing neurons to
spatial integration in mouse V1. J. Neurosci., 33:11145 --
11154.
19
O'Donnell, C. and van Rossum, M. C. (2014). System-
atic analysis of the contributions of stochastic voltage
gated channels to neuronal noise. Front Comput Neu-
rosci, 8:105.
Ohshiro, T., Angelaki, D. E., and DeAngelis, G. (2013). A
normalization model of multisensory integration ac-
counts for distinct forms of cross-modal and within-
modal cue integration by cortical neurons. Program No.
360.19. 2013 Neuroscience Meeting Planner. San Diego,
CA: Society for Neuroscience, 2013. Online.
Okun, M. and Lampl, I. (2008).
Instantaneous correla-
tion of excitation and inhibition during ongoing and
sensory-evoked activities. Nat. Neurosci., 11:535 -- 537.
Ozeki, H., Finn, I. M., Scha(cid:128)er, E. S., Miller, K. D., and Fer-
Inhibitory stabilization of the cortical
ster, D. (2009).
network underlies visual surround suppression. Neu-
ron, 62:578 -- 592.
Pehlevan, C. and Sompolinsky, H. (2014). Selectivity and
sparseness in randomly connected balanced networks.
PLoS ONE, 9:e89992.
Polat, U., Mizobe, K., Pe(cid:130)et, M. W., Kasamatsu, T., and
Norcia, A. M. (1998). Collinear stimuli regulate visual
responses depending on cell's contrast threshold. Na-
ture, 391:580 -- 584.
Poulet, J. F. and Petersen, C. C. (2008). Internal brain state
regulates membrane potential synchrony in barrel cor-
tex of behaving mice. Nature, 454:881 -- 885.
Priebe, N., Mechler, F., Carandini, M., and Ferster, D.
(2004). The contribution of spike threshold to the di-
chotomy of cortical simple and complex cells. Nat. Neu-
rosci., 7:1113 -- 22.
Priebe, N. J. and Ferster, D. (2008).
Inhibition, spike
threshold and stimulus selectivity in primary visual
cortex. Neuron, 57:482 -- 497.
Renart, A., de la Rocha, J., Bartho, P., Hollender, L., Parga,
N., Reyes, A., and Harris, K. D. (2010). The asyn-
chronous state in cortical circuits. Science, 327:587 -- 590.
Reynolds, J. H. and Chelazzi, L. (2004). A(cid:130)entional modu-
lation of visual processing. Annu. Rev. Neurosci., 27:611 --
647.
Rosenbaum, R., Smith, M. A., Kohn, A., Rubin, J. E., and
Doiron, B. (2017). The spatial structure of correlated
neuronal variability. Nat. Neurosci., 20:107 -- 114.
Rubin, D. B., Van Hooser, S. D., and Miller, K. D. (2015).
The stabilized supralinear network: A unifying circuit
motif underlying multi-input integration in sensory
cortex. Neuron, 85:402 -- 417.
Sadagopan, S. and Ferster, D. (2012). Feedforward ori-
gins of response variability underlying contrast invari-
ant orientation tuning in cat visual cortex. Neuron,
74:911 -- 923.
Sadeh, S. and Ro(cid:130)er, S. (2015). Orientation selectivity in
inhibition-dominated networks of spiking neurons: ef-
fect of single neuron properties and network dynamics.
PLoS Comput. Biol., 11:e1004045.
Sanzeni, A., Akitake, B., Goldback, H., Leedy, C., Brunel,
N., and Histed, M. (2019). Inhibition stabilization is a
widespread property of cortical networks. biorxiv, doi:
h(cid:130)ps://doi.org/10.1101/656710.
Sato, T. K., Hausser, M., and Carandini, M. (2014). Dis-
tal connectivity causes summation and division across
mouse visual cortex. Nat. Neurosci., 17:30 -- 32.
Sceniak, M., Ringach, D. L., Hawken, M., and Shapley,
R. (1999). Contrast's e(cid:128)ect on spatial summation by
macaque v1 neurons. Nature Neurosci., 2:733 -- 739.
Schneidman, E., Freedman, B., and Segev, I. (1998).
Ion
channel stochasticity may be critical in determining
the reliability and precision of spike timing. Neural
Comput, 10:1679 -- 1703.
Schoonover, C. E., Tapia, J. C., Schilling, V. C., Wimmer, V.,
Blazeski, R., Zhang, W., Mason, C. A., and Bruno, R. M.
(2014). Comparative strength and dendritic organi-
zation of thalamocortical and corticocortical synapses
onto excitatory layer 4 neurons. J. Neurosci., 34:6746 --
6758.
Schwabe, L., Ichida, J. M., Shushruth, S., Mangapathy, P.,
and Angelucci, A. (2010). Contrast-dependence of sur-
round suppression in Macaque V1: experimental test-
ing of a recurrent network model. Neuroimage, 52:777 --
792.
Rosenbaum, R. and Doiron, B. (2014). Balanced networks
of spiking neurons with spatially dependent recurrent
connections. Physical Review X, 4(2):021039.
Sengpiel, F., Blakemore, C., and Sen, A. (1997). Character-
istics of surround inhibition in cat area 17. Exp. Brain
Res., 116:216 -- 228.
20
Shadlen, M. N. and Newsome, W. T. (1998). The variable
discharge of cortical neurons: implications for connec-
tivity, computation, and information coding.
J. Neu-
rosci., 18:3870 -- 3896.
Shu, Y., Hasenstaub, A., and McCormick, D. A. (2003).
Turning on and o(cid:128) recurrent balanced cortical activity.
Nature, 423:288 -- 293.
Shushruth, S., Ichida, J. M., Levi(cid:130), J. B., and Angelucci, A.
(2009). Comparison of spatial summation properties
of neurons in macaque V1 and V2.
J. Neurophysiol.,
102:2069 -- 2083.
So(cid:129)ky, W. and Koch, C. (1993). The highly irregular firing
of cortical-cells is inconsistent with temporal integra-
tion of random EPSPS. J. Neurosci., 13:334 -- 350.
Song, X. M. and Li, C. Y. (2008). Contrast-dependent
and contrast-independent spatial summation of pri-
mary visual cortical neurons of the cat. Cerebral Cortex,
18:331 -- 336.
Stevens, C. and Zador, A. (1998).
Input synchrony and
the irregular firing of cortical neurons. Nat. Neurosci.,
1:210 -- 217.
Tan, A. Y., Chen, Y., Scholl, B., Seidemann, E., and Priebe,
N. J. (2014). Sensory stimulation shi(cid:129)s visual cortex
from synchronous to asynchronous states. Nature,
509:226 -- 229.
Troyer, T. W. and Miller, K. D. (1997). Physiological gain
leads to high ISI variability in a simple model of a cor-
tical regular spiking cell. Neural Comput., 9:971 -- 983.
Tsodyks, M. V. and Sejnowski, T. (1995). Rapid state
switching in balanced cortical network models. Net-
work, 6:111 -- 124.
Tsui, J. M. and Pack, C. C. (2011). Contrast sensitivity
of MT receptive field centers and surrounds. J. Neuro-
physiol., 106:1888 -- 1900.
van Vreeswijk, C. and Sompolinsky, H. (1996). Chaos
in neuronal networks with balanced excitatory and in-
hibitory activity. Science, 274:1724 -- 1726.
van Vreeswijk, C. and Sompolinsky, H. (1998). Chaotic
balanced state in a model of cortical circuits. Neural
Computation, 10:1321 -- 1371.
Wang, S., Miller, K. D., and van Hooser, S. B. (2019). Com-
bined visual and pa(cid:130)erned optogenetic stimulation of
ferret visual cortex reveals that cortical circuits respond
to independent inputs in a sublinear manner. Cosyne
Abstracts 2019, Lisbon, Portugal, (to appear):1 -- 15.
Wehr, M. and Zador, A. M. (2003). Balanced inhibition
underlies tuning and sharpens spike timing in auditory
cortex. Nature, 426:442 -- 446.
Wilent, W. B. and Contreras, D. (2005).
Stimulus-
dependent changes in spike threshold enhance feature
selectivity in rat barrel cortex neurons.
J. Neurosci.,
25:2983 -- 2991.
Wu, G. K., Arbuckle, R., Liu, B. H., Tao, H. W., and Zhang,
L. I. (2008). Lateral sharpening of cortical frequency
tuning by approximately balanced inhibition. Neuron,
58:132 -- 143.
Wu, G. K., Li, P., Tao, H. W., and Zhang, L. I. (2006).
Nonmonotonic synaptic excitation and imbalanced in-
hibition underlying cortical intensity tuning. Neuron,
52:705 -- 715.
Xue, M., Atallah, B. V., and Scanziani, M. (2014). Equal-
izing excitation-inhibition ratios across visual cortical
neurons. Nature, 511:596 -- 600.
Yizhar, O., Fenno, L. E., Prigge, M., Schneider, F., David-
son, T. J., O'Shea, D. J., Sohal, V. S., Goshen, I., Finkel-
stein, J., Paz, J. T., Stehfest, K., Fudim, R., Ramakrish-
nan, C., Huguenard, J. R., Hegemann, P., and Deis-
seroth, K. (2011). Neocortical excitation/inhibition bal-
ance in information processing and social dysfunction.
Nature, 477:171 -- 178.
Zhou, M., Liang, F., Xiong, X. R., Li, L., Li, H., Xiao, Z., Tao,
H. W., and Zhang, L. I. (2014). Scaling down of balanced
excitation and inhibition by active behavioral states in
auditory cortex. Nat. Neurosci., 17:841 -- 850.
21
|
1610.04844 | 1 | 1610 | 2016-10-16T11:34:59 | Complete coverage of space favors modularity of the grid system in the brain | [
"q-bio.NC"
] | Grid cells in the entorhinal cortex fire when animals that are exploring a certain region of space occupy the vertices of a triangular grid that spans the environment. Different neurons feature triangular grids that differ in their properties of periodicity, orientation and ellipticity. Taken together, these grids allow the animal to maintain an internal, mental representation of physical space. Experiments show that grid cells are modular, i.e. there are groups of neurons which have grids with similar periodicity, orientation and ellipticity. We use statistical physics methods to derive a relation between variability of the properties of the grids within a module and the range of space that can be covered completely (i.e. without gaps) by the grid system with high probability. Larger variability shrinks the range of representation, providing a functional rationale for the experimentally observed co-modularity of grid cell periodicity, orientation and ellipticity. We obtain a scaling relation between the number of neurons and the period of a module, given the variability and coverage range. Specifically, we predict how many more neurons are required at smaller grid scales than at larger ones. | q-bio.NC | q-bio |
Complete coverage of space favors modularity
of the grid system in the brain
A. Sanzeni,1, 2 V. Balasubramanian,3 G. Tiana,4 and M. Vergassola2
1Department of Physics, University of Milan and INFN, Via Celoria 13, 20133 Milano, Italy
2Department of Physics, University of California San Diego, La Jolla, CA 92093-0374, USA
3David Rittenhouse Laboratory, University of Pennsylvania, Philadelphia, PA 19104, USA
4Centre for Complexity & Biosystems and Department of Physics,
University of Milan and INFN, University of Milan, via Celoria 16, 20133 Milano, Italy
Grid cells in the entorhinal cortex fire when animals that are exploring a certain region of space
occupy the vertices of a triangular grid that spans the environment. Different neurons feature
triangular grids that differ in their properties of periodicity, orientation and ellipticity. Taken
together, these grids allow the animal to maintain an internal, mental representation of physical
space. Experiments show that grid cells are modular, i.e.
there are groups of neurons which
have grids with similar periodicity, orientation and ellipticity. We use statistical physics methods
to derive a relation between variability of the properties of the grids within a module and the
range of space that can be covered completely (i.e. without gaps) by the grid system with high
probability. Larger variability shrinks the range of representation, providing a functional rationale
for the experimentally observed co-modularity of grid cell periodicity, orientation and ellipticity.
We obtain a scaling relation between the number of neurons and the period of a module, given
the variability and coverage range. Specifically, we predict how many more neurons are required at
smaller grid scales than at larger ones.
I.
INTRODUCTION
Classical behavioral experiments show that the navi-
gation of mammals relies on an internal representation
of space called a "cognitive map" [1]. Research on the
neural basis of this internal representation started with
the discovery of place cells in the hippocampus of rats,
neurons that have their activity controlled by the phys-
ical position occupied by the animal [2]. The discovery
of place cells generated an extensive investigation of the
spatial representation system in the brain, which led to
the discovery of grid cells [3], as well as different types
of neurons whose activity codes for head direction [4],
speed [5], and borders of the environment [6, 7] (see [8]
for a recent review). The discovery of cells that consti-
tute a positioning system in the brain was the motivation
for the Nobel Prize in Physiology awarded in 2014.
One of the most striking elements composing the cog-
nitive map is in the entorhinal cortex (EC) [3], where
grid cells respond when the animal occupies one of the
vertices of a triangular grid that tessellates space.
It
is widely believed that these neurons provide a metric
for the spatial representation system, since their relation
with physical position does not reshuffle in different en-
vironments, unlike what happens for place cells where
"remapping"' occurs [9].
Grid cells are organized in modules -- grids in a mod-
ule are clustered around a discrete period which increases
along the dorso-ventral axis of the EC [10, 11]. Grids in
a module also share similar orientations and ellipticities
while varying in spatial phase [11]. Experimentally, there
is a power-law relation between the periodicities of dif-
ferent modules -- a rationale for the power law was given
in [12, 13].
The term "module" for grid cells differs fundamentally
from the same term used in the context of brain (or city)
networks [14]. There, neurons correspond to the nodes of
a network whose edges correspond to axonal connections
among the neurons. Modularity refers to the formation of
clusters of nodes that are more densely connected among
themselves than to nodes in other modules. The reason
for modularity in these networks is that edges have costs
that scale with their length, so that spatial aspects are
important and commonly lead to cluster formation. By
contrast, the triangular lattices in the EC grid system de-
scribe firing patterns of individual grid cell neurons as an
animal explores the environment. In other words, there
is no physical edge between vertices of the triangular fir-
ing lattices of grid cells, and no cost associated to their
length. Hence, physical proximity is not relevant for grid
cells and has no bearing on the problem of explaining the
modular organization of grid cells in the EC.
Why is the grid system modular? The key point un-
derlying our arguments in the sequel is that behavioral
deficits in orientation and navigation result if the neural
representation has gaps, i.e. complete coverage of space is
a fundamental requirement for the cognitive map to func-
tion. Specifically, we use statistical physics methods to
show that variability in period, orientation or ellipticity
randomizes the relative phases of firing fields, and leads
to failure of spatial coverage. Larger variability entails a
smaller physical range that can be covered without gaps
(which would lead to behavioral deficits). Hence, opti-
mizing spatial coverage gives a functional argument for
reduced variability and for the observed co-modularity
of grid cells. We also predict a scaling law relating the
period and number of neurons in a module.
II. RESULTS
A. A model of grid cells' activity
For our specific purpose of analyzing efficient coverage
of space, the firing field of grid cells can be simplified as
follows. After thresholding for noise, the smooth lumps
formed by firing fields are treated as being uniformly ac-
tive inside a localized region and inactive outside (Fig. 1).
Noise and firing inhomogeneity inside the active region
only degrade the uniformity of coverage relative to this
model. Thus, treating firing fields as step functions al-
lows us to derive bounds on how well a given grid archi-
tecture can cover space.
Specifically, we represent the activity of grid cells as
(cid:19)
(cid:18)
(cid:88)
n,m∈Z
a(x) =
χ
φ + R(θ) [nv + mu)] − x
(cid:96)/2
,
(1)
where x is the vector locating the position of the an-
in two dimensions, v = λ1(cos(β), sin(β)) and
imal
u = λ2(1, 0) are the elementary vectors that generate
the grid, (cid:96) is the diameter of a firing field, and n and m
are integers indexing the vertices of the grid. R(θ) is an
overall rotation of the grid by an angle θ, the angle β
describes the relative rotation of the grid basis vector v
relative to u, and the phase φ represents a shift with re-
spect to a reference point. The activity of an equilateral,
unrotated triangular grid has λ1 = λ2 = λ, β = π/3 and
θ = 0. The set of the six vertices defined by the triplet
u, v, u − v and their opposite vectors forms an hexagon
that can be inscribed into an ellipse. The ratio between
the axes of the ellipse defines the ellipticity of the grid
( = 1 for equilateral grids). Hereafter, we study isosceles
grids, where the relation cos β = 1/√1 + 32 holds, but
our conclusions hold generally (see Appendix A). Finally,
for the purpose of analyzing coverage we take χ = 1 when
its argument is < 1 and χ = 0 otherwise, i.e. we are in-
terested in whether a neuron is active or not at a given
point (disregarding its strength of activity).
In a module, grid cells with similar spacing have similar
orientation, ellipticity and firing field size [11]. However,
parameters of the grids have an appreciable variability,
which we quantify using experimental data reported in
Stensola et al. [11] as follows. For each animal where
the distribution of grid parameters is available, we fit the
data with a sum of Gaussians. For each module, we used
one Gaussian for the period (mean λ, standard devia-
tion σλ) and one for the orientation (mean θ, standard
deviation σθ.) The two standard deviations are roughly
constant in the various modules.
Indeed, the Pearson
correlation coefficient is 0.21 between σλ and λ and 0.28
between σθ and λ. The standard deviation of the grid pe-
riod is about 6 cm. Assuming 10 modules and that the
smallest is about 40cm, the ratio σλ/λ goes from 0.01
to 0.15. The standard deviation of the orientation in a
module is about 0.03 rad. In the literature we were not
able to find the distribution of ellipticity in the popu-
2
lation within a single module. We know that ellipticity
also has a modular structure and that across a population
(all modules) the mean ellipticity is around 1.16 ± 0.003
[11]. In the following analysis we assume a standard de-
viation in ellipticity in the range 0.01-0.15, i.e. similar
to variability in grid spacing. Finally, we fixed the ratio
between the firing field width and the grid spacing at the
experimental value λ/(cid:96) ∼ 1.63 ± 0.04 [15].
FIG. 1. Deformations in grid parameters induce de-
phasing. (A) Spatial activity of a grid cell (a) and its pos-
sible deformations: dilation (b), rotation (c) and ellipticity
transformation (d). The activity after transformation (green)
is superimposed on the reference activity (gray). (B) Using
a set of grid parameters we divide space into unit cells repre-
sented by gray hexagons. A neuron with the same set of grid
parameters (a) has constant relative phase (black arrows) in
each unit cell. A neuron with different grid parameters (b)
has variable phases in different unit cells, e.g.
the grid is
rotated as in panel (A.c) and the center of the unit cell is
covered by a firing field in U1 but not in U2.
B. Dephasing and decorrelation of neuronal
activity
In order to cover an environment with grid cells, there
must be at least one active neuron at each point. The
average orientation θ, ellipticity and period λ within
a module define a tessellation of the plane into periodic
unit cells. Perfect periodicity would imply that once a
unit cell is covered, all of space is covered. However,
perfect periodicity is broken by the variability discussed
above, which results in deviations from the average grid.
Indeed, as shown in Fig. 1B, the pattern of firing fields
changes across unit cells, a phenomenon that we call "de-
phasing". Here we characterize this effect by computing
the correlation coefficient between the number of neurons
that are active at the center of two unit cells.
given by n(x) =(cid:80)N
The number of neurons active at a spatial point x is
i=1 ai(x), where ai is the spatial ac-
tivity of the i−th neuron given by Eq. (1) and N is the
number of neurons in the system. Consider a set of neu-
rons whose grid parameters are drawn from Gaussian dis-
tributions with standard deviations σλ, σθ and σ. We
θλ'l'cλlλ1λ2cU1U2ababdABcθ compute numerically the correlation (normalized to 1 for
coincident points) between the numbers of neurons n(x)
and n(y) active at different points x and y, by averag-
ing over statistical realizations (Fig. 2). The correlation
declines systematically with the separation in the grid
lattice. The corresponding correlation length L (defined
as the distance at which the correlation drops to 1/e)
decreases with the variance in the parameters of the grid
cells (Fig. 2) and is in the meter scale for the smallest
modules, which is within the behavioral range of a few
tens of meters found in rats [16 -- 18].
We can understand the asymptotic behavior of the cor-
relation function of the number of neurons active at two
spatial points at large separations as follows. We are
assuming that grid cells in a module fire independently.
Therefore, the correlation function of the number of neu-
rons active at x and y, ρ(n(x), n(y)), is equal to the cor-
relation function of the activity of a single generic neuron
a(x), averaged over the distribution of grid parameters,
ρ(a(x), a(y)). The mean activity of a single neuron is
obtained from Eq. (1) by averaging over the grid param-
eters. This quantity can be written as
(cid:90)
Qφ(y, x) ≡
a(y)a(x) dPθdPλdP ,
(3)
where Qφ(y, x) is the joint probability distribution that
a neuron is active both at y and at x. The distribution
depends parametrically on φ. The joint probability can
be computed as
Qφ(y, x) = Qφ(yx)Qφ(x) ,
(4)
where Qφ(yx) is the conditional probability that a neu-
ron is active at y if it is active at x (for a given φ). The
quantity Qφ(x) is the probability that a neuron is active
at x, again for a given φ.
(cid:104)a(x)(cid:105) =
a(x) dPθdPλdPdPφ ,
(2)
where dP(∗) represents the probability distribution for
the parameter (∗). As discussed above, orientation, pe-
riod and ellipticity of the grids follow a Gaussian distri-
bution whilst the spatial phase is uniformly distributed
in a unit cell. To compute the integral, we divide space
into unit cells and consider the center of the cell contain-
ing x as a reference point for the phase of the grid φ.
Because φ is uniformly distributed in the unit cell, and
because χ = 1 within the firing field and χ = 0 outside,
the integral over φ is a constant equal to the ratio be-
tween the area of a firing field and that of a unit cell,
i.e. π/2√3 ((cid:96)/λ)2. The remaining integrals are equal to
unity and we finally obtain (cid:104)a(x)(cid:105) = π/2√3 ((cid:96)/λ)2.
In order to compute the correlation function we need to
determine the quantity (cid:104)a(y)a(x)(cid:105). which can be written
as
(cid:90)
(cid:90)
(cid:104)a(y)a(x)(cid:105) =
Qφ(y, x)dPφ ;
3
In order
to evaluate the conditional probability
Qφ(yx), we divide space into unit cells using the mean
grid parameters. We consider a neuron with φ = (0, 0),
i.e. with a firing field centered at the origin, and analyze
the evolution of its phase φn in the unit cells centered
at y = (nλ, 0), n = 0, 1, . . . .
If the grid cell has the
same grid properties as the average grid, its phase will
be invariant, i.e. φn = φ, hence Qφ(yx) = 1 and the
correlation function will be a constant that does not de-
pend y. If there is variability, the phase of the grid will
be randomly distributed in the two-dimensional area of
the unit cell centered at (nλ, 0) as n increases. It follows
that Qφ(yx) → π/2√3 ((cid:96)/λ)2; (cid:104)a(y)a(x)(cid:105) → (cid:104)a(x)(cid:105)2
and the correlation asymptotically goes to zero as shown
in Fig. 2.
In Appendix B we discuss the behavior of the correla-
tion function in the absence of orientation variance. This
analysis is not relevant to describe the biological sys-
tem, where orientation variance is estimated to be about
0.03 rad, but constrains the definition of the correlation
length. In particular, we show that the threshold used
to define the correlation length ought to be in the range
[0.28, 1], which includes our choice of a threshold equal
to 1/e.
C. Coverage drives modularity
In order to cover an environment with a set of grid
cells, there has to be at least one active neuron at every
point. The correlation length L characterizes the scale
beyond which the numbers of active neurons become ap-
proximately independent. A region of size R is thereby
decomposed in R2/L2 regions whose coverage probabil-
ities are roughly independent of each other.
If each of
these is covered with probability p, the probability P of
covering the whole environment is
log(P ) = γ
R2
L2 log(p) ,
(5)
where γ is a constant that depends on the geometry of
the system.
The probability p of covering a correlation volume of
linear size L as a function of the probability of covering
a unit cell puc, was obtained numerically as follows. We
computed the covering probability of a circular environ-
ment of radius R using sets of grid cells characterized
by different puc. Results of the simulations are shown in
Fig. 3. For every value of the radius R the logarithm of
the covering probability rescaled over log(puc) does not
depend on puc (Fig. 3B). It follows that the probability
p of covering a correlation volume of linear size L can be
expressed as
log(p) = K log(puc) ,
(6)
where K is a function that depends only on L/λ for di-
mensional reasons.
4
FIG. 3. Covering probability of a correlation volume
of linear size L. (A) Covering probability P of a circular en-
vironment of linear size R computed numerically for different
environment size and covering probability of the unit cell puc.
The different values of puc have been obtained using sets of
grid cells made of a different number of neurons. (B) Results
of panel (A) collapse on a common curve when their loga-
rithm is rescaled by log(puc), justifying the functional form
introduced in Eq. 6. (C) Numerical covering probability p of
a correlation volume of linear size L have been used to obtain
an empirical description of the function K (L/λ) described in
Eq. 6 (red dots). The best fit (black line) is given by the
function K (x) = 0.73 + 13 log(x). Simulations parameters
are λ/(cid:96) =1.63, σθ = 0.04, σ = 0. In panels (A-B) we fixed
σλ/λ = 0.08 whilst the number of neurons N is 30 (magenta,
squares), 40 (light blue, circles), 50 (brown, stars). In panel
(C) we fixed N = 30 and σλ/λ has been varied to span the
different values of L observed in the biological system as de-
scribed in Fig 2E.
FIG. 4. The covering probability P decreases with the
variance of grid parameters and the size of the envi-
ronment. P is computed for circular environments of radius
R and for different grid variances. Colors represent differ-
ent values of one of the variances (orientation (A), ellipticity
(B), period (C)) while the other two are fixed (σλ/λ = 0.03,
σ = 0(A), σλ/λ = 0.02, σθ = 0.02 (B), σθ = 0.02, σ = 0
(C)). The case with no noise is in black. Numerical results
(colored symbols) match theoretical predictions (continuous
line) obtained by Eq. (8). The number of neurons N = 30.
more efficient, because with the same number of neurons,
and hence a fixed puc, it will have fewer gaps. Since the
correlation length decreases if the standard deviations
increase, we conclude that coverage drives modularity --
grid cells with similar period should have similar orien-
tations and ellipticities as observed experimentally [11].
FIG. 2. Variability of grid parameters induces decor-
relation in grid cells activity. The correlation coefficient
of the number of neurons active at the center of two unit cells
along the line x = nλ(1, 0), n ∈ {0, 1, . . .} for different values
of the standard deviations in the parameters of the grid cells.
In the plots x = x. Colors in the first row represent differ-
ent values of one of the variances (orientation (A), ellipticity
(B), period (C)) while the other two are fixed (σλ/λ = 0.03,
σ = 0(A), σλ/λ = 0.02, σθ = 0.02 (B), σθ = 0.02, σ = 0
(C)). The case with no noise is represented in black. Larger
variances lead to a rapid decrease in the correlation as a func-
tion of separation. (See Appendix B for details on panel C.)
The correlation length depends on three variances which we
varied in pairs obtaining contour plots (σθ = 0.01 (D), σ = 0
(E), σλ/λ = 0.01 (F)). The white points in panel (E) cor-
respond to the values of the standard deviations measured
in [11].
Over a range of grid variances that includes the ex-
perimentally measured values, we found that L/λ (cid:46) 20
(Fig. 2). In this range, we found numerically that
K (x) = c1 + c2 log(x)
(7)
(c1 = 0.73, c2 = 13) gives a good description of the data
(see Fig. 3C) .
Combining Eqs. (5), (6) and (7), we obtain :
(cid:18) L
(cid:19)
log(P ) = γ
R2
L2
K
λ
log(puc) .
(8)
To test this estimate, we numerically analyzed the cover-
ing probability of a circular environment of radius R by
N neurons whose grid parameters are drawn from Gaus-
sian distributions. We then checked if every point in the
environment is covered by at least one grid cell and we
averaged over realizations. Fig. 4 confirms the validity of
Eq. (8), with a proportionality constant γ = 0.804.
Fig. 4 and Eq. (8) show that the covering probability
of a region increases with the correlation length. In this
sense, a set of grid cells with a larger correlation length is
251020σθ=0.01σλ/λ00.10.10σσθ=0.0σλ/λ00.10.10σθ=0.0100.10.10σσλ/λσ0102010−210−1100x/λρ(n(x),n(0))01020x/λ01020x/λABσλ/λ=0.03,σ=0σλ/λ=0.02,σθ=0.02σ=0,σθ=0.02C=σθ0.02, 0.04, 0.06=σ0.0, 0.07, 0.15=0.02, 0.04, 0.12σλ/λDEF100101102100101102103104R/λlog(P)/log(p )100101102R/λ−log(P)10−610−410−2100-log(puc)=10−6-log(puc)= 410−5-log(puc)=10−35101510203040L/λucK=log(p)/log(p )ucABC0204000.20.40.60.81R/λP02040R/λ02040R/λABσλ/λ=0.03,σ=0σλ/λ=0.02,σθ=0.02σ=0,σθ=0.02C=σθ0.02, 0.04, 0.06=0.0, 0.07, 0.15=0.02, 0.04, 0.12σλ/λσ D. Gaps decline exponentially with the number of
neurons
We now quantify how the number of neurons N in a
module affects the probability of covering a range R. The
dependence on N in Eq. (8) occurs through the factor
puc. The random distribution of phases of grid cells [3]
dictates an exponential dependence between the proba-
bility puc and the number of neurons N .
Indeed, con-
sider N neurons that cover a unit cell of a d-dimensional
grid with a single gap. An additional neuron added with
a random phase will fail to overlap the gap with some
probability h < 1. If we add Q additional neurons inde-
pendently, the probability that they all miss the gap is
hQ, i.e. the probability of gap persistence declines expo-
nentially with the number of added neurons. Subleading
terms are captured by analyzing partial coverage with
each additional neuron (see Appendix C).
Thus,
for a large number of neurons in a two-
dimensional grid module, we expect that puc ∼ 1 −
exp (−αN ), where α is a positive constant that depends,
by dimensional analysis, on the ratio (cid:96)/λ. In the oppo-
site limit, when N is smaller than the area of the unit cell
divided by the area of the firing field, coverage cannot be
achieved and puc = 0, as confirmed numerically in Fig. 5.
In summary, the estimate for the probability P of cov-
ering a two dimensional circular region of radius R is
log(P ) = γ
R2
L2
K (L/λ) log(1 − eF(N )) ,
(9)
where the function F (N ) behave as just discussed, which
is validated by numerical simulations (Figs. 4, 5). On
the one hand, the probability of gaps in coverage declines
exponentially with N . On the other hand, the probability
of gaps in coverage of a range R increases exponentially
as (R/L)2K (L/λ), where L decreases as the variability
in a module increases. Hereafter, we balance these two
effects to estimate the number of neurons required to
cover space in modules of different mean periods.
E. Prediction: smaller period modules need more
neurons
Eq. (9) gives the relation between the number of neu-
rons Ni and the parameters of the i-th module. Since the
different modules vary systematically in their period, this
relation predicts an associated variation in the number of
neurons.
Assume that an animal encodes position within a re-
gion of size R2 that is common to all the modules, and
that the probability of covering space is the same at all
scales. As we showed above, the probability of gaps in
coverage declines exponentially with the number of neu-
rons, and the coefficient in the exponent depends on the
ratio (cid:96)/λ between the grid field width and the period.
It is established experimentally that this ratio is fixed
5
FIG. 5. The covering probability P increases with the
number of neurons N . (A) The numerical computation of
P for a unit cell (black dots) is combined with Eq. (9) to
predict P for an environment of size R/λ = 20 (black line).
Results of numerical simulations are in green. The function
1 − P asymptotically decays as exp(−αN ) with α ≈ 0.4 (red
the environment size R
line).
for different N . Results of the simulations (colored symbols)
match predictions (black lines) by Eq. (9). Parameters are
λ/(cid:96) =1.63, σθ = 0.04, σλ/λ = 0.08, σ = 0.
(B) The probability P vs.
dicted fraction of neurons in a given module, Ni/(cid:80)
among modules [10, 11]. Thus we can evaluate the pre-
i Ni,
where the denominator is a sum over modules, and Ni is
obtained by inverting Eq. (9).
The results of this prediction and a comparison with
the extant experimental data are shown in Fig. 6. The
theoretical predictions are given for a variety of ranges
and coverage probabilities, with the grid periods and
variabilities fixed from experimental data. Qualitatively,
the theory predicts for any choice of parameters that the
number of neurons should decline with the period of the
module, as also suggested by the data.
Responses from 4 -- 5 modules spanning up to 50% of
the dorsoventral extent of mEC feature a smallest period
of about 40cm and a ratio of 1.42 between consecutive
scales [11]. This suggests that there should be about 10
modules in total in the rat grid system with a maximum
period of about 10m. Fitting our theoretical predictions
to experimental data [11], we found that a range of a few
tens of meters can indeed be covered with a high cover-
age probability in the range 80-90%. Within the range
of parameters that allows this coverage in our model,
we predict a decrease of about 50 − 70% in the num-
ber of neurons between the first and the tenth module
(Fig. 6). Experimental uncertainties and possible biases
in recording from harder-to-reach modules with larger
periods, prevent stringent fits. Nevertheless, our theory
robustly states that the number of neurons should decline
with the period of the grid module.
III. DISCUSSION
A striking experimental observation about the grid sys-
tem in the entorhinal cortex is that it is organized in
discrete modules that share similar periods, orientations
and ellipticities [11]. Given this modular structure, the
AB02040N1−Punit cellR/λ=20100101102R/λ−log(P)N=30N=40N=5010−610−410−210010100−5 6
activity of grid cells with larger periods [24] : attractor
models indeed predict that networks with smaller num-
bers of neurons will drift more. Further data is needed to
confirm these indirect lines of evidence. Comprehensive
recordings from many grid modules are challenging be-
cause modules of a given period are not strictly localized
anatomically, and because ventral regions are harder to
record from. But such data will greatly illuminate mod-
els of the functional logic of the grid system, and will
further test our quantitative predictions.
ACKNOWLEDGMENTS
VB was supported by the Fondation P.G. de Gennes
at the ENS, Paris; by NSF grant PHY-1066293 at the
Aspen Center for Physics; and by NSF PoLS grant PHY-
1058202. VB and MV were supported by NSF Grant
PHY11-25915 at the KITP, Santa Barbara.
Appendix A: Covering probability of non-isosceles
grids
In the main text we analyzed the covering probability
of the grid system assuming isosceles grids; in this Section
we show that our results hold also in the case of general
grids.
The triangular lattice defining the spatial activity of
a grid cell is determined by linear combinations of two
elementary vectors v and u. The reference frame can be
chosen to have the vector u coinciding with the x-axis,
i.e. u = λ2(1, 0) and the vector v = λ1(cos(β), sin(β)),
where β is the angle formed by the two elementary vec-
tors (which can be restricted to the first quadrant). The
two positive numbers λ2 and λ1 are the moduli of the
two elementary vectors. The set of the six vertices de-
fined by the triplet u, v, u − v and their opposite vec-
tors, forms an hexagon that can be inscribed into an el-
lipse centered at the origin, whose general equation is
Ax2 + 2Bxy + Cy2 = 1 (see Fig. S1A). The (inverse
squared) length of the two axes of the ellipse is deter-
mined by the eigenvalues of the quadratic form and their
orthogonal directions are determined by the correspond-
ing eigenvectors.
An alternative parametrization of the ellipse is given
by : 1) the direction δ of the axes of the ellipse with
respect to the axes of the reference frame ; 2) the ratio
between the length of the two axes (i.e. the ellipticity
of the grid as defined in [11]) ; 3) the length λ/√ of the
axis parallel to the x-axis when δ = 0 (the other axis has
length √ λ). By requiring that the ellipse pass through
the three independent vertices u, v, u− v, we obtain the
relations
λ2 = λ√
F1
;
√
3
λ1 sin β =
2 λ2F1 ;
F1 ≡ cos2 δ + 1
F2 ≡ 1 − √3(cid:0) − 1
sin2 δ ;
(cid:1) sin δ cos δ ,
λ1 cos β = λ2F2
;
2
(A1)
FIG. 6. Number of neurons required for coverage de-
creases with the spatial period. (A) We used experimen-
tal data in [11] to estimate the number of recorded neurons vs
their spatial period (circles). We applied k -- means clustering
to identify four modules according to the spatial period. For
each module we computed the associated number of neurons
and plotted the mean and the standard deviation of the num-
ber of neurons N normalized over the number N1 of neurons
for the first modulus (black squares and lines). Theoretical
predictions given by Eq. (9) obtained with different values
of R and P (lines) are compatible with experimental data.
(B) We extrapolated the number of neurons over ten mod-
ules for different values of R, P . Simulation parameters are
λ/(cid:96) = 1.63, σθ = 0.04, σ = 0 and σλ = 6cm.
geometric progression of grid periods can be shown to
minimize the number of neurons required to provide a
specified spatial resolution [12, 13]. However, why would
a modular architecture be necessary in the first place?
In this paper, we have shown that efficient coverage of
space favors modularity.
To study how variability in the grid parameters within
a module would affect the probability of holes in cover-
age, we simply asked whether each neuron did or did not
fire above threshold at a given location. Alternatively,
we could sum firing profiles of grid cells to assess how
grid variability affects homogeneity of the population fir-
ing across space. Again, the key variable would be the
correlation in the expected number of action potentials
at each point in space. The overall probability of cov-
erage would be determined by a product of factors over
each correlation volume, leading to the same conclusion.
We chose to analyze coverage because any grid coding
scheme, e.g., [12, 13, 19 -- 23], requires neurons to be ac-
tive at each point in space. Thus, we view our approach
as setting a minimal requirement for a functioning grid
system for encoding location. Our model predicted that
there would be fewer neurons in modules with larger peri-
ods. We compared our theory with the actual numbers of
neurons recorded across modules, which should be taken
with caution because of potential biases in the recording
methods, especially for deeper structures in the brains.
Some additional evidence for a decrease of neurons with
the period of modules stems from the relatively smaller
size of the ventral entorhinal cortex (which is enriched
in large periods) relative to the dorsal region. Indirect
evidence also comes from the larger drifts seen in the
AB50100150012λ (cm)N/N105100.40.60.81moduleN/N1R=10m, P=0.8R=30m, P=0.8R=10m, P=0.9R=30m, P=0.9 which provide a general mapping between the free pa-
rameters λ, , δ of the ellipse and the free parameters λ1,
λ2, β of the vectors u and v. Ellipses with δ = 0 have
axes aligned with the coordinate system. Elementary
algebra shows that this condition corresponds to isosce-
les triangles with v = u − v, i.e. 2λ1 cos β = λ2 or
cos β = 1/√1 + 32. The special case of = 1 fixes
λ1 = λ2 = λ and cos β = 1/2, i.e. corresponds to equilat-
eral triangles. Note that the direction δ is related only to
the deformation of the hexagon defined by the elemen-
tary vectors and its variations do no affect the orientation
θ of the grid.
We generalize the analysis of the main text to cases
where the axes of the ellipse are not aligned with the
coordinate system (δ (cid:54)= 0), which generally corresponds
to scalene triangles. We choose a parametrization where
, δ and λ vary independently. The upshot is that the
results presented in the main text still hold. Specifically,
for fixed σλ/λ and σθ we compute the correlation coef-
ficient between the number of neurons that are active
at the center of two unit cells, as described in the main
text. Fig. S1 illustrates the results of the simulations for
different values of σ and σδ using general grids. As for
the other sources of variability, the correlation decreases
rapidly with the separation between the two centers and
the correlation length decreases as σ and σδ increase.
Finally, Fig. S1D also shows that the covering probabil-
ity conforms to the relation (5) presented in the main
text.
Appendix B: Analysis of the correlation function
with no orientation variance
In the main text we showed that, as long as there is
some variability in the orientation, the correlation func-
tion of the number of neurons active at two spatial points
tends to zero as the distance increases. Here we discuss
the asymptotic behavior of the correlation function in the
case without orientation variance in the grid parameters.
The derivation of the asymptotic correlation has been
outlined in Section II B; the absence of orientation vari-
ance affects the computation of the conditional probabil-
ity as follows. We divide space into unit cells using the
mean grid parameters and indicate with φn the phase in
the unit cells centered at y = (nλ, 0), n = 0, 1, . . . . In
the case in which the grid has the same orientation as
the mean grid but different period (λ(cid:48)
(cid:54)= λ), the phase
φn will gradually shift along the x−axis as n increases
but it will always belong to a one dimensional surface
with fixed y component equal to φy
0. For large n the
phase becomes randomly distributed along the segment
[(n − 1/2)λ, (n + 1/2)λ]. The resulting probability that
the point y is covered, given that x = (0, 0) is covered,
depends on φy
0. In particular, a grid that covers the point
0)2 be-
x will have an intersection of length 2
tween its firing field and the segment [−1/2λ, +1/2λ].
(cid:113)
(cid:96)2/4 − (φy
7
FIG. S1. Variability in ellipticity and direction δ in-
duces decorrelation of activity and decreases the cov-
ering probability. We perform the same analysis as in the
main text for the more general case δ (cid:54)= 0. (A) Represen-
tation of the grid activity with δ (cid:54)= 0. The six firing fields
defined by the triplet u, v, u − v and their opposite vectors
(green circles) have centers belonging to an ellipses (curved
red line), the ellipse axes (straight red lines) are rotated ro-
tated by an angle δ respect to the vector u aligned with the
x-axis. (B) The Pearson correlation coefficient is computed as
in Fig. 2 (σλ/λ = 0.01, σθ = 0.01, σ = 0.06). (C) Correlation
length for different values of σ and σδ computed as in Fig. 2
(σλ/λ = 0.01, σθ = 0.01).
(D) Covering probability com-
puted as in Fig. 4 (σλ/λ = 0.01, σθ = 0.01, σ = 0.06). We
find that increasing variability reduces the correlation length
and decreases the covering probability. Eq. (8) correctly de-
scribes the covering probability as a function of variance and
environment size.
Because of the previous argument, for large n this in-
terval will be uniformly distributed along the segment
[(n − 1/2)λ, (n + 1/2)λ] so it will cover the point y with
probability
(cid:113)
0)2
(cid:96)2/4 − (φy
λ
2
Qφ(yx) =
.
(B1)
Furthermore, the probability to have a neuron active at
ABCDσδ=0.03σδ=0.06σδ=0.10σθ=0.01,σλ/λ,σ=0.01=0.06σδ=0.03σδ=0.06σδ=0.10010203010−2100x/λρ(n(x),n(0))0204000.51R/λPσθ=0.01,σλ/λ,σ=0.01=0.06uvv-uδσθ=0.01,σλ/λ=0.01σδσL/λ00.050.100.050.10.153510150.15 x is
Qφ(x) =
(cid:40)
0
1
0 )2 + (φy
0)2 > ((cid:96)/2)2 ,
if (φx
(φx
0 )2 + (φy
0)2 ≤ ((cid:96)/2)2 .
8
(B2)
Combining the previous results we obtain (cid:104)a(y)a(x)(cid:105) →
√
3(cid:96)3
4
9λ3 . Hence, if the orientation variance is zero the corre-
lation coefficient between two distant points reaches the
asymptotic value
ρ(a(∞), a(0)) =
8(cid:96)
√
3πλ − π(cid:96)2
2
√
1 − π(cid:96)2
2
3λ2
3λ2
.
(B3)
This has been confirmed numerically in Fig. S2. The
same argument holds if ellipticity variance is present.
In the main text we defined the correlation length of
a grid system as the distance at which the correlation
in the number of active cells falls below the threshold
1/e. Results of the present Section qualify the range in
which this threshold could be chosen. Indeed, for a given
(cid:96)/λ, Eq. (B3) gives the asymptotic value of the corre-
lation function when no variance in the orientation is
in the biological system (cid:96)/λ ≈ 1/1.63 and
present, e.g.
the asymptotic value of the correlation is about 0.28 (see
Fig. S2B). Because of the effect described above, if the
threshold used to define the correlation length is chosen
below this value, the correlation length will depend only
on the orientation variance. Hence, in order to assess the
length scale of correlation that is affected by the variance
in all the grid parameters, a threshold slightly above the
asymptotic value should be chosen (in the main text we
used 1/e.) This choice is relevant because it captures
the dependence of the covering probability on the vari-
ance in the grid parameters. In fact, if our definition of
the correlation length is used to analyze the covering of
a system, the analytical results obtained from our ap-
proach are in agreement with direct numerical analysis
performed with (see Fig. 4) and without (see Fig. S2D)
orientation variance.
Appendix C: Covering probability of a unit cell
We discuss the covering probability of a unit cell in one
dimension. We take each grid cell to be active in intervals
of length (cid:96), regularly spaced with centers at distance λ
apart. A unit cell is given by an interval of length λ.
The covering probability of a unit cell by N grid cells is
analogous of that of covering a region of length λ by N
intervals of length (cid:96) with periodic boundary conditions
on the region. This probability distribution has been
computed analytically in [25] and reads
(cid:18)
(cid:19)
(cid:96)
λ
N(cid:88)
k=0
(cid:18)N
(cid:19)
(−1)k
k
P
N,
=
f (k) ,
(C1)
otherwise. For large N this relation reduces to the result
where f (k) = (1 − k(cid:96)/λ)N−1 if k(cid:96) < λ and f (k) = 0
used in the main text P = 1 − N (1 − (cid:96)/λ)N−1.
FIG. S2. The correlation length depends on the be-
havior above the asymptotic value of the correlation.
(A) When there is no variance in grid orientations, the corre-
lation function between the number of active cells at two lo-
cations, ρ(n(x), n(0)), approaches a nonzero asymptotic value
that depends on (cid:96)/λ. The solid lines indicate the theoretical
prediction of the asymptotic values from Eq. (B3). Simula-
tion parameters are σλ/λ = 0.07, σθ = 0, σ = 0. (B) The
asymptotic value of the correlation in the absence of orien-
tation variance is predicted by Eq. (B3) (black line). Rep-
resentative values corresponding to the three curves in panel
A are marked by the colored symbles. Note that there is a
maximum value in the asymptotic correlation as a function of
l/λ. (C) The correlation decreases faster when the variance
in the period increases ((cid:96)/λ = 0.61, σθ = 0, σ = 0). (D) Nu-
merical simulations (colored symbols) determine the covering
probability of the environment for the different variances in
the grid parameters and environment sizes. The numerics are
accurately predicted by Eq. (9) of the main text (solid lines)
in which the correlation length for a grid system was assessed
as the distance at which ρ in panel C decreased to 1/e. This
threshold is always larger than the asymptotic value of the
correlation (see main text).
The proof of (C1) is presented below for the sake of
completeness. Because we have periodic boundary con-
ditions on the region to be covered, we can regard it as
a circle of unit length and we can take the intervals of
length ζ = (cid:96)/λ to be arcs on this circle. The arcs are la-
belled by their order of occurrence in the anti-clockwise
01020100σθ=0,σ=0,σλ/λ=0.07x/λρ(n(x),n(0))=0.12=0.61=0.8600.5100.10.20.30.4ρ(n(∞),n(0))01020100x/λρ(n(x),n(0))σλ/λ=0.02σλ/λ=0.07σλ/λ=0.120204000.51R/λPc(R)σθ=0,σ=0,=0.6110-110-1ABCD direction around the circle, starting by convention from
the north pole. The arcs are identified by their initial
position ; there is a gap after the r-th arc if the distance
between the initial positions of the r-th and the r + 1-th
arcs is larger than the size of the arcs. For convenience,
we rigidly translate all the arcs so that the first one is
positioned at the north pole -- this convention does not
affect the probability of coverage.
Consider N random arcs of length ζ on the circle.
Draw k arbitrary arcs from this set (say (r1, r2, . . . rk),
with k ≤ N ). Let f (k) be the probability that each arc
in this randomly selected subset is followed by a gap, ir-
respective of the state (followed by a gap or not) of all the
other arcs. From the f (k)'s, the probability (C1) of leav-
ing no gaps is computed as follows. First, let Q(ng, nu) be
the probability that ng prescribed arcs are each followed
by a gap and nu prescribed arcs are each not followed by
a gap, with the rest of the N arcs in unspecified states.
Then, Q(ng, 1) = f (ng) − f (ng + 1) because f (ng) in-
cludes the probability that the extra arc might be gapped
or ungapped, while f (ng + 1) subtracts the probability
that the extra arc is in fact gapped. By a similar reason-
ing we obtain Q(ng, 2) = Q(ng, 1) − Q(ng + 1, 1) and so
on recursively up to ng + nu = N . Simple algebra shows
then that the probability of leaving no gaps is
P
N,
= Q(0, N ) =
f (k) .
(C2)
(cid:18)
(cid:19)
(cid:96)
λ
N(cid:88)
k=0
(cid:18)N
(cid:19)
(−1)k
k
9
The formula (C2) leaves us to determine the expres-
sion of f (k), which is done as follows. When an arc, say
r, is followed by a gap, we rigidly shift backward (clock-
wise) all the following arcs up to the last (N -th) by an
amount ζ. Because the r-th arc is followed by a gap, the
state of all the arcs other than r is not affected by this
backward shift and we are left with a final region of size
ζ that does not contain any initial position of the arcs
(see Fig. S3). Note that whether the N -th arc is gapped
or ungapped before this shift corresponds to whether or
not the last arc partially overlaps with the final region
of size ζ. The probability of distributing N − 1 initial
positions of the arcs (other than the first one fixed at the
origin) in a region of size 1 − ζ gives f (1) = (1 − ζ)N−1.
The reasoning for f (2) is similar. If the two prescribed
gapped arcs are r1 and r2 > r1, we first shift backward
by ζ all the arcs following r1 and then again by ζ those
following r2. We are then left with a final unoccupied
region of size 2ζ. The crucial point is that the state of all
the arcs other than r1 and r2 is again unaffected. We can
then compute f (2) = (1 − 2ζ)N−1 as the probability of
distributing N−1 initial positions of the arcs in the avail-
able region of size 1 − 2ζ. Generalizing the reasoning to
k arcs gives the expression f (k) = (1 − kζ)N−1, provided
the total length of the arcs is smaller than the length of
the circumference, i.e. kζ ≤ 1, otherwise f (k) = 0. That
completes the proof and yields Eq. (C1).
[1] E. C. Tolman, Psych. Review 55, 189 (1948).
[2] J. D. J. O'Keefe, Brain Research 34, 171 (1971).
[3] T. Hafting, M. Fyhn, S. Molden, M.-B. Moser, and E. I.
Moser, Nature 436, 801 (2005).
[4] J. B. Ranck, Society for Neuroscience Abstracts 10
(1984).
[5] E. Kropff, J. E. Carmichael, M.-B. Moser,
and E. I.
Moser, Nature 523, 419 (2015).
[6] F. Savelli, D. Yoganarasimha, and J. J. Knierim, Hip-
pocampus 18, 1270 (2008).
[7] T. Solstad, C. N. Boccara, E. Kropff, M.-B. Moser, and
E. I. Moser, Science 322, 1865 (2008).
[8] E. I. Moser, Y. Roudi, M. P. Witter, and C. Kentros,
Nature Publishing Group 15, 466 (2014).
[9] R. Muller and J. Kubie, Journal of Neuroscience 7, 1951
(1987), cited By 651.
[10] C. Barry, R. Hayman, N. Burgess, and K. J. Jeffery,
Nature neuroscience 10, 682 (2007).
[11] H. Stensola, T. Stensola, T. Solstad, K. Froland, M.-B.
Moser, and E. I. Moser, Nature 492, 72 (2012).
[12] X.-x. Wei, J. Prentice, and V. Balasubramanian, eLife ,
e08362 (2015).
[13] A. Mathis, A. V. Herz, and M. Stemmler, Neural Comp.
24, 2280 (2012).
[14] M. Barth´elemy, Physics Reports 499, 1 (2-11).
[15] L. M. Giocomo, S. a. Hussaini, F. Zheng, E. R. Kandel,
M.-B. Moser, and E. I. Moser, Cell 147, 1159 (2011).
[16] D. E. Davis, J. T. Emlen, and A. W. Stokes, Journal of
Mammalogy , 207 (1948).
FIG. S3. Shift of the arcs following the gap affects
only the covering of the arc that is followed by the
gap. (A) Circle of unit length with seven arcs of which only
the initial position has been represented (red lines at angle
rk). Each arc has length ζ = 1/4 and the 5-th arc is followed
by a gap (green region). (B) the 6-th and 7-th arc have been
rotated clockwise by an angle π/2 that corresponds to an arc
of length ζ. The rotation shifts the gap to the region before
the first arc and does not affect the state of the 6-th and 7-th
arc.
[17] S. E. Braun, Journal of Mammalogy 66, 1 (1985).
[18] N. A. Slade and R. K. Swihart, Journal of Mammalogy
64, 580 (1983).
[19] I. R. Fiete, Y. Burak, and T. Brookings, J. Neurosci.
28, 6858 (2008).
[20] Y. Burak and I. R. Fiete, PLoS Comp. Bio. 5, e1000291
(2009).
[21] S. Sreenivasan and I. R. Fiete, Nature 14, 1330 (2011).
[22] M. Stemmler, A. Mathis,
and A. V. Herz, bioRxiv
AB7r6r5r4r3r2r1r7r6r5r4r3r2r1r (2015).
86, 827 (2015).
[23] D. Bush, C. Barry, D. Manson, and N. Burgess, Neuron
[25] W. L. Stevens, Ann. Eugenics 9, 315 (1939).
87, 507 (2015).
[24] K. Hardcastle, S. Ganguli, and L. M. Giocomo, Neuron
10
|
1604.02404 | 1 | 1604 | 2016-04-08T17:15:29 | Comparison of Network Analysis Approaches on EEG Connectivity in Beta during Visual Short-Term Memory Binding Tasks | [
"q-bio.NC"
] | We analyse the electroencephalogram signals in the beta band of working memory representation recorded from young healthy volunteers performing several different Visual Short-Term Memory (VSTM) tasks which have proven useful in the assessment of clinical and preclinical Alzheimer's disease. We compare network analysis using Maximum Spanning Trees (MSTs) with network analysis obtained using 20% and 25% connection thresholds on the VSTM data. MSTs are a promising method of network analysis negating the more classical use of thresholds which are so far chosen arbitrarily. However, we find that the threshold analyses outperforms MSTs for detection of functional network differences. Particularly, MSTs fail to find any significant differences. Further, the thresholds detect significant differences between shape and shape-colour binding tasks when these are tested in the left side of the display screen, but no such differences are detected when these tasks are tested for in the right side of the display screen. This provides evidence that contralateral activity is a significant factor in sensitivity for detection of cognitive task differences. | q-bio.NC | q-bio | Comparison of Network Analysis Approaches on EEG Connectivity in
Beta during Visual Short-Term Memory Binding Tasks
Keith Smith1,2, student member, IEEE, Hamed Azami1, student member, IEEE,
Javier Escudero1, member, IEEE, Mario A. Parra2, John M. Starr2
Abstract -- We analyse the electroencephalogram signals in
the beta band of working memory representation recorded from
young healthy volunteers performing several different Visual
Short-Term Memory (VSTM) tasks which have proven useful
in the assessment of clinical and preclinical Alzheimer's disease.
We compare network analysis using Maximum Spanning Trees
(MSTs) with network analysis obtained using 20% and 25%
connection thresholds on the VSTM data. MSTs are a promising
method of network analysis negating the more classical use of
thresholds which are so far chosen arbitrarily. However, we find
that the threshold analyses outperforms MSTs for detection
of functional network differences. Particularly, MSTs fail to
find any significant differences. Further, the thresholds detect
significant differences between shape and shape-colour binding
tasks when these are tested in the left side of the display screen,
but no such differences are detected when these tasks are tested
for in the right side of the display screen. This provides evidence
that contralateral activity is a significant factor in sensitivity
for detection of cognitive task differences.
I. INTRODUCTION
Alzheimer's Disease (AD) is the most common form of
dementia in the world and, due to the increasing age of
the population, the number of people affected is likely to
dramatically increase in the years to come. Visual Short-Term
Memory Binding (VSTMB) tasks are potentially useful in
the detection of AD [1][2]. AD patients perform significantly
worse at shape-colour binding tasks than shape only tasks,
whereas healthy old adults show no such diminished ability
[3]. Further, no significant impairment is found in ability
of VSTMB tasks due to non-AD dementias [1] and major
depression [2], suggesting specificity of impairment to AD.
Still,
there are many different and subtle factors which
may affect cognitive tasks performance. Thus, we need to
understand better the neurophysiological correlates of brain
activity during these tasks in order to work towards a rigorous
framework for assisting AD detection at preclinical stages.
6
1
0
2
r
p
A
8
]
.
C
N
o
i
b
-
q
[
1
v
4
0
4
2
0
.
4
0
6
1
:
v
i
X
r
a
*This work was partially supported by the Engineering and Physical
Sciences Research Council. MAP was awarded an MRC Centenary Early
Career Awards #MRC-R42552. This work was conducted within the context
of The University of Edinburgh Centre for Cognitive Ageing and Cognitive
Epidemiology, part of the cross council Lifelong Health and Wellbeing Ini-
tiative (MR/K026992/1). MAP work is currently supported by Alzheimer's
Society, Grant #AS-R42303.
1Keith Smith, Hamed Azami
are with
the
Institute of Digital Communications, School of Engineering,
University of Edinburgh, King's Buildings, Edinburgh, UK, EH9
[email protected], [email protected],
3JL.
[email protected]
Javier Escudero
and
2Keith Smith, Mario A. Parra and John M. Starr are with Alzheimer Scot-
land Dementia Research Centre, University of Edinburgh, 7 George Square,
9JZ. [email protected],
Edinburgh,
[email protected]
EH8
Analysis of Electroencephalogram (EEG) signals is par-
ticularly relevant for clinical detection of brain pathology in
large at-risk populations, as required for AD. In this study
we analyse EEG signal data of young healthy volunteers
performing four distinct VSTM tasks, two of which involve
memorising only shapes and the other two are VSTMB tasks,
involving joint shape-colour binding memorisation. Here, we
use network theory analysis in order to uncover differences
affecting ability in these tasks.
Network theory is a widely applied analytical framework
for studying interdependent phenomena [4] which is fast
becoming a standard approach for complementing functional
connectivity analysis in the brain [5][6]. A network consists
of a set of nodes with connections formed between them.
Network analysis is naturally suited to applications in EEG
connectivity analysis where the electrodes form a bijective
mapping with nodes in a network and the connections
are defined by a similarity or dependency measure applied
between pairs of EEG signals. The values acquired from
these measures are in the form of an nxn weighted adjacency
matrix for a network with n nodes. Applying the similarity
measure between nodes i and j obtains the ith row and jth
column entry of the adjacency matrix [4]. The networks
obtained from similarity measures are complete, meaning
connections exist between all possible pairs of nodes, and
weighted, where measures give magnitudes between 0 and
1. Subjecting the resulting matrices to a threshold is desirable
since it simplifies analysis and discards many uninformative
low-weight connections. However, there is currently no ob-
jective threshold for binarising the networks. This leads to
study-by-study differences in choices leading to different and
sometimes conflicting results [7].
A promising branch for unbiased network analysis is the
Maximum Spanning Tree (MST) [8][9]. A spanning tree
is an acyclic, simple, connected, sub-network that connects
to every node in the network and the maximum spanning
tree is a spanning tree such that the sum of the weights
of the connections in the original network included in the
spanning tree are maximised [10]. This representation, when
analysed in simulated networks, has been shown to be robust
to underlying changes in the network and it overcomes the
problem of threshold bias [9].
In this study we aim to use network analysis of EEG
signals to look at differences occurring in brain activity when
performing similar, but distinct visual short-term memory
cognitive tasks. We want to understand more about these
tasks in order to eventually test their application in detection
of AD. In comparing the MSTs with standard threshold
techniques, we wish to test suitability of MSTs for detecting
changes in functional brain networks.
II. MATERIALS
A. Subjects
EEG signals were recorded for 23 healthy young volun-
teers participating in different VSTM tasks. Of the volun-
teers, five were left-handed and eight were women. Written
consent was given by all subjects and the study was approved
by the Psychology Research Ethics Committee, University of
Edinburgh.
B. Tasks
We consider a subset of four tasks of a larger study
involving eight tasks. A schematic diagram of the tasks is
shown in Fig. 1 which also gives examples of the uncommon
types of objects being probed. The positions of the objects
were randomised separately for study and test displays to
ensure that position was not a factor in memorisation. Each
participant completed 8 practice trials followed by 170 test
trials for each of the tasks. In half of the trials the objects
of the study display were the same as the objects of the
test display, while in the other half they were different. The
test was then to decipher whether or not the objects in the
study and test displays were the same which the volunteers
indicated by pressing buttons with both hands. There were
three objects in each hemisfield of the display screen. There
are four distinct tasks distinguished by two binary conditions:
i) single feature shape tested in left Hemisphere Response
(HPR), ii) single feature shape tested in right HPR, iii) shape-
colour binding tested in left HPR, iv) shape-colour binding
tested in right HPR. Due to the contralateral behaviour of
the brain, left and right hemisfield tests correspond to right
hemisphere response (HPR) and left HPR, respectively.
C. Recordings
Only the trials with correct responses and no serious
artefacts were kept since incorrect responses would not
inform on memory binding activity. In a few cases, no useful
data was available for a volunteer performing one of the tasks
resulting in an unequal number of volunteers per task.
Fig. 1. Structure of the VSTM tasks
We analysed the working memory representation, consist-
ing of study display and maintenance periods, since this is
found to be where differences in brain activity are most
prevalent [12]. The EEG data was collected using NeuroScan
version 4.3. This consisted of epochs with length of 1
second with -0.2 seconds of pre-stimulus recordings at 250
samples per second. A bandpass of 0.01-80 Hz was used in
recording. Forty EEG channels were recorded from common
EEG sites, the majority of which were international 10/20
sites. Only thirty channels were kept for our purposes. The
ten discarded channels consisted of four ocular channels,
two linked mastoid reference channels and four which were
discarded due to systematic noise (T5, T6, FT9 & FT10). Fig.
2 shows an abstract simplification of the EEG electrodes.
III. METHODS
A. Signal Processing
Pre-processing, frequency analysis and connectivity anal-
ysis were performed using FieldTrip [13]. First, the 30 chan-
nels were re-referenced using an average reference which is
more electrophysiologically silent [14]. Frequency analysis
was then implemented from 0 seconds onwards using the
multi-taper method with Slepian sequences and 2 Hz spec-
tral smoothing. A one second zero-padding was applied to
achieve 0.5 Hz resolution and the data was partitioned into
five frequency bands. We focus here on β (12.5 -32 Hz)
due to its utility for AD detection [15] and sensory/memory
integration [16]. After this, the debiased, Weighted Phase-
Lag Index (dWPLI) [17], an improved form of the Phase-
Lag Index [18] for small sample sizes, was applied to obtain
one connectivity matrix per trial. This similarity measure
was chosen for its robustness to volume conduction effects
as well as its ability to measure time-lagged inter-signal
dependence which we assume is important for inter-regional
communication in the brain [14].
B. Network Theory
1) Maximum Spanning Trees: From the resulting adja-
cency matrices, the MSTs were computed using an algorithm
based on Kruskal's algorithm [19]. This adds the strongest
weights in the network, as binary values, one by one to
an empty nxn matrix. At each step this MST matrix is
checked for cycles. If no cycles are present, the next strongest
connection is added and the algorithm continues. If a cycle
is created from adding a connection,
that connection is
discarded before the algorithm continues. Cycles are checked
using the property that, for a simple graph, G,
G contains a cycle ⇐⇒ 0.5T race(L) ≥ Rank(L) + 1,
where L is the Laplacian matrix of G. Once the MST matrix
has n − 1 connections with no cycles, i.e. is a spanning tree
for the underlying network, the algorithm stops.
2) Threshold Binarised Networks: Threshold Binarised
Networks (TBN) computed from thresholds keeping the 20%
and 25% of strongest connections, rounded to the closest
integer, were obtained for analysis and comparisons. Such
conditions. A paired t-test was also run for shape only vs.
shape-colour binding in left HPR and shape only vs. shape-
colour binding in right HPR. We report here only on p-values
which are significant at the 5% level.
Notably, there were no significant differences found in
the statistical analysis for the MST networks in the whole
network measures.
1) Two-way ANOVA: For the 20% TBNs, significant
interaction, p = 0.0100, was found for L between the left-
right and shape-binding conditions. For the 25% Threshold,
significant difference was found in the Left-Right grouping
for C at p = 0.0184, but none for interaction between Left-
Right HPR and Shape-Bind conditions as found in the 20%
threshold case, particularly the L p-value was above 0.1.
2) Paired t-test: For the 20% TBNs, in shape vs. binding
conditions for left and right HPR separately, it was found
that every metric found significant differences in the right
HPR conditions: L, p = 0.0243; C, p = 0.0330; Maximum
Degree, p = 0.0405. In contrast, there was no significant
difference found in the left HPR conditions. For the 25%
TBNs, again significant differences were found in shape vs.
binding in the right HPR condition with none present in the
left HPR condition. These were for C at p = 0.0080 and
Maximum Degree at p = 0.0018.
Since we assume that the functional connectivity networks
during different task performances are different in reality,
these results suggest that the MST is not a suitable method
for detecting differences at the level of VSTM task perfor-
mance in EEG signals. It may be that 30 electrodes is too
few and that higher density EEG would prove beneficial for
MST analysis. We found that some cases of noisy activity of
a single node for a patient during one of the tasks proved to
be a real problem for the MSTs. This resulted in cases with
very high degree nodes, in some cases even hub like networks
with one 29-degree node connected to twenty nine 1-degree
nodes. Noise removal attempts could be implemented to
suppress this, however this would only introduce the arbitrary
choices of threshold that the MST is put forward to avoid.
the MSTs may underrepresent strong re-
gional activity in the network by the simple fact that no
cycles and thus no clustering is allowed in the MST. Such
activity may be vital in understanding the differences in
activity between different VSTM task conditions. Likewise,
MSTs may overrepresent weak activity in the network where
connections which are much weaker than others may make
it into the MST by virtue of the fact that the node which it
connects to the network is underactive during the tasks.
Furthermore,
TBNs prove more robust
to noise simply due to the
greater number of connections, and thus information, present.
However, differences in results were apparent even when
a fairly small increase of threshold was implemented, fur-
ther exemplifying the threshold problem. All three metrics
showed significant differences in right HPR in the 20% case,
whereas only C and Maximum Degree showed significant
differences in the 25% case. Yet, the 25% values were more
significant
than the 20% values. Some of these may be
explained by the fact that L converges to 1 as connections
Fig. 2. Simplified Map of channels.
thresholds are chosen suitably low to enhance comparability
with MSTs.
3) Whole Topology Network Measures: We computed
several common network measures for the whole topology
of the networks. Correlations exist between the diameter
and leaf fraction of the MST with the characteristic path
length, L, and the local clustering coefficient, C, of TBNs,
respectively [9]. We computed these values alongside the
values of the maximum degree for both MSTs and TBNs.
The Diameter of an MST is the longest shortest path
between all pairs of nodes. This has an intuitive link to the
characteristic path length, L, from standard network analysis,
defined as the average shortest path length in the network,
since a smaller diameter implies that there is a shorter route
from any node to any other. The Leaf Fraction is the fraction
of 1-degree nodes in the MST. The link to the local clustering
coefficient, C, of standard networks, defined as average over
all nodes of the probability that any two nodes each sharing
a connection with a given node also share a connection,
is less intuitive. Though it can partially be seen in that a
high leaf fraction alludes to dense pockets of connectivity
which implies higher clustering. The Maximum Degree of
the network is the highest degree in the network where the
degree of a node is the number of connections which that
node shares in the network.
4) Node-specific Network Measures: Noticing that,
in
terms of activity related to different tasks, the differences
between networks may be quite subtle and so more obvious
at a node-specific level, we also computed the eigenvector
centrality. This measure gives high values for both high
degree nodes and nodes connected to high degree nodes, thus
measuring the importance of that node in the connectivity of
the network [4]. This was computed for nodes O1, O2, P3
and P4 (see Fig. 2), since the parietal and occipital regions
are instrumental regions for these tasks [12]. All of the
measures were computed either using the Brain Connectivity
Toolbox (BCT) [20] or else using straightforward calcula-
tions in MATLAB.
IV. RESULTS AND DISCUSSION
A. Whole Topology results
A metric value was computed for the trial-average network
of every volunteer during each of the tasks. These were then
subjected to a two-way ANOVA test to look for differences
between left and right HPR conditions, shape only and shape-
colour binding conditions, and interaction between these
increase and C and maximum degree converge to 0 as
connections decrease.
We also adjusted our p-values by implementing the false
detection rate at q = 0.05. We note that the matter of depen-
dency of different measures in networks for a given threshold
is somewhat unclear, thus discretion in interpretation must
be advised. Results indicated that only the values for C and
maximum degree in the 25% TBNs are reliable results of a
significant difference.
B. Node-Specific Results
The p-values from paired t-tests were evaluated for Eigen-
vector Centrality in nodes P3, P4, O1, and O2 for shape vs.
shape-colour binding in the left and right HPR separately.
1) MSTs: Similarly to the global network measures, no
significant differences were found for the MST nodes.
2) 20% TBNs: For the 20% threshold, there was sig-
nificant difference for P3 and O2 in the right HPR with
p = 0.0481 and p = 0.0417, respectively. Interestingly,
a significant difference was also found in the left HPR
condition for P4 with p = 0.0064.
3) 25% TBNs: For the 25% Threshold, significant dif-
ference was found in right HPR for P3 with p = 0.0443.
Signficant difference was found in left HPR for P4 with
p = 0.0227 and also in O2 with p = 0.0182. The results
found here agree for P3 and P4 but conflict with O2, again
exemplifying threshold bias.
Node specific metrics may be more helpful for detecting
differences in left HPR. A more in depth analysis on a
broader range of nodes would be helpful
to understand
these values and aid towards the framework of network
theory as a tool for understanding interdependent functional
activity in the brain during cognitive tasks. Interestingly
left HPR conditions show a difference in P4, in the right
hemisphere of the brain, providing evidence that activity
may be interdependent between hemispheres regardless of
directed stimulation.
V. CONCLUSIONS
We found differences in network topologies of functional
connectivity between performance of shape only and shape-
colour binding tasks for right hemisphere response. This
contrasted with a lack of evidence of differences for left
hemisphere response, although at the level of node-specific
measures some evidence was found for this. The right pari-
etal region, node P4, was sensitive to differences in left HPR.
Contralateral activity plays an important role in the VSTM
tasks. Particularly, tasks for right hemisphere response may
prove more sensitive to short-term memory as a potential test
for AD. We also found that MSTs were unable to pick up on
the differences clearly present in TBNs. Because of the small
sample size of volunteers, these findings would benefit from
more studies before conclusions can be drawn. The next step
pertaining to MST sensitivity would be to keep information
of weights for the connections in the MST to keep more
information of the importance of connections. Further, an
unbiased threshold is in development [21].
REFERENCES
[1] S.D. Sala, M.A.Parra, K. Fabi, S.Luzzi, S. Abrahams, "Short-term
memory binding is impaired in AD but not in non AD dementias ",
Neuropsychologia, 50: 833-840, 2012.
[2] M.A. Parra, S. Abrahams, R.H. Logie, S.D. Sala, "Visual short-term
memory binding in AD and depression", Journal of Neurology, 257:
1160-1169, 2010.
[3] M.A. Parra, S. Abrahams, K. Fabi, R.H. Logie, S. Luzzi, S. D. Sala,
"Short-term memory binding deficits in Alzheimer's Disease", Brain,
doi:10.1093/brain/awp036, 10 pages, 2009.
[4] M.E.J. Newman, "Networks", Oxford University Press, Oxford, 2010.
[5] E. Bullmore, O. Sporns, "Complex brain networks: graph theoretical
analysis of structural and functional systems", Nature, 10:186-198,
2009.
[6] D. Papo, M. Zanin, J.A. Pineda-Pardo, S. Boccaletti, J.M. Buld´u,
"Functional brain networks: great expectations, hard times, and the
big leap forward", Phil. Trans. R Soc. B, 369(1653): 20130525, 2014.
[7] B. Tijms et al., "Alzheimer's disease: connecting findings from graph
theoretical studies of brain networks ", Neurobiology of Aging, 34:
2023-2036, 2013.
[8] K. Ciftci, "Minimum Spanning Tree reflects the alterations of the
default mode network during Alzheimer's Disease", Annals of Biomed-
ical Engineering, 39(5): 1493-1504, 2011.
[9] P. Tewarie et al., "The minimum spanning tree: an unbiased method
for brain network analysis ", Neuroimage, 104: 177-188, 2015.
[10] C.J. Stam, P.Tewarie, E. Van Dellen, E.C.W.vanStraaten, A. Hille-
brand, P. Van Mieghem, "The Trees and the Forest: Characterization of
complex brain networks with minimum spanning trees", International
Journal of Psychophysiology, 92: 129-138,2014.
[11] M. Vourkas, E. Karakonstantaki, P. Simos, V. Tsirka, M. Antonakakis,
M. Vamvoukas, C.Stam, S. Dimitriadis, S. Micheloyannis, "Simple
and difficult mathematics in children, A minimum spanning tree EEG
network analysis", Neuroscience Letters, 576:28-33, 2014.
[12] M.A. Parra, S. Della Sala, R.H. Logie, A.M. Morcom, "Neural
correlates of shape-color binding in visual working memory", Neu-
ropsychologia, 52(0): 27-36, 2014.
[13] R. Oostenveld, P. Fries, E. Maris and J-M. Schoffelen, "FieldTrip:
Open Source Software for Advanced Analysis of MEG, EEG, and
Invasive Electrophysiological Data", Computational Intelligence and
Neuroscience, Volume 2011, 156869, 9 pages, 2011.
[14] E. van Diessen, T. Numan, E. van Dellen, A.W. van der Kooi, M.
Boersma, D. Hofman, R. van Lutterveld, B.W. van Dijk, E.C.W. van
Straaten, A. Hillebrand, C.J. Stam, "Opportunities and methodological
challenges in EEG and MEG resting state functional brain network re-
search", Clinical Neurophysiology, doi:10.1016/j.clinph.2014.11.018,
2014.
[15] C.J. Stam, Y. van der Made, Y.A. Pijnenburg, & P. Scheltens, "EEG
synchronization in mild cognitive impairment and Alzheimer's dis-
ease", Acta Neurol Scand, 108(2): 90-96, 2003.
[16] A.K. Engel, D. Senkowski, T.R. Schneider, "Multisensory Integration
through Neural Coherence", In: Murray MM, Wallace MT, editors,
"The Neural Bases of Multisensory Processes", Boca Raton (FL): CRC
Press; 2012.
[17] M. Vinck, R. Oostenveld, M. van Wingerden, F. Battaglia, C.M.A.
Pennartz, "An improved index of phase-synchronization for electro-
physiological data in the presence of volume-conduction, noise and
sample-size bias", NeuroImage,55:1548-1565, 2011.
[18] C.J. Stam, G. Nolte, A. Daffertshofer,"Phase Lag Index: Assessment
of Functional Connectivity From Multi Channel EEG and MEG With
Diminished Bias From Common Sources", Human Brain Mapping,
28:1178-1193, 2007.
[19] J. B. Kruskal, Jr., "On the Shortest Spanning Subtree of a Graph and
the Traveling Salesman Problem", American Mathematical Society,
7(1):48-50, 1956.
[20] M. Rubinov, O. Sporns, "Complex network measures of brain connec-
tivity: Uses and interpretations", NeuroImage, 52:1059-1069, 2010.
[21] K. Smith, H. Azami, M. A. Parra, J. M. Starr, J.Escudero, "Cluster-
Span Threshold: An Unbiased Threshold for Binarising Weighted
Complete Networks in Functional Connectivity Analysis", Accepted
for presentation at the 37th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society.
|
1811.04698 | 3 | 1811 | 2018-12-21T03:22:32 | Topology of the mesoscale connectome of the mouse brain | [
"q-bio.NC"
] | The wiring diagram of the mouse brain has recently been mapped at a mesoscopic scale in the Allen Mouse Brain Connectivity Atlas. Axonal projections from brain regions were traced using green fluoresent proteins. The resulting data were registered to a common three-dimensional reference space. They yielded a matrix of connection strengths between 213 brain regions. Global features such as closed loops formed by connections of similar intensity can be inferred using tools from persistent homology. We map the wiring diagram of the mouse brain to a simplicial complex (filtered by connection strengths). We work out generators of the first homology group. Some regions, including nucleus accumbens, are connected to the entire brain by loops, whereas no region has non-zero connection strength to all brain regions. Thousands of loops go through the isocortex, the striatum and the thalamus. On the other hand, medulla is the only major brain compartment that contains more than 100 loops. | q-bio.NC | q-bio |
Topology of the mesoscale connectome of the mouse brain
Pascal Grange
Xi'an Jiaotong-Liverpool University, Department of Mathematical Sciences, Suzhou, China
[email protected]
Abstract
The wiring diagram of the mouse brain has recently been mapped at a mesoscopic scale in the Allen
Mouse Brain Connectivity Atlas. Axonal projections from brain regions were traced using green
fluoresent proteins. The resulting data were registered to a common three-dimensional reference
space. They yielded a matrix of connection strengths between 213 brain regions. Global features such
as closed loops formed by connections of similar intensity can be inferred using tools from persistent
homology. In this paper the wiring diagram of the mouse brain is mapped to a simplicial complex
(filtered by connection strength), and generators of the first homology group are computed. Some
regions, including nucleus accumbens, are connected to the entire brain by loops. Thousands of loops
go through the isocortex, the striatum and the thalamus. On the other hand, medulla is the only
major brain compartment that contains more than 100 loops.
1
Introduction
The Allen Mouse Brain Connectivity Atlas [1] has filled a major gap in the knowledge of
neuroanatomy by providing a brain-wide map of the connectome of the mouse brain.
Adeno-associated viral vectors expressing enhanced green fluorescent protein were injected into the
mouse brain, allowing to trace axonal projections. The scale of these experiments is mesoscopic [2], in
the sense that injections target groups of neurons belonging to brain regions assumed to be
homogeneous. The microscopic scale would correspond to mapping individual synapses where
neurons make contact. The resulting three-dimensional fluorescent traces were registered to brain
regions defined in the hierarchical Allen Reference Atlas [3].
These results allowed the construction of the first inter-region connectivity model of the mouse
brain. It takes the form of a connectivity matrix, whose rows and columns correspond to a brain
region, and whose entries model the connection strengths between pairs of regions defined by classical
neuroanatomy. The derivation of the entries of of the connectivity matrix assumed homogeneity of
brain regions and additivity of connection strengths. Projection densities correspond to axons
highlighted by tracers, hence the projection densities from several different sources sum to produce
projection density in a given region. This additivity assumption was used to address experimental
data in which an injection of viral tracer infected several neighbouring brain regions.
The construction of the model from injection data (possibly infecting several regions for each
injection) assumes that projections are homogeneous in each region, and that projections from
different sources add to produce projection density to a given region. The entries of the connectivity
matrix were observed to span a 105-fold range, and are approximately fit by a log-normal distribution.
1/19
The connectivity matrix is naturally mapped to a graph that models the wiring diagram of the
mouse brain. Some features of the wiring diagram such as node degree and clustering coefficient have
been observed to be reproduced by scale-free networks [4] and small-world networks [5], but none of
these models was found to fit all observed features of the wiring diagram. This begs for further
modelling taking into account the global properties of the connectivity map, together with the large
range of intensities mapped by connection strengths.
On the other hand, the recent computational developments of algebraic topology [6 -- 9] have given
rise to spectacular applications to data analysis in biological sciences, including obstructions to
phylogeny in viral evolution [10], and brain networks and neural correlations [11 -- 15]. In this paper we
will apply techniques from algebraic topology, a branch of mathematics characterizing properties of
spaces that are invariant by continuous deformation, to work out global properties of the wiring
diagram of the mouse brain.
We will first review the presentation of the Allen Mouse Brain Connectivity Atlas in the form of a
matrix modelling connection strengths between brain regions. We will define a mapping from these
data to a filtered simplicial complex and explain why the generators of the first homology group are
in one-to-one correspondence with closed loops in the mesoscale connectome. The anatomy of these
loops will then be analysed by grouping loops according to the major brain compartments they
intersect. Moreover, we will compare the fraction of the brain reached from each region by loops, to
the fraction reached by direct connections encoded in the connectivity matrix.
2 Materials and methods
2.1 Connectivity strengths in the mouse brain, local and global properties
Each image series in the Allen Mouse Connectivity Atlas corresponds to an injection of tracer,
followed by sectioning and and imaging of the brain1. In [1], registered data from 469 image series
were combined in order to estimate an inter-region connectivity matrix, presented in matrix form2:
C(r, r(cid:48)) =(cid:8)connection strength from region labelled r to region labelled r(cid:48)(cid:9) ,
1 ≤ r, r(cid:48) ≤ R.
(1)
The size of this connectivity matrix is R = 213, and each of the indices in {1, . . . , R} corresponds to a
brain region, defined in the Allen Reference Atlas (ARA, [3, 16]).
The connectivity matrix captures the local structure of the connectome at mesocopic scale, as it
estimates the strength of direct connections from a region to other regions. To capture global features
of the connectome, we would like to identify closed circuits constructed from these connections.
Algebraic topology formalises this notion: an irreducible closed loop in a topological space (a loop
that cannot be shrunk to a point by continuous deformations) is a loop that is not the boundary of a
disc in the topological space. It is therefore a one-dimensional cycle that is not the boundary of a
two-dimensional object. The family of such objects in a topological space is invariant by continuous
deformation, and has a group structure. It is called the first homology group of the topological space,
and denoted by H1.
If we repeat this reasoning in dimension zero, we obtain the more familiar notion of connected
component: elements of H0 are zero-dimensional objects that cannot be joined by a path drawn on
the topological space. They correspond to distinct connected components. More generally, the
1 © 2014 Allen
Institute
for Brain
Science.
Allen Connectivity Atlas.
Available
from:
http://connectivity.brain-map.org/.org
2The entries of the matrix C correspond to the left-hand side of Fig. 3 in [1].
2/19
elements of the homology group Hk are objects of dimension k that are not boundaries of
(k + 1)-dimensional objects in a topological space, and therefore formalise the notion of hole.
Moreover, the algebraic structure of homology groups allows us to count independent objects in each
of them: the number of generators of the homology group Hk is called the k-th Betti number and
denoted by bk. The Betti numbers b0 and b1 are the number of connected components and the
number of independent loops respectively.
2.2 Mapping the matrix of connection strengths to a filtered simplicial complex
The connectivity matrix C is not symmetric, because projections from a given brain region to other
brain regions are oriented (just as axons are). Let us map the connectivity matrix to a weighted
graph with R vertices corresponding to brain regions, and weighted edges corresponding to the entries
of the connectivity matrix. Edges with two different orientations and different weights can exist
between pairs of vertices. Let us apply a decreasing function to the entries of the connectivity matrix
and define for instance:
d(r, s) = − log (C(r, s) + ) ,
(2)
where is a positive regulator chosen to be smaller than the minimum connection strength, so that
entries of d are bounded. The families of loops worked out in this paper do not depend on the choice
of (in practice we took = 0.01minr,sC(r, s), resulting in a maximum entry of 42 for d). This
operation results in an approximately normal distribution of entries in the matrix d. The quantity
d(r, s) has been observed in [1] to be positively correlated to the spatial distance between region
labelled r and region labelled s. However, the entries of d are not quite distances because d is not
symmetric.
We propose to map brain regions to a graph with 2R vertices as follows. Each brain region,
labelled r, is mapped to a pair of vertices (sr, tr), where sr represents the sources of action potentials
(the axons of the cell bodies in the region, that conduct axon potentials), and tr the targets of action
potentials (the dendrites in the region, that receive action potentials). An edge is drawn between
vertices tr and sr for each r in {1, . . . , R}. With this doubling prescription we declare edges between
targets and sources to be present in each region. We are going to look for closed paths in the wiring
diagram of the brain, constructed from the non-zero entries C(sr, tr) of the connection matrix. Direct
approaches to persistent homology for asymmetric networks have been proposed, based on the
Dowker complex, see [17], but the present doubling prescription takes advantage of the mesoscopic
scale of the connectome data.
Given the family of 2R labelled vertices we have just described, we can construct a family of
graphs and work out the families of independent loops for each of them by executing the following
pseudo-code:
Consider a fixed f ∈]0, max1≤r<s≤Rd(r, s)] (referred to as a filtration value).
Draw an edge between any source labelled sa and any target labelled tb such
1.
2.
that d(sa, tb) ≤ f .
3.
Work out generators of the homology groups H0(f ) (connected components) and
H1(f ) (independent loops) in the resulting graph.
The family of graphs (depending on the parameter f ) is called a simplicial complex. The third
step of the procedure uses techniques of simplicial homology, implemented in JavaPlex [18, 19].
Simplices in our case consist of vertices (zero-simplices, corresponding to brain regions), and edges
(one-simplices, corresponding to axons connecting two brain regions). The result of the above
procedure depends on the value of the filtration value f . At f = 0 we have as many connected
3/19
(a)
(b)
Fig 1. A toy model with three regions. (A) Each of the regions is doubled at filtration value
zero, resulting in a source vertez and a target vertex. The filtered complex start with three connected
components and no loop. For increasing using values of the filtration parameters, edges are drawn
from to reflect projections from the nodes in the family (sr)1≤r≤3 to nodes in the family (sr)1≤r≤3.
(B) The graph gets connected at filtration value f = 2 (resulting in b0(f ) = 1 for f ≥ 2). A loop
appears at f = 3, the corresponding generator of the first homology group is
[s1, t2] + [t2, s2] + [s2, t1] + [t1, s1]. It goes through the regions labelled 1 and 2.
4/19
s1s2s3t1t2t3s1s2s3t1t2t3s1s2s3t1t2t3s1s2s3t1t2t3f=0f=1f=2f=3 components as points, and no loop. The number of connected components decreases when f
increases. It reaches 1 for some value of the filtration value. When f is increased beyond this value,
edges may still be drawn, possibly changing the number of loops.
When f reaches the maximum entry of the matrix d, the simplicial complex cannot be changed
anymore by increasing the filtration value. If we repeat the above procedure for all values of the
entries of the matrix d, arranged in increasing order, we can work out for each feature F (which can
be a connected component or a loop) an interval of filtration values [fmin(F ), fmax(F )] for which the
feature F is present. These intervals can be drawn for all existing features F of a fixed dimension,
yielding graphs of barcode form, one per dimension of feature (see Fig. 1B). Features persistent over
a longer interval are thought less likely to be due to noise [20]. In particular, we will restrict ourselves
to cycles that persist from their first appearance (at a given filtration value) to the maximum entry of
the matrix d. The results of the above procedure are sketched of Fig. 1 for a toy model consisting of
0 3 4
2 0 4
4 1 0
). The pseudo-code
three regions (and connection strengths assumed to be mapped to d =
is executed for each unique value of the entries of d, in ascending order.
2.3 Neuroanatomy of cycles
2.3.1 Three-dimensional presentation of brain regions
To analyse the anatomy of persistent cycles, we used the voxelised version of the Allen Reference
Atlas at a spatial resolution of 25 microns 3. This three-dimensional grid contains numerical labels
encoding neuroanatomy at the finest level compatible with the resolution (the voxelised atlas consists
of Vtot (cid:39) 3.1 × 107 cubic voxels, with 677 distinct regions). These numerical labels are associated to
the names of brain regions accroding to the hierarchical annotation of the ARA 4. For each of the
R = 213 regions in the mesoscale model of the connectome (Eq. 1), we resolved the numerical label
and the corresponding voxels. If the region has descendants in the hierarchical annotation, we lump
together the voxels corresponding to these descendants, by annotating them with the numerical label
of the region. After this step, the voxelised atlas contains 361 distinct regions. The regions included
in the mesoscale model of the connectome therefore do not quite span the entire brain. Let us denote
by Vr the number of voxels in region labelled r:
Vr = {v, voxel labelled v ∈ region labelled r} ,
1 ≤ r ≤ R.
The brain regions corresponding to the rows of the connectivity matrix contain
R(cid:88)
V =
Vr (cid:39) 2.4 × 107 voxels,
(3)
(4)
r=1
about 79 percent of the total volume Vtot of the brain. In the rest of this paper we will disregard the
brain regions that are not included in the matrix of connectivity strengths, and the volume V will be
referred to as the volume of the brain.
Grouping brain regions at a coarse hierarchical level (according to the ARA [3, 16]) yields the
following major brain compartments which will sometimes be designated by acronyms for
3 See download instructions and code snippets in MATLAB and Python on the Allen Brain Atlas data portal:
http://help.brain-map.org/display/mouseconnectivity/API#API-DownloadAtlas3-DReferenceModels
4The hierarchical system of annotation containing names of brain regions and numerical ids is available online at
http://api.brain-map.org/api/v2/structure graph download/1.json
5/19
presentation purposes: Isocortex, olfactory areas (OLF), cortical subplate (CTXsp), striatum (STR),
pallidum (PAL), thalamus (TH), hypothalamus (HY), midbrain (MB), medulla (MY), cerebellar
cortex (CBX). These major brain compartments will be referred to as big regions.
2.3.2 Fraction of the brain connected by loops to a given region
Given a brain region labelled r and a filtration value f , let us denote by Cr(f ) the family of
generators of the first homology group at filtration value f that go through the region r:
Cr(f ) = {c ∈ H1(f ), r ∈ c} ,
(5)
and by Cr(f ) the set of vertices through which the elements of Cr(f ) go:
Cr(f ) = {s ∈ {1, . . . , R},∃c ∈ H1(f ), s ∈ c} .
(6)
In the toy model of Fig. 1 with three brain regions, the sets C1(2) and C3(3) are empty. The set C1(3)
consists of one loop, corresponding to Cr(f ) = {1, 2}.
We can calculate the volume of the regions connected to the region r by loops, as a fraction of the
total number of voxels in the brain:
φr(f ) =
1
V
Vs.
(7)
where V is the total number of voxels defined in Eq. 4, and Vs is the number of voxels in region
labelled s. As the volumes of brain regions are not all equal, we can define an alternative measure by
the fraction of the total number of regions connected to region labelled r by loops:
(cid:88)
s∈Cr(f )
(cid:12)(cid:12)(cid:12)Cr(f )
(cid:12)(cid:12)(cid:12) .
νr(f ) =
1
R
(8)
(9)
(10)
For a region labelled r that is connected by loops to all brain regions, for some value of the filtration
f , we will have φr(f ) = νr(f ) = 1.
It is natural to compare these ratios to the analogous ratios obtained by taking into account only
direct connections between pairs of brain regions, as encoded by the connectivity matrix C. Let us
denote these quantities by φ(c)
number of regions:
r (f ) for the fraction of volume and by ν(c)
r (f ) for the fraction of the
R(cid:88)
R(cid:88)
s=1
s=1
φ(c)
r (f ) =
1
V
ν(c)
r (f ) =
1
R
Vsmax (1(C(r, s) > f ), 1(C(s, r) > f )) ,
max (1(C(r, s) > f ), 1(C(s, r) > f )) ,
where 1 denotes the indicator function. By construction the four fractions defined in Eqs 7,8,9,10 are
growing functions of the filtration value f . There can be regions in the family Cr(f ) that are not
directly connected to region r based on the inter-region connectivity matrix. There can also be
regions with projections to or from region r that are not involved in any closed cycle going through
region r at the given filtration value f . There is therefore no a priori solidarity between the quantities
φ and ν, based on loops, and their analogues φ(c) and ν(c), based on direct connections.
6/19
2.4 Brain-wide maps of loops going through a given brain region
To sum up the connections by loops between two given brain regions, we can calculate the sum of
connection strengths at which cycles connecting these two regions appear, weighted by the numbers
of such cycles. As the generators of the first homology groups are independent, this is consistent with
the additivity assumption of connection strengths between brain regions. For two regions labelled r
and s we introduce
ω(r, s) =
e−f 1(s ∈ Cr(f )),
where the sum is over the distinct filtration values (the distinct entries of the matrix d defined in Eq.
2), and the exponential function comes from Eq. 2 relating connection strengths to the distance used
to defined filtration values.
{1, . . . , R}, we can define a brain-wide map of connection strengths to region labelled r, by defining
for voxel labelled v:
Fixing the region label r and allowing the region label s to take all the possible values in
Wr(v) =
e−f 1(s(v) ∈ Cr(f )),
where s(v) is the index (in {1, . . . , R}) of the brain region to which voxel labelled v belongs. The
quantity Wr maps a voxel to a real number, and can be visualised as a heat map for instance.
To estimate how localised a brain-wide map of connection strengths is, we can compute the
f
Kullback -- Leibler divergence from the above-defined profile and the uniform brain-wide density. If we
denote by the label of the region to which voxel labelled v belongs, this divergence is expressed as
(cid:88)
f
(cid:88)
(11)
(12)
V(cid:88)
v=1
KL(r) =
1
V
Wr(v) log(Wr(v)),
(13)
where V is the total number of voxels defined in Eq. 4, and we used the fact that Wr/V is a
normalised density function. This divergence induces a ranking of brain regions.
3 Results
3.1 Persistent cycles
The bar code corresponding to the persistent generators of the first two homology groups is shown in
Fig. 2. The maximum Betti numbers are too large for the individual bars to be visible, as they are on
Fig. 1B. By construction the Betti number b0 starts at R = 213. The Betti number b1 starts at 0 and
the plateau at filtration value 25 in the bar code of the first homology group corresponds to
b1 = 16, 529. At filtration values larger than fconn = 7, all brain regions are in the same connected
component. Moreover, from the growth profile of the bar code of the first homology group, we can see
that thousands of persistent cycles appear at filtration values lower than fconn (b1 = 8, 230 at
f = fconn).
3.2 Loops highlight the cortico-striato-thalamic network
To obtain a coarse picture of the anatomy of cycles, let us work out which family of big regions (as
defined in Section 2.3.1) is intersected by each loop. We can rank the resulting families of big regions
by decreasing value of prevalence. For an example of the brain regions in a loop appearing at filtration
7/19
Fig 2. Bar code of the complex based on the inter-region connectivity matrix. The second
plateau that appears at filtration value f = 42 corresponds to − log , where is the regulator
introduced in Eq. 2. The shape of the barcode and the families of generators of H1 and H0 would be
unchanged if went to zero.
value f = 2, together with the names of the corresponding big regions, see Table 1. This loop appears
at a low filtration value and we will see that the family of big regions it intersects (isocortex,
striatum, thalamus and midbrain) is frequent among loops in the wiring diagram of the mouse brain.
For definiteness we did this calculation at two special filtration values:
- the filtration value fconn = 7 at which the Betti number b0 falls to 1.
- the maximum filtration value (fmax = 37), above which no extra loops appear (before the value
− log = 42 introduced in Eq. 2 to provide a cut-off to the plots).
We found that 334 distinct families of big regions occur, but the most frequent 27 families account
for 50% of the loops. The most prevalent families of big regions are presented in Table 2 (where they
are ordered by decreasing prevalence at filtration value f = 7). We notice that midbrain is
represented in most of the rows of this table5, and that the three most frequent combinations
(accounting for more than 10% of all the loops) go through six or more big regions. The family of
four big regions corresponding to the regions in the loop presented in Table 1 is the fourth most
5Midbrain is also represented in most of the loops that appear at low filtration values f < 3.
8/19
Table 1. Brain regions in a circuit appearing at filtration value f = 2.
Big region
Brain region
Isocortex
Anterior cingulate area, dorsal part
Isocortex
Primary auditory area
Striatum
Caudoputamen
Medial geniculate complex, dorsal part
Thalamus
Medial geniculate complex, ventral part Thalamus
Superior colliculus, motor related
Midbrain
Six regions appear, belonging to four distinct big brain regions.
prevalent family of big regions in loops appearing at filtration values lower than f = 7, with 249
distinct loops (resp. fifth, f = 37 and 328 loops). If we focus on small families of big regions (with
three regions or fewer), we observe that the most prevalent is the cortico-striato-thalamic family (7th
row of the table, with 183 loops at f = 7). Highlighting occurrences of isocortex, striatum and
thalamus in colour in Table 1 (in red, green and blue respectively), we notice that the first 17 rows
present at least one of these colours (in 14 cases associated with midbrain, highlighted in brown).
The next small family consists of isocortex and striatum (16th rank, with 106 loops), followed at rank
18 by medulla. With 104 loops, medulla contains more cycles than the cortico-thalamic family
(ranked 25 with 80 loops). Moreover, medulla is the only single big region to support more than 10
cycles (see Table 3, which shows that only 147 cycles are confined to a single big region, out of which
134 appear before the filtration value 7).
To assess whether the dominance of medullar loops among loops confined to a single big region
could have been guessed based on the matrix of connectivity strengths, let us calculate how much of
the connection strengths from its sub-regions project within the same region. For each big region
labelled b, let us define:
(cid:80)R
(cid:80)R
(cid:80)R
(cid:80)R
CintraCtot
(b) =
r=1
s=1 1(B(r) = b)1(B(s) = b)C(r, s)
r=1
s=1 1(B(r) = b)C(r, s)
,
(14)
where the symbol 1(B(r) = b) equals 1 if the region labelled r is part of the big region labelled b (for
instance we can read from the third row of Table 1 that 1(B(Caudoputamen) = Striatum) = 1). The
ratio Cintra/Ctot is the conditional probability of a connection from within the big region b to project
within region b. The sorted values of this ratio are plotted on Fig. 3. The average value is 31%, and
the values range from 6% (thalamus) to 68% (cerebellar cortex, which is found to contain only one
loop, see Table 3B).
The fact that 80 percent of the cycles confined to a single big region are in medulla could not have
been guessed based on the connectivity matrix only: 47% of connections from medullar regions
project to medullar regions, which is above average, but the medulla is only ranked fourth among 12
big regions by our measure of conditional probability.
3.3 Loops can connect brain regions to the entire brain
On average, loops connect a given brain region to a larger domain in the brain than direct
projections. This is true for both the volumetric and the counting measures defined in Eqs 9 and 10,
and at all filtration values. This can be observed on Fig. 4, where the following averages across all
9/19
Fig 3. Fraction of inter-region connection strengths supported in the big region of origin. Acronyms
of big regions read as follows: OLF = olfactory areas, HPF = hippocampal formation, CTXsp =
cortical subplate, STR = striatum, PAL = pallidum, TH = thalamus, HY = hypothalamus, MB =
midbrain, MY = medulla, CBX = cerebellar cortex.
brain regions are plotted:
(cid:104)φ(f )(cid:105) : =
(cid:104)φ(c)(f )(cid:105) =
R(cid:88)
R(cid:88)
i=1
i=1
1
R
1
R
φr(f ),
(cid:104)ν(f )(cid:105) =
1
R
R(cid:88)
i=1
φ(c)
r (f ),
(cid:104)ν(c)(f )(cid:105) =
1
R
νr(f ),
R(cid:88)
i=1
ν(c)
r (f ).
(15)
Moreover, all four averages reach an asymptote, and the ratios of asymptotic values do not depend
heavily on the choice of measure, as:
(cid:104)φ(25)(cid:105) (cid:39) 1.39(cid:104)φ(c)(25)(cid:105) (cid:39) 87%,
and (cid:104)ν(25)(cid:105) (cid:39) 1.37(cid:104)ν(c)(25)(cid:105) (cid:39) 80%.
(16)
However, the maximum values of the fractions φr and νr are highly heterogeneous among brain
regions. The heat map of Fig. 5 shows the fraction of the brain φr(f ), with the region labels r
grouped by big regions (and ordered within each big region by decreasing value of the maximum
reached at large filtration value). In particular, 18 regions reach the entire brain through connections
by loops. Their names are shown on Table 4, together with the big regions to which they belong
(midbrain is the most represented big region, with 4 regions). The ansiform lobule, which is part of
the cerebellar cortex, is connected to the entire brain by loops, and is the brain region connected to
the largest volume in the brain based on the connectivity matrix (it is ranked first by the measure
10/19
Fig 4. Fraction of the brain reached by loops, averaged across all regions in the connectivity atlas.
φ(c)). However, not all the regions appearing in Table 4 are that strongly connected based on the
connectivity matrix. For instance the posterior amygdalar nucleus is ranked 173 out of 213 by the
measure ψ(c)).
3.4 Brain-wide density of connections by loops to a given region
The region for which Wr has lowest Kullback -- Leibler divergence from a uniform profile is the nucleus
accumbens. The region with highest Kullback -- Leibler divergence from a uniform profile is the
flocculus, a region in the cerebellar cortex. Heat maps of the densities of loops for both regions are
shown on Fig. 6. They are projected on sagittal, coronal and axial planes. For each of these regions
the volume can be presented either as the sum of values of voxels (Figs 6A,C) or as maximal-intensity
projections (Figs 6B,D). On maximal-intensity projections of the density Wr, the voxels belonging to
region labelled r have the largest value by construction (because all loops go through this region).
Maximal-intensity projections can exhibit at most R different values (they are piecewise constant on
each brain region). On the other hand, projections of sums of the density Wr may have higher values
in voxels that do not belong to region labelled r, depending on the direction of projection. Even
though nucleus accumbens is connected to the entire brain by loops (meaning all the voxels in Figs.
6A,B correspond to non-zero values), the plots still appear rather shallow, and the strength of
connection by loops is still far from uniform, even in the region that minimises the divergence from a
uniform brain-wide profile.
11/19
Fig 5. Heat maps of the fraction of the brain reached by loops φr(f ), as a function of the
filtration value (horizontal axis). The rows correspond to brain regions (grouped by big regions
according to coarse neuroanatomy). The subregions in each big region are ordered by decreasing
value of the asymptote φr(35) .
12/19
Table 2. Families of big regions containing loops.
Number
of loops
(f = 7)
Rank
(out
of
334)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
410
268
265
249
248
199
183
165
141
135
130
130
122
115
107
106
105
104
96
85
85
84
Cumulated
percent-
age of
loops
(f = 7)
6%
9%
12%
15%
19%
21%
23%
25%
27%
29%
31%
32%
34%
35%
37%
38%
39%
40%
42%
43%
44%
45%
Number
of loops
(f = 37)
600
614
455
328
346
283
222
255
175
175
190
208
149
163
135
117
139
115
108
104
111
118
Cumulated
percent-
age of
loops
(f = 37)
4%
8%
11%
13%
15%
17%
19%
20%
21%
22%
24%
25%
26%
27%
28%
29%
29%
30%
31%
32%
32%
33%
Big regions intersected
Cortical Subplate, Hippocampal Formation,
Hypothalamus, Isocortex, Midbrain, Striatum,
Thalamus
Cortical Subplate, Hypothalamus, Isocortex,
Midbrain, Striatum, Thalamus
Cortical Subplate, Hypothalamus, Isocortex,
Midbrain, Pallidum, Striatum, Thalamus
Isocortex, Midbrain, Striatum, Thalamus
Hypothalamus, Isocortex, Midbrain, Pallidum,
Striatum, Thalamus
Hypothalamus, Isocortex, Midbrain, Striatum,
Thalamus
Isocortex, Striatum, Thalamus
Cortical Subplate, Isocortex, Midbrain,
Striatum, Thalamus
Hypothalamus, Isocortex, Midbrain, Thalamus
Hypothalamus, Midbrain, Pallidum, Thalamus
Cortical Subplate, Hippocampal Formation,
Hypothalamus, Midbrain, Striatum, Thalamus
Cortical Subplate, Hippocampal Formation,
Hypothalamus, Isocortex, Midbrain, Pallidum,
Striatum, Thalamus
Cortical Subplate, Isocortex, Striatum
Hypothalamus, Isocortex, Midbrain, Pallidum,
Thalamus
Cortical Subplate, Hypothalamus, Isocortex,
Midbrain, Pallidum, Striatum
Isocortex, Striatum
Cortical Subplate, Hippocampal Formation,
Hypothalamus, Midbrain, Olfactory Areas,
Striatum, Thalamus
Medulla
Hypothalamus, Midbrain, Thalamus
Hypothalamus, Midbrain, Pallidum
Hypothalamus, Isocortex, Midbrain, Pallidum,
Striatum
Medulla, Midbrain, Pallidum
The families of big regions are ordered by decreasing number of loops they contain (at filtration value f = 7).
13/19
Table 3. Tables of single big regions containing loops,
ordered by decreasing number of loops.
Number of loops Fraction (%) Big region label
104
8
8
6
5
3
(A)
Medulla
Hippocampal Formation
Hypothalamus
Olfactory Areas
Isocortex
Midbrain
78
6
6
5
4
3
Number of loops Fraction (%) Big region label
115
9
8
6
5
3
1
(B)
Medulla
Hypothalamus
Hippocampal Formation
Olfactory Areas
Isocortex
Midbrain
Cerebellar Cortex
79
7
6
5
4
3
1
(A) For filtration value f = 7, there is a total of 134 cycles. (B) For
filtration value f = 37, there are 147 loops, and just one more
region represented.
Table 4. Table of brain regions connected to all brain regions by loops.
Name of region
Anterior cingulate area, dorsal part
Big region
Isocortex
Hippocampal Formation Entorhinal area, lateral part
Cortical Subplate
Cortical Subplate
Striatum
Striatum
Striatum
Pallidum
Thalamus
Thalamus
Hypothalamus
Hypothalamus
Midbrain
Midbrain
Midbrain
Midbrain
Cerebellar Cortex
On average these brain regions are connected to 72% of the volume of the brain by direct connections, which
is largest than the average value (cid:104)φ(c)(cid:105) (cid:39) 63%. However, 5 regions correspond to values below this average.
φ(c)
r (25)(%) Rank by φ(c), out of 213
85
81
81
50
68
55
59
77
71
74
70
73
75
56
66
59
98
Basolateral amygdalar nucleus
Posterior amygdalar nucleus
Nucleus accumbens
Caudoputamen
Medial amygdalar nucleus
Globus pallidus, internal segment
Medial geniculate complex, dorsal part
Peripeduncular nucleus
Lateral hypothalamic area
Subthalamic nucleus
Central linear nucleus raphe
Midbrain reticular nucleus
Periaqueductal gray
Superior colliculus, motor related
Ansiform lobule
13
31
29
173
73
148
123
40
64
52
67
56
49
138
87
125
1
14/19
(a)
(b)
(c)
(d)
Fig 6. Heat map of the weighted regions connected by loops to given brain regions,
projected in the sagittal, coronal and axial directions.
sums of intensities. (B) Nucleus accumbens, projection of maximal intensities. (C) Flocculus,
projection of sums of intensities. (D) Flocculus, projection of maximal intensities. For visualisation
purposes, both hemispheres are shown, even though the results are based on the ipsilateral
connectivity matrix only.
(A) Nucleus accumbens, projection of
15/19
4 Discussion
Persistent homology reveals global properties of the mouse connectome at the mesoscopic scale. The
high prevalence of cortico-striato-thalamic loops is a strong consistency check between the mesoscale
experimental approach [1, 2] and the present topological method, as the
cortico-striato-thalamo-cortical circuit is known as a major circuit regulating complex behaviours (for
a review of its involvement in the Tourette sydrome see [21]).
The global nature of topological tools reveals connections between regions that are not linked by
direct projections. Moreover, the lists of generators of the first homology groups allow to count the
number of independent loops for a given threshold in connection strengths. The relative abundance of
closed loops in the medulla is perhaps best interpreted in terms of the life-sustaining automatic
functions regulated by the medulla. The corresponding circuits are topologically insulated from
circuits involved in conscious behaviours or learning [22].
Axons, in addition to conducting action potentials, transport biological molecules [23, 24].
Integration with other brain-wide data sets, in particular gene-expression data [25, 26], has already
shown that connected regions in the adult rodent brain have increased similarity in gene-expression
profiles, and sets of genes most correlated to connectivity have been identified (see [27] for results
based on connectivity data in the rat brain). It would be interesting to investigate to which extent
these correlations persist across the loops that we identified in the mouse brain. Moreover, spatial
gene expression patterns have been related to cortico-striatal functional networks [28]. Loops that are
as persistent as the cortico-striatal ones yield natural candidates for places where to look for such
genetic signatures of functional networks. Moreover, the brain-wide coverage of the gene-expression
atlas [25, 26] allows to estimate spatial densities of cell types known by their transcriptional
activity [29, 30, 33], which yields neuroanatomical predictions relevant to brain disorders [31, 32]. The
present work could lead to insights into the correlation structure between the local density of loops
and the spatial density of cell types. Another direction of research is the conservation of patterns
across species [34 -- 36], even though gene-expression and connectivity data do not have the same brain
coverage as in the mouse atlas.
We have worked out the persistent loops of the wiring diagram of the brain based on ipsilateral
projections only, leaving out the issue of contralateral proections. Contralateral projections have been
mapped (see [1], right-hand side of Fig. 4A), but they do not allow to work out closed loops because
the missing projections would involve injecting tracers into the left hemisphere, and mapping the
contralateral projections again. However, the known contralateral connection strengths have been
found to be significantly weaker than ipsilateral ones. One can therefore conjecture than loops going
through brain regions in different hemispheres do not dramatically change the global features of the
wiring diagram of the brain.
There are surprisingly few cycles confined to only one of of the major brain regions (or big
regions), but most of them develop before the brain becomes fully connected. Moreover, medulla is by
far the region with the most confined loops.
5 Acknowledgments
This work is supported by the Research Center for Precision Medicine, HT-URC,
Xi'an Jiaotong-Liverpool University, Suzhou, China.
16/19
References
1. Oh SW, Harris J, Ng L, Winslow B, Cain N, Mihalas S, et al. A mesoscale connectome of the
mouse brain. Nature. Nature Publishing Group; 2014;508: 207 -- 14. pmid:24695228.
2. Bohland JW, Wu C, Barbas H, Bokil H, Bota M, Breiter H C, et al. A proposal for a
coordinated effort for the determination of brainwide neuroanatomical connectivity in model
organisms at a mesoscopic scale. 2009; PLoS computational biology, 5(3), e1000334.
3. Dong HW . The Allen reference atlas: A digital color brain atlas of the C57Bl/6J male mouse.
John Wiley & Sons Inc; 2009.
4. Barab´asi AL, Albert R. Emergence of scaling in random networks. 1999; Science, 286(5439),
509 -- 512.
5. Watts DJ, Strogatz SH. Collective dynamics of 'small-world' networks. Nature. 1998
Jun;393(6684):440.
6. Edelsbrunner H, Letscher D, Zomorodian A. Topological persistence and simplification.
InFoundations of Computer Science, 2000. Proceedings. 41st Annual Symposium on 2000 (pp.
454-463). IEEE.
7. Edelsbrunner H, Harer J. Computational topology: an introduction. American Mathematical
Soc.; 2010.
8. Zomorodian A, Carlsson G. Computing persistent homology. Discrete & Computational
Geometry. 2005 Feb 1;33(2):249 -- 74.
9. Wasserman L. Topological data analysis. Annual Review of Statistics and Its Application. 2018
Mar 7;5:501 -- 32.
10. Chan JM, Carlsson G, Rabadan R. Topology of viral evolution. Proceedings of the National
Academy of Sciences. 2013 Oct 29:201313480.
11. Curto C, Giusti C, Marku K, Pastalkova E, Itskov V. Pairwise correlation graphs from
hippocampal population activity have highly non-random, low-dimensional clique topology.
BMC neuroscience. 2013 Jul;14(1):P182.
12. Petri G, Expert P, Turkheimer F, Carhart-Harris R, Nutt D, Hellyer PJ, Vaccarino F.
Homological scaffolds of brain functional networks. Journal of The Royal Society Interface.
2014 Dec 6;11(101):20140873.
13. Giusti C, Pastalkova E, Curto C, Itskov V. Clique topology reveals intrinsic geometric
structure in neural correlations. Proceedings of the National Academy of Sciences. 2015 Nov
3;112(44):13455-60.
14. Sizemore AE, Giusti C, Kahn A, Vettel JM, Betzel RF, Bassett DS. Cliques and cavities in the
human connectome. Journal of computational neuroscience. 2018 Feb 1;44(1):115-45.
15. Saggar M, Sporns O, Gonzalez-Castillo J, Bandettini PA, Carlsson G, Glover G, Reiss AL
(2018). Towards a new approach to reveal dynamical organization of the brain using
topological data analysis. Nature communications, 9(1), 1399.
17/19
16. Swanson L (2004). Brain maps: structure of the rat brain. Gulf Professional Publishing.
17. Chowdhury S, M´emoli F (2018). A functorial Dowker theorem and persistent homology of
asymmetric networks. Journal of Applied and Computational Topology, 2(1-2), 115 -- 175.
18. Adams H, Tausz A, Vejdemo-Johansson M (2014, August). JavaPlex: A research software
package for persistent (co) homology. In International Congress on Mathematical Software (pp.
129 -- 136). Springer, Berlin, Heidelberg.
19. Adams H, Tausz A, JavaPlex tutorial , available from
http://www.math.duke.edu/~hadams/research/javaplex tutorial.pdf
20. Collins A, Zomorodian A, Carlsson G, Guibas LJ (2004). A barcode shape descriptor for curve
point cloud data. Computers & Graphics, 28(6), 881 -- 894.
21. Robertson MM, Eapen V, Singer HS, Martino D, Scharf JM, Paschou P, et al. (2017). Gilles de
la Tourette syndrome. Nature reviews Disease primers, 3, 16097.
22. Rybak IA, Shevtsova NA, Paton JFR, Dick TE, John WS, Morschel M, et al. (2004). Modeling
the ponto-medullary respiratory network. Respiratory physiology & neurobiology, 143(2-3),
307 -- 319.
23. Paus T, Pesaresi M, French L (2014). White matter as a transport system. Neuroscience, 276,
117 -- 125.
24. Josh Huang Z, Zeng H (2013). Genetic approaches to neural circuits in the mouse. Annual
review of neuroscience, 36, 183-215.
25. Lein ES, Hawrylycz MJ, Ao N, Ayres M, Bensinger A, Bernard A, Boe AF, Boguski MS,
Brockway KS, Byrnes EJ, Chen L. Genome-wide atlas of gene expression in the adult mouse
brain. Nature. 2007 Jan;445(7124):168.
26. Ng L, Bernard A, Lau C, Overly CC, Dong HW, Kuan C, Pathak S, Sunkin SM, Dang C,
Bohland JW, Bokil H. An anatomic gene expression atlas of the adult mouse brain. Nature
neuroscience. 2009 Mar;12(3):356.
27. French L, Pavlidis P. Relationships between gene expression and brain wiring in the adult
rodent brain. PLoS computational biology. 2011 Jan 6;7(1):e1001049.
28. Anderson KM, Krienen FM, Choi EY, Reinen JM, Yeo BT, Holmes AJ. Gene expression links
functional networks across cortex and striatum. Nature communications. 2018 Apr 12;9(1):1428.
29. Grange P, Bohland JW, Okaty BW, Sugino K, Bokil H, Nelson SB, Ng L, Hawrylycz M, Mitra
PP. Cell-type -- based model explaining coexpression patterns of genes in the brain. Proceedings
of the National Academy of Sciences. 2014 Mar 25:201312098.
30. Grange P, Hawrylycz M, Mitra PP. Computational neuroanatomy and co-expression of genes in
the adult mouse brain, analysis tools for the Allen Brain Atlas. Quantitative Biology. 2013 Mar
1;1(1):91 -- 100.
31. Menashe I, Grange P, Larsen EC, Banerjee-Basu S, Mitra PP. Co-expression profiling of autism
genes in the mouse brain. PLoS computational biology. 2013 Jul 25;9(7):e1003128.
18/19
32. Grange P, Menashe I, Hawrylycz M. Cell-type-specific neuroanatomy of cliques of
autism-related genes in the mouse brain. Frontiers in computational neuroscience. 2015 May
29;9:55.
33. Bohland JW, Bokil H, Pathak SD, Lee CK, Ng L, Lau C, Kuan C, Hawrylycz M, Mitra PP.
Clustering of spatial gene expression patterns in the mouse brain and comparison with classical
neuroanatomy. Methods. 2010 Feb 1;50(2):105-12.
34. Sporns O. The human connectome: a complex network. Annals of the New York Academy of
Sciences. 2011 Apr;1224(1):109-25.
35. Stafford JM, Jarrett BR, Miranda-Dominguez O, Mills BD, Cain N, Mihalas S, Lahvis GP,
Lattal KM, Mitchell SH, David SV, Fryer JD. Large-scale topology and the default mode
network in the mouse connectome. Proceedings of the National Academy of Sciences. 2014 Dec
30;111(52):18745-50.
36. Hawrylycz M, Miller JA, Menon V, Feng D, Dolbeare T, Guillozet-Bongaarts AL, et al.
Canonical genetic signatures of the adult human brain. Nat Neurosci. Nature Publishing
Group, a division of Macmillan Publishers Limited. All Rights Reserved.; 2015;18: 1832 -- 1844.
19/19
|
1701.00096 | 4 | 1701 | 2017-03-10T17:06:53 | Loss of brain inter-frequency hubs in Alzheimer's disease | [
"q-bio.NC"
] | Alzheimer's disease (AD) causes alterations of brain network structure and function. The latter consists of connectivity changes between oscillatory processes at different frequency channels. We proposed a multi-layer network approach to analyze multiple-frequency brain networks inferred from magnetoencephalographic recordings during resting-states in AD subjects and age-matched controls. Main results showed that brain networks tend to facilitate information propagation across different frequencies, as measured by the multi-participation coefficient (MPC). However, regional connectivity in AD subjects was abnormally distributed across frequency bands as compared to controls, causing significant decreases of MPC. This effect was mainly localized in association areas and in the cingulate cortex, which acted, in the healthy group, as a true inter-frequency hub. MPC values significantly correlated with memory impairment of AD subjects, as measured by the total recall score. Most predictive regions belonged to components of the default-mode network that are typically affected by atrophy, metabolism disruption and amyloid-beta deposition. We evaluated the diagnostic power of the MPC and we showed that it led to increased classification accuracy (78.39%) and sensitivity (91.11%). These findings shed new light on the brain functional alterations underlying AD and provide analytical tools for identifying multi-frequency neural mechanisms of brain diseases. | q-bio.NC | q-bio | Loss of brain inter-frequency hubs in Alzheimer's disease
J. Guillona,b, Y. Attalc, O. Colliota,b, V. La Cortee,f, B. Duboisd, D. Schwartzb, M.
Chavezb, F. De Vico Fallania,b,∗
bCNRS UMR-7225, Sorbonne Universites, UPMC Univ Paris 06, Inserm U-1127, Institut du cerveau
aInria Paris, Aramis project-team, 75013, Paris, France
et la moelle (ICM), Hopital Pitie-Salpetriere, 75013, Paris, France
cmyBrain Technologies, Paris, France
dDepartment of Neurology, Institut de la Memoire et de la Maladie dAlzheimer - IM2A, Paris, France
eInstitute of Psychology, University Paris Descartes, Sorbonne Paris Cite, France
fINSERM UMR 894, Center of Psychiatry and Neurosciences, Memory and Cognition Laboratory,
Paris, France
Abstract
Alzheimer's disease (AD) causes alterations of brain network structure and function. The
latter consists of connectivity changes between oscillatory processes at different frequency
channels. We proposed a multi-layer network approach to analyze multiple-frequency
brain networks inferred from magnetoencephalographic recordings during resting-states
in AD subjects and age-matched controls.
Main results showed that brain networks tend to facilitate information propagation
across different frequencies, as measured by the multi-participation coefficient (M P C).
However, regional connectivity in AD subjects was abnormally distributed across fre-
quency bands as compared to controls, causing significant decreases of M P C. This
effect was mainly localized in association areas and in the cingulate cortex, which acted,
in the healthy group, as a true inter-frequency hub.
M P C values significantly correlated with memory impairment of AD subjects, as
measured by the total recall score. Most predictive regions belonged to components of
the default-mode network that are typically affected by atrophy, metabolism disruption
and amyloid-β deposition. We evaluated the diagnostic power of the M P C and we
showed that it led to increased classification accuracy (78.39%) and sensitivity (91.11%).
These findings shed new light on the brain functional alterations underlying AD
and provide analytical tools for identifying multi-frequency neural mechanisms of brain
diseases.
Keywords: MEG, Brain connectivity, Multilayer networks, Neurodegenerative diseases
∗Corresponding author.
Email [email protected]
7
1
0
2
r
a
M
0
1
]
.
C
N
o
i
b
-
q
[
4
v
6
9
0
0
0
.
1
0
7
1
:
v
i
X
r
a
1. Introduction
Recent advances in network science has allowed new insights in the brain organization
from a system perspective. Characterizing brain networks, or connectomes, estimated
from neuroimaging data as graphs of connected nodes has not only pointed out impor-
tant network features of brain functioning - such as smallworldness, modularity, and
regional centrality - but it has also led to the development of biomarkers quantifying
reorganizational mechanisms of disease (1). Among others, Alzheimer's disease (AD),
which causes progressive cognitive and functional impairment, has received great atten-
tion by the network neuroscience community (1–3). AD is histopathologically defined by
the presence of amyloid-β plaques and tau-related neurofibrillary tangles, which cause
loss of neurons and synapses in the cerebral cortex and in certain subcortical regions (2).
This loss results in gross atrophy of the affected regions, including degeneration in the
temporal and parietal lobe, and parts of the frontal cortex and cingulate gyrus (4).
Structural brain networks, whose connections correspond to inter-regional axonal
pathways are therefore directly affected by AD because of connectivity disruption in sev-
eral areas including cingulate cortices and hippocampus (5, 6). A decreased number of
fiber connections eventually lead to a number of network changes on multiple topological
scales. At larger scales, AD brain networks estimated from diffusion tensor imaging (DTI)
showed increased characteristic path length as compared to healthy subjects leading to
a global loss of network smallworldness (2, 7). Similar topological alterations have been
also documented in resting-state brain networks estimated from functional magnetic res-
onance imaging (fMRI) (8), as well as from magneto/electroencephalographic (M/EEG)
signals, the latter ones often reported within the alpha frequency range (8−13 Hz) which
is typically affected in AD (9–11). On smaller topological scales, structural brain net-
work studies have demonstrated a loss of connector hubs in temporal and parietal areas
that correlates with cognitive decline (2, 12, 13). In addition, higher-order association
regions appear to be affected in functional brain networks inferred from fMRI (2, 14) and
MEG signals, the latter showing a characteristic loss of parietal hubs in higher (> 14 Hz)
frequency ranges (15, 16).
Graph analysis of brain networks has advanced our understanding of the organiza-
tional mechanisms underlying human cognition and disease, but a certain number of
issues still remain to be addressed (17, 18). For example, conventional approaches ana-
lyze separately brain networks obtained at different frequency bands, or in some cases,
they simply focus on specific frequencies, thus neglecting possible insights of other spec-
tral contents on brain functioning (17). However, several studies have hypothesized and
reported signal interaction or modulations between different frequency bands that are
supportive of cognitive functions such as memory formation (19–21). Moreover, recent
evidence shows that neurodegenerative processes in AD do alter functional connectivity
in different frequency bands (16, 22, 23). How to characterize this multiple information
from a network perspective still remains poorly explored. Here, we proposed a multi-layer
network approach to study multi-frequency connectomes as networks of interconnected
layers, containing the connectivity maps extracted from different bands. Multi-layer net-
work theory has been previously used to synthesize MEG connectomes from a whole
population (24), characterize temporal changes in dynamic fMRI brain networks (12),
and integrating structural information from multimodal imaging (fMRI, DTI) (25, 26).
Its applicability to multi-frequency brain networks has been recently illustrated in fMRI
2
connectomes for which, however, the frequency ranges of interest remains quite limited
(27).
We focused on source-reconstructed MEG connectomes, characterized by rich fre-
quency dynamics, that were obtained from a group of AD and control subjects in eyes-
closed resting-state condition. We hypothesized that the atrophy process in AD would
lead to an altered distribution of regional connectivity across different frequency bands
and we used the multiplex participation coefficient to quantify this effect both at global
and local scale (28). We evaluated the obtained results, which provide a novel view of the
brain reorganization in AD, with respect to standard approaches based on single-layer
network analysis and flattening schemes (29). Finally, we tested the diagnostic power of
the measured brain network features to discriminate AD patients and healthy subjects.
2. Methods
2.1. Experimental design and data pre-processing
The study involved 25 Alzheimer's diseased (AD) patients (13 women) and 25 healthy
age-matched control (HC) subjects (18 women). All participants underwent the Mini-
Mental State Examination (MMSE) for global cognition (30) and the Free and Cued
Selective Reminding Test (FCSRT) for verbal episodic memory (31–33). Specifically, we
considered the Total Recall (TR) score - given by the sum of the free and cued recall
scores - which has been demonstrated to be highly predictive of AD (34).
Inclusion criteria for all participants were:
i) age between 50 and 90; ii) absence
of general evolutive pathology; iii) no previous history of psychiatric diseases; iv) no
contraindication to MRI examination; v) French as a mother tongue. Specific criteria
for AD patients were: i) clinical diagnosis of Alzheimer's disease; ii) Mini-Mental State
Examination (MMSE) score greater or equal to 18. Magnetic resonance imaging (MRI)
acquisitions were obtained using a 3T system (Siemens Trio, 32-channel system, with
a 12-channel head coil). The MRI examination included a 3D T1-weighted volumetric
magnetization-prepared rapid-gradient echo (MPRAGE) sequence with 1mm isotropic
resolution and the following parameters:
repetition time (TR)=2300 ms, echo time
(TE)=4.18ms, inversion time (TI)=900 ms, matrix=256x256. This sequence provided
a high contrast-to-noise ratio and enabled excellent segmentation of high grey/white
matter.
The magnetoencephalography (MEG) experimental protocol consisted in a resting-
state with eyes-closed (EC). Subjects seated comfortably in a dimly lit electromagnet-
ically and acoustically shielded room and were asked to relax and fix a central point
on the screen. MEG signals were collected using a whole-head MEG system with 102
magnetometers and 204 planar gradiometers (Elekta Neuromag TRIUX MEG system)
at a sampling rate of 1 000 Hz and on-line low-pass filtered at 330 Hz. The ground elec-
trode was located on the right shoulder blade. An electrocardiogram (EKG) Ag/AgCl
electrodes was placed on the left abdomen for artifacts correction and a vertical elec-
trooculogram (EOG) was simultaneously recorded. Four small coils were attached to
the participant in order to monitor head position and to provide co-registration with the
anatomical MRI. The physical landmarks (the nasion, the left and right pre-auricular
points) were digitized using a Polhemus Fastrak digitizer (Polhemus, Colchester, VT).
We recorded three consecutive epochs of approximately 2 minutes each. All subjects
gave written informed consent for participation in the study, which was approved by
3
the local ethics committee of the Pitie-Salpetriere Hospital. Signal space separation was
performed using MaxFilter (35) to remove external noise. We used in-house software
to remove cardiac and ocular blink artifacts from MEG signals by means of principal
component analysis. We visually inspected the preprocessed MEG signals in order to
remove epochs that still presented spurious contamination. At the end of the process,
we obtained a coherent dataset consisting of three clean preprocessed epochs for each
subject.
2.2. Source reconstruction, power spectra and brain connectivity
We reconstructed the MEG activity on the cortical surface by using a source imaging
technique (36, 37). We used the FreeSurfer 5.3 software (surfer.nmr.mgh.harvard.edu) to
perform skull stripping and segment grey/white matter from the 3D T1-weighted images
of each single subject (38, 39). Cortical surfaces were then modeled with approximately
20000 equivalent current dipoles (i.e., the vertices of the cortical meshes). We used
the Brainstorm software (40) to solve the linear inverse problem though the wMNE
(weighted Minimum Norm Estimate) algorithm with overlapping spheres (41). Both
magnetometer and gradiometer, whose position has been registered on the T1 image
using the digitized head points, were used to localize the activity over the cortical surface.
The reconstructed time series were then extracted from 148 regions of interest (ROIs)
defined by the Destrieux atlas (42).
We computed the power spectral density (PSD) of the ROI signals by means of the
Welch's method; we chose a 2 seconds sliding Hanning window, with a 25% overlap. The
number of FFT points was set to 500 for a frequency resolution of 0.5Hz. We estimated
functional connectivity by calculating the spectral coherence between each pair of ROI
signals (43). As a result, we obtained for each subject and epoch, a connectivity matrix of
size 148× 148 where the (i, j) entry contains the value of the spectral coherence between
the signals of the ROI i and j at a frequency f .
We then averaged the connectivity matrices within the following characteristic fre-
quency bands (44, 45): δ (2-4 Hz); θ (4-8 Hz); α = α1 (8-10.5 Hz) and α2 (10.5-13 Hz);
β = β1 (13-20 Hz) and β2 (20-30 Hz); γ (30-45 Hz). We further averaged the resulting
connectivity matrices across epochs to obtain our raw individual brain networks whose
nodes were the ROIs (n = 148) and links, or edges, were the spectral coherence values.
2.3. Single-layer network analysis
In order to cancel the weakest noisy connections, we thresholded the values in the
connectivity matrices and retained the same number of links in each brain network at
every frequency band, or layer. We considered six representative connection density
thresholds corresponding to an average node degree k = {1, 3, 6, 12, 24, 48}. These values
cover the density range [0.007, 0.327] which contains the typical density values used in
complex brain network analysis (17, 18, 46). The resulting sparse brain networks, or
graphs, were represented by adjacency matrices A, where the aij entry indicates the
presence or absence of a link between nodes i and j.
2.3.1. Participation coefficient
Given a network partition, the local participation coefficient (P Ci) of a node i mea-
sures how evenly it is connected to the different clusters, or modules of the network (47).
4
Nodes with high participation coefficients are considered central hubs as they allow for
the information exchange among different modules. The global participation coefficient
P C of a network at layer λ is then given by the average of the P Ci values:
N(cid:88)
i=1
P C [λ] =
1
n
P C [λ]
i =
1
n
(cid:34)
N(cid:88)
1 − M [λ](cid:88)
i=1
m=1
(cid:33)2(cid:35)
(cid:32) k[λ]
i,m
k[λ]
i
,
(1)
where k[λ]
i,m is the number of weighted links from the node i to the nodes of the module
m of the layer λ. By construction, P C ranges from 0 to 1. Here, the partition of the
networks into modules was obtained by maximizing the modularity function as defined
by (48).
2.3.2. Flattened networks
We also computed the participation coefficients for brain networks obtained by flat-
tening the frequency layers into a single overlapping or aggregated network (28). In an
overlapping network, the weight of an edge oij corresponds to the number of times that
the nodes i and j are connected across layers:
oij =
a[λ]
ij ,
(2)
(cid:88)
λ
In an aggregated network, the existence of an edge indicates that nodes i and j are
connected in at least one layer:
aij =
(cid:26) 1
0
if ∃λ : a[λ]
otherwise
ij (cid:54)= 0
,
(3)
Notice that, by construction, flattened networks do not preserve the original connec-
tion density of the single layer networks.
2.4. Multi-layer network analysis
We adopted a multi-layer network approach to integrate the information from brain
networks at different frequency bands, while preserving their original structure. We built
for each subject a multiplex network (Fig. 1a,b) where different layers correspond to
different frequency bands and each node in one layer is virtually connected to all its
counterparts in all the other layers.
Without loss of generality, if we consider the standard neurophysiological frequency
bands, the resulting supra-adjacency matrix A is given by the following intra-layer of
adjacency matrices on the main diagonal:
A = {A[δ], A[θ], A[α], A[β], A[γ]},
(4)
where A[λ] is adjacency matrix of the frequency layer λ. By construction, the inter-layer
adjacency matrices of multiplexes are intrinsically defined as identity matrices.
5
N(cid:88)
i=1
N(cid:88)
i=1
(cid:34)
1 −(cid:88)
(cid:32)
λ
(cid:33)2(cid:35)
2.4.1. Multi-participation coefficient
We considered the multi-layer version of the local participation coefficient M P Ci to
measure how evenly a node i is connected to the different layers of the multiplex (28).
This way, nodes with high multi-participation coefficients are considered central hubs as
they would allow for a better information exchange among different layers. The global
multi-participation coefficient is then given by the average of the M P Ci values:
M P C =
1
n
M P Ci =
1
n
M
M − 1
N LP [λ]
i
,
(5)
i = k[λ]
where N LP [λ]
i /oi, stands for node-degree layer proportion, which measures the
percentage of the total number of links (i.e. in all layers) of node i that are in layer λ. By
construction, if nodes tend to concentrate their connectivity in one layer, the global multi-
participation coefficient tends to 0; on the contrary, if nodes tend to have the same number
of connections in every layer, the M P C value tends to 1 (Fig. 1c). Hence, a node with a
high M P C has the potential to facilitate communication across layers. The Matlab code
for the computation of the M P C is freely available at https://github.com/devuci/BNT.
We also used the standard coefficient of variation CVi to measure the dispersion of
the degree of a node i across layers. A global coefficient of variation CV is then obtained
by averaging the CVi values across all the nodes (Supplementary Text).
2.5. Statistical analysis
We first analyzed network features on global topological scales in order to detect
statistical differences between AD and HC subjects at the whole system level. Only for
those conditions (e.g., frequency bands) that resulted significantly different on the global
scale, we also assessed possible group-differences on the local topological scale of single
nodes. This hierarchical approach allowed us to associate brain network differences on
multiple topological scales (49). For global network features, we used a non-parametric
permutation t-test to assess statistical differences between groups, with a significance
level of 0.05. For local network features, we applied a correction for multiple comparisons
by computing the rough false discovery rate (FDR) (50, 51). In both cases, surrogate
data were generated by randomly exchanging the group labels 10 000 times.
To test the ability of the significant brain network properties to predict the cogni-
tive/memory impairment of AD patients, we used the non-parametric Spearman's corre-
lation coefficient R. We set a significance level of 0.05 for the correlation of global network
features, with a FDR correction in the case of multiple comparisons (local features).
2.6. Classification
We used a classification approach to evaluate the discriminating power of the local
brain network properties which resulted significantly different in the AD and HC group.
Because we did not know in advance which were the most discriminating features, we
tested different combinations.
In particular, for each local network property, we first
ranked the respective ROIs according to the p-values returned by the between-group
statistical analysis (see previous section). For each subject s, we then tested different
6
feature vectors obtained by concatenating, one-by-one, the values of the network features
extracted from the ranked ROIs. The generic feature vector cs reads:
cs = [g1, ..., gk]
(6)
where gk is a generic local network feature and k is a rank that ranges from 1 (the
most significant ROI) to the total number of significant ROIs. When different network
properties were considered (e.g., P C and M P C), we concatenated the respective cs
feature vectors allowing for all the possible combinations.
To quantify the separation between the feature vectors of AD and HC subjects, we
used a Mahalanobis distance classifier. We applied a repeated 5-fold cross-validation
procedure where we randomly split the entire dataset into a training set (80%) and a
testing test (20%). This procedure was eventually iterated 10 000 times in order to obtain
more accurate classification rates. To assess the classification performance we computed
the sensitivity (Sens), specificity (Spec) and accuracy (Acc), defined respectively as the
percentage of AD subjects correctly classified as AD, the percentage of HC subjects
classified as HC and the total percentage of subjects (AD and HC) properly classified.
We also computed the receiver operating characteristic (ROC) curve and its area under
the curve (AU C) (52).
3. Results
Power analysis of source-reconstructed MEG signals confirmed the characteristic
changes in the oscillatory activity of AD subjects compared to HC subjects (Fig. 2a)
(53–56). Significant alpha power decreases were more evident in the parietal and occip-
ital regions (Z < −2.58), while significant delta power increases (Z > 2.58) were more
localized in the frontal regions of the cortex (Fig. 2b).
3.1. Reduced gamma inter-modular connectivity
As expected the value of the connection density threshold had an impact on the
network differences between groups. For the sake of simplicity, we selected the first
threshold for which we could observe a significant group difference for both single- and
multi-layer analysis. The obtained results determined the choice of a representative
threshold, common to all the brain networks, corresponding to an average node degree
k = 12 (Fig. S1).
We first evaluated the results from the single-layer analysis. By inspecting the global
participation coefficient P C, we reported in the gamma band a significant decrease of
inter-modular connectivity in AD as compared to HC (Z = −2.50, p = 0.017; Fig. 3a
inset). This behavior was locally identified in association ROIs including temporal and
parietal areas (p < 0.05, FDR corrected; Fig. 3a; Tab. 2). No other significant differences
were reported in other frequency bands or in flattened brain networks (Fig. S1).
3.2. Disrupted inter-frequency hub centrality
Then we assessed the results from the multi-layer analysis. Both AD and HC subjects
exhibited high global multi-participation coefficients (M P C > 0.9), suggesting a general
propensity of brain regions to promote interactions across frequency bands. However,
7
such tendency was significantly lower in AD than HC subjects (Z = −2.24, p = 0.028;
Fig. 3b inset). This loss of inter-frequency centrality was prevalent in association ROIs
including temporal, parietal and cingulate areas, and with a minor extent in motor areas
(p < 0.05, FDR corrected; Fig. 3b; Tab. 2).
Among those regions, the right cingulate cortex was classified as the main inter-
frequency hub as revealed by the spatial distribution of the top 25% M P C values in
the HC group (Fig. 4a). In HC subjects the connectivity of this region across bands,
as measured by the node degree layer proportion N LP , was relatively stable (Kruskal-
Wallis test, χ2 = 10.79, p = 0.095), while it was significantly altered in AD subjects
(Kruskall-Wallis test, χ2 = 14.98, p = 0.020). In particular, the AD group exhibited
a remarkably reduced alpha2 connectivity and increased theta connectivity (Fig. 4b).
Similar results were also reported for the left cingulate cortex (AD: χ2 = 11.89, p = 0.064;
HC: χ2 = 6.98, p = 0.323), although it was not significant in terms of M P C differences
(Fig. 3b; Tab. 2).
3.3. Diagnostic power of brain network features
We adopted a classification approach to evaluate the power of the most significant
local network properties in determining the state (i.e., healthy or diseased) of each indi-
vidual subject. The best results were achieved neither when we considered single-layer
features (i.e., P C [γ]
i ) nor when we considered multi-layer features (M P Ci) (respectively,
first column and row of panels in Fig. 5a).
Instead, a combination of the two most
significant features gave the best classification in terms of accuracy (Acc = 78.39%)
and area under the curve (AU C = 0.8625) (Fig. 5a,b). While the corresponding speci-
ficity was not particularly high (Spec = 65.68%), the sensitivity was remarkably elevated
(Sens = 91.11%).
3.4. Relationship with cognitive and memory impairment
We finally evaluated the ability of the significant brain network changes to predict
the cognitive and memory performance of AD subjects. We first considered the results
from single-layer analysis. We found a significant positive correlation between the global
participation coefficient P C in the gamma band and the MMSE score (R = 0.4909,
p = 0.0127; Fig. 6a). Then we considered the results from multi-layer analysis. We
reported a higher significant positive correlation between the global multi-participation
coefficient M P C and the TR score (R = 0.5547, p = 0.0074; Fig. 6c). These relationships
were locally identified in specific ROIs including parietal, temporal and cingulate areas
of the default mode network (DMN) (57) (p < 0.05, FDR corrected; Fig. 6b,d; Tab. 3).
4. Discussion
Graph analysis of brain networks have been largely exploited in the study of AD
with the aim to extract new predictive diagnostics of disease progression. Typical ap-
proaches in functional neuroimaging, characterized by oscillatory dynamics, analyze brain
networks separately at different frequencies thus neglecting the available multivariate
spectral information. Here, we adopted a method to formally take into account the topo-
logical information of multi-frequency connectomes obtained from source-reconstructed
MEG signals in a group of AD and healthy subjects during EC resting states.
8
Main results showed that, while flattening networks of different frequency bands at-
tenuates differences between AD and HC populations, keeping the multiplex nature of
MEG connectomes allow to capture higher-order discriminant information. AD subjects
exhibited an aberrant multiplex brain network structure that significantly reduced the
global propensity to facilitate information propagation across frequency bands as com-
pared to HC subjects (Fig. 3b, inset). This could be in part explained by the higher
variability of the individual node degrees across bands (Fig. S2).
Such loss of inter-frequency centrality was mostly localized in association areas as
well as in the cingulate cortex (Fig. 3b; Tab. 2), which resulted the most important hub
promoting interaction across bands in the HC group (Fig. 4a). Because all these areas are
typically affected by AD atrophy (4) we hypothesize that the anatomical withering might
have impacted the neural oscillatory mechanisms supporting large-scale brain functional
integration. Notably, the significant alteration of the connectivity across bands observed
in the cingulate cortex could be ascribed to typical M/EEG connectivity changes observed
in AD, such as reduced alpha coherence (54–56, 58) (Fig. 4b). We also found a significant
decrease in the primary motor cortex (right precentral gyrus). While previous studies
have identified this specific region as a connector hub in human brain networks (2), its
role in AD still needs to be clarified in terms of node centrality's changes with respect
to healthy conditions.
While flattening network layers represents in general an oversimplification, analyzing
single layers can still be a valid approach that is worth of investigation. Because the M P C
is a pure multiplex quantity, we considered the conceptually akin version for single-layer
networks, the standard participation coefficient P C, which evaluates the tendency of
nodes to integrate information from different modules, rather than from different layers
(28, 47). AD patients exhibited lower inter-modular connectivity in the gamma band with
respect to HC subjects (Fig. 3a; Tab. 2) that was localized in association areas including
frontal, temporal, and parietal cortices (Fig. 3a; Tab. 2). Damages to these regions can
lead to deficits in attention, recognition and planning (59). Our results support the
hypothesis that AD could include a disconnection syndrome (60–62). Furthermore, they
are in line with previous findings showing P C decrements in AD, although those declines
were more evident in lower frequency bands and therefore ascribed to possible long-range
low-frequency connectivity alteration (2, 15).
Put together, our findings indicated that AD alters the global brain network orga-
nization through connection disruption in several association regions, which play im-
portant roles in sensory processing by integrating information from other cortical re-
gions through high-frequency channels (63–67). Notably, we showed that the global loss
of inter-modular interactions in the gamma band is paralleled by a diffused decrease
of inter-frequency centrality. Future studies, involving recordings of limbic structures
and/or stimulation-based techniques, should elucidate whether these two distinct reor-
ganizational processes are truly independent or linked through possible cross-frequency
mechanisms which are known to be essential for normal memory formation (68–70).
As a confirmation of the complementary information carried out by the multi-layer
approach, we reported an increased classification accuracy when combining the local P C
and M P C features. The observed diagnostic power is in line with previous accuracy
values obtained with standard graph theoretic approaches (around 80%) but exhibits
slightly higher sensitivity (> 90%), which is often desired to avoid false negatives (71–75).
Other approaches should determine if and to what extent the use of more sophisticated
9
machine learning algorithms, or the inclusion of basic connectivity features (76–78) and
different imaging modalities (79), can lead to higher classification performance and better
diagnosis (2).
Previous works have documented relationships between brain network properties and
neuropsychological measurements in AD, suggesting a potential impact for monitoring
disease progression and for the development of new therapies (7, 8, 10, 72, 80, 81). This is
especially true for the standard P C which has exhibited stronger correlations and larger
between-group differences (2). In line with this prediction, we also reported significant
correlations between the MMSE cognitive scores and the P C values of the AD patients
in the gamma band (Fig. 6a). An even stronger correlation was found, however, for the
global M P C values and the TR scores (Fig. 6b, Tab. 3). Recent studies suggest that TR
scores could be more specific for AD (82, 83) as compared to MMSE scores which could
be biased by differences in years of education, lack of sensitivity to progressive changes
occurring with AD, as well as fail in detecting impairment caused by focal lesions (84).
Locally, the regions whose M P C correlated with TR were part of the default-mode
network (DMN) (Tab. 3), which is heavily involved in memory formation and retrieval
(57, 85). According to recent hypothesis, these areas are directly affected by atrophy
and metabolism disruption, as well as amyloid-β deposition (86, 87). Put together, our
results suggest that AD symptoms related to episodic memory losses could be determined
by the lower capacity of strategic DMN association areas to let information flow across
different frequency channels.
Methodological considerations
We estimated brain networks by means of spectral coherence, a connectivity measure
widely used in the electrophysiological literature because of its simplicity and relatively
intuitive interpretation (88). While this measure is known to suffer from possible volume
conduction effects, recent evidence showed that source reconstruction techniques, like
the one we adopted here, could at least mitigate this bias (89) and generate connectivity
patterns consistent within and between subjects (90). In a separate analysis, we used the
imaginary coherence as a candidate alternative to eliminate volume conduction effects
(91). We demonstrated that while no significant between-group differences could be
obtained in terms of M P C (data not shown here), the spatial distribution of the M P C
values was very similar to that observed in the brain networks obtained with the spectral
coherence, especially for the internal regions along the longitudinal fissure (Fig. S3).
Differently from other multiplex network quantities, such as those based on paths and
walks (92), the M P C has the advantage to not depend on the weights of the inter-layer
links which, in general, are difficult to estimate or to assign from empirically obtained bi-
ological data. This is especially true in network neuroscience where, so far, the strength
of the inter-layer connections is parametric and subject to arbitrariness (27) or esti-
mated through measures of cross-frequency coupling (21) whose biological interpretation
remains still to be completely elucidated (20).
5. Conclusions
We proposed a multi-layer network approach to characterize multi-frequency brain
networks in Alzheimer's disease. The obtained results gave new insights into the neu-
ral deterioration of Alzheimer's disease by revealing an abnormal loss of inter-frequency
10
centrality in memory-related association areas as well as in the cingulate cortex. Lon-
gitudinal studies, including prodromal mild cognitive impairment subjects, will need to
assess the predictive value of this new information as a potential non-invasive biomarker
for neurodegenerative diseases.
Acknowledgments
We are grateful to F. Battiston for his useful comments and suggestions. This work
has been partially supported by the program "Investissements d'avenir" ANR-10-IAIHU-
06. FD acknowledges support from the "Agence Nationale de la Recherche" through
contract number ANR-15-NEUC-0006-02. The content is solely the responsibility of
the authors and does not necessarily represent the official views of any of the funding
agencies.
References
[1] C. J. Stam, Modern network science of neurological disorders, Nat Rev Neurosci 15 (10) (2014)
683–695. doi:10.1038/nrn3801.
[2] B. M. Tijms, A. M. Wink, W. de Haan, W. M. van der Flier, C. J. Stam, P. Scheltens, F. Barkhof,
Alzheimer's disease: Connecting findings from graph theoretical studies of brain networks, Neuro-
biol. Aging 34 (8) (2013) 2023–2036. doi:10.1016/j.neurobiolaging.2013.02.020.
[3] C. J. Stam, Use of magnetoencephalography (MEG) to study functional brain networks in neu-
rodegenerative disorders, Journal of the Neurological Sciences 289 (1–2) (2010) 128–134. doi:
10.1016/j.jns.2009.08.028.
[4] G. L. Wenk, Neuropathologic changes in Alzheimer's disease, J Clin Psychiatry 64 Suppl 9 (2003)
7–10.
[5] S. E. Rose, F. Chen, J. B. Chalk, F. O. Zelaya, W. E. Strugnell, M. Benson, J. Semple, D. M.
Doddrell, Loss of connectivity in Alzheimer's disease: An evaluation of white matter tract integrity
with colour coded MR diffusion tensor imaging, J. Neurol. Neurosurg. Psychiatr. 69 (4) (2000)
528–530.
[6] Y. Zhou, J. H. Dougherty, K. F. Hubner, B. Bai, R. L. Cannon, R. K. Hutson, Abnormal connec-
tivity in the posterior cingulate and hippocampus in early Alzheimer's disease and mild cognitive
impairment, Alzheimers Dement 4 (4) (2008) 265–270. doi:10.1016/j.jalz.2008.04.006.
[7] C.-Y. Lo, P.-N. Wang, K.-H. Chou, J. Wang, Y. He, C.-P. Lin, Diffusion tensor tractography
reveals abnormal topological organization in structural cortical networks in Alzheimer's disease, J.
Neurosci. 30 (50) (2010) 16876–16885. doi:10.1523/JNEUROSCI.4136-10.2010.
[8] E. J. Sanz-Arigita, M. M. Schoonheim, J. S. Damoiseaux, S. A. R. B. Rombouts, E. Maris,
F. Barkhof, P. Scheltens, C. J. Stam, Loss of 'Small-World' Networks in Alzheimer's Disease:
Graph Analysis of fMRI Resting-State Functional Connectivity, PLOS ONE 5 (11) (2010) e13788.
doi:10.1371/journal.pone.0013788.
[9] C. J. Stam, W. de Haan, A. Daffertshofer, B. F. Jones, I. Manshanden, A. M. v. C. van Walsum,
T. Montez, J. P. A. Verbunt, J. C. de Munck, B. W. van Dijk, H. W. Berendse, P. Scheltens,
Graph theoretical analysis of magnetoencephalographic functional connectivity in Alzheimer's dis-
ease, Brain 132 (1) (2009) 213–224. doi:10.1093/brain/awn262.
[10] W. de Haan, Y. A. Pijnenburg, R. L. Strijers, Y. van der Made, W. M. van der Flier, P. Scheltens,
C. J. Stam, Functional neural network analysis in frontotemporal dementia and Alzheimer's disease
using EEG and graph theory, BMC Neuroscience 10 (2009) 101. doi:10.1186/1471-2202-10-101.
[11] F. Miraglia, F. Vecchio, P. M. Rossini, Searching for signs of aging and dementia in EEG through
network analysis, Behavioural Brain Research 317 (2017) 292–300. doi:10.1016/j.bbr.2016.09.
057.
[12] D. S. Bassett, N. F. Wymbs, M. A. Porter, P. J. Mucha, J. M. Carlson, S. T. Grafton, Dynamic
reconfiguration of human brain networks during learning, PNAS 108 (18) (2011) 7641–7646. doi:
10.1073/pnas.1018985108.
11
[13] N. A. Crossley, A. Mechelli, J. Scott, F. Carletti, P. T. Fox, P. McGuire, E. T. Bullmore, The hubs
of the human connectome are generally implicated in the anatomy of brain disorders, Brain 137 (8)
(2014) 2382–2395. doi:10.1093/brain/awu132.
[14] R. L. Buckner, J. Sepulcre, T. Talukdar, F. M. Krienen, H. Liu, T. Hedden, J. R. Andrews-Hanna,
R. A. Sperling, K. A. Johnson, Cortical hubs revealed by intrinsic functional connectivity: Mapping,
assessment of stability, and relation to Alzheimer's disease, J. Neurosci. 29 (6) (2009) 1860–1873.
doi:10.1523/JNEUROSCI.5062-08.2009.
[15] W. de Haan, W. M. van der Flier, T. Koene, L. L. Smits, P. Scheltens, C. J. Stam, Disrupted
modular brain dynamics reflect cognitive dysfunction in Alzheimer's disease, NeuroImage 59 (4)
(2012) 3085–3093. doi:10.1016/j.neuroimage.2011.11.055.
[16] M. M. Engels, C. J. Stam, W. M. van der Flier, P. Scheltens, H. de Waal, E. C. van Straaten,
Declining functional connectivity and changing hub locations in Alzheimer's disease: An EEG
study, BMC Neurol 15. doi:10.1186/s12883-015-0400-7.
[17] F. De Vico Fallani, J. Richiardi, M. Chavez, S. Achard, Graph analysis of functional brain networks:
Practical issues in translational neuroscience, Phil. Trans. R. Soc. B 369 (1653) (2014) 20130521.
doi:10.1098/rstb.2013.0521.
[18] E. Bullmore, O. Sporns, Complex brain networks: Graph theoretical analysis of structural and
functional systems, Nat. Rev. Neurosci. 10 (3) (2009) 186–198. doi:10.1038/nrn2575.
[19] R. T. Canolty, R. T. Knight, The functional role of cross-frequency coupling, Trends in Cognitive
Sciences 14 (11) (2010) 506–515. doi:10.1016/j.tics.2010.09.001.
[20] V. Jirsa, V. Muller, Cross-frequency coupling in real and virtual brain networks, Front Comput
Neurosci 7. doi:10.3389/fncom.2013.00078.
[21] M. J. Brookes, P. K. Tewarie, B. A. E. Hunt, S. E. Robson, L. E. Gascoyne, E. B. Liddle, P. F.
Liddle, P. G. Morris, A multi-layer network approach to MEG connectivity analysis, NeuroImage
132 (2016) 425–438. doi:10.1016/j.neuroimage.2016.02.045.
[22] F. J. Fraga, T. H. Falk, P. A. M. Kanda, R. Anghinah, Characterizing Alzheimer's Disease Severity
via Resting-Awake EEG Amplitude Modulation Analysis, PLoS One 8 (8). doi:10.1371/journal.
pone.0072240.
[23] K. J. Blinowska, F. Rakowski, M. Kaminski, F. De Vico Fallani, C. Del Percio, R. Lizio, C. Babiloni,
Functional and effective brain connectivity for discrimination between Alzheimer's patients and
healthy individuals: A study on resting state EEG rhythms, Clin Neurophysioldoi:10.1016/j.
clinph.2016.10.002.
[24] Y. Ghanbari, L. Bloy, V. Shankar, J. C. Edgar, T. P. L. Roberts, R. T. Schultz, R. Verma, Function-
ally driven brain networks using multi-layer graph clustering, Med Image Comput Comput Assist
Interv 17 (Pt 3) (2014) 113–120.
[25] T. Simas, M. Chavez, P. R. Rodriguez, A. Diaz-Guilera, An algebraic topological method for mul-
timodal brain networks comparisons, Front Psychol 6. doi:10.3389/fpsyg.2015.00904.
[26] F. Battiston, V. Nicosia, M. Chavez, V. Latora, Multilayer motif analysis of brain networks,
arXiv:1606.09115 [cond-mat, physics:physics, q-bio]arXiv:1606.09115.
[27] M. De Domenico, S. Sasai, A. Arenas, Mapping multiplex hubs in human functional brain network,
arXiv:1603.05897 [cond-mat, physics:physics, q-bio]arXiv:1603.05897.
[28] F. Battiston, V. Nicosia, V. Latora, Structural measures for multiplex networks, Phys. Rev. E 89 (3)
(2014) 032804. doi:10.1103/PhysRevE.89.032804.
[29] M. De Domenico, A. Sol´e-Ribalta, E. Cozzo, M. Kivela, Y. Moreno, M. A. Porter, S. G´omez,
A. Arenas, Mathematical Formulation of Multilayer Networks, Phys. Rev. X 3 (4) (2013) 041022.
doi:10.1103/PhysRevX.3.041022.
[30] M. F. Folstein, S. E. Folstein, P. R. McHugh, "Mini-mental state". A practical method for grading
the cognitive state of patients for the clinician, J Psychiatr Res 12 (3) (1975) 189–198.
[31] H. Buschke, Cued recall in Amnesia, Journal of Clinical Neuropsychology 6 (4) (1984) 433–440.
doi:10.1080/01688638408401233.
[32] E. Grober, H. Buschke, H. Crystal, S. Bang, R. Dresner, Screening for dementia by memory testing,
Neurology 38 (6) (1988) 900–903.
[33] B. Pillon, B. Deweer, Y. Agid, B. Dubois, Explicit memory in Alzheimer's, Huntington's, and
Parkinson's diseases, Arch. Neurol. 50 (4) (1993) 374–379.
[34] M. Sarazin, C. Berr, J. De Rotrou, C. Fabrigoule, F. Pasquier, S. Legrain, B. Michel, M. Puel,
M. Volteau, J. Touchon, M. Verny, B. Dubois, Amnestic syndrome of the medial temporal type
identifies prodromal AD: A longitudinal study, Neurology 69 (19) (2007) 1859–1867. doi:10.1212/
01.wnl.0000279336.36610.f7.
[35] S. Taulu, J. Simola, Spatiotemporal signal space separation method for rejecting nearby interference
12
in MEG measurements, Phys. Med. Biol. 51 (7) (2006) 1759. doi:10.1088/0031-9155/51/7/008.
[36] B. He, Brain electric source imaging: Scalp Laplacian mapping and cortical imaging, Crit Rev
Biomed Eng 27 (3-5) (1999) 149–188.
[37] S. Baillet, J. J. Riera, G. Marin, J. F. Mangin, J. Aubert, L. Garnero, Evaluation of inverse methods
and head models for EEG source localization using a human skull phantom, Phys Med Biol 46 (1)
(2001) 77–96.
[38] B. Fischl, D. H. Salat, E. Busa, M. Albert, M. Dieterich, C. Haselgrove, A. van der Kouwe, R. Kil-
liany, D. Kennedy, S. Klaveness, A. Montillo, N. Makris, B. Rosen, A. M. Dale, Whole brain seg-
mentation: Automated labeling of neuroanatomical structures in the human brain, Neuron 33 (3)
(2002) 341–355.
[39] B. Fischl, D. H. Salat, A. J. W. van der Kouwe, N. Makris, F. S´egonne, B. T. Quinn, A. M. Dale,
Sequence-independent segmentation of magnetic resonance images, Neuroimage 23 Suppl 1 (2004)
S69–84. doi:10.1016/j.neuroimage.2004.07.016.
[40] F. Tadel, S. Baillet, J. C. Mosher, D. Pantazis, R. M. Leahy, F. Tadel, S. Baillet, J. C. Mosher,
D. Pantazis, R. M. Leahy, Brainstorm: A User-Friendly Application for MEG/EEG Analy-
sis, Brainstorm: A User-Friendly Application for MEG/EEG Analysis, Computational Intelli-
gence and Neuroscience, Computational Intelligence and Neuroscience 2011, 2011 (2011) e879716.
doi:10.1155/2011/879716,\%002010.1155/2011/879716.
[41] F.-H. Lin, T. Witzel, S. P. Ahlfors, S. M. Stufflebeam, J. W. Belliveau, M. S. Hamalainen, Assessing
and improving the spatial accuracy in MEG source localization by depth-weighted minimum-norm
estimates, NeuroImage 31 (1) (2006) 160–171. doi:10.1016/j.neuroimage.2005.11.054.
[42] C. Destrieux, B. Fischl, A. Dale, E. Halgren, Automatic parcellation of human cortical gyri and
sulci using standard anatomical nomenclature, Neuroimage 53 (1) (2010) 1–15. doi:10.1016/j.
neuroimage.2010.06.010.
[43] G. C. Carter, Coherence and time delay estimation, Proceedings of the IEEE 75 (2) (1987) 236–255.
doi:10.1109/PROC.1987.13723.
[44] C. J. Stam, A. M. van Cappellen van Walsum, Y. A. L. Pijnenburg, H. W. Berendse, J. C. de
Munck, P. Scheltens, B. W. van Dijk, Generalized synchronization of MEG recordings in Alzheimer's
Disease: Evidence for involvement of the gamma band, J Clin Neurophysiol 19 (6) (2002) 562–574.
[45] C. Babiloni, R. Ferri, D. V. Moretti, A. Strambi, G. Binetti, G. Dal Forno, F. Ferreri, B. Lanuzza,
C. Bonato, F. Nobili, G. Rodriguez, S. Salinari, S. Passero, R. Rocchi, C. J. Stam, P. M. Rossini,
Abnormal fronto-parietal coupling of brain rhythms in mild Alzheimer's disease: A multicentric
EEG study, Eur. J. Neurosci. 19 (9) (2004) 2583–2590. doi:10.1111/j.0953-816X.2004.03333.x.
[46] M. Rubinov, O. Sporns, Complex network measures of brain connectivity: Uses and interpretations,
NeuroImage 52 (3) (2010) 1059–1069. doi:10.1016/j.neuroimage.2009.10.003.
[47] R. Guimer`a, L. A. N. Amaral, Cartography of complex networks: Modules and universal roles, J
Stat Mech 2005 (P02001) (2005) P02001–1–P02001–13. doi:10.1088/1742-5468/2005/02/P02001.
[48] M. E. J. Newman, Finding community structure in networks using the eigenvectors of matrices,
Phys. Rev. E 74 (3) (2006) 036104. doi:10.1103/PhysRevE.74.036104.
[49] F. De Vico Fallani, S. Clausi, M. Leggio, M. Chavez, M. Valencia, A. G. Maglione, F. Ba-
biloni, F. Cincotti, D. Mattia, M. Molinari, Interhemispheric Connectivity Characterizes Corti-
cal Reorganization in Motor-Related Networks After Cerebellar Lesions, Cerebellumdoi:10.1007/
s12311-016-0811-z.
[50] Y. Benjamini, Y. Hochberg, Controlling the False Discovery Rate: A Practical and Powerful Ap-
proach to Multiple Testing, Journal of the Royal Statistical Society. Series B (Methodological) 57 (1)
(1995) 289–300.
[51] J. H. Zar, Biostatistical Analysis, Prentice Hall PTR, 1999.
[52] T. Hastie, R. Tibshirani, J. Friedman, The Elements of Statistical Learning, Springer Series in
Statistics, Springer New York, New York, NY, 2009.
[53] C. Babiloni, G. Binetti, E. Cassetta, D. Cerboneschi, G. Dal Forno, C. Del Percio, F. Ferreri,
R. Ferri, B. Lanuzza, C. Miniussi, D. V. Moretti, F. Nobili, R. D. Pascual-Marqui, G. Rodriguez,
G. L. Romani, S. Salinari, F. Tecchio, P. Vitali, O. Zanetti, F. Zappasodi, P. M. Rossini, Mapping
distributed sources of cortical rhythms in mild Alzheimer's disease. A multicentric EEG study,
Neuroimage 22 (1) (2004) 57–67. doi:10.1016/j.neuroimage.2003.09.028.
[54] J. Jeong, EEG dynamics in patients with Alzheimer's disease, Clin Neurophysiol 115 (7) (2004)
1490–1505. doi:10.1016/j.clinph.2004.01.001.
[55] J. Dauwels, F. Vialatte, A. Cichocki, Diagnosis of Alzheimer's Disease from EEG Signals:
doi:10.2174/
Where Are We Standing?, Current Alzheimer Research 7 (6) (2010) 487–505.
156720510792231720.
13
[56] R. Wang, J. Wang, H. Yu, X. Wei, C. Yang, B. Deng, Power spectral density and coherence analysis
of Alzheimer's EEG, Cogn Neurodyn 9 (3) (2015) 291–304. doi:10.1007/s11571-014-9325-x.
[57] R. L. Buckner, J. R. Andrews-Hanna, D. L. Schacter, The Brain's Default Network, Annals of the
New York Academy of Sciences 1124 (1) (2008) 1–38. doi:10.1196/annals.1440.011.
[58] C. J. Stam, B. F. Jones, I. Manshanden, A. M. van Cappellen van Walsum, T. Montez, J. P. A.
Verbunt, J. C. de Munck, B. W. van Dijk, H. W. Berendse, P. Scheltens, Magnetoencephalographic
evaluation of resting-state functional connectivity in Alzheimer's disease, NeuroImage 32 (3) (2006)
1335–1344. doi:10.1016/j.neuroimage.2006.05.033.
[59] D. Purves, G. J. Augustine, D. Fitzpatrick, L. C. Katz, A.-S. LaMantia, J. O. McNamara, S. M.
Williams (Eds.), Neuroscience, 2nd Edition, Sinauer Associates, 2001.
[60] R. C. Pearson, M. M. Esiri, R. W. Hiorns, G. K. Wilcock, T. P. Powell, Anatomical correlates of the
distribution of the pathological changes in the neocortex in Alzheimer disease, Proc. Natl. Acad.
Sci. U.S.A. 82 (13) (1985) 4531–4534.
[61] S. E. Arnold, B. T. Hyman, J. Flory, A. R. Damasio, G. W. Van Hoesen, The topographical and
neuroanatomical distribution of neurofibrillary tangles and neuritic plaques in the cerebral cortex
of patients with Alzheimer's disease, Cereb. Cortex 1 (1) (1991) 103–116.
[62] M. Catani, D. H. Ffytche, The rises and falls of disconnection syndromes, Brain 128 (Pt 10) (2005)
2224–2239. doi:10.1093/brain/awh622.
[63] W. H. Miltner, C. Braun, M. Arnold, H. Witte, E. Taub, Coherence of gamma-band EEG activity
as a basis for associative learning, Nature 397 (6718) (1999) 434–436. doi:10.1038/17126.
[64] T. J. Buschman, E. K. Miller, Top-down versus bottom-up control of attention in the prefrontal and
posterior parietal cortices, Science 315 (5820) (2007) 1860–1862. doi:10.1126/science.1138071.
[65] M. Siegel, T. H. Donner, R. Oostenveld, P. Fries, A. K. Engel, Neuronal Synchronization along
the Dorsal Visual Pathway Reflects the Focus of Spatial Attention, Neuron 60 (4) (2008) 709–719.
doi:10.1016/j.neuron.2008.09.010.
[66] G. G. Gregoriou, S. J. Gotts, H. Zhou, R. Desimone, High-frequency, long-range coupling between
prefrontal and visual cortex during attention, Science 324 (5931) (2009) 1207–1210. doi:10.1126/
science.1171402.
[67] J. F. Hipp, A. K. Engel, M. Siegel, Oscillatory synchronization in large-scale cortical networks
predicts perception, Neuron 69 (2) (2011) 387–396. doi:10.1016/j.neuron.2010.12.027.
[68] R. T. Canolty, E. Edwards, S. S. Dalal, M. Soltani, S. S. Nagarajan, H. E. Kirsch, M. S. Berger,
N. M. Barbaro, R. T. Knight, High Gamma Power Is Phase-Locked to Theta Oscillations in Human
Neocortex, Science 313 (5793) (2006) 1626–1628. doi:10.1126/science.1128115.
[69] N. Axmacher, M. M. Henseler, O. Jensen, I. Weinreich, C. E. Elger, J. Fell, Cross-frequency coupling
supports multi-item working memory in the human hippocampus, PNAS 107 (7) (2010) 3228–3233.
doi:10.1073/pnas.0911531107.
[70] R. Goutagny, N. Gu, C. Cavanagh, J. Jackson, J.-G. Chabot, R. Quirion, S. Krantic, S. Williams,
Alterations in hippocampal network oscillations and theta–gamma coupling arise before Aβ over-
production in a mouse model of Alzheimer's disease, Eur J Neurosci 37 (12) (2013) 1896–1902.
doi:10.1111/ejn.12233.
[71] Y. Li, Y. Wang, G. Wu, F. Shi, L. Zhou, W. Lin, D. Shen, Alzheimer's Disease Neuroimaging
Initiative, Discriminant analysis of longitudinal cortical thickness changes in Alzheimer's disease
using dynamic and network features, Neurobiol. Aging 33 (2) (2012) 427.e15–30. doi:10.1016/j.
neurobiolaging.2010.11.008.
[72] J. Wang, X. Zuo, Z. Dai, M. Xia, Z. Zhao, X. Zhao, J. Jia, Y. Han, Y. He, Disrupted functional brain
connectome in individuals at risk for Alzheimer's disease, Biol. Psychiatry 73 (5) (2013) 472–481.
doi:10.1016/j.biopsych.2012.03.026.
[73] C.-Y. Wee, P.-T. Yap, W. Li, K. Denny, J. N. Browndyke, G. G. Potter, K. A. Welsh-Bohmer,
L. Wang, D. Shen, Enriched white matter connectivity networks for accurate identification of MCI
patients, Neuroimage 54 (3) (2011) 1812–1822. doi:10.1016/j.neuroimage.2010.10.026.
[74] C.-Y. Wee, P.-T. Yap, D. Zhang, K. Denny, J. N. Browndyke, G. G. Potter, K. A. Welsh-Bohmer,
L. Wang, D. Shen, Identification of MCI individuals using structural and functional connectivity
networks, Neuroimage 59 (3) (2012) 2045–2056. doi:10.1016/j.neuroimage.2011.10.015.
[75] B. Horwitz, J. B. Rowe, Functional biomarkers for neurodegenerative disorders based on the network
paradigm, Progress in Neurobiology 95 (4) (2011) 505–509. doi:10.1016/j.pneurobio.2011.07.
005.
[76] D. Dai, H. He, J. Vogelstein, Z. Hou, Network-Based Classification Using Cortical Thickness of
AD Patients, in: D. Hutchison, T. Kanade, J. Kittler, J. M. Kleinberg, F. Mattern, J. C. Mitchell,
M. Naor, O. Nierstrasz, C. Pandu Rangan, B. Steffen, M. Sudan, D. Terzopoulos, D. Tygar, M. Y.
14
Vardi, G. Weikum, K. Suzuki, F. Wang, D. Shen, P. Yan (Eds.), Machine Learning in Medical
Imaging, Vol. 7009, Springer Berlin Heidelberg, Berlin, Heidelberg, 2011, pp. 193–200.
[77] J. Shao, N. Myers, Q. Yang, J. Feng, C. Plant, C. Bohm, H. Forstl, A. Kurz, C. Zimmer, C. Meng,
V. Riedl, A. Wohlschlager, C. Sorg, Prediction of Alzheimer's disease using individual structural
connectivity networks, Neurobiol Aging 33 (12) (2012) 2756–2765. doi:10.1016/j.neurobiolaging.
2012.01.017.
[78] L. Zhou, Y. Wang, Y. Li, P.-T. Yap, D. Shen, (adni), the Alzheimer's Disease Neuroimaging Initia-
tive, Hierarchical Anatomical Brain Networks for MCI Prediction: Revisiting Volumetric Measures,
PLOS ONE 6 (7) (2011) e21935. doi:10.1371/journal.pone.0021935.
[79] Z. Dai, C. Yan, Z. Wang, J. Wang, M. Xia, K. Li, Y. He, Discriminative analysis of early Alzheimer's
disease using multi-modal imaging and multi-level characterization with multi-classifier (M3), Neu-
roImage 59 (3) (2012) 2187–2195. doi:10.1016/j.neuroimage.2011.10.003.
[80] N. Shu, Y. Liang, H. Li, J. Zhang, X. Li, L. Wang, Y. He, Y. Wang, Z. Zhang, Disrupted topo-
logical organization in white matter structural networks in amnestic mild cognitive impairment:
Relationship to subtype, Radiology 265 (2) (2012) 518–527. doi:10.1148/radiol.12112361.
[81] C. J. Stam, B. F. Jones, G. Nolte, M. Breakspear, P. Scheltens, Small-world networks and functional
connectivity in Alzheimer's disease, Cereb. Cortex 17 (1) (2007) 92–99. doi:10.1093/cercor/
bhj127.
[82] E. Grober, A. E. Sanders, C. Hall, R. B. Lipton, Free and cued selective reminding identifies
very mild dementia in primary care, Alzheimer Dis Assoc Disord 24 (3) (2010 Jul-Sep) 284–290.
doi:10.1097/WAD.0b013e3181cfc78b.
[83] L. Velayudhan, S.-H. Ryu, M. Raczek, M. Philpot, J. Lindesay, M. Critchfield, G. Livingston,
Review of brief cognitive tests for patients with suspected dementia, Int Psychogeriatr 26 (8) (2014)
1247–1262. doi:10.1017/S1041610214000416.
[84] T. N. Tombaugh, N. J. McIntyre, The mini-mental state examination: A comprehensive review, J
Am Geriatr Soc 40 (9) (1992) 922–935.
[85] R. A. Sperling, B. C. Dickerson, M. Pihlajamaki, P. Vannini, P. S. LaViolette, O. V. Vitolo,
T. Hedden, J. A. Becker, D. M. Rentz, D. J. Selkoe, K. A. Johnson, Functional Alterations in
Memory Networks in Early Alzheimer's Disease, Neuromolecular Med 12 (1) (2010) 27–43. doi:
10.1007/s12017-009-8109-7.
[86] R. L. Buckner, A. Z. Snyder, B. J. Shannon, G. LaRossa, R. Sachs, A. F. Fotenos, Y. I. Sheline,
W. E. Klunk, C. A. Mathis, J. C. Morris, M. A. Mintun, Molecular, structural, and functional char-
acterization of Alzheimer's disease: Evidence for a relationship between default activity, amyloid,
and memory, J. Neurosci. 25 (34) (2005) 7709–7717. doi:10.1523/JNEUROSCI.2177-05.2005.
[87] M. D. Greicius, G. Srivastava, A. L. Reiss, V. Menon, Default-mode network activity distinguishes
Alzheimer's disease from healthy aging: Evidence from functional MRI, Proc Natl Acad Sci U S A
101 (13) (2004) 4637–4642. doi:10.1073/pnas.0308627101.
[88] R. Srinivasan, W. R. Winter, J. Ding, P. L. Nunez, EEG and MEG coherence: Measures of func-
tional connectivity at distinct spatial scales of neocortical dynamics, J. Neurosci. Methods 166 (1)
(2007) 41–52. doi:10.1016/j.jneumeth.2007.06.026.
[89] J.-M. Schoffelen, J. Gross, Source connectivity analysis with MEG and EEG, Hum Brain Mapp
30 (6) (2009) 1857–1865. doi:10.1002/hbm.20745.
[90] G. L. Colclough, M. W. Woolrich, P. K. Tewarie, M. J. Brookes, A. J. Quinn, S. M. Smith, How
reliable are MEG resting-state connectivity metrics?, Neuroimage 138 (2016) 284–293. doi:10.
1016/j.neuroimage.2016.05.070.
[91] G. Nolte, O. Bai, L. Wheaton, Z. Mari, S. Vorbach, M. Hallett, Identifying true brain interaction
from EEG data using the imaginary part of coherency, Clinical Neurophysiology 115 (10) (2004)
2292–2307. doi:10.1016/j.clinph.2004.04.029.
[92] S. Boccaletti, G. Bianconi, R. Criado, C. I. del Genio, J. G´omez-Gardenes, M. Romance, I. Sendina-
Nadal, Z. Wang, M. Zanin, The structure and dynamics of multilayer networks, Physics Reports
544 (1) (2014) 1–122. doi:10.1016/j.physrep.2014.07.001.
15
Figures and tables
Figure 1: Multi-frequency brain networks. Panel a) shows five representative networks extracted from
typical frequency bands. b) Procedure to construct a multi-frequency network by virtually connecting
the homologous brain nodes among frequency layers. c) Inter-frequency node centrality. A two-layer
multiplex is considered for the sake of simplicity. The blue node acts as an inter-frequency hub (i.e.,
multi-participation coefficient M P C = 1) as it allows for a balanced information transfer between layer
α and γ; the red node, who is disconnected in layer α, blocks the information flow and has M P C = 0.
16
Figure 2: Spectral analysis of MEG signals. a) Power spectrum density (PSD) for a representative
occipital sensor before source reconstruction. Each line corresponds to a subject. Bold lines show the
group-averaged values in the Alzheimer's disease group (AD) and in the healthy control group (HC).
b) Statistical PSD group differences. Z-scores are obtained using a non-parametric permutation t-test.
Results are represented both as sensor and source space.
17
Figure 3: Network analysis of brain connectivity. a) Inter-modular centrality. Statistical brain maps of
group differences for local participation coefficients P Ci in the gamma band. Only significant differences
are illustrated (p < 0.05, FDR corrected). The labels same ranks are used as labels. The inset shows
the results for the global P C; vertical bars stand for group-averaged values while error bars denote
standard error means. In both cases, Z-scores are computed using a non-parametric permutation t-test.
b) Inter-frequency centrality. Statistical brain maps of group differences for local multi-participation
coefficients M P Ci. The inset shows the results for the global M P C; same conventions as in a).
18
Figure 4: Inter-frequency hub centrality distribution. a) The median values of local multi-participation
coefficients (M P Ci) are shown over the cortical surface for the healthy group. Only the top 25% is
illustrated for the sake of visualization. The corresponding list of ROIs is illustrated in the horizontal bar
plot. b) Group-median values of the node-degree layer proportion (N LPi) for the right and left cingulate
cortex. The grey line corresponds to the expected value if connectivity were equally distributed across
frequency bands (N LPi = 1/7).
19
Figure 5: Classification performance of brain network features. a) Matrices show the classification
rates (accuracy=Acc, specificity=Spec, sensitivity=Sens, area under the curve=AUC) corresponding to
the combination of the most significant P C[γ]
and M P Ci network features, respectively on the rows
and columns of each matrix. Black squares highlight the highest accuracy rate and the corresponding
specificity, sensitivity and AUC. b) Scatter plots show the Mahalanobis distance of each subject from
the AD and HC classes. Separation lines (y = x: equal distances) are drawn in grey. Red circles stand
for Alzheimer's disease (AD) subjects , blue ones for healthy controls (HC). The bottom right plot shows
the ROC curve associated with the best network features configuration. The optimal point is marked
by a green circle.
i
20
Figure 6: Correlation between brain network properties and cognitive/memory scores. a) Scatter plot
of the global participation coefficient in the gamma band (P C[γ]) and the mini-mental state examina-
tion (MMSE) score of AD subjects (Spearman's correlation R = 0.4909, p = 0.0127). b) Correlation
brain maps of the local participation coefficient in the gamma band (P C[γ]
) and the mini-mental state
examination (MMSE) score of AD subjects. Only significant R values are illustrated (p < 0.05, FDR
corrected). c) Scatter plot of the global multi-participation coefficient (P C) and the total recall (TR)
score of AD subjects (Spearman's correlation R = 0.5547, p = 0.0074). d) Correlation brain maps of
the local multi-participation coefficient (M P Ci) and the total recall (TR) score of AD subjects. Only
significant R values are illustrated (p < 0.05, FDR corrected).
i
21
Age
MMSE
FR
TR
Control (HC) Alzheimer (AD)
70.8 (9.1)
28.2 (1.4)
31.5 (6.6)
46.3 (1.5)
73.5 (9.4)
23.2 (3.6)
14.9 (6.5)
33.9 (10.0)
p-value
0.3142
< 10−5
< 10−5
< 10−5
Table 1: Characteristics, cognitive and memory scores of experimental subjects. Mean values and
standard deviations (between parentheses) are reported. The last column shows the p-values returned by
a non-parametric permutation t-tests with 10 000 realizations. MMSE = mini-mental state examination
score; TR = total recall memory test score (/48); FR = free recall memory test (/48).
22
Index Rank ROI label
Lat Fis-ant-Horizont L
Pole temporal R
P C[γ]
i
M P Ci
1
2
3 G front inf-Triangul L
S temporal transverse L
4
5 G pariet inf-Supramar L
1 G precentral R
2 G front inf-Opercular R
S oc middle and Lunatus L
3
4 G pariet inf-Supramar L
S interm prim-Jensen L
5
S temporal transverse R Temporal
6
7
S pericallosal R
Cortex
Frontal
Temporal
Frontal
Temporal
Parietal
Motor
Motor
Occipital
Parietal
Parietal
Limbic
Z score
-3.6507
-2.8642
-2.4562
-2.3887
-2.3820
-3.4735
-2.5239
-2.4582
-2.4860
-2.3708
-2.3996
-2.3041
p-value
0.0007
0.0063
0.0198
0.0207
0.0222
0.0006
0.0127
0.0138
0.0142
0.0147
0.0191
0.0203
Table 2: Statistical group differences for local brain network properties. ROI labels, abbreviated accord-
ing to the Destrieux atlas, are ranked according to the resulting p-values. The same ranks are used as
labels in Fig. 3. ROIs highlighted in bold belong to the default mode network (DMN).
23
Correlation Rank ROI label
Lat Fis-ant-Vertical R
S interm prim-Jensen R
P C[γ]
i
- MMSE
M P Ci - TR
S pericallosal R
Lat Fis-ant-Horizont L
S collat transv post L
S circular insula ant L
1
2 G occipital sup L
3
4 G and S cingul-Ant R
5
6 G and S transv frontopol R
1
2
3
4 G parietal sup R
5
6
7
8
9 G and S occipital inf L
Cortex
Frontal
Occipital
Parietal
Limbic
Limbic
Frontal
Frontal
Occipital
Frontal
Parietal
Frontal
Temporal
Frontal
Temporal
Occipital
Occipital
10 G occipital sup R
11 G postcentral L
Sensory
12 G pariet inf-Supramar R Parietal
Parietal
13
14
Parietal
15
Temporal
S orbital lateral R
Pole temporal L
S orbital lateral L
S temporal sup R
S subparietal R
S interm prim-Jensen L
S temporal inf L
R coeff.
0.5480
0.5005
0.4948
0.4864
0.4735
0.4585
0.6915
0.6706
0.6214
0.6061
0.5920
0.5739
0.5462
0.5457
0.5368
0.5208
0.5191
0.5151
0.5066
0.4915
0.4869
p-value
0.0046
0.0108
0.0119
0.0137
0.0168
0.0212
0.0004
0.0006
0.0020
0.0028
0.0037
0.0052
0.0085
0.0086
0.0100
0.0130
0.0133
0.0142
0.0161
0.0202
0.0216
Table 3: Correlations of local brain network properties and cognitive/memory scores. ROI labels,
abbreviated according to the Destrieux atlas, are ranked according to the resulting p-values. ROIs
written in bold belong to the default mode network (DMN).
24
The global coefficient of variation is given by averaging CVi values across all the
Supplementary Material
Supplementary Text
nodes:
CV =
N(cid:88)
i=1
1
n
N(cid:88)
σ[·]
k[·]
ki
CVi =
1
n
(S1)
is the
i
where σ[·]
ki
mean value.
is the standard deviation of the degree of node i across layers and k[·]
i=1
i
Differently from M P C, CV tends to 0 when the links of the nodes tend to evenly
distribute across layers, and give higher values when they rather tend to be concentrated
in one layer or, more in general, differently distributed across layers.
Supplementary Figures
Figure S1: Statistical differences between global brain network properties of AD and HC subjects. These
figures illustrate the p-values resulting from the permutation t-tests as a function of the average node
degree k used to threshold the layers of the multi-frequency brain networks. In panel a), we show the
p-values for multi-layer and flattened analysis whereas in panel b) the p-values resulting from single-layer
analysis.
25
Figure S2: This figure shows the global coefficient of variation (CV ): first the difference between the pop-
ulations as an inset plot (p = 0.0521) and the correlation with the global multi-participation coefficient
(M P C) as a main plot (p < 10−15, R = −0.9742).
26
Figure S3: Inter-frequency hub centrality distribution for brain networks obtained with imaginary co-
herence. a) The median values of local multi-participation coefficients (M P Ci) are shown over the
cortical surface for the healthy group. Only the top 25% is illustrated for the sake of visualization.
The corresponding list of ROIs is illustrated in the horizontal bar plot. b) Group-median values of the
node-degree layer proportion (N LPi) for the right and left cingulate cortex. The grey line corresponds
to the expected value if connectivity were equally distributed across frequency bands (N LP = 1/7).
27
|
1501.01144 | 2 | 1501 | 2016-01-02T06:32:29 | Airborne Ultrasonic Tactile Display BCI | [
"q-bio.NC",
"cs.HC"
] | This chapter presents results of our project, which studied whether contactless and airborne ultrasonic tactile display (AUTD) stimuli delivered to a user's palms could serve as a platform for a brain computer interface (BCI) paradigm. We used six palm positions to evoke combined somatosensory brain responses to implement a novel contactless tactile BCI. This achievement was awarded the top prize in the Annual BCI Research Award 2014 competition. This chapter also presents a comparison with a classical attached vibrotactile transducer-based BCI paradigm. Experiment results from subjects performing online experiments validate the novel BCI paradigm. | q-bio.NC | q-bio |
Airborne Ultrasonic Tactile Display BCI
Katsuhiko Hamada1∗, Hiromu Mori2,†, Hiroyuki Shinoda1
and Tomasz M. Rutkowski2,3,†,‡
1 The University of Tokyo, Tokyo, Japan
2 Life Science Center of TARA, University of Tsukuba, Tsukuba, Japan
3 RIKEN Brain Science Institute, Wako-shi, Japan
[email protected] & http://bci-lab.info/
Abstract
This chapter presents results of our project, which studied whether contactless and
airborne ultrasonic tactile display (AUTD) stimuli delivered to a user's palms could serve
as a platform for a brain computer interface (BCI) paradigm. We used six palm positions
to evoke combined somatosensory brain responses to implement a novel contactless tactile
BCI. This achievement was awarded the top prize in the Annual BCI Research Award 2014
competition. This chapter also presents a comparison with a classical attached vibrotac-
tile transducer-based BCI paradigm. Experiment results from subjects performing online
experiments validate the novel BCI paradigm.
1 Introduction
State -- of -- the -- art brain computer interfaces (BCIs) are usually based on mental, visual or au-
ditory paradigms, as well as body movement imagery paradigms, which require extensive user
training and good eyesight or hearing. In recent years, alternative solutions have been proposed
to make use of the tactile modality [1, 2, 3] to enhance BCI efficiency. The concept reported
in this chapter further extends the brain's somatosensory channel by applying a contactless
stimulus generated with an airborne ultrasonic tactile display (AUTD) [4]. This is an expanded
version of a conference paper published by the authors [5].
The rationale behind the use of the AUTD is that, due to its contactless nature, it allows
for a more hygienic application, avoiding the occurrence of skin ulcers (bedsores) in patients
in a locked -- in state (LIS). The AUTD permits a less complex application of the BCI for the
caregivers comparing to classical attached vibrotactile transducers' setups.
This chapter reports very encouraging results with AUTD -- based BCI (autdBCI) compared
to the classical paradigm using vibrotactile transducer -- based oddball (P300 response -- based)
somatosensory stimulus (vtBCI) attached to the user's palms [3].
The rest of the chapter is organized as follows. The next section introduces the methods
used in the study. The results obtained in online experiments with 13 healthy BCI users are
then discussed. Finally, conclusions are drawn and directions for future research are outlined.
2 Methods
Thirteen male volunteer BCI users participated in the reported in this chapter experiments.
The users' mean age was 28.54, with a standard deviation of 7.96 years. The experiments were
∗K. Hamada is currently with DENSO Corporation, Japan.
†The author was supported in part by the Strategic Information and Communications R&D Promotion
‡The corresponding author.
Program no. 121803027 of The Ministry of Internal Affairs and Communication in Japan.
1
performed at the Life Science Center of TARA, University of Tsukuba, at the University of
Tokyo and at RIKEN Brain Science Institute, Japan. The online (real-time) EEG autdBCI
and vtBCI paradigm experiments were conducted in accordance with the WMA Declaration of
Helsinki - Ethical Principles for Medical Research Involving Human Subjects and the procedures
were approved and designed in agreement with the ethical committee guidelines of the Faculty
of Engineering, Information and Systems at University of Tsukuba, Japan (experimental per-
mission 2013R7).
The AUTD stimulus generator produced vibrotactile contactless stimulation of the human
skin via the air using focused ultrasound [4, 6]. The effect was achieved by generating an
ultrasonic radiation static force produced by intense sound pressure amplitude (a nonlinear
acoustic phenomenon). The radiation pressure deformed the surface of the skin on the palms,
creating a virtual touch sensation. An array of ultrasonic transducers mounted on the AUTD
(see Figure 1) created the focused radiation pressure at an arbitrary focal point by choosing a
phase shift of each transducer appropriately (the so -- called phased array technique). Modulated
radiation pressure created a sensation of tactile vibration similar to the one delivered by classical
vibrotactile transducers attached to the user's palms, as shown in Figure 2. The AUTD device
developed by the authors [4, 6] (see Figure 1) adhered to ultrasonic medical standards and did
not exceed the permitted skin absorption levels (approximately 40 times below the permitted
limits). The effective vibrotactile sensation was set to 50 Hz [6] to match with tactile skin
mechanoreceptors' frequency characteristics and the notch filter that EEG amplifiers use for
power line interference rejection.
As a reference, in the second vtBCI experiment, contact vibrotactile stimuli were also applied
to locations on the users' palms via the transducers HIHX09C005-8. Each transducer in the
experiments was set to emit a square acoustic frequency wave at 50 Hz, which was delivered
from the ARDUINO micro -- controller board with a custom battery -- driven and isolated power
amplifier and software developed in -- house and managed from a MAX 6 visual programming
environment.
The two experiment set-ups above are presented in Figures 2 and 3. Two types of experi-
ments were performed with the volunteer healthy users. Experiments with the target paralyzed
users are planned as a followup of the current pilot project. Psychophysical experiments with
foot -- button -- press responses were conducted to test uniform stimulus difficulty levels from re-
sponse accuracy and time measurements. The subsequent tactile oddball online BCI EEG
experiments evaluated the autdBCI paradigm efficiency and allowed for a comparison with
the classical skin contact -- based vtBCI reference. In both the above experiment protocols, the
users were instructed to spell sequences of six digits representing the stimulated positions on
their palms. The training instructions were presented visually by means of the BCI2000 [7]
and MAX 6 programs with the numbers 1 − 6 representing the palm locations as depicted in
Figure 2.
The EEG signals were recorded with the g.USBamp amplifier system from g.tec Medical
Engineering GmbH, Austria, using 16 active g.LADYbird electrodes. The electrodes were at-
tached to the head locations: Cz, Pz, P3, P4, C3, C4, CP5, CP6, P1, P2, POz, C1, C2,
FC1, FC2, and FCz, as in the 10/10 extended international system. The ground electrode was
attached to the FPz position, and the reference was attached to the left earlobe. No electro-
magnetic interference was observed from the AUTD or vibrotactile transducers operating with
frequencies notch -- filtered together with power line interference from the EEG. The EEG sig-
nals captured were processed online with an in -- house extended BCI2000 -- based application [7],
using a stepwise linear discriminant analysis (SWLDA) classifier [8] with features drawn from
0 − 800 ms ERP intervals decimated by a factor of 20.
Figure 1: The AUTD array with ultrasonic transducers used to create the contactless tactile
pressure sensation.
Figure 2: User's palms with attached vibrotactile transducers used in vtBCI experiments. Each
stimulus location reflects a different digit.
Figure 3: A user during the autdBCI experiment with both palms placed under the AUTD
array with ultrasonic transducers.
The stimulus length and inter -- stimulus -- interval were set to 400 ms, and the number of
epochs to average was set to 15. The EEG recording sampling rate was set at 512 Hz, and
the high and low pass filters were set at 0.1 Hz and 60 Hz, respectively. The notch filter to
remove power line interference removed activity between 48 ∼ 52 Hz. Each user performed
three experiment runs (randomized 90 targets and 450 non-targets each). As feedback, the
spelled numbers (palm position assigned digits as in Figure 2) were shown on a display to the
user.
3 Results
The grand mean averaged evoked responses to targets and non -- targets are depicted together
with standard error bars in Figure 4 and as matrices with an area under the curve (AUC)
analysis for feature separability in Figure 5. The BCI six digit sequences spelling accuracy
analyses for both experiments for the various averaging options are summarized in Figure 6. The
chance level was 16.6%. The mean six digit sequence spelling accuracies for 15-trial averaged
ERPs were 63.8% and 69.4% for autdBCI and vtBCI, respectively. The maximum accuracies
Figure 4: The autdBCI (blue - targets; red - non -- targets) and vtBCI (green - targets; black -
non -- targets) grand mean averaged ERP responses, together with standard error bars.
were 78.3% and 84.6% respectively. The differences were not significant, supporting the concept
of using autdBCIs. However, a single trial classification offline analysis of the collected responses
resulted with the best obtained accuracies of 83.0% for autdBCI and 53.8% for vtBCI, leading
to a possible 19.2 bit/min and 7.9 bit/min, respectively.
In the case of the autdBCI, only a single user's results were bordering on the level of
chance, and four subjects attained 100% (10 trials averaging). On average, lower accuracies
were obtained with the classical vtBCI, with which three users bordered on the level of chance,
and only one user scored 100% accuracy level in SWLDA -- classified averaged responses.
Figure 5: The autdBCI grand mean averaged ERP responses, shown as matrix plots for targets
in the top panel; non -- targets in the middle; and area under the curve (AUC) of the response
discriminability analysis (AUC > 0.5 marks the discriminable latencies).
4 Conclusions
This study demonstrates results obtained with a novel six -- command -- based autdBCI paradigm.
We compared the results with a classical vibrotactile transducer stimulus -- based paradigm. The
experiment results obtained in this study confirm the validity of the contactless autdBCI for
interactive applications and the possibility to further improve the results through single trial --
based SWLDA classification.
The EEG experiment with our paradigm confirms that contactless (airborne) tactile stim-
uli can be used to create six command -- based BCIs in real time. A short demo with online
application of the paradigm to robotic arm control is available on YouTube [9].
The results presented offer a step forward in developing and validating novel neurotechnology
applications. Since most users did not achieve very high accuracy during online BCI operation,
Figure 6: Averaged autdBCI and vtBCI spelling accuracy across six digits, colour coded for
each user with standard error bars.
especially with only a few trials, the current paradigm obviously requires improvement and
modification. These requirements determine the major lines of study for future research.
However, even in its current form, the proposed autdBCI can be regarded as a practical so-
lution for LIS patients (locked into their own bodies despite often intact cognitive functioning),
who cannot use vision or auditory-based interfaces due to sensory or other disabilities. The
reported autdBCI project was awarded The BCI Annual Research Award 2014 for "A fasci-
nating new idea never explored before," according to the Chairman of the Jury for the Annual
BCI Research Award 2014, Prof. Gernot R. Mueller-Putz from the Institute for Knowledge
Discovery, Graz University of Technology, Austria.
References
[1] G. Muller-Putz, R. Scherer, C. Neuper, and G. Pfurtscheller, "Steady-state somatosensory evoked
potentials: suitable brain signals for brain-computer interfaces?" Neural Systems and Rehabilitation
Engineering, IEEE Transactions on, vol. 14, no. 1, pp. 30 -- 37, March 2006.
[2] A.-M. Brouwer and J. B. F. Van Erp, "A tactile P300 brain-computer interface," Frontiers
http://www.frontiersin.org/
[Online]. Available:
in Neuroscience, vol. 4, no. 19, 2010.
neuroprosthetics/10.3389/fnins.2010.00019/abstract
[3] H. Mori, Y. Matsumoto, S. Makino, V. Kryssanov, and T. M. Rutkowski, "Vibrotactile stimulus
frequency optimization for the haptic BCI prototype," in Proceedings of The 6th International
Conference on Soft Computing and Intelligent Systems, and The 13th International Symposium
on Advanced Intelligent Systems, Kobe, Japan, November 20-24, 2012, pp. 2150 -- 2153. [Online].
Available: http://arxiv.org/abs/1210.2942
[4] T. Iwamoto, M. Tatezono, and H. Shinoda, "Non-contact method for producing tactile sensation
using airborne ultrasound," in Haptics: Perception, Devices and Scenarios, ser. Lecture Notes
in Computer Science, M. Ferre, Ed. Springer Berlin Heidelberg, 2008, vol. 5024, pp. 504 -- 513.
[Online]. Available: http://dx.doi.org/10.1007/978-3-540-69057-3 64
[5] K. Hamada, H. Mori, H. Shinoda, and T. M. Rutkowski, "Airborne ultrasonic tactile display brain-
computer interface paradigm," in Proceedings of the 6th International Brain-Computer Interface
Conference 2014, G. Mueller-Putz, G. Bauernfeind, C. Brunner, D. Steyrl, S. Wriessnegger, and
R. Scherer, Eds. Graz University of Technology Publishing House, 2014, pp. Article ID 018 -- 1 -- 4.
[Online]. Available: http://castor.tugraz.at/doku/BCIMeeting2014/bci2014 018.pdf
[6] K. Hamada, "Brain-computer interface using airborne ultrasound tactile display," Master Thesis,
The University of Tokyo, Tokyo, Japan, March 2014, in Japanese.
[7] G. Schalk and J. Mellinger, A Practical Guide to Brain -- Computer Interfacing with BCI2000.
Springer-Verlag London Limited, 2010.
[8] D. J. Krusienski, E. W. Sellers, F. Cabestaing, S. Bayoudh, D. J. McFarland, T. M.
the P300
[Online]. Available:
Vaughan, and J. R. Wolpaw, "A comparison of classification techniques
speller," Journal of Neural Engineering, vol. 3, no. 4, p. 299, 2006.
http://stacks.iop.org/1741-2552/3/i=4/a=007
for
[9] T. M. Rutkowski, "The autdBCI and a robot control (the winner project of The BCI Annual
Research Award 2014)," YouTube video. [Online]. Available: http://youtu.be/JE29CMluBh0
|
1806.02888 | 1 | 1806 | 2018-06-07T20:20:07 | Correspondence of Deep Neural Networks and the Brain for Visual Textures | [
"q-bio.NC",
"cs.CV"
] | Deep convolutional neural networks (CNNs) trained on objects and scenes have shown intriguing ability to predict some response properties of visual cortical neurons. However, the factors and computations that give rise to such ability, and the role of intermediate processing stages in explaining changes that develop across areas of the cortical hierarchy, are poorly understood. We focused on the sensitivity to textures as a paradigmatic example, since recent neurophysiology experiments provide rich data pointing to texture sensitivity in secondary but not primary visual cortex. We developed a quantitative approach for selecting a subset of the neural unit population from the CNN that best describes the brain neural recordings. We found that the first two layers of the CNN showed qualitative and quantitative correspondence to the cortical data across a number of metrics. This compatibility was reduced for the architecture alone rather than the learned weights, for some other related hierarchical models, and only mildly in the absence of a nonlinear computation akin to local divisive normalization. Our results show that the CNN class of model is effective for capturing changes that develop across early areas of cortex, and has the potential to facilitate understanding of the computations that give rise to hierarchical processing in the brain. | q-bio.NC | q-bio | Correspondence of Deep Neural Networks and the
Brain for Visual Textures
Md Nasir Uddin Laskar
Luis G Sanchez Giraldo
Odelia Schwartz*
Computational Neuroscience Lab, Dept. of Computer Science, University of Miami, FL, USA
E-mail: {nasir, lgsanchez, odelia}@cs.miami.edu
1
8
1
0
2
n
u
J
7
]
.
C
N
o
i
b
-
q
[
1
v
8
8
8
2
0
.
6
0
8
1
:
v
i
X
r
a
Abstract-Deep convolutional neural networks (CNNs) trained
on objects and scenes have shown intriguing ability to predict
some response properties of visual cortical neurons. However,
the factors and computations that give rise to such ability,
and the role of intermediate processing stages in explaining
changes that develop across areas of the cortical hierarchy, are
poorly understood. We focused on the sensitivity to textures as a
paradigmatic example, since recent neurophysiology experiments
provide rich data pointing to texture sensitivity in secondary but
not primary visual cortex. We developed a quantitative approach
for selecting a subset of the neural unit population from the
CNN that best describes the brain neural recordings. We found
that the first two layers of the CNN showed qualitative and
quantitative correspondence to the cortical data across a number
of metrics. This compatibility was reduced for the architecture
alone rather than the learned weights, for some other related
hierarchical models, and only mildly in the absence of a nonlinear
computation akin to local divisive normalization. Our results
show that the CNN class of model
is effective for capturing
changes that develop across early areas of cortex, and has the
potential to facilitate understanding of the computations that give
rise to hierarchical processing in the brain.
The tremendous progress in machine learning has shown
that deep CNNs trained on image classification are remarkably
good at object and scene recognition [1], [2]. Although CNNs
([3], [4], [1],[5]) are only crudely matched to the hierarchical
structure of the brain, such models have been intriguingly able
to predict some aspects of cortical visual processing [6], [7],
[8], [9], [10], [11], [12], [13], [14]. Here we focus on a ques-
tion that has been less explored, namely understanding how
the visual representation changes hierarchically across low
layers of the artificial network, in comparison to early cortical
areas. By considering low layers with fewer transformational
stages, we seek to get a better handle on some fundamental
questions that are not well understood: What makes the CNN
effective in capturing cortical data and when does it break?
What computations are important? What is the importance of
supervised training versus the architecture itself? The CNN
class of model under consideration includes linear and non-
linear computations that are widely used in modeling neural
systems, such as convolution, rectification, pooling of neural
units, and local (divisive) response normalization. We also
compare to other related hierarchical models [15], [16], [17],
[18].
We focus on the transformation between primary visual
cortex (V1) and secondary visual cortex (V2) as a paradigmatic
example. The changes in representation between V1 and V2,
and the computations that give rise to such changes, are not
well understood. Recent experimental neurophysiology studies
in macaque (and humans) have shown compelling analyses
that cortical area V2, but not area V1, is sensitive to naturalistic
textures [19], [20]. We focus on this data because it is currently
the best test for distinguishing V1 from V2, and it provides
a rich data set that could be compared to hierarchical models
(Fig. 1a). Since cortex develops sensitivity to textures across
the first two neural areas, we asked if the first two layers
of hierarchical models such as the CNN can develop similar
sensitivity, and what factors (Fig. 1b) are critical in doing so.
Textures are ubiquitous in natural scenes. Apart from their
representation in V2, there is a rich history of studying texture
representation in higher visual areas such as V4 [21], [22],
[23], [24], [25], [26], and texture perception in psychophysics
[27], [28], [29], [30], [31], [32]. Visualization of image fea-
tures in the CNN reveals that the second layer but not the
first layer has some texture selectivity [1], [5]. However, there
has been little work on this or other hierarchical models for
understanding if and how texture representation in a population
of such units relates to cortical measurements across early
visual areas.
Our results show that the CNN class of model is effective for
We first showed that the CNN model has some qualitative
correspondence between the first two layers of the CNN and
the biological cortical data, across a number of metrics. Similar
to the cortical data, the CNN developed more sensitivity to
the textures versus noise at the second layer. We also found
some differences in the strength of effects between the CNN
and the brain. To quantitatively compare the CNN model and
data, we developed an approach for systematic quantification
by selecting (Fig. 1c) a population of neural units from the
CNN that best describe the primate brain recordings. We found
quantitative correspondence between the CNN and the cortical
data at the population level.
Moreover, we found that this correspondence was reduced
when incorporating random weights rather than the weights
learned from images, for some other hierarchical models in
the literature which did not develop as much selectivity to
the textures versus noise in the second layer, and with only
a mild influence of a nonlinear computation known as local
response normalization in the neural networks community
(loosely matched to cross-orientation divisive normalization
in V1 [34], [33]).
2
(a) Brain vs CNN model
(b) CNN with stages of L2 computations
(c) Unit selection
Fig. 1: Simplified cartoon of hierarchical processing in the visual cortex and CNN. (a, top row): Cortical visual processing in
the brain. Here we focus only on areas V1 and V2 of the ventral stream. V1 has spatially oriented receptive fields (filters),
but receptive fields in V2 are not yet clearly understood (hence puzzle symbol). For V1, one orientation is shown in the front;
behind are spatially overlapping filters at different orientations. In addition to linear filtering, V1 undergoes further nonlinear
computations such as pooling and divisive normalization. The full linear nonlinear (LN) processing at each stage is not exactly
known (hence the question mark). We do not depict processing by the retina and LGN prior to V1, nor additional transformations
after V2 (see [9] Fig. 1). (a, bottom row): Processing in CNNs. LN indicates linear and nonlinear transformations. The red
box in the center represents the receptive field, the portion of the input visible to the unit. The weights of the filter banks
in each layer (L1 and L2; higher layers not shown) are chosen based on supervised discriminative learning on images. We
then ask whether there is correspondence between the representation in the CNN layers and in the cortical areas in the brain,
especially between V2 and L2, although we also explore other layers of the CNN. (b) CNN with detailed intermediate linear
and nonlinear computations in L2, from which we analyze the selectivity of the output at each stage. After convolution, the
nonlinear transformations in the AlexNet [1] include a ReLU (Rectified Linear) non-linearity, max pooling, and local response
normalization loosely matched to divisive normalization models of cross-orientation suppression in V1 [33]. (c) Schematic of
selection of 103 units from the brain recordings (e.g., from area V2). We select same number of units from the CNN (e.g.,
from layer L2) to find correspondence.
capturing changes across early areas of the cortical hierarchy.
This more broadly presents the opportunity to go beyond
demonstrating compatibility, to teasing out the computations
that are important for hierarchical representation and process-
ing in the brain. Our approach can be more widely applied
to other related architectures, computational building blocks,
stimuli, and neural areas.
I. RESULTS
The main simulations we ran followed the biological neuro-
physiology experiments with texture images in [19] and [20].
We used CaffeNet, a variant of AlexNet [1], a popular deep
CNN model widely applied in computer vision and neuro-
science. Here we refer to the network as AlexNet. We chose
AlexNet as our base model, because it includes computations
that are loosely matched to visual cortex, such as pooling
and local (divisive) response normalization. In addition, the
3
(a) L1 103 units
(b) L1, 8 × 8 spatial region all units
(c) L1 units
(d) L2 103 units
(e) L2, 4 × 4 spatial region all units
(f) L2 units
Fig. 2: Qualitative comparison of the CNN with the brain neurophysiology data. (a, b, d and e) t-SNE visualization of CNN
responses to natural textures. Each point represents a sample and each color represents a family. The CNN L2 units are able
to better separate out the texture categories than L1. This indicates that L2 units are more selective to the high order texture
properties of different categories. This is comparable with Fig. 4 in [20] for the biological neurophysiology V1 and V2 data,
but requiring in the CNN more than 103 units to obtain similar separation levels as the recorded V2 population. (c and f)
Distribution of variance ratio after one-way ANOVA analysis of the CNN L1 and L2 responses across samples and families
from randomly selected 103 units. L1 responses are more variable across samples in the same family. L2 responses are more
variable across families than L1. Geometric mean variance ratios are shown with triangle marker which are 0.18 in L1 and
0.29 in L2 and the difference is significant (p < 0.003, t test on the log variance ratio). This trend of L2 versus L1 units is
qualitatively similar to the properties of V2 neurons shown in [20], though the variance ratios are higher for the cortical data.
receptive field size ratio can be controlled roughly to match the
V1 to V2 ratio. Our approach was to keep a single base model,
and then manipulate the architecture in various ways. We later
also considered some other hierarchical architectures in the
literature. The CNN model was trained on the ILSVRC2012
ImageNet dataset, a popular large-scale image database [35].
We then presented each of the texture images to the resulting
model.
We took the layer outputs after pooling and normalization,
referring to them as L1, L2 and so forth. To get a better
handle on where exactly in the neural network compatibility
with V2 first emerges, we also considered layer outputs at
all other intermediate points of L2 in the network (Fig. 1b).
This allowed us to better understand how the computational
building blocks (e.g., convolution, pooling, normalization) in
the CNN may give rise to the differences observed in texture
selectivity between V1 and V2; in other words, at what point
in the CNN there is a transition from V1-like behavior to V2-
like behavior.
A. Texture generation and neurophysiology data
The neurophysiology data for V1 and V2 is described in
[19] and [20], with recordings from macaque monkeys. We
used the synthetic textures of [19], which were generated from
a set of 15 real texture images. Each synthetic texture image-
was generated using the approach of Portilla and Simoncelli
[36]. We refer to each set of textures generated from the same
source image as family and all the images within the same
family as samples. Naturalistic textures for a given family
were generated each with a different random seed. Spectrally
matched noise images (which we denote noise images) were
generated by randomizing the phase of the synthetic images.
The noise images have the same spatial frequency distribution
of energy as the original ones, but lack the differences in
higher order statistics.
Overall,
the image set
included 15 samples from each
family, resulting in 225 texture and 225 noise images. We
downsampled the textures so that the effective portion of the
image that the CNN units are sensitive to is equated to the
receptive field size of the cortical neurons. For more detail on
10−1100100101Varianceratio0.00.20.4ProportionofunitsL110−1100100101Varianceratio0.00.20.4ProportionofunitsL2 this process and the texture generation, see Methods.
For the data comparison, we largely focused on a population
metric of modulation index which is indicative of selectivity
to the textures versus the noise. We initially made a quali-
tative comparison between the biology and model, and then
developed a way to quantitatively compare the macaque data
to the CNN model units. We also considered other qualitative
comparisons between the CNN and the cortical data.
B. Qualitative correspondence of the CNN to the neuro-
physiology cortical data
We equated the number of CNN neural units in our simu-
lations to the number of units in the neural population as in
[19] and [20] (102 V1 and 103 V2 neurons). For the CNN, we
considered the total number of units as the number of channels
in a given layer, times a center 2 × 2 spatial neighborhood
(see Methods). We then randomly selected 103 neural units as
shown in Fig. 1c.
We first asked whether there is qualitative correspondence
between the CNN model and the experimental neural cortical
data, and compared these with a number of metrics. This then
prompted the quantitative approach in the next section.
1) Visualization of the CNN population selectivity:
To first gain intuition that L1 and L2 differ in their texture
representation, we visualized the unit population activity.
We transformed the responses of CNN layers from a high-
dimensional space (where dimensionality is the number of
units in the given CNN layer) to a 2-dimensional space to
visualize the unit response properties of L1 and L2. We used
the Barnes-Hut t-distributed stochastic neighbor embedding (t-
SNE) [37], [38] algorithm to achieve this visualization. t-SNE
is a technique for dimensionality reduction that tries to model
small pairwise distances to capture local data structures in the
low-dimensional space.
In Fig. 2a, 2b, 2d and 2e each point results from embedding
an image represented by a high-dimensional response vector
into two dimensions. Therefore, we have a total of 225 points,
that come from the same number of images from 15 texture
categories. Each point represents the population response to
a texture sample, and samples belonging to a same family
share the same color. L1 responses are more scattered and do
not group the same texture family together. This is apparent
both when randomly choosing 103 units (Fig. 2a); and when
considering all units in an 8×8 spatial neighborhood (Fig. 2b),
amounting to a total of 3072 units. In the L2 response space,
samples from the same texture family group more tightly
together. This is less apparent when using 103 random units
in L2 (Fig. 2d), but becomes striking when considering all
units in a 4 × 4 spatial neighborhood (Fig. 2e), amounting to
a total of 2048 units. L2 therefore better separates the texture
responses than L1, qualitatively similar to what has been
shown for V2 versus V1 in the biological data [20]. However,
a larger number of units form the CNN are required to match
the texture discrimination capabilities of the V2 population.
2) Tolerance across versus within texture family: We
also found that, with the same number of units, L2 responses
4
are more tolerant than L1 to the variations in image features
across samples within a texture family, qualitatively similar to
the V2 versus V1 data (Fig. 2c, 2f).
The variance ratio is calculated by taking the geometric
mean of the ratio of variance across-family to the variance
across-samples, with a large value indicating the high tolerance
of neurons to the statistical variation of samples in the same
family. More specifically, we ran a one-way ANOVA analysis
of the CNN L1 and L2 unit responses across samples and fam-
ilies, for a set of randomly selected 103 units. We found that
the L2 variance ratio (0.29) is significantly greater (p < 0.003,
t-test on the log variance ratio) than L1 (0.18). This gives
an indication that L2 unit responses show stronger variability
toward different families and higher tolerance to the samples
within a same family. These observations are qualitatively
consistent with the neurophysiology results reported in [20]
for V2 versus V1, although the variance ratios are higher in
the cortical data than in the CNN (0.63 in V2 versus 0.4 in
V1). This indicates that the cortical neurons show stronger
variability towards families than the CNN. A note of caution in
comparing the results is that the cortical data analysis included
a nested ANOVA with consideration of the noise that occurs
from stimulus repeats. In the CNN, there was no source of
noise for stimulus repeats and so we did not consider this
component in the ANOVA.
3) Differentiating L2 units from L1 in CNNs: To further
show the distinction between L1 and L2 for texture sensitivity,
we followed the approach in [19] for V1 and V2, to compute a
modulation index metric. The modulation index captures the
differential response between textures and noise. As for the
data, the mean modulation index was computed for each of
the 15 texture families, resulting in 15 mean modulation index
values. This was done for each of the network layers, L1 and
L2. We computed the modulation index from the responses
of all samples from each family, both natural and noise, and
averaged over the number of neural units in the respective
layer.
The modulation index M is defined as the difference of
responses from the textures to noise for each neural unit,
divided by their sum, according to (1):
M =
rna − rno
rna + rno
,
(1)
where, rna and rno are the responses to naturalistic textures
and noise respectively. Fig. 3a shows the average modulation
index for all texture categories in the CNN, both for L1 (red)
and L2 (blue).
Averages are obtained from 10000 repeats, where at each
iteration, we randomly select 103 neural units and compute
the modulation index.
High modulation index of a population of neural units
towards a family means that this group of units are highly sen-
sitive to this specific family; hence they show high differential
response. Since first and second order statistics are matched
for both natural and noise images, a differential response also
means that units latch onto higher order statistical properties
of the stimuli.
5
(a) CNN model
(b) Brain neurophysiology data
Fig. 3: Modulation index across families averaged over neural units. (a) Data from the CNN. L2 (blue) units have higher
modulation index than L1 (red) and hence higher differential response to the textures versus the noise. Light gray area shows
the expected modulation due to chance, 2.5th and 97.5th percentiles of the null distribution of modulation. Corresponding
textures and spectrally matched noise of the modulations are shown at the top. (b) Data from the biological neurophysiology
experiments of V2 and V1 (Fig. 2(e) in [19]).
As a control to examine the influence of the CNN archi-
tecture alone, instead of using the model weights that resulted
from training on the ImageNet database, we generated random
weights (in the interval [−1, 1]) for the L1 and L2 layers and
averaged over 100 iterations. While keeping the L1 weights
as trained, randomization only in the L2 units decreased the
average modulation index to 0.05 (from 0.18). The difference
between the L1 and L2 modulation index still remained
significant (p < 0.000003, t-test). We found that randomizing
both L1 and L2 significantly decreased the sensitivity of the
L2 units to the textures versus the noise and yielded an average
modulation index of -0.02 and 0.01 respectively for L1 and
L2. This lead to no significant difference between the L1 and
L2 modulation index (p = 0.0025, t-test on the magnitude).
We can see that, irrespective of texture types, L2 loses its
distinctive behavior in the complete absence of the trained
weights and shows similar modulation to L1. This indicates
that the selectivity we report in L2 is non-trivial and the CNN
model with trained weights corresponds better to the biological
data than the CNN architecture itself.
Fig. 4: Modulation index comparison between the V2 data
and a set of 103 randomly selected units from the CNN L2.
Random selection of units does not fit the neurophysiology
data well (Euclidean error 0.37), prompting us to explore
various subset selection methods for selecting units from the
population. V2 data was collected in [19]. For visual clarity,
we plot icons of the actual textures representing each family,
but the mean modulation index was calculated for responses
to the synthesized textures and noise images.
We found that the average modulation index of L1 neural
units is near zero and the modulation index of L2 is substan-
tially higher than L1. The diversity in modulation index for
the different texture families is shown in Fig. 3, for both the
CNN (Fig. 3a) and the biological data (Fig. 3b). The average
modulation index of L2 (0.18) is higher than L1 (-0.04). The
difference between the indexes of L1 and L2 is significant
(p < 0.0000005, t-test considering signs; p < 0.00001, t-test
ignoring signs and considering only the magnitudes) and is
qualitatively comparable, but stronger than the biological data
(V1: ≈ 0.00 and V2: ≈ 0.12 [19]). More specifically, Fig. 3a
shows that the L2 modulation is more drastic in some texture
families than others, as also observed for the V2 data [19].
However, the rank order of the textures was different between
the CNN and the biological data (prompting our quantitative
subset selection approaches below).
C. Systematic quantification of the CNN to the visual
brain data
We have thus far shown some qualitative correspondence
between the CNN and the biological cortical data. Fig. 4
illustrates that for a random selection of CNN units, indeed
this correspondence is only qualitative. The mean modulation
indices for the various textures in the V2 cortical data versus
the L2 in the CNN are correlated but clearly do not lie on
a straight line. We wondered whether there exists a set of
103 CNN units that can well fit the cortical data. That is,
rather than considering all units in the CNN or some random
selection of units, we posited that perhaps a subset of the units
could better explain the subset of experimentally recorded V2
neurons. For the remainder of the paper, we focus on this
question with respect to the modulation index metric, capturing
sensitivity to the textures versus the noise.
151015Texturefamilies−0.10.00.10.3ModulationindexL2L1151015Texturefamilies−0.10.00.10.3V2V10.00.10.20.3ModulationinBiology,V20.00.10.20.3ModulationinCNN,L2E=0.37 6
(a) L1, subset greedy
(b) L1, full population
(c) L1, subset regularized
(d) L2, greedy
(e) L2, full population
(f) L2, subset regularized
Fig. 5: CNN response fits to the V2 neurophysiology data, comparing the L1 and L2 fits. We consider various fitting strategies.
(a-c) L1 fits with greedy subset selection (Euclidean error 0.61), full population optimally weighted (error 0.30), and regularized
subset selection (error: 0.61) fitting approaches. (d-f) L2 fits with greedy (error: 0.03), full population optimally weighted (error:
0.01), and regularized subset selection (error: 0.08) approaches. L1 neural units can not fit the V2 data well (first row). L2
units yield a close fit to the V2 data in all methods (second row). V2 data was collected in [19]. Note that we plot icons of the
actual textures representing each family, but the fits to the mean modulation index are based on responses to the synthesized
texture images versus the noise.
We note that the subset selection aspect of this problem
makes it different from a standard regression and from ap-
proaches we are aware of for fitting neural data, which often
take a weighted average of the units [8], [7]. Why search for
a subset of CNN units that can fit the cortical modulation
index data? Our rationale was that finding such a subset
would suggest
there is some
overlap between the CNN units and the cortical neurons in
their representation for the texture versus the noise stimuli, as
explained below.
the population level,
that at
To show a systematic quantification, we probed the CNN
to select a subset of 103 neural units that is most consistent
with the cortical neurophysiology experiments (Fig. 1c). We
considered several approaches for subset selection. First, we
employed a greedy technique, which we call subset greedy,
to choose a subset of 103 units that best match the data from
the brain. Briefly, from the set of all possible neural units,
the greedy approach chooses the first unit with the closest
euclidean distance to the V2 mean modulation index data;
then the second unit is added to this subset so as to minimize
the euclidean distance; and so on until a total of 103 units are
chosen (see Methods).
For comparison to our greedy fitting approach, we also
applied an optimal weighted average or full population ap-
proach. The full population approach finds a weighted sum of
the neural units (under the constraint that the weights sum
to 1) that is the closest in squared Euclidean distance to
the experimental data. Notice that the weighted average may
include all available units and weight units differently. The
greedy approach is, in contrast, an approximation that finds a
subset of 103 units with equal weights that best matches the
neural data.
Our rationale is that the full population approach shows
the best fit one can obtain with units from a given layer.
However, it does not show an actual population representation
that matches the data; it only reveals a linear transform of the
representation. The suboptimal method chooses a subset of
103 units and thus uses the actual CNN representation. This
subset selection approach is therefore more comparable to the
analysis in the cortical experiments, in which the modulation
index is computed as an equally weighted average of the
neural units. In addition to the greedy approach, we also
applied another suboptimal model selection technique, which
is a regularized version of the full population fit that selects
103 units and we termed it as subset regularized. See Methods
section for more details about the model fitting techniques.
In the next sections, we show results for fitting the CNN
neural population to the V2 texture data with these approaches.
We find that the L2 population can well fit the V2 data, but
that the L1 population provides a poor fit. We then proceed to
0.00.10.20.30.00.10.20.3ModulationinCNNE=0.610.00.10.20.30.00.10.20.3E=0.300.00.10.20.30.00.10.20.3E=0.610.00.10.20.3ModulationinBiology,V20.00.10.20.3ModulationinCNNE=0.030.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.010.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.08 7
(a) L1, subset greedy
(b) L1, full population
(c) L1, subset regularized
(d) L2, greedy
(e) L2, full population
(f) L2, subset regularized
Fig. 6: Same CNN model fits as in Fig. 5 comparing L1 to L2, but with cross-validation. (a-c) L1 fits with greedy subset
selection (error: 0.63), full population optimally weighted (error: 0.59), and regularized subset selection (error: 0.62) fitting
approaches. (d-f) L2 fits with greedy (error: 0.22), full population optimally weighted (error: 0.19), and regularized subset
selection (error: 0.23) approaches. L1 neural units can not fit the V2 data well (first row). The cross-validated fits maintain
a similar trend to Fig. 5, with the best correspondence for the L2 units (second row). This indicates the generalizability of
our different fitting methods. V2 data was collected in [19]. Note that we plot icons of the actual textures representing each
family, but the fits to the mean modulation index are based on responses to the synthesized texture images versus the noise.
cross-validated fits, showing that this main result holds when
we train the subset selection on one set of texture and noise
images and test on the left out images. Finally, we show
cross-validation results for a wider range of neural network
manipulations.
1) L2 population fits are well matched to the V2 data
compared to L1: In this section, we discuss the fitting results
for the full data set. In the next section, we then discuss the
equivalent results for the cross-validation.
We found that a subset of 103 L2 neural units exist that
provide a good fit to the V2 neurophysiology data (Fig. 5;
second row). In contrast, all three fitting approaches showed
that for the L1 units, no such set exists that can fit the V2
data well (Fig. 5; first row). We found that even the full
population optimally weighted L1 fit (Fig. 5b) could not fit
the V2 biological data well. This indicated that the second
layer, but not the first layer of the CNN, is better matched
to the V2 data in terms of the sensitivity to textures versus
spectrally matched noise.
We quantified the fits with the Euclidean error distance
E between the mean modulation indices in the neurophysi-
ological data versus the modulation indices obtained from the
CNN for each family. A smaller Euclidean distance indicates
a better fit to the V2 data and higher correspondence to the
brain (see Methods section for details). The rationale behind
using the Euclidean distance as a measure of correspondence
is that it is directly related to the root mean squared error
(MSE) up to a normalization constant. Our optimal weighted
and subset regularized fits are done in terms of squared
Euclidean distance, which for the optimal fitting method makes
the error and regularization terms work at similar scales. In
the subset greedy approach, MSE and Euclidean (and even
squared Euclidean) distances indicate the same outcome.
We obtained Euclidean errors of 0.03, 0.01 and 0.08 for the
L2 subset greedy, full population and subset regularized ap-
proaches respectively. In contrast, the L1 fits yielded Euclidean
errors of 0.61, 0.30 and 0.61 for the three fitting approaches.
This poor L1 fit did not change if we took the outputs from
any other stage in the first layer. We therefore found that the
correspondence of the CNN with V2 for texture sensitivity
emerges in the second layer of the CNN after the (ReLU)
rectification, but that no stage in the first layer (even after
pooling and normalization) could account for the V2 texture
sensitivity data.
As we have seen in Fig. 5, the L1 units do not fit the V2
data, even using the full population. We also explored the L1
fit of the CNN unit responses to the V1 data. Both the subset
greedy and full population selection could find a set of units
from L1 that fits the V1 data well. In addition, we considered
the size of the spatial neighborhood from which we chose the
0.00.10.20.30.00.10.20.3ModulationinCNNE=0.630.00.10.20.30.00.10.20.3E=0.590.00.10.20.30.00.10.20.3E=0.620.00.10.20.3ModulationinBiology,V20.00.10.20.3ModulationinCNNE=0.220.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.190.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.24 units. In the greedy approach, selecting L1 units from a center
4× 4 spatial area fit V1 better (error 0.02) than choosing them
from a 2 × 2 neighborhood (error 0.14). At the same time,
selection from a larger 4 × 4 pool of L1 units did not rescue
the V2 data fit (error 0.44). This indicated that L1 can fit the
V1 data much better than the V2 data. The converse was not
true: That is, both L1 and L2 could fit the V1 modulation
index data well, possibly since L2 inherits some properties
from L1.
We wondered if the main result was peculiar to the [1]
network that was pre-trained on the ImageNet database. We
therefore trained the CNN on another popular large scale
image database known as the Places365 database [39]. We
found a similar trend, namely that the L2 units were more
selective than the L1 units. We obtained L1 and L2 Euclidean
errors of (greedy selection: 0.63 vs 0.10; full population: 0.49
vs 0.01; and subset regularized selection: 0.62 vs 0.11).
2) Cross-validating the L2 population fits: We have thus
far fit the full set of texture data with the CNN. We found
it interesting that the L2 but not the L1 units could well
fit the V2 data. Indeed, L1 could not well fit the V2 data,
even though the training and test data sets were the same.
To test the generalization capability of our method, we ran
cross-validation fits.
We obtained better cross validation results by extending the
image dataset to a total of 225 texture and noise images for
each family. We learned the population (e.g., of 103 units)
using 224 texture and noise images from each family for
the training, and made a prediction of the mean modulation
index for the left-one-out set of 15 images (see Methods).
Fig. 6 shows the exact same fits as for Fig. 5, but with
the cross-validation test results. This reveals the same main
trend as we found for the fitting of the full
texture data
without cross-validation: L2 (second row) but not L1 (first
row) could provide a good fit
to the V2 data. Euclidean
errors for the subset greedy method were 0.20 and 0.22 for
the training and test predictions, respectively. Considering the
whole population, we obtained train and test errors of 0.15 and
0.19; and for the regularized fits we obtained errors of 0.20
and 0.23 respectively. Note that the train and test errors were
rather close. The train errors were higher than the fitting errors
for the original texture images (Fig. 5), probably because of
the added texture images to which we were fitting.
These fitting errors were all lower than the random selection
of 103 units in the population that we examined earlier
(compare to Fig. 4; Euclidean error 0.37). In addition to the
Euclidean error, we also considered the explained variance
(R2): this was 0.60 for the subset greedy, 0.54 for the subset
regularized, and 0.70 for the full population. In contrast, the
explained variance for the random population of 103 units was
0.40. Note that this represents a lower bound, since we are not
considering the variability due to samples in a family, nor are
we taking into account variability in the experiments due to
stimulus repeats.
In the remainder of the paper, we report cross-validated
results for all conditions. Fitting without cross-validation re-
sulted in comparable results, but in some cases lead to potential
8
over-fitting (for instance, the almost perfect fit of L2 to the V2
data). This was particularly the case for the full population fit,
which has the added luxury of using any number of (weighted)
units in the population. For this reason, we report all remaining
results for the cross-validated fits.
3) CNN population fits are reduced for the architecture
alone, with random rather than learned weights: Given
the good correspondence of the L2 neural units to the V2
data, we wondered at what point this fit breaks or can be
reduced. In particular, we wanted to address the following:
Is training on natural images important for the CNN model
to develop texture selectivity? What is the effect of the CNN
architecture itself? Can the CNN architecture already develop
texture selectivity, just by stacking up multiple layers?
We therefore considered a control of taking random weights
for the CNN layers,
instead of weights learned from the
supervised model trained on images. We generated random
weights in the interval of [−1, 1] for the L1 and L2 CNN
layers. We then asked how well the L2 layer (with the random
rather than learned weights) can fit the V2 data. We repeated
this process and took the average of 10 iterations from the
layer responses. We found that taking random weights in the
CNN resulted in a larger error and therefore a reduced fit,
compared to the trained network. This can be seen in Fig. 7.
The Euclidean errors were 0.52, 0.50, and 0.53, respectively
for the greedy subset, full population, and regularized subset
techniques.
Fig. 8 shows a more comprehensive summary of the cross-
validated fitting errors across a range of random weight
controls. The figure first summarizes the main results on the
trained weights, before showing the results for the various
randomized conditions. As described in the previous sub-
section, the trained L1 neural units (layer 1) exhibited the
largest fitting errors among all layers and controls, meaning
they resulted in a poor fit (hence little correspondence) to
the neurophysiology V2 data (see also Fig. 6a, 6b, 6c). The
trained L2 neural units exhibited the lowest fitting errors in
all fitting techniques (greedy subset 0.22, full population 0.19,
and regularized subset 0.23), meaning that L2 achieved better
correspondence to the V2 data (6d, 6e, 6f).
We exhaustively explored a range of controls for randomiz-
ing the CNN layer weights (Fig. 8), and therefore considering
the influence of the architecture alone. Overall, assigning
random weights to the network layers increased the cross-
validated fitting error. Randomizing only layer 2 weights and
keeping the trained L1 weights (rand 2) lead to comparatively
much better fits than randomizing both layer 1 and layer
2 weights (rand 12). The rand 2 fits were much closer to
the L2 trained model, indicating that training the first layer
alone went a long way in obtaining a good fit; but was still
significantly worse than the L2 trained (p < 0.000002 in
greedy; p < 0.00007 in optimal; p < 0.00002 in regularized;
one sample t-test).
We wondered if deeper random architectures could lead
to better correspondence with the brain data. We therefore
considered randomizing layer 1, 2 and 3 weights (rand 123)
and randomizing layer 1, 2, 3 and 4 weights (rand 1234).
9
(a) L1L2 random, subset greedy
(b) L1L2 random, full population
(c) L1L2 random, subset regularized
Fig. 7: V2 fits of the CNN architecture with random weights (rather than weights learned from natural images), cross validated.
Both the L1 and L2 weights are randomly selected (hence denoted as L1L2 random). (a) Greedy subset selection fits (error:
0.52). (b) full population optimally weighted fits (error: 0.50). (c) regularized subset selection fits (error: 0.53). Randomization
of the CNN weights of L1 and L2 leads to reduced compatibility of L2 with the V2 data (compare to Fig. 6).
For these conditions, we fit the outputs of layer 3 and 4
respectively, to the data. The goal here was to see if stacking
more random layers helped in obtaining a better fit to the
data. However, the error remained high even when we stacked
together 4 layers (compare rand 12 with rand 123 and rand
1234). Therefore, a deeper random network did not rescue the
fit.
We also asked what happens if the trained weights within
each filter are shuffled to destroy any spatial correlations. This
maintains the distribution of the trained weights in each of the
filters. To test this, we spatially shuffled the trained weights
for each of the filters in both layers (shuffle 12). We found that
this resulted in a better fit than the randomized counterpart,
but still remained rather poor (compare rand 12 vs shuffle
12). In this case, the fitting errors stayed slightly lower than
the random because the network might benefit from having
the weights come from the same distribution as the trained
(albeit that the weights are scrambled). Nevertheless, the error
remained high, even in the shuffled version.
These factors give a clear indication that training the model
on image classification leads to a better correspondence with
the brain texture selectivity data. Even one trained layer is
able to influence the subsequent layer(s) to gain some cortical
correspondence.
On one hand, this reveals the necessity of training the deep
learning models on the natural image dataset beforehand to
achieve a better match to the V2 texture sensitivity data (and
also high recognition accuracies for that matter). Other studies
have also indicated the necessity of model
training [10].
Indeed, the primate brain is also "trained" on natural scene
data (albeit not necessarily in the same supervised manner).
Nevertheless, it is interesting that the architecture alone can
partly account for the data (see also [40], [41]).
4) Effect of normalization and different CNN computa-
tions in texture selectivity: An important question regarding
the CNN is how the various computations influence the
compatibility of the model to the data.
AlexNet includes a local response normalization in L1 and
L2, whereby each unit response is divisively normalized by the
responses of 5 units (including the self) that spatially overlap.
This loosely mimics the divisive cross-orientation suppression
in cortical V1 neurons [33]. We therefore wondered: What
happens to the selectivity if we ignore the local normalization
of L1 (norm1) and L2 (norm2) layers altogether?
We first considered removing normalization from the CNN
with random L2 weights, following the approach of the
previous subsection. We found that the local normalization
had mild impact on the compatibility with the biological data.
This can be seen in Fig. 8 (rand 2 vs rand 2 noN; p < 0.001
in greedy, p = 0.70 in full population, which was high and
did not pass significance, and p < 0.009 in subset regularized;
independent sample t-test).
A more direct way to tease apart the different computations
(conv, ReLU, pool, norm) involved in L2, is to consider the
intermediate outputs of the CNN trained on the ImageNet
database as in Fig. 1b, and to quantify the impact of each
of these on the compatibility with the V2 data. This gives
us a sense of how much each of the computations contribute
to capturing the high-order statistics. We therefore generated
outputs from each of the points in L2.
Outputs from conv2 (i.e., after only the convolution in the
second layer) had high fitting errors. This is because the
response from the conv layers can be negative before the
ReLU. We found that compatibility to the V2 data starts to
develop already after the rectification (i.e., ReLU2) stage (with
Euclidean errors of subset greedy 0.33, full population 0.31,
subset regularized 0.30). The fitting errors after pooling were
(subset greedy 0.23; full population 0.20; subset regularized
0.24). After the local normalization (i.e., the point in L2 that
we initially referred to in all our measurements), the fitting
errors were (subset greedy 0.22; full population 0.19; subset
regularized 0.23). The main improvement in the fit appeared
to be at the L2 pooling stage.
As a third way to gauge the importance of local nor-
malization, we retrained AlexNet on the ImageNet database,
but with the local normalization layers removed. From an
object recognition perspective, removing the normalization
layers in the CNN model decreased the accuracy with a
small margin (from 57.0% to 55.71%), echoing previous
observations. Fitting errors with and without normalization
0.00.10.20.3ModulationinBiology,V20.00.10.20.3ModulationinCNNE=0.520.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.500.00.10.20.3ModulationinBiology,V20.00.10.20.3E=0.53 10
Fig. 8: Summary of the CNN model fitting errors to the V2 data (cross-validated), including various model manipulations. rand
2 denotes applying random weights only for L2 and keeping the trained weights for L1; rand 12 denotes applying random
weights for both L1 and L2; rand 123 denotes random weights for L1, L2 and L3; and so on. shuffle 12 denotes shuffling the
weights in L1 and L2; and noN denotes no normalization in the CNN layers. Errors are measured in Euclidean distance on
225-fold (leave-one-out) cross-validation. Errors in random and shuffled weights are averaged over 10 iterations and error bars
show their standard deviation. L1 units (layer 1) cannot fit the V2 data well (see also Fig. 6; first row). Errors are minimum
in L2 (layer 2), indicating that the second layer of the CNN could already well match the cortical V2 data for the texture
sensitivity (see also Fig. 6; second row). Randomizing the weights in the CNN overall increases the fitting errors and hence
reduces the compatibility for the texture data. This indicates the importance of training the CNNs on natural scenes to develop
texture selectivity, resulting in a better match to the V2 data. See a more detailed discussion about the different randomized
conditions in the main text.
were: subset greedy: 0.22 vs 0.30; full population: 0.19 vs
0.27; and subset regularized: 0.24 vs 0.30. This indicated
that the trained model with local normalization resulted in a
better fit than the model trained without normalization. Taken
together, our results suggest that normalization had only a mild
role in improving compatibility.
5) Fitting higher layers of the CNN: So far, our main
focus has been when selectivity to textures first develops,
strongly differentiating L1 from L2. Nevertheless, it is in-
teresting to consider how this changes across higher levels
of the deep network. The results for Euclidean error are
summarized in Table I. An increase in error might be expected
at higher layers, similar to the observation that area V4 better
discriminates synthesized "jumbled image" textures than area
IT [25], [42]. However, for higher layers of the CNN beyond
L2, the error remained lower. Overall, the main change in error
was between L1 and L2, compared to the smaller changes for
the higher layers.
D. Selectivity in other hierarchical model architectures
We have thus far focused on the AlexNet CNN, and obtained
a good fit to V2 with L2 (but not L1) units. Our main approach
was to take a single popular base model and its computational
building blocks, and examine how manipulating certain pieces
impacts the compatibility of the CNN model with the brain
V2 data.
In this section, we consider several other popular models in
the literature. We note that we are taking these models "off the
shelf" as is usually done in the literature, but that the various
models differ from each other in many ways (e.g., the exact ar-
chitecture and Linear Nonlinear operations, the receptive field
size in each layer, and the number of units). Nevertheless, this
exercise is informative for considering changes that develop
across the first and second layers.
1) Analysis of the VGG network layers: We have focused
in detail on one particular popular deep CNN, AlexNet [1]. We
also applied our methodology, using the same input images as
in AlexNet, to the VGG16 [43] network (which we denote
VGG),
trained on the same ImageNet dataset. We found
similar trends in the VGG layers as we have seen in AlexNet.
However, we found that the VGG develops selectivity more
gradually than AlexNet. Starting from L1, the fit kept improv-
ing, with the largest reduction between L3 and L2 (similar
to the L2 to L1 comparison in AlexNet). The Euclidean
errors quantify this trend (Table I). In particular, between
L2 and L3, the errors reduced from 0.42 to 0.15 in greedy,
from 0.33 to 0.12 with the full population, and from 0.43
to 0.17 in the subset regularized technique. The fit for L4
remained similar to L3, with some increase in error for L5
layer 1 layer 2 rand 2 rand 12 rand 123 rand 1234shuffle 12 rand 2 noN0.00.20.40.60.7CNN fitting errors to V2subset greedyfull populationsubset regularized TABLE I: Euclidean error distances between the model fits and V2 neurophysiology data, in all layers of AlexNet and VGG
net (cross-validated). In AlexNet, the most significant reduction in fitting errors (hence increase in correspondence with V2)
happens in L2; in the VGG the same happens in L3.
11
AlexNet
VGG Net
Approach
Subset greedy
Full population
Subset regularized
L1
0.63
0.59
0.62
L2
0.22
0.19
0.24
L3
0.17
0.11
0.24
L4
0.09
0.07
0.14
L5
0.16
0.09
0.19
L1
0.63
0.61
0.62
L2
0.42
0.33
0.43
L3
0.15
0.12
0.17
L4
0.10
0.06
0.20
L5
0.21
0.03
0.35
(a) Fitting errors
(b) Modulation index
Fig. 9: HMAX and ScatNet model fitting errors and mean modulation index. (a) Euclidean errors for the HMAX and ScatNet
models, cross-validated. Both of the models show the same trend for the second versus the first layer fits, but have higher
errors (hence reduced fit) in comparison to the CNN L2 layer. (b) Average modulation indices computed for many random
draws of 103 units show that the second layer in both models has reduced selectivity to the textures vesus the noise (low or
negative mean modulation index values) relative to the cortical data. In contrast, the CNN has higher sensitivity to the textures
versus the noise already in the second layer (see Fig. 3a).
with the subset selection methods. Since the VGG units have
a smaller receptive size than AlexNet, we wondered if further
downsampling the textures to obtain a more realistic RF size
match between L2 and V2, would improve the result. However,
downsampling further showed unreasonably high fitting errors,
perhaps because in this case the images become too blurry and
partially lose their texture properties.
In sum, for the VGG network, compatibility with the texture
data first emerged in L3, and the change between L2 and L3
was the most striking. This means that L3 in VGG becomes
more comparable with L2 in AlexNet (and to V2 in the brain).
This is probably because the increase of the RF size in AlexNet
is more rapid than in the VGG. L2 RF size in AlexNet is
39 × 39 (with stride = 2 in L1), but the same in the VGG net
is only 16 × 16; in contrast, the RF size in L3 of the VGG
net is 44 × 44. This is also consistent with the observation
that the modulation index for the texture images computed
over the Portilla and Simoncelli parameters [36] is weaker
and more variable at small sizes (Corey Ziemba, personal
communication; [44]). The VGG also does not include nor-
malization, but based on our manipulations with AlexNet, we
believe that its effect is minimal for the trained network in
terms of compatibility with the V2 data. Overall, we found
that middle layers in the VGG showed better compatibility
with the biology.
2) Analysis of HMAX and ScatNet: We have thus far
focused on the CNN class of models. We wanted to further
ask if differences develop across the first two layers for some
related hierarchical visual models, such as HMAX [18], [17]
and ScatNet [15], [16], and how their selectivity compares to
the CNN model.
HMAX is a popular model [18], [17], (see also [45])
motivated by the hierarchical organization of visual cortex.
HMAX builds an increasingly complex and invariant feature
representations in the upper layers and has been shown to be
robust in recognizing both shape and texture-based objects.
We presented the same texture stimulus ensembles and gen-
erated the response from both layers, named C1 and C2 (see
Methods). We found that the HMAX model layer 2 fit the
V2 data better than the HMAX layer 1. However, compared
to the AlexNet CNN, the Euclidean errors for HMAX were
considerably higher (greedy: 0.53 vs 0.22; full population: 0.44
vs 0.19; subset regularized: 0.52 vs 0.24) as shown in Fig. 9a.
Can we further pinpoint what is different in the HMAX se-
lectivity to the textures versus the noise? Our subset selection
approach finds a population of units that can best fit the data.
However, it doesn't give an indication of the selectivity of the
population on average to the textures versus the noise. We
C1C2M1M20.00.51.0Fitting errors to V2HMAXScatNetsubset greedyfull populationsubset regularizedHMAXScatNet−0.100.000.15Average modulation indexlayer 1layer 2 therefore considered as a complementary approach, randomly
selecting 103 units 10000 times from the HMAX model,
similar to what we did for AlexNet in our initial analysis.
We found that the mean modulation index for the first two
layers of HMAX respectively were (C1: -0.04 and C2: -0.01;
see Fig. 9b). This negative mean modulation index means that
on average, a random selection of 103 units in the HMAX was
more selective to the noise than to the textures in the second
layer. This was quite different from the V2 cortical data (e.g.,
-0.05 in the HMAX versus 0.12 in the V2 data). This is also
in contrast to the CNN, which on average was more selective
to the textures than the noise in L2, more similar to the V2
data (0.18 versus 0.12 in the V2 data). The CNN model was
therefore more compatible to the cortical data.
We also considered another network model, called scattering
convolutional network or ScatNet [15], [16], that computes
invariant representations by wavelet transforms and pooling
(see Methods). This network is able to discriminate textures by
incorporating higher order moments. As with the other models,
we found that the second layer of ScatNet fits the V2 data
better than the first layer of ScatNet (Fig. 9). However, in
comparison, the CNN L2 units achieved better compatibility
to the biological cortical data than the ScatNet layer 2 (Fig.
5b and 5e). The Euclidean error distance in ScatNet layer 2
was higher than in the CNN L2 (greedy: 0.39 vs 0.22; full
population: 0.39 vs 0.19; subset regularized: 0.39 vs 0.24).
We also found that the average modulation index of randomly
selected 103 units was (M1 -0.14 and M2 0.05; see Fig. 9b).
This indicates the reduced sensitivity of ScatNet units towards
textures versus spectrally matched noise.
Our observation is that
the CNN overall shows better
correspondence with the V2 data compared to HMAX and
ScatNet across the first two layers. Here the receptive field
size is not a limiting factor as in the VGG, since the size
develops more rapidly. It is important to realize that the CNN
model differs from HMAX and ScatNet across many factors.
First, these models differ in the approach to training. The CNN
is trained on natural images in a supervised discriminative
manner. The universal filters in the second layer of the HMAX
model are obtained by extracting prototypes from a set of
natural images. ScatNet uses an energy preservation property.
One possibility is that the discriminative training of the CNN
(as opposed to image prototyping or preservation) encourages
faster development of sensitivity to the higher order statistics
that distinguish the texture images from the noise. In addition,
both the HMAX and ScatNet models we used consisted of two
layers and the learning is not influenced by later layers, but
AlexNet is deeper and has eight layers that are included in
the supervised training; this may pose an advantage in the
learning.
However, the amount of features learned is significantly
outnumbered in the other models compared to the CNN.
Whereas for the AlexNet CNN model we used 512 units (2×2
neighborhood, 128 filters) in L2, the HMAX has 3200 units
(8 patch sizes, 400 prototypes), and the ScatNet has 1536
units (2 × 2 neighborhood, 384 filters). This may pose an
advantage for the other models in selecting a population subset
that best fits the data. There are also differences in the exact
12
computational building blocks. However, all models include
a pooling operation, with the max pooling in both HMAX
and the CNN. All models also include a linear stage, with
convolution in both the ScatNet and the CNN.
II. DISCUSSION
A number of related hierarchical models in the machine
learning and neuroscience literature include similar basic com-
putational building blocks stacked together, namely, convolu-
tion, spatial pooling, and sometimes local response normal-
ization. This paper is motivated by the intriguing question:
What makes CNNs, which are only very crudely matched to
the brain architecture, able to capture some aspects of cortical
processing surprisingly well?
We specifically focused on texture processing in early areas
of visual cortex, for which [19] have compellingly shown
develops in V2 but not in V1. This provided a rich data set to
quantitatively compare the CNN and other related classes of
hierarchical models. We found that L2 (but not L1) of AlexNet
could well fit the V2 data. We also showed some qualitative
similarities for texture selectivity between the first two layers
of the CNN and the first two cortical visual areas.
We were interested in the question of when this compatibil-
ity first emerged, since in cortex there is a clear difference
between the first two cortical neural areas for the texture
stimuli. For a number of models we examined (AlexNet [1],
VGG16 Net [43], HMAX [17], ScatNet [15]), we found that
the details of the architecture and training made a difference
regarding the compatibility of the model with the V2 data, and
how quickly this developed across layers of the hierarchical
network.
We presented initial versions of this work and the ability of
deep neural networks to qualitatively capture some of the V2
versus V1 texture data in abstract form [46]. Ziemba et al. have
shown that a descriptive model of V2 can capture some of the
qualitative results on the increased sensitivity to naturalistic
textures in abstract and thesis form [47], [48]. Zhuang et al.
showed increased sensitivity to textures versus noise in higher
layers of deep neural networks, and related this to sparsity
[49]. The work described here, in contrast, asks at what point
texture selectivity first emerges in the CNN layers, and focuses
on both qualitatively comparing and quantitatively fitting the
experimental data to changes that develop across the first two
cortical areas.
What factors are important for better compatibility between
the CNN and the cortical V2 data already in the second layer?
We found that training on natural images was necessary for
the model to develop compatibility with the cortical data.
Various manipulations of random or shuffled weights could
partly account for the modulation index data, but lead to a
reduced fit between the CNN model and the V2 data relative
to the learned weights. This indicated that the architecture by
itself was not sufficient to obtain good compatibility. However,
interestingly, having a trained first layer but random weights
in the second layer, resulted in a good fit. The importance of
retaining the trained weights (rather than random weights) in
the first layer may be because the filters need to be matched to
the frequencies and orientations that appear more often in the
natural images in order to pick out the higher order structure
of the textures in the subsequent layer.
In addition, we found that the receptive field size needed to
develop sufficiently fast (comparing AlexNet and the VGG),
and that the local normalization in Alexnet only had a limited
role in obtaining good fits to the texture data. Both the HMAX
and ScatNet followed a similar trend of better compatibility
in the second versus the first layer, but overall did not fit the
V2 data as well, and showed less sensitivity to the textures
versus the noise in the second layer. This may be due to
the discriminative learning in the CNN, versus the image
prototyping or preservation of energy in the other models.
Although the AlexNet CNN showed compatibility with the
V2 data, it too had some deviations from the cortical data.
For instance, in the qualitative results in Fig. 2 it is intriguing
to see that, with the same amount of neural units, the brain
V2 outperforms the CNN L2 at grouping together different
texture categories. In addition, as indicated by the modulation
index, the CNN on average was more selective to the textures
versus the noise than the V2 population. Its rank order of
the selectivity to the textures on average also deviated from
the data. This suggests that
the CNN still has room for
improvement
in terms of capturing the cortical data. One
possibility for improvement in future work is incorporating
more realistic models of surround normalization (see, e.g., the
range of surround normalization models used for modeling V1
data in [50]).
Local response normalization is a computation prevalent in
visual cortex [34]. It's possible that the limited role of normal-
ization in obtaining compatibility with the V2 data is due to the
homogeneous nature of the textures. It is possible that divisive
normalization will play a more important role in capturing
data for non homogeneous images. Therefore, future work
should test compatibility with V2 data over a broader range of
natural stimuli and tasks. Future work should also incorporate
other well known forms of cortical divisive normalization (e,g.,
surround normalization) into CNNs and consider their role in
capturing cortical data across the hierarchy.
In addition, adding more units per layer may also play
a role, and we found that choosing from a larger spatial
neighborhood could improve the CNN fit. However, larger
spatial neighborhoods did not reveal a good fit for the L1
units in the AlexNet, and the first layer of all models was
not compatible with the V2 data. This also resonates with
the original texture model of [36] that was actually used to
generate the experimental stimuli; although the model does
not have an explicit V2 unit representation, the textures are
generated by joint statistics between V1 model units,
i.e.
by a 2-layer model. Though we have not exhausted all the
possible hierarchical models, our method is pragmatic enough
to be applied to any hierarchical models to test and find
correspondence with the brain neurophysiology data.
While we have focused on texture selectivity in V2, there is
room to explore compatibility of CNNs with changes across
the early cortical hierarchy for other stimulus properties. For
instance, biological studies have found selectivity in V2 to
conjunctions of orientations, and to figure ground [51], [52],
13
[53], with some aspects addressed in computational models
of V2 [54], [55], [56], [57], [58] [59]. There may not be one
unique CNN architecture that explains the neural data, but
we believe that testing across fairly early visual areas (e.g.,
V1 and V2) with less stacked computations, and for a wider
range of stimuli and tasks, can facilitate understanding of the
critical factors and computations. Beyond area V2, studies
have also shown that CNN units are compatible with shape
tuning properties in visual area V4 [12].
In the quantitative comparisons between the modeling and
data, we developed approaches for subset selection. These
were more appropriate for the given cortical data than a
weighted average of the neural population, which is typically
used in fitting data. This is because the subset approach more
faithfully represented the data analysis, which included an
equally weighted average modulation index. This approach
also allowed us to ask the question about whether there exists
a population of units in the CNN that can well represent the
experimental data. We therefore chose a neural population
from the representation itself rather than a linear transform
of the representation. By finding a subset of neural units that
are most compatible with the data, it may be possible in the
future to drive new experiments in which stimuli are generated
from this population of model units and tested on the data.
This may be applied more generally in the future to modeling
other data sets and neural areas.
On one hand, our results add to the intriguing findings that
CNNs trained on natural images have some compatibility with
biological data, and moreover we found that this holds across
low levels of the cortical hierarchy. But we believe that our
approach goes beyond showing compatibility, by providing
a direction for manipulating these early layers and teasing
apart what aspects of CNNs (and other related hierarchical
models), the training, and computational building blocks are
most critical. This creates the opportunity for more discussion
and systematic study of the various building blocks of deep
networks, and opens the door to answer the long standing
research question about correspondence between primate and
machine vision.
III. ACKNOWLEDGMENTS
We are grateful to Adam Kohn, Ruben Coen Cagli, and
Corey Ziemba for discussions and very helpful comments on
the manuscript; to Eero Simoncelli and Corey Ziemba for
discussions and providing us with the original texture dataset
used in the experiments; and to David Grossman and Ariel
Lavi for discussions during their REU research. This work was
supported by a Google Faculty Research Award, the National
Science Foundation (grant 1715475), and a hardware donation
from NVIDIA.
A. Texture generation
IV. METHODS
For the CNN simulations, we used the same ensemble of
synthetic texture images as in [19] and [20]. The synthetic
images were grayscale images of size 320×320 and generated
from an original set of 15 texture images. From each original
texture, multiple synthetic texture images that matched the
statistics of the original image were generated. Naturalistic
textures for a given family were generated each with a dif-
ferent random seed, using an iterative process of constraining
Gaussian white noise images to have similar marginal and joint
statistical properties of the original textures [36]. Spectrally
matched noise images were generated by randomizing the
phase, i.e. computing the Fourier transform and inverse Fourier
transform after phase randomization. From the 15 original
textures, we have 15 different samples from each family,
resulting in a total of 225 images of naturalistic textures,
and 225 images of spectrally matched noise, as used in [19]
and [20]. For cross-validation, we generated extra images per
texture family from the model of [36].
B. Matching receptive fields with the physiology data
We wanted to match as much as possible the spatial extent
of the images that the model receptive field (versus the typical
experimental neuron) is sensitive to. The input images were
size 256 × 256, and the average receptive field size for V2
was approximately 150 × 150 (with the V1 receptive field
approximately half that size). The receptive field size of units
in AlexNet is 39× 39 for L2 and 15 for L1. We downsampled
the input images by a factor of 4 to obtain images of size 64×
64, so that the effective size of the L2 receptive field was closer
to the neurons recorded from in the experiment. We could not
get an exact match, due to the constraint of downsampling by
factors of 2. For HMAX, we downsampled only by a factor of
2 to obtain images of size 128 × 128, since the C1 receptive
fields in HMAX can be up to 49 × 49 pixels. For ScatNet,
the filter sizes are determined relative to the input images, so
there was no need to downsample. We also originally ran the
whole set of simulations without downsampling the images at
all, and the results remained qualitatively similar, except that
there were light improvements in the compatibility to the data
with the appropriate downsampling.
Following the experiments, we contrast normalized the
images before feeding them to the networks (CNN, ScatNet,
Hmax). The luminance (l) is given by the mean pixel intensity
of the downsampled image (Id). The contrast (c) is given by
the standard deviation. The contrast normalized images (In)
are then defined as follows:
Id − l
c
In = α
+ β,
(2)
where the desired contrast α defines the range of the input
pixels and the desired luminance β defines the intensity
centered around the range. We use 0.22 as the value of α
and since the desired luminance is gray, we use 0.5 as the
value for β.
C. Deep CNN models for texture simulations
trained on natural
In our simulations, we mostly used the pre-trained AlexNet
model,
images and specifically on the
ILSVRC 2012 ImageNet [35] dataset. We also re-trained the
network on ImageNet, or on the Places365 database [39],
yielding similar results. We also contrasted this with an
14
equivalent model architecture that included random weights (in
the interval [−1, 1]) rather than pre-trained weights. AlexNet
consists of five convolution followed by three fully connected
layers. The first and second convolution layers are followed
by local response normalizations and max-pooling. We used
CaffeNet which is a variant of AlexNet where normalization
follows the pooling. We refer to this as AlexNet for conve-
nience. We examined the outputs from the first and second
normalization layers (along with a more exhaustive examina-
tion of other layers), and compared them to the experimental
data for V1 and V2 neuron outputs.
We used a modified version of AlexNet by changing the
stride at L1 from 4 to 2. This allowed us to significantly reduce
the receptive field size in L2 (from 67×67 to 39×39; with L1
of size 15×15), making it more comparable with the biological
ratio of V2 to V1 RF size. This modification also matched
the experimental data better in our simulations. In addition,
we simulated the response of the first 48 (instead of 96) L1
filters as they are the ones that show orientation selectivity;
the remaining are more color selective. These 48 filters are
the input to the first 128 (out of 256) channels in layer 2,
so we considered these first 128 filters from L2. We focused
mostly on the center four (2 × 2) spatial positions from each
of these selected channels. We tested our method on larger
neighborhoods and obtained qualitatively similar results.
We later considered the VGG16 network [43], trained on
the same ImageNet dataset as AlexNet. VGG16 is a 16 layer
network stacked with multiple (usually 2 or 3) convolution
layers with 3 × 3 filters and then followed by a pooling layer.
We examined outputs from those five pooling layers, which
we refer to as L1 through L5.
D. Other hierarchical models for texture simulations
We also fit the texture data to the HMAX model [17].
We generated responses from both layers, named C1 and
C2. The C1 layer includes a range of receptive field sizes,
with the largest approximately 49 by 49 pixels. The original
HMAX model pools over the full extent of the image for
L2; we reduced this to a 128 by 128 pixel region so that
the receptive field size ratio between a C2 and C1 unit is
more comparable to the receptive field size ratio between
V2 and V1. To compute the C2 responses, we employed the
universal set of units provided with the HMAX author's code,
available at [60]. Furthermore, since the original code outputs
the squared norm of the distance to the prototypes, we added
the exponential function and selected a scaling factor (denoted
as β in [17]) of 1 to resemble the Gaussian like tuning of the
neural responses.
We further fit the texture data to the scattering convolutional
network or ScatNet [15], [16], that computes translation and
deformation invariant representations by wavelet transforms
and modulus pooling in a hierarchical fashion. We used the
code provided by the authors, available at [61].
E. Local response normalization in CNN
Local response normalization plays an important role in
hierarchical object recognition models. Local response nor-
malization is used in both layer 1 and layer 2 of the AlexNet
and has been shown to improve the recognition accuracy. It
is loosely related to cross-orientation suppression in the brain,
by normalizing the responses of groups of units with spatially
overlapping receptive fields (5 in the case of AlexNet). If ai
x,y
is the rectified linear activation at the (x, y) position in each
i-th channel (or unit), then the normalized response bi
x,y is
defined by [1]
(cid:18)
bi
x,y =
ai
x,y
(cid:80)
min(N−1,i+m/2)
j=max(0,i−m/2)
k + α
(aj
x,y)2
(cid:19)β
(3)
where m is the size of the normalization neighborhood, N is
the total number of channels in the layer. Constants k, m, α, β
are the hyper-parameters with the default values of 2, 5, 10−4,
and 0.75 respectively.
Normalization is done across the spatially overlapping unit
activations across channels. Each 1 × 1 response is selected
and normalized with corresponding values of all the channels
across the channel dimension.
From a machine learning perspective, local response nor-
malization is specifically useful to normalize the unbounded
activations coming from the ReLU (Rectified Linear) non-
linearity. It detects high-frequency features with large re-
sponses and penalizes the responses which are uniformly large
in a local neighborhood. It is a type of regularization that
encourages competition amongst units in the network.
F. CNN population fitting and subset selection approaches
We considered the total number of units in the CNN as
the number of units in a given layer (e.g., 48 for layer 1 and
128 for layer 2), times a center 2-by-2 spatial neighborhood.
The rationale was that experimental data can be collected
for receptive fields at different spatial positions. We chose a
2-by-2 spatial neighborhood, and did not find a significant
difference when exploring larger spatial neighborhoods. We
selected 103 units from L2 and 102 units from L1, to match
the population numbers in the neurophysiology experiments
of [19]. Before starting the subset selection procedure, we
removed from consideration the CNN neural units that had
zero response to any family. This amounted to 432 units out
of a total of 512 units from which we selected the subset of
size 103.
1) Subset greedy approach: We consider the subset
greedy technique known as forward selection to choose a
subset of 103 units that best match the data from the cortical
neurons. In the greedy approach, the goal is to each time select
another neural unit so as to minimize the Euclidean distance
between the neural data and the CNN model modulation
indices, until we have chosen 103 units from the available
CNN layer population. The approach is greedy, because it
optimizes the selection of the next unit as best it can given the
current set of units. However, it does not guarantee a globally
optimal solution.
Formally, let t be a 15-dimensional vector containing the
average modulation indices per texture family from the 103
recorded neurons in the physiological experiments, A be the
15
set of n CNN neural units, and mj the average modulation
indices per texture family of the jth simulated unit in A.
Starting from S (0) = ∅, the greedy algorithm adds a unit to
the current set of selected units that minimizes the squared
Euclidean distance between the biological data and model
average modulation indices:
S (k) = S (k−1)(cid:91)
argmin
j∈A\S (k−1)
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)t − 1
k
(cid:88)
j(cid:48)∈S (k−1)(cid:83) j
(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)(cid:13)2
2
mj(cid:48)
.
The above procedure is repeated until the desired number of
units is obtained. In particular, we stop at 103 units.
2) Optimal weighted average or full population: In this
approach, we find a weighted average of the set of simulated
units that best fit the biological data, by solving a constrained
optimization problem. The constraint guarantees that the sum
of the weights add to 1. This approach does obtain an optimal
solution, and therefore shows the best one can do. However,
note that it does not as faithfully capture the analysis of the
neural data, since for the data analysis an equally weighted
average of 103 neural responses give rise to the modulation
index data.
Formally, we have:
minimize
subject to wi ≥ 0, i = 1, ..., n and
(cid:107)M w − t(cid:107)2
w
2
n(cid:88)
wi = 1,
(4)
i=1
where M = [m1, m2,··· , mn]
is the matrix of average
modulation index values for all families computed according
to (1), and w is the vector of weights for all the n simulated
neurons. In terms of Euclidean distance, this is the best fit that
could be attained by considering a weighted average on the
simulated units. Nevertheless, note that the solution need not
be sparse since there is no mechanism forcing the weights to
become zero, and the disparity of the weight values can be
hard to interpret.
3) Subset regularized average followed by threshold:
As noted above, for the optimal weighted average, weights
wi can be very disparate. Since we seek to select a subset
of the simulated units whose regular average (all weights are
equal) follows closely the physiological experiments, we relax
the selection problem by solving a regularized version of the
optimization problem (5), as follows:
(cid:107)M w − t(cid:107)2
2 + λ(cid:107)w(cid:107)2
minimize
subject to wi ≥ 0, i = 1, ..., n and
w
2
(5)
n(cid:88)
wi = 1,
i=1
where λ > 0 is the trade-off parameter that promotes weight
equalization. For λ = 0, which is equivalent
to solving
(4), we found that only 14% of the simulated neural units
have the weights wi ≥ 2e−3 with only a handful of them
i wi = 1. As λ
containing large values that account for (cid:80)n
increases, the regularization term pushes the weights towards
the center of the simplex. For instance, for λ = 0.8, we
found that approximately 40% of the neural units have weights
wi ≥ 2e−3. The subset of units is selected by applying a
threshold to the estimated weights, as proposed in [62] and
then choosing the 103 units with the highest weights. However,
a main difference from [62] is that our two-stage procedure is
applied to the solution of (5) instead of (4). This approach also
yields an excellent fit to the V2 data for L2 units, as shown
in Fig. 5 (third row, middle column). We used λ = 0.8 for all
fits; lower λ increased the fitting error but did not alter the
trends (and vice versa).
G. Cross-validation
For the cross-validation, we extended the image dataset.
In the original data, there were only 15 images generated in
each family. We therefore extended the set by generating 210
additional images (texture and noise) from each of the texture
families. We optimized the learning, assuming that each group
of 15 images, out of the 225 in each family, should yield an
average modulation index that is as close as possible to the
mean modulation index for that family in V2. Therefore, for
225 images in each family, we randomly divided the images
into groups of 15. This yielded 15 data points per family, and
a total of 225 data points for all 15 families. We then applied
a 225 fold leave-one-out technique, training on 224 points and
leaving out one point (corresponding to leaving out one set of
15 images). We thus learned the population (e.g., of 103 units
in the greedy subset selection method) with the 224 training
data points, and made a prediction of the mean modulation
index for the left-one-out set of 15 images.
H. Euclidean distance as the measurement of correspon-
dence
To quantify the error between the CNN model and the
neurophysiology data, we use the Euclidean distance met-
ric. We calculate the Euclidean distance between the values
computed from the CNNs and the neurophysiology V2 data.
For the CNNs, the computed values include the modulation
indices obtained from the various fitting and subset selection
techniques, or predicted modulation index in the case of the
cross validation.
Given two vectors x = {x1, x2, ··· , xn} and y =
{y1, y2, ··· , yn} are two points in Euclidean n-space, the
Euclidean distance d(x, y) is computed using the 2-norm, as
follows:
(cid:118)(cid:117)(cid:117)(cid:116) n(cid:88)
d(x, y) =
(yi − xi)2.
(6)
i=1
In all our experiments, n is usually 15, the number of texture
categories, as we take the average over samples and/or units.
Lower Euclidean distances indicate a better fit of the model to
the V2 data, and therefore higher correspondence of the model
to the brain.
16
References
[1] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification
with deep convolutional neural networks," in Advances in neural infor-
mation processing systems (NIPS), pp. 1097–1105, 2012.
[2] Y. LeCun, Y. Bengio, and G. Hinton, "Deep learning," Nature, vol. 521,
pp. 436–444, 2015.
[3] Y. LeCun, B. Boser, J. S. Denker, D. Henderson, R. E. Howard, W. Hub-
bard, and L. D. Jackel, "Handwritten digit recognition with a back-
propagation network," in Advances in neural information processing
systems (NIPS), 1989.
[4] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning
applied to document recognition," Proceedings of the IEEE, vol. 86,
no. 11, pp. 2278–2324, 1998.
[5] M. Zeiler and R. Fergus, "Visualizing and understanding convolutional
networks," in European Conference on Computer Vision (ECCV), 2014.
[6] S.-M. Khaligh-Razavi and N. Kriegeskorte, "Deep supervised, but not
unsupervised, models may explain IT cortical representation," PLoS
Comput Biol, vol. 10, no. 11, p. e1003915, 2014.
[7] D. L. Yamins, H. Hong, C. F. Cadieu, E. A. Solomon, D. Seibert, and
J. J. DiCarlo, "Performance-optimized hierarchical models predict neural
responses in higher visual cortex," Proceedings of the National Academy
of Sciences, vol. 111, no. 23, pp. 8619–8624, 2014.
[8] N. Kriegeskorte, "Deep neural networks: A new framework for modeling
biological vision and brain information processing," Annual Review of
Vision Science, vol. 1, pp. 417–446, 2015.
[9] D. L. Yamins and J. J. DiCarlo, "Using goal-driven deep learning models
to understand sensory cortex," Nature Neuroscience, vol. 19, no. 3,
pp. 356–365, 2016.
[10] R. M. Cichy, A. Khosla, D. Pantazis, A. Torralba, and A. Oliva, "Deep
neural networks predict hierarchical spatio-temporal cortical dynamics
of human visual object recognition reveals hierarchical correspondence,"
arXiv preprint arXiv:1601.02970, 2016.
[11] G. Umut and v. G. Marcel A. J., "Deep neural networks reveal a gradient
in the complexity of neural representations across the ventral stream,"
The Journal of Neuroscience, vol. 35, no. 27, pp. 10005–10014, 2015.
[12] D. Pospisil, A. Pasupathy, and W. Bair, "Comparing the brain's repre-
sentation of shape to that of a deep convolutional neural network," in
The First International Workshop on Computational Models of the Visual
Cortex: Hierarchies, Layers, Sparsity, Saliency and Attention, ACM, 5
2016.
[13] C. F. Cadieu, H. Hong, D. L. K. Yamins, N. Pinto, D. Ardila, E. A.
Solomon, N. J. Majaj, and J. J. DiCarlo, "Deep neural networks rival the
representation of primate IT cortex for core visual object recognition,"
PLoS Compt Biol., vol. 10, no. e1003963, 2014.
[14] S. A. Cadena, G. H. Denfield, E. Y. Walker, L. A. Gatys, A. S. Tolias,
M. Bethge, and A. S. Ecker, "Deep convolutional models improve
images," bioRxiv,
predictions of macaque V1 responses to natural
p. 201764, 2017.
[15] J. Bruna and S. Mallat, "Invariant scattering convolution networks,"
Transactions on Pattern Analysis and Machine Intelligence (PAMI),
vol. 35, no. 8, pp. 1872–1886, 2013.
[16] L. Sifre and S. Mallat, "Rotation, scaling and deformation invariant
scattering for texture discrimination," in Computer Vision and Pattern
Recognition (CVPR), pp. 1233–1240, 2013.
[17] T. Serre, L. Wolf, S. Bileschi, M. Riesenhuber, and T. Poggio, "Robust
object recognition with cortex-like mechanisms," IEEE Trans. on Pattern
Analysis and Machine Intelligence, vol. 29, no. 3, pp. 411–426, 2007.
[18] M. Riesenhuber and T. Poggio, "Hierarchical models of object recog-
nition in cortex," Nature Neuroscience, vol. 2, no. 11, pp. 1019–1025,
1999.
[19] J. Freeman, C. M. Ziemba, D. J. Heeger, E. P. Simoncelli, and J. A.
Movshon, "A functional and perceptual signature of the second visual
area in primates," Nature Neuroscience, vol. 16, no. 7, pp. 974–981,
2013.
[20] C. M. Ziemba, J. Freeman, A. Movshon, and E. P. Simoncelli, "Se-
lectivity and tolerance for visual texture in macaque V2," Proc of Natl
Academy of Science (PNAS), vol. 113, no. 22, 2016.
[21] W. H. Merigan, "Cortical area V4 is critical for certain texture dis-
is not dependent on attention," Visual
crimination, but
Neuroscience, vol. 31, no. 23, pp. 8543–8555, 2000.
this effect
[22] A. Hanazawa and H. Komatsu, "Influence of the direction of elemental
luminance gradients on the responses of V4 cells to textured surfaces,"
Journal Neuroscience, vol. 21, pp. 4490–4497, 2001.
[23] F. Arcizet, C. Jouffrais, and P. Girard, "Natural textures classification
in area V4 of the macaque monkey," Exp Brain Research, vol. 189,
pp. 109–120, 2008.
[24] A. S. Nandy, T. O. Sharpee, J. H. Reynolds, and J. F. Mitchell, "The fine
structure of shape tuning in area V4," Neuron, vol. 78, pp. 10780–10793,
2013.
[25] G. Okazawa, S. Tajima, and H. Komatsu, "Image statistics underlying
natural texture selectivity of neurons in macaque V4," Proceedigs of
National Academy of Science (PNAS), vol. 112, no. 4, 2015.
[26] G. Okazawa, S. Tajima, and H. Komatsu, "Gradual development of
visual texture-selective properties between macaque areas V2 and V4,"
Cerebral Cortex, vol. 27, no. 10, pp. 4867–4880, 2016.
[27] B. Julesz, "Experiments in the visual perception of texture," Scientific
American, vol. 232, no. 4, pp. 34–43, 1975.
[28] H. Tamura, S. Mori, and T. Yamawaki, "Textural features corresponding
to visual perception," IEEE Transactions on Systems, Man, and Cyber-
netics, vol. 8, no. 6, pp. 460–473, 1978.
[29] B. Julesz, "Textons, the elements of texture perception and their inter-
actions," Nature, vol. 290, pp. 91–97, 1981.
[30] J. R. Bergen and E. H. Adelson, "Early vision and texture perception,"
Nature, vol. 333, no. 6171, pp. 363–364, 1988.
[31] J. Malik and P. Perona, "Preattentive texture discrimination with early
vision mechanisms," J. of the Optical Society of America. A, Optics and
image science, vol. 7, no. 5, pp. 923–932, 1990.
[32] T. S. A. Wallis, C. M. Funke, A. S. Ecker, L. A. Gatys, F. A. Wichmann,
and M. Bethge, "A parametric texture model based on deep convolutional
features closely matches texture appearance for humans," Journal of
Vision, vol. 17-5, no. 12, 2017.
[33] D. J. Heeger, "Normalization of cell responses in cat striate cortex,"
Visual neuroscience, vol. 9, no. 02, pp. 181–197, 1992.
[34] M. Carandini and D. J. Heeger, "Normalization as a canonical neural
computation," Nature Reviews Neuroscience, vol. 13, no. 1, pp. 51–62,
2012.
[35] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma,
Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and
L. Fei-Fei, "ImageNet Large Scale Visual Recognition Challenge,"
International Journal of Computer Vision (IJCV), vol. 115, no. 3,
pp. 211–252, 2015.
[36] J. Portilla and E. P. Simoncelli, "A parametric texture model based on
joint statistics of complex wavelet coefficients," International Journal
of Computer Vision (IJCV), vol. 40, no. 1, pp. 49–71, 2000.
[37] L. van der Maaten, "Accelerating t-SNE using tree-based algorithms,"
Journal of Machine Learning Research (JMLR), vol. 15, pp. 3221–3245,
2014.
[38] L. van der Maaten and G. E. Hinton, "Visualizing data using t-SNE,"
Journal of Machine Learning Research (JMLR), vol. 9, pp. 2579–2605,
2008.
[39] B. Zhou, A. Lapedriza, A. Khosla, A. Oliva, and A. Torralba, "Places:
A 10 million image database for scene recognition," IEEE Transactions
on Pattern Analysis and Machine Intelligence, 2017.
[40] K. Jarrett, K. Kavukcuoglu, M. Ranzato, and Y. LeCun, "What is the best
multi-stage architecture for object recognition?," in Computer Vision,
2009 IEEE 12th International Conference on, pp. 2146–2153, IEEE,
2009.
[41] A. Saxe, P. W. Koh, Z. Chen, M. Bhand, B. Suresh, and A. Y. Ng,
"On random weights and unsupervised feature learning," in Proceedings
of the 28th international conference on machine learning (ICML-11),
pp. 1089–1096, 2011.
[42] N. C. Rust and J. J. DiCarlo, "Selectivity and tolerance ("invariance")
both increase as visual information propagates from cortical area V4 to
IT," Journal of Neuroscience, vol. 30, no. 39, pp. 12978–12995, 2010.
[43] K. Simonyan and A. Zisserman, "Very deep convolutional networks for
17
large-scale image recognition," in International Conference on Learning
Representations (ICLR), 2015.
[44] C. Ziemba, R. Perez, E. Simoncelli, and J. Movshon, "Selectivity of
contextual modulation in macaque V1 and V2," in Annual Meeting,
Neuroscience, Nov 2017.
[45] K. Fukushima, "Neocognitron: A self-organizing neural network model
for a mechanism of pattern recognition unaffected by shift in position,"
Biological cybernetics, vol. 36, no. 4, pp. 193–202, 1980.
[46] M. N. U. Laskar, L. G. S. Giraldo, and O. Schwartz, "Deep learning
textures," in Computational and
captures V2 selectivity for natural
Systems Neuroscience (Cosyne), Abstract, 2017.
[47] C. M. Ziemba, R. L. Goris, J. A. Movshon, and E. P. Simoncelli,
"Characterizing receptive field selectivity in area v2," in MODVIS
Workshop, VSS abstract, 2015.
[48] C. M. Ziemba, Neural representation and perception of naturalistic
image structure. PhD thesis, Center for Neural Science, New York
University, New York, NY, May 2016.
[49] C. Zhuang, Y. Wang, D. Yamins, and X. Hu, "Deep learning predicts
correlation between a functional signature of higher visual areas and
sparse firing of neurons," Front. Comput. Neuroscience, vol. 11, no. 100,
2017.
[50] R. Coen-Cagli, A. Kohn, and O. Schwartz, "Flexible gating of contex-
tual modulation during natural vision," Nature Neuroscience, vol. 18,
pp. 1648–1655, 2015.
[51] M. Ito and H. Komatsu, "Representation of angles embedded within
in area V2 of macaque monkeys," The Journal of
contour stimuli
Neuroscience, vol. 24, pp. 3313–3324, 2004.
[52] Y. El-Shamayleh and J. A. Movshon, "Neuronal responses to texture-
defined form in macaque visual area V2," The Journal of Neuroscience,
vol. 31, no. 23, pp. 8543–8555, 2011.
[53] H. Zhou, H. S. Friedman, and R. Von Der Heydt, "Coding of border
ownership in monkey visual cortex," The Journal of Neuroscience,
vol. 20, no. 17, pp. 6594–6611, 2000.
[54] D. J. Heeger, E. P. Simoncelli, and J. A. Movshon, "Computational
models of cortical visual processing," Proceedigs of National Academy
of Science (PNAS), vol. 93, no. 4, pp. 623–627, 1996.
[55] H. Lee, C. Ekanadham, and A. Ng, "Sparse deep belief net model for
visual area V2," in Advances in neural information processing systems,
pp. 873–880, 2008.
[56] R. Coen-Cagli and O. Schwartz, "The impact on midlevel vision of
statistically optimal divisive normalization in V1," Journal of Vision,
vol. 13, no. 8, 2013.
[57] H. Hosoya and A. Hyvarinen, "A hierarchical statistical model of natural
images. explains tuning properties in V2," The Journal of Neuroscience,
vol. 35, no. 29, pp. 10412–10428, 2015.
[58] L. Zhaoping, "Border ownership from intracortical interactions in visual
area V2," Neuron, vol. 47, no. 1, pp. 143–153, 2005.
[59] J. Hegde and D. C. V. Essen, "Selectivity for complex shapes in primate
visual area V2," Journal of Neuroscience, vol. 20, no. RC61, 2000.
[60] http://maxlab.neuro.georgetown.edu/hmax.html#code, 2007.
accessed 23-April-2018].
[Online;
[61] http://www.di.ens.fr/data/software/scatnet/download/, 2013. [Online; ac-
cessed 23-April-2018].
[62] P. Li, S. Sundar Rangapuram, and M. Slawski, "Methods for Sparse and
Low-Rank Recovery under Simplex Constraints," ArXiv e-prints, May
2016.
|
1708.05282 | 1 | 1708 | 2017-08-01T17:33:58 | Extracranial estimation of neural mass model parameters using the Unscented Kalman Filter | [
"q-bio.NC"
] | Data assimilation, defined as the fusion of data with preexisting knowledge, is particularly suited to elucidating underlying phenomena from noisy/insufficient observations. Although this approach has been widely used in diverse fields, only recently have efforts been directed to problems in neuroscience, using mainly intracranial data and thus limiting its applicability to invasive measurements involving electrode implants. Here we intend to apply data assimilation to non-invasive electroencephalography (EEG) measurements to infer brain states and their characteristics. For this purpose, we use Kalman filtering to combine synthetic EEG data with a coupled neural-mass model together with Ary's model of the head, which projects intracranial signals onto the scalp. Our results show that using several extracranial electrodes allows to successfully estimate the state and parameters of the neural masses and their interactions, whereas one single electrode provides only a very partial and insufficient view of the system. The superiority of using multiple extracranial electrodes over using only one, be it intra- or extracranial, is shown over a wide variety of dynamical behaviours. Our results show potential towards future clinical applications of the method. | q-bio.NC | q-bio |
Extracranial estimation of neural mass model
parameters using the Unscented Kalman Filter
Lara Escuain-Poole1, Jordi Garcia-Ojalvo2,*, Antonio J. Pons1,*
1 Department of Physics, Polytechnic University of Catalonia, Terrassa, Spain
2 Department of Health and Experimental Sciences, Pompeu Fabra University,
Barcelona, Spain
* [email protected], [email protected]
Abstract
Data assimilation, defined as the fusion of data with preexisting knowledge, is
particularly suited to elucidating underlying phenomena from noisy/insufficient
observations. Although this approach has been widely used in diverse fields, only
recently have efforts been directed to problems in neuroscience, using mainly in-
tracranial data and thus limiting its applicability to invasive measurements involv-
ing electrode implants. Here we intend to apply data assimilation to non-invasive
electroencephalography (EEG) measurements to infer brain states and their char-
acteristics. For this purpose, we use Kalman filtering to combine synthetic EEG
data with a coupled neural-mass model together with Ary's model of the head,
which projects intracranial signals onto the scalp. Our results show that using
several extracranial electrodes allows to successfully estimate the state and pa-
rameters of the neural masses and their interactions, whereas one single electrode
provides only a very partial and insufficient view of the system. The superior-
ity of using multiple extracranial electrodes over using only one, be it intra- or
extracranial, is shown over a wide variety of dynamical behaviours. Our results
show potential towards future clinical applications of the method.
Author Summary
To completely understand brain function, we will need to integrate experimental
information into a consistent theoretical framework. Invasive techniques as EcoG
recordings, together with models that describe the brain at the mesoscale, provide
valuable information about the brain state and its dynamical evolution when com-
bined with techniques coming from control theory, such as the Kalman filter. This
method, which is specifically designed to deal with systems with noisy or imper-
fect data, combines experimental data with theoretical models assuming Bayesian
inference. So far, implementations of the Kalman filter have not been suited for
non-invasive measures like EEG. Here we attempt to overcome this situation by
introducing a model of the head that allows to transfer the intracranial signals
produced by a mesoscopic model to the scalp in the form of EEG recordings. Our
results show the advantages of using multichannel EEG recordings, which are
extended in space and allow to discriminate signals produced by the interaction
of coupled columns. The extension of the Kalman method presented here can be
expected to expand the applicability of the technique to all situations where EEG
1/23
recordings are used, including the routine monitoring of illnesses or rehabilitation
tasks, brain-computer interface protocols, and transcranial stimulation.
Introduction
After several decades studying its morphology and dynamics [1], the basic mech-
anisms that describe the functioning of the brain are still far from being com-
pletely understood. There are different reasons that explain this arduous route
towards understanding this organ. First, the neurons that form the brain are very
diverse morphologically [2] and dynamically [3]. Second, these neurons are con-
nected to each other in extremely large numbers and forming very complex net-
works [4], whose structural characteristics are still mostly unknown. And third,
brain dynamics are very irregular and complex [5, 6]. The opposed views of an
essentially noisy brain and a deterministic brain exhibiting chaotic activity have
been often contrasted. On the one hand there is multiple evidence, both theoret-
ical and experimental, that justifies a stochastic view of the brain [7, 8]. On the
other hand, other studies reveal deterministic, or rather reproducible, dynamical
behaviour [9,10] both at the microscopic scale [11] and at the mesoscale recorded
by electroencephalograms (EEG) or magnetoencephalograms (MEG) [12]. The
reality is probably a combination of the two views. The fact that the brain receives
continuous external inputs from the sensory system also makes its dynamical and
experimental interpretation more complex because, even though experiments are
designed to minimize uncontrolled inputs, they cannot completely rule them out.
Another important limitation for studying the brain is that experimental record-
ings (such as EEG or fRMI) are almost always indirect reflections of the underly-
ing neural activity [13].
A way of facing the complexities described above is by systematically com-
paring the experimental observations of brain activity with mathematical models
based on specific hypotheses, which can thereby be validated or disproven. Mod-
elling cerebral activity has been attempted both with top-down and bottom-up ap-
proaches [14–17]. Many of these theoretical models are simplifications that cap-
ture the basic ingredients of brain dynamics, while others are detailed accounts of
the dynamics of neurons that necessarily forgo the description of the whole brain.
In that context, a more feasible scale of study is the mesoscopic scale [18–24].
Many of the modern experimental techniques record information coming from
populations of neurons organized in so-called cortical columns. Neural mass
models describe mathematically the activity of these populations using reason-
ably simple equations [25]. These models can describe both the intrinsic oscilla-
tory behaviour recorded at the mesoscale or event-related responses [26, 27] with
morphologically plausible assumptions for their construction.
In all modeling strategies, however, identifying realistic values for the pa-
rameters of the model is a challenging task. One way to address this problem
is by integrating experimental information into the models using Bayesian in-
ference [28–32]. This strategy has started to be pursued by using Kalman fil-
tering to integrate experimental data at both the microscopic scale of neuronal
networks [33] and the mesoscopic scale of neural mass models [34]. This data
assimilation approach is based on the fact that neuronal activity is highly noisy,
and allows to estimate both the state and the parameters of the theoretical model
using the experimental data available. The method has been used to estimate,
for example, the effective connectivity that characterizes epileptic seizures on a
patient-specific basis (see [35] and references therein). Kalman filtering has also
been used to analyze the suppression of epileptic seizures in coupled neural mass
2/23
models [36], and the induction of the anaesthetised state by drugs [37, 38]. But
these studies use mainly invasive intracranial signals, and it would be desirable to
extend them to non-invasive extracraneal measurements such as EEG. Intracra-
nial signals can be translated into EEG signals in a forward manner [39, 40], and
in the opposite direction, solving the inverse problem allows to infer intracranial
signals from EEG recordings [41–43]. In this paper we are going to adopt this
approach to extend the Kalman filtering technique, by including a model of the
head that transfers intracranial signals onto the scalp.
Methods
To obtain a reliable estimation of the state and the dynamics of the brain, we
require a biologically inspired mathematical model of its dynamics, experimen-
tal data (as non-invasive as possible), and the means of fusing both sources of
information together. In this paper, we use Jansen and Rit's model [25, 44] to
represent the dynamical evolution of the cortical structures. For our sets of ex-
perimental data we model EEG in silico using, again, Jansen and Rit's model
together with a model of the head. Finally, we use the Unscented Kalman Filter
as our data assimilation algorithm to estimate jointly the state and the parameters
of the model [45–47].
Mesoscopic neural mass model
Jansen and Rit's model [25,44] describes the mesoscopic activity of a population
of neurons [48, 49], providing a good compromise between physiological real-
ism and computational simplicity. This model simplifies the neuronal diversity
of a cortical column in three interacting populations: pyramidal neurons, excita-
tory interneurons, and inhibitory interneurons. The larger pyramidal population
excites both groups of interneurons, which in turn feed back into the pyramidal
cells. In our approximation, the pyramidal population is also driven by excitatory
noise from distant areas of the brain and by neighbouring columns.
The dynamics of each population rely on two different transformations. The
first converts the average density of incoming action potentials into an average
post-synaptic membrane potential (excitatory or inhibitory). It takes the form of
a second-order differential equation for excitatory inputs,
xe(t) + 2a xe(t) + a2xe(t) = Aa ue(t),
and for inhibitory inputs,
xi(t) + 2b xi(t) + b2xi(t) = Bb ui(t),
(1)
(2)
where ue,i(t) and xe,i(t) are the input and output of the transformations, respec-
tively, A and B are the amplitudes of the excitatory and inhibitory post-synaptic
potentials, and a and b are the lumped representations of the sums of the re-
ciprocal of the time constant of the passive membrane, and all other spatially
distributed delays in the dendritic network.
The second transformation converts the net average membrane potential of
the population, v, into an average firing rate, and is described by the following
sigmoid function:
(3)
3/23
Sigm(v) =
2e0
1 + er(v0−v)
where e0 is the maximum firing rate of the population, r controls the slope of
the sigmoid, and v0 is the post-synaptic potential for which a 50% firing rate is
obtained.
We model the brain as a system of Nd coupled cortical columns (dipole sources)
with the addition of noise. The following equations define our model for each cor-
tical column i:
xi
0(t) + 2a xi
0(t) =Aa Sigm[xi
0(t) + a2xi
2(t)],
Ki j Sigm(x j
1(t − τi j) − x j
(4)
2(t − τi j))
xi
1(t) + 2a xi
1(t) + a2xi
1(t) =Aa
2(t) + 2b xi
xi
2(t) + b2xi
2(t) =Bb
Nd(cid:88)
1(t) − xi
j=1
p(t) + k
(cid:16)
+ C2 Sigm[C1xi
0(t)]
C4 Sigm[C3xi
0(t)]
,
,
(cid:17)
(5)
(6)
where C1 to C4 are connectivity constants that govern the interactions between
populations, p(t) is a noisy external input, and the summation term includes the
delayed input from other coupled cortical columns. k modulates the strength of
the coupling, K is the adjacency matrix, and τi j is the delay with which column i
receives the signal of column j. Table 1 provides the descriptions and values of
these parameters. The electrical activity detected by the electrodes on the scalp
is originated by the weighted sum of the averaged membrane potential of the
pyramidal cells of all the cortical columns, xi(t) = xi
1(t) − xi
2(t) [50].
Table 1. Description and default values of the parameters for the system of
neural masses. See the Results section for details of the configuration of each
numerical experiment. Here, PC refers to pyramidal cells, EI to excitatory
interneurons, II to inhibitory interneurons, EPSP to excitatory post-synaptic
potential, and IPSP to inhibitory post-synaptic potential.
Param. Description
A
B
a
b
C1
C2
C3
C4
e0
v0
r
k
K
τ
p
EPSP amplitude
IPSP amplitude
Rate constant for the excitatory population*
Rate constant for the inhibitory population*
Strength of synaptic connections from PC to EI
Strength of synaptic connections from II to PC
Strength of synaptic connections from PC to II
Strength of synaptic connections from EI to PP
Maximum firing rate of the population
Mean threshold of the population
Steepness of the sigmoidal transformation
Coupling constant
Adjacency matrix
Delay
External input
Value
3.25 mV
22.00 mV
100 s−1
50 s−1
135
108
33.75
33.75
2.5 s−1
6 mV
0.56 mV−1
10
Ki j = 1, i (cid:44) j
Ki, j = 0, i = j
According
to distance [51]
200 s−1
*Lumped representation of the sum of the reciprocal of the time constant of
passive membrane and all other spatially distributed delays.
4/23
Extracranial data generation
The main contribution of this paper is the use of multichannel extracranial data
to obtain information about the neuronal populations inside the brain using data
assimilation. To accomplish this, we use synthetic EEG data generated in silico
using Jansen and Rit's model and Ary's head model. To that end, we transform the
output x(t) of the neural masses to EEG signals z(t) in the electrodes (see Fig. 1).
This transformation is mediated by a lead field matrix [40], which builds on the
basic idea of calculating the electric potential caused by a dipole source [13] on
a three-layer isotropic hemisphere [39, 52]. The lead field matrix also contains
information about the geometry of the problem (e.g., locations of cortical columns
and electrodes) and about the electrophysiology of the head (e.g., conductivities
of the different tissues). The following equations show the potential Ve,i on an
e [53] , caused by the dipole qi(t) = xi(t) qi generated
electrode e, located at re
by the cortical column i, located at ri
In these equations,
e = 1, . . . , Ne, where Ne is the total number of electrodes, and i = 1, . . . , Nd,
where Nd is the total number of dipoles:
q and oriented as qi.
Ve,i(re
e; ri
q, qi) (cid:117) v1(re
e; µ1ri
q, ρ1qi) + v2(re
e; µ2ri
q, ρ2qi) + v3(re
e; µ3ri
q, ρ3qi),
(cid:17) · qi,
(cid:17) · qi,
(cid:17) · qi.
(7)
(8)
(9)
(10)
(11)
(cid:16)
(cid:16)
(cid:16)
where vectors are typeset in bold and
1 − ce,i,1
(ce,i,1
1 − ce,i,2
(ce,i,2
1 − ce,i,3
(ce,i,3
e; ri
e; ri
e; ri
In these expressions,
q, qi) =
q, qi) =
q, qi) =
v1(re
v2(re
v3(re
2
2
2
e · ri
(re
e · ri
(re
e · ri
(re
2
q))ri
q))ri
q))ri
q + ce,i,1
q + ce,i,2
q + ce,i,3
2
2
(ri
q)2re
e
q)2re
(ri
e
(ri
q)2re
e
2
(cid:32)
de,i · ri
q
(de,i)3
2
+
1
,
de,i − 1
(cid:33)
(cid:17)
re
e
,
.
+
de,i + re
e
e, ri
reΓ(re
q)
e)2 − (ri
q · re
e)
ce,i,s
1
=
ce,i,s
2
=
Γ(re
e, ri
1
4πσs(ri
q)2
1
4πσs(ri
q) = de,i(cid:16)
(de,i)3
q)2
re
ede,i + (re
σs is the tangential conductivity of each surface [52] and ρs and µs are the Berg
parameters relative to it [54] (see Table 2). de,i is the distance between the dipole
i and the electrode e under consideration.
Table 2. Values of the Berg parameters for the three surfaces [52,54].
surface 3
1.0
0.4421
0.3561
parameter
Tangential conductivity σs
Berg parameter ρs
Berg parameter µs
surface 2
0.0125
0.7687
0.2389
surface 1
1.0
0.9901
0.0659
Table 3. Cartesian coordinates of the dipoles used throughout the study.
dipole 1
dipole 2
dipole 3
x
0.1688
0.3766
0.6622
y
0.2242
-0.8520
-0.2242
z
0.2597
0.2597
-0.1948
5/23
Figure 1. Extracranial data generation and illustration of Ary's model of
the head. The light and dark red arrows indicate dipole sources, and the
electrodes are shown as grey and black rectangles. The elements in the cartoon
illustrate how all the signals produced by the cortical columns (represented with
the solid red line in the top left panel) are transformed into an electrode reading
(shown in black dots in the top right panel) through the lead field matrix. In this
drawing, as in Equations (7-11), rq is the distance from the origin to the cortical
column under consideration; re is the distance from the origin to the electrode;
and d is the distance from the cortical column to the electrode. The placement of
the arrows here is for illustration purposes only; in our study, the cortical
columns are placed on the surface of the brain, close to the skull.
The Unscented Kalman Filter for data assimilation
The Unscented Kalman Filter (UKF) is our algorithm of choice to bring together
the dynamical state of the model and the experimental data. It is a standard tool
in the field of systems and control engineering, and has been shown to be both
computationally efficient and robust even when dealing with stochastic nonlinear
systems [55].
In order to simultaneously estimate the state and parameters of
the model described by Eqs. (4)-(6), we regard it as a discrete-time state-space
dynamical system of the following form:
0, x1
1, x1
xk+1 = F (xk) + vk
(12)
zk = H (xk) + wk
(13)
2 , θ1, . . . , θNp) ∈ Rnx is the state vector (related to
where x = (x1
the variables and parameters of the model), with θp being the parameters to esti-
mate, which obey the equations θp = 0. z ∈ Rnz is the measurement vector (our in
silico EEG readings). v and w are uncertainty terms that account for process noise
and measurement noise, respectively, with Gaussian distributions p(v) ∼ N(0, Q)
and p(w) ∼ N(0, R), respectively. F is obtained with a numerical implementation
0, . . . , xNd
2, x2
6/23
rerqd of Eqs. (4)-(6), as described below. H relates the state to measurement space. In-
terestingly, this basic part of the Kalman filter is in our case implemented by the
skull, the effect of which is represented by the lead field matrix, based on Ary's
head model and introduced above.
The UKF is a recursive predictor-corrector-type algorithm that aims to min-
imise the mean square error of the estimated states and parameters over time. For
each time step it calculates a prediction of the state and parameters of the system,
which is corrected when the information from a measurement is incorporated.
The amount of confidence given to the model and measurement is quantified by
the Kalman gain K, which is calculated at each time step based on prediction
covariances as well as model and measurement error covariances (Q and R, re-
spectively). For more details on the implementation of the filter, the reader is
referred to S1 Appendix and to Refs. [45–47, 56].
Generation of in silico datasets
For this paper three different in silico datasets were generated. We consider both
simulated electrocorticography (ECoG, intracortical) and electroencephalogra-
phy (EEG, extracranial) readings (using Ary's model in the latter case). All
datasets used the same locations [57] for cortical columns and electrodes, as
shown in Figs. 6 to 8. The strength of the coupling was set at a medium value
so that the cortical columns have an effect on one another without fully synchro-
nizing behaviours, and the configurations of the couplings are as shown in Fig. 2.
Table 1 shows representative values for the parameters used in all analyses unless
otherwise specified. In this paper we focus on estimating the amplitudes A of the
EPSPs of the different cortical columns, and therefore we choose values for these
amplitudes that produce signals that reflect various dynamic regimes that we wish
to explore. The numerical solver used to generate the in silico time series was the
Heun algorithm [58] with a time step of ∆t = 1 ms; the length of the data is of
100 s in all cases.
Figure 2. The two cortical column motifs used in this paper. Unidirectionally
coupled cortical columns have no backflow, and bidirectionally coupled columns
are coupled all-to-all. See Table 3.
To set the matrices Q and R-which reflect the quality of measurement and
model, and which crucially affect the output of the filter-, we used our knowl-
edge of the characteristics of the data to fix an initial guess [59], then adjusted it
to meet performance criteria.
7/23
132132ab For each of the experiments we conducted 50 realizations of each estimation
with different initial conditions; therefore, all the figures show averages of the
50 estimations, unless otherwise specified. The initial conditions for state and
parameter estimations were randomly generated; the parameters, however, were
constrained to deviate no more than 90% of their actual value as an initial as-
sumption.
Results
In order to compare the performance of the extra- and intracranial approaches to
Kalman filtering, we have analyzed three different cortical column configurations,
each using one of the two motifs shown in Fig. 2. Where relevant, two different
types of estimations have been used: intracranial and extracranial. Intracranial
estimation uses experimental data that would have hypothetically been obtained
from an electrocorticography, that is, using a single intracortical electrode, and is
estimated with the data provided by a single location -in other words, the direct
output of Jansen and Rit's model. Extracranial estimation, on the other hand, em-
ploys experimental data originated from EEG recordings, using several electrodes
placed on the skull, and is implemented here with the projection on the head of
the model output. We now discuss the three different dipole configurations that
we have considered.
Three unidirectionally coupled cortical columns
The first study was performed with the cortical columns coupled unidirectionally
(panel (a) of Fig. 2), as described in [59]. The parameters were set to standard
values [25] for the three cortical columns (see Table 1), except for the first column,
in which A1 was set to 3.58 mV to make it hyperexcitable. The coupling constant
was set to a medium value, large enough for the cortical columns to have a visible
effect on each other but not so large that they will fully synchronize and lock
their dynamics. In this case, information flows unidirectionally because of the
way the cortical columns are coupled [59]. As can be seen in the lower panels of
Fig. 3, the first cortical column has a random spiking activity, due to the increased
value of A and the presence of noise [60]. Due to the architecture of the coupling,
cortical column 1 causes cortical columns 2 and 3 to spike also, when otherwise
they would have simply fluctuated around their resting level.
The upper panels of Fig. 3 show the intracortical and extracranial estimations
of A for the three cortical columns. The estimation for A1 of the first column
converges to its correct value, with both the intra- and extracortical approaches.
This was to be expected, since the first cortical column receives no inputs from
other elements of the system. In contrast, the intracortical estimations for cortical
columns 2 and 3 converge to values significantly higher than their actual value of
3.25 mV. We conjecture that this is caused by the spiking of these two cortical
columns, which as mentioned above is due to the influence of cortical column
1. Multi-channel extracranial information, in contrast, allows to see the complete
picture of the coupled cortical columns and treat them as a single composed sys-
tem, contrary to the partial picture obtained from the information provided by
the single intracranial recordings. Therefore, estimation is better when using ex-
tracranial information with several electrodes, as shown in the upper panels of the
figure. The lower panels of Fig. 3 show the estimation of the state. The UKF
shows great efficacy when the estimation is extracranial, but performs poorly in
8/23
the case of intracortical estimation (with the exception of cortical column 1, be-
cause it has no input from other cortical columns).
Figure 3. Intracranial and extracranial fittings with propagated excitation
along unidirectionally coupled cortical columns. The upper panels show the
estimation of parameter A, and the lower panels show the estimations of the
observed states. The solid lines show the averages of the 50 realisations of the
estimation, and the shadowed areas indicate the standard deviation. The actual
values of the A parameters are A1 = 3.58 mV, A2 = 3.25 mV, and A3 = 3.25 mV,
the other parameters being set to standard values (Table 1). All three cortical
columns received an external input, p, in the shape of Gaussian white noise with
mean 90 s−1 and standard deviation 20 s−1. The coupling constant was set to
k = 10. The measurements were corrupted with noise of mean 0 and standard
deviation 100 mV for extracranial measurements and 5 mV for intracortical
measurements. Except for cortical column 1, with intracortical data the filter
converges to a much higher value than the target, whereas with extracranial data
the filter converges to a value which is accurate. In the lower panels it is shown
that extracranial estimations of the state are also accurate, whereas intracortical
estimations fail to reproduce the spikes correctly.
Three bidirectionally coupled cortical columns: coarse param-
eter estimation
The second experiment aims to explore the filter's possibilities in more extreme
situations. The three cortical columns are located as in the previous section, but
coupled bidirectionally (panel (b) of Fig. 2). Additionally, the maximum am-
plitudes of the excitatory PSPs are set to A1 = 4.25 mV, A2 = 10.00 mV, and
A3 = 3.25 mV. These values were chosen to force the three cortical columns to
be in very different dynamical regimes: cortical column 1 operates in a spiking
regime; cortical column 2 oscillates with alpha frequency but with an amplitude
similar to that of the spikes; and cortical column 3 oscillates in a more standard
regime, as described in [25].
9/23
3.03.43.84.2020406080Ai(mV)time (s)dipole 1xtracortical, 15 electrodesintracortical, 1 electrodeactual value020406080time (s)dipole 2extracortical, 15 electrodesintracortical, 1 electrodeactual value020406080100time (s)dipole 3extracortical, 15 electrodesintracortical, 1 electrodeactual value-4412205052545658state (mV)time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value5052545658time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value505254565860time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value Moderate intracortical measurement noise. Figure 4 shows again the perfor-
mance obtained using the experimental data from a set of extracraneal electrodes
compared to using individual intracortical electrodes for each cortical column. In
this case we show the 50 realisations of each filtering, without showing the aver-
age. The extracranial data for this experiment were corrupted with a measurement
noise of zero mean and standard deviation 100 s−1; the intracortical data were cor-
rupted with a measurement noise of standard deviation 5 s−1 in order to maintain
similar levels of signal-to-noise ratio.
As shown in Fig. 4, the intracortical parameter estimations do not approximate
very well the target value. In particular, the estimations of A for cortical column
2 converge to three different values depending on the initial conditions. The state
estimation follow the actual state of the system closely only for cortical column
1. The situation is very different when with extracranial electrodes, where all 50
realisations of the estimations converge with much more precision to the correct
values for both state and parameters (with the exception of A2, which still tends
to lower values in a very small quantity of the realisations). Again, extracranial
performance is better, in general, to intracortical.
Figure 4. Intracranial and extracranial fittings for coarse parameter
estimation in the case of bidirectional coupling. As in the previous figure, the
upper panels show the estimation of A for each cortical column and the lower
panels show the estimations of the observed states. The results are shown here
without averaging. The actual values of the amplitudes of the EPSPs are
A1 = 4.25 mV, A2 = 10.00 mV, and A3 = 3.25 mV; the rest of the parameters
were set to standard values (Table 1). The external input p for each of the three
cortical columns is of mean 200 s−1 and standard deviation 100 s−1. The
coupling constant was set to k = 5. The intracortical measurements were
corrupted with noise of mean 0 and standard deviation 5 mV, while the noise in
the extracranial measurements is of standard deviation 100 mV. Extracranial
estimations of the parameters are both faster and more accurate than intracortical
estimations; this applies also to the state, whose dynamics are more faithfully
reproduced using multi-electrode extracranial estimation (as shown in the lower
panels).
10/23
2468020406080Ai(mV)time (s)dipole 1extracortical, 15 electrodesintracortical, 1 electrodeactual value26101418020406080time (s)dipole 2extracortical, 15 electrodesintracortical, 1 electrodeactual value23456020406080100time (s)dipole 3extracortical, 15 electrodesintracortical, 1 electrodeactual value-101030505051state (mV)time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value5051time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value505152time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value High intracortical measurement noise. The difference between intracranial
and extracranial estimation is even larger for higher measurement noise (Fig. 5).
In this case, the amount of noise in the intracortical data was set to the same
value as the noise in the extracranial data. The value of R was tuned to reflect the
increase in measurement noise, but the intracortical estimations failed to obtain
the correct values for the parameters and reproduce the state.
Figure 5. Intracranial and extracranial fittings for coarse parameter
estimation, with a higher amount of intracortical measurement noise. The
upper panels show the estimation of the EPSPs for each cortical column and the
lower panels show the estimations of the observed states. The results are shown
here without averaging. The actual values of the amplitudes of the EPSPs are
A1 = 4.25 mV, A2 = 10.00 mV, and A3 = 3.25 mV; the rest of the parameters
were set to standard values (Table 1). The external input p for each of the three
cortical columns is of mean 200 s−1 and standard deviation 100 s−1. The
coupling constant was set to k = 5. The intracortical measurements were
corrupted with noise of mean 0 and standard deviation 100 mV-about an order
of magnitude higher than the noise in the previous graph-, while the noise in
the extracranial measurements is of standard deviation 100 mV. Extracranial
estimations of the parameters are also faster and more accurate than intracortical
estimations, more markedly so in this case; as to the state, in this more extreme
case, the intracortical estimation does not mimic the evolution of the system in
any way.
Using one single extracranial electrode. Using the same dataset, we aimed
to investigate the outcome of using each extracranial electrode individually [61],
as opposed to using the complete subset as until now. Therefore we used each
electrode separately to estimate the state and parameters of the complete system,
with 50 realisations of the estimation for each electrode. By doing so, we show
that the quality of the estimations is strongly dependent on the relative positions
of sources and electrodes.
In Figs. 6 to 8 we present the results for the estimation of parameter A of each
11/23
0246810020406080Ai(mV)time (s)dipole 1extracortical, 15 electrodesintracortical, 1 electrodeactual value048121620020406080time (s)dipole 2extracortical, 15 electrodesintracortical, 1 electrodeactual value13579020406080100time (s)dipole 3extracortical, 15 electrodesintracortical, 1 electrodeactual value-101030505051state (mV)time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value5051time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value505152time (s)extracortical, 15 electrodesintracortical, 1 electrodeactual value of the three cortical columns separately. The histograms show the distribution of
the 50 estimations of A using each electrode, placed in the respective position of
the electrode in question. Vertical colored lines in the histograms mark the value
of the three A parameters being estimated (one in each figure). The histograms
show a strong dependence on space of the quality of the estimations. As a general
trait, the estimations are better when the electrodes are near the cortical column
whose value of A is being estimated, whereas the more distant electrodes show a
wider distribution of final values for the parameter.
Figure 6. Distribution of 50 realizations of A estimations from a single
electrode for cortical column 1 (solid red circle). The histograms are placed at
the location of the corresponding measuring electrode, and the location of the
three cortical columns generating the activity are shown with colored circles
(with the full circle corresponding to the column whose value of A is being
estimated in this figure). Vertical lines with the same colors as the circles mark
the corresponding actual A values. The distributions tend to be narrowest in the
vicinities of cortical column 1. Nevertheless, they do not group around the target
value of A1 = 4.25 mV (vertical red line), as they should, but around that of
A3 = 3.25 mV (vertical blue line).
In Fig. 6 the distribution of the estimations of A1 are shown. The distributions
tend to be narrowest in the vicinities of the cortical column whose A value is
being estimated. However, it is noteworthy that the histograms obtained from the
observations in distant electrodes tend to group not around the actual value of
A1 = 4.25 mV (red vertical line), but of A3 = 3.25 mV (blue vertical line). This
result suggests that the algorithm is unable to distinguish the origin of the EEG
activity when sources and electrodes are distant from each other.
12/23
Figure 7 shows the results of the estimation of A2 (actual value shown by
vertical green lines), revealing wider distributions in general, which indicates a
stronger dependence on initial conditions. Although it is true that the electrodes
near cortical column 2 perform better in estimating A for that column, the dif-
ference with more distant electrodes is not as large as for the estimates of A for
cortical columns 1 and 3.
Figure 7. Distribution of 50 realizations of A estimations from a single
electrode for cortical column 2 (solid green circle). The distributions here are
wider than for A1 and A3, although they still tend to be more accurate near the
cortical column (solid green circle) and group around the target value of
A2 = 10.00 mV (vertical green line).
Finally, Fig. 8 shows the performance of each electrode when A3 is being
estimated (actual value shown by vertical blue lines in the figure). Interestingly,
even the electrodes located at the far left of the figure lead to a good estimate of A,
comparable to that coming from the electrodes in the far right, which are closer
to column 3 and could therefore be expected to provide a much more accurate
estimation.
However accurate some of the single electrodes' estimations are, using the
complete set of 15 electrodes invariably yields better results. This is because, in
Kalman filtering, combining many sources of information always improves the
final estimation, even if some of the sources are inaccurate or incomplete [62].
13/23
Figure 8. Distribution of 50 realizations of A estimations from a single
electrode for cortical column 3 (solid blue circle). As in the two previous
figures, the distributions for the electrodes closest to the source (solid blue circle)
are narrow, grouping around the correct value (A3 = 3.25 mV, vertical blue line).
Surprisingly, the electrodes in the far left also give rise to narrow distributions.
Three bidirectionally coupled cortical columns: fine parameter
estimation
In the previous section, the value of A of one of the cortical columns was much
larger than the other two. We now consider the same coupling motif, but with
values of the A parameter that are much closer together in value: A1 = 3.58 mV,
A2 = 3.25 mV, and A3 = 3.10 mV. The purpose of this test was to ascertain
whether the filter could differentiate between parameters with smaller differences
in value. This ability is very important if we expect to use the technique in clin-
ical applications. Fig. 9 shows the extracranial estimation of the A parameters
using the complete subset of 15 electrodes. The estimations converge to the ac-
tual values with enough accuracy as to give hopes of using the filter in a clinical
setting.
Discussion
The most important limitation of current data assimilation processes in neuro-
science is that the appropriate experimental recordings are usually intracranial.
Using Kalman filtering to fit these data to neural mass models shows promise in
several contexts and applications. Here we have modified this type of approach
14/23
Figure 9. Extracranial fit with parameters close together in value. The
estimations of the amplitude of the EPSPs of the three cortical columns are
shown after averaging over 50 realizations (solid lines); the shadowed areas
indicate the standard deviation. The actual values of the amplitudes of the EPSPs
are A1 = 3.58 mV, A2 = 3.25 mV, and A3 = 3.10 mV; the rest of the parameters
were set to standard values (Table 1). The external input p for each of the three
cortical columns is of mean 200 s−1 and standard deviation 100 s−1. The
coupling constant was set to k = 5. The noise in the extracranial measurements
is of standard deviation 100 mV. The estimation of the parameters is fairly
accurate.
by extending the base neural mass model with a head model, with the aim of
integrating non-invasive experimental recordings. We have explored the limita-
tions and advantages of our model using in silico data in very well controlled
conditions. Even though we keep the exploration of the technique using real EEG
experimental data in mind, here we bring forth a proof of concept by performing
several experiments that address different aspects of the method.
In this paper we have considered a system comprised of three cortical columns,
modelled according to Jansen and Rit's equations and coupled following two dif-
ferent motifs. The cortical columns are all driven by a noisy input. The signal
from the cortical columns is then transferred to the skull, after which it is cor-
rupted by noise to simulate electrode readings from EEG. These are then used
to estimate the amplitude of the excitatory post-synaptic potentials using the Un-
scented Kalman Filter.
The first study involves three columns that are coupled unidirectionally with
no backflow. The first cortical column is made hyperexcitable by increasing the
excitatory post-synaptic potential to A1 = 3.58 mV; this cortical column causes
the second cortical column and, indirectly, the third to modify their behavior as a
result of the spiking of the first. For the intracranial estimations, single intracorti-
cal electrodes measured the evolution of the three cortical columns independently;
for the extracranial estimations, 15 extracranial electrodes were used simultane-
ously. Applying the Kalman filter to the extracranial data provided a good estima-
tion of the A parameters; the intracortical measurements, however, yielded mixed
results. The estimation for cortical column 1 was accurate, whereas for cortical
columns 2 and 3 the estimation was above the target value. We attribute this to the
fact that columns 2 and 3 are excited by column 1, which spikes due to a higher
value of A. As a consequence, when independently evaluated, the estimation is
higher than the actual value. Therefore we suggest that one intracranial electrode
15/23
2.53.03.54.04.50204060801003.103.253.58Ai(mV)time (s)dipole 1dipole 2dipole 3actual values provides only a partial view of the system, and thus cannot capture the behaviors
of all three cortical columns and the interactions between them; the use of many
electrodes provides a more complete view of the system.
Next we considered a situation in which the dipoles were coupled bidirec-
tionally in an all-to-all configuration. The A parameters were chosen such as
to cause different dynamic behaviors in the cortical columns: A1 = 4.25 mV,
A2 = 10.00 mV, and A3 = 3.25 mV. Three types of fitting via Kalman filtering
were performed, using (i) independent intracortical recordings of single cortical
columns were filtered, (ii) the complete subset of 15 extracranial electrodes, and
(iii) single extracranial electrodes. The intracortical data were corrupted with two
different levels (medium and high) of measurement noise. For both cases, the
multi-electrode extracraneal estimation surpasses the intracortical results in both
speed and quality; the difference, however, is more marked in the presence of
higher measurement noise in the intracortical recordings. The results for the sin-
gle electrodes show a significant influence of space on the quality of the estima-
tions, in the sense that estimations of electrodes close to the source are relatively
accurate, and electrodes further away from the source might not allow to discrim-
inate the source of the information correctly, or might completely fail to represent
the system.
Finally, we considered the situation of an identical cortical column configu-
ration -in terms of situation and coupling-, except for the values of the EPSPs
of the cortical columns: A1 = 3.58 mV, A2 = 3.25 mV, and A3 = 3.10 mV.
This dataset was filtered only extracranially, with the purpose of evaluating the
filter's ability to discriminate parameter values within narrower ranges. The re-
sults were accurate, which is promising in views of applying the algorithm in a
clinical setting.
Taken as a whole, our results show that, independently of the need to explore
more realistic situations, extracranial EEG recordings constitute a good candidate
to be used together with neural mass models and Kalman filters, provided the
method is extended with a head model. Using non-invasive techniques in these
processes widens the applications of Kalman-based data assimilation methods in
neuroscience.
The adaptation of the method to a specific possible applications deserves its
own exploration. The possibility of determining parameters of cerebral dynam-
ics in a non-invasive manner would allow us to study, for instance, the origins
of the variability in EEG recordings. It would also enable exploring automatic
biometric-based user recognition systems [63] and, through single-patient charac-
terization, tracking the changes in brain dynamics due to aging [64,65], and mon-
itoring the evolution of diseases [66]. The possibility of tracking the evolution
of brain states during motor imagery-control [67] or task-switching control [68]
is also open. Besides, a good description of the brain state would allow the ef-
ficient control of epilepsy [69], a good performance in brain-machine interface
tasks [70], and the detection and control of transcranial brain stimulation [71].
Rehabilitation tasks [72, 73] may also benefit from the possibility of monitoring
brain states reliably.
Supporting Information
S1 Appendix
The Unscented Kalman Filter (UKF) algorithm.
UKF is a predictor-corrector algorithm that estimates the state and parameters
at a given time step k in two phases. The first one predicts the state based solely on
16/23
the dynamical information of the system, i.e., the model. The second incorporates
a measurement with which to correct the first estimation.
The first step of the algorithm involves computing the expectation of the state
and of the state covariance at time instant k + 1, known as the a priori estimation.
For this we use a numerical implementation (using Heun's solver) of Jansen and
Rit's model of a cortical column [25, 44], as described in the Methods section.
The nature of the nonlinearities of this model prevents us from using a simple
linearization approach to propagating the statistics of the state variables across
the transformation. Therefore, we incorporate the unscented transform (UT) in
our formulation of the Kalman filter, which, instead of attempting to propagate a
distribution through the nonlinearity, first propagates a series of deterministically
chosen points through the nonlinearity and then recovers the statistical informa-
tion of the distribution from these.
Therefore, the a priori estimation of the state, x−
k , is obtained as follows,
beginning with the calculation and projection of the 2n + 1 (where n is the state
size) sigma points,
Σk−1,0 = xk−1
Σk−1,i = xk−1 +
(cid:16)(cid:112)
Σk−1,i = xk−1 −(cid:16)(cid:112)
(cid:17)
(cid:17)
(19)
(20)
17/23
(n + λ)Pk−1
(n + λ)Pk−1
i = 1, ..., n
i = n + 1, ..., 2n
,
i
i−n
,
(14)
where Pk−1 is the estimated state covariance matrix for the previous time step.
This continues with the condensation of the projected sigma points into the a
priori state estimate:
X∗
kk−1 = f (Σk−1, uk−1)
x−
k =
i X∗
Wm
i,kk−1
2L(cid:88)
2L(cid:88)
i=0
P−
k =
Wcov
i
[X∗
i,kk−1 − x−
k ][X∗
i,kk−1 − x−
k ]T + Q
(15)
(16)
(17)
i=0
where Q is the state error covariance and Wm and Wcov are the weight vectors,
defined as
Wm
0
=
Wcov
0
=
Wm
i
= Wcov
i
=
λ
n + λ
λ
+ 1 − α2 + β
n + λ
1
2(n + λ) , i = 1, ..., 2n
(18)
In Eqs. 14 and 18, α, β and κ are scaling factors, and λ is calculated as λ =
α2(n + κ) − n. α, the primary scaling factor, determines the spread of the sigma
points around the mean and is usually set between 0.001 and 1. β contains prior
information about the distribution of x; for Gaussian distributions, its optimal
value is 2. κ, the tertiary scaling parameter, is usually set to 0 [55].
ment, y−
Finally, the sigma points are redrawn [45] and the estimation of the measure-
k , is calculated:
Υkk−1 = H[Σkk−1]
y−
k =
Wm
i
Υi,kk−1
2L(cid:88)
i=0
The use of a measurement to correct the state estimation implies the mapping
of the a priori estimate onto the measurement space for comparison. It is worth-
while to note that, in our case, this transformation is a linear matrix H which
relates the state of the cortical columns to an EEG reading. See Extracranial data
generation section for details.
The second step of the algorithm corrects the a priori estimation of state and
covariance by using the information available from the most recent measurement
(in our case, an EEG reading). The impact of the measurement is determined by
the Kalman gain Kk, which essentially expresses the level of confidence on the
accuracy of the model and the level of noise in the data.
Pykyk
=
Wcov
i
[Υi,kk−1 − y−
k ][Υi,kk−1 − y−
k ]T + R
2L(cid:88)
2L(cid:88)
i=0
[Xi,kk−1 − x−
k ][Υi,kk−1 − y−
k ]T
i
=
i=0
Wcov
Pxkyk
Kk = Pxkyk P−1
ykyk
xk = x−
Pk = P−
k + Kk(zk − y−
k )
k − Kk Pykyk KT
k
(21)
(22)
(23)
(24)
(25)
where Pykyk is the predicted measurement covariance, Pxkyk is the state-measurement
cross-covariance, R is the measurement error covariance, and zk is the measure-
ment for the current time step.
Acknowledgments
This work was partially supported by the Catalan Government (AGAUR grant
FI-DGR 2014-2017) and by the Spanish Ministry of Economy and Competitive-
ness and FEDER (project FIS2015-66503). JGO also acknowledges support from
the the Catalan Government (project 2014SGR0947), the ICREA Academia pro-
gramme, and from the "Mar´ıa de Maeztu" programme for Units of Excellence in
R&D (Spanish Ministry of Economy and Competitiveness, MDM-2014-0370).
References
1. Yuste R. From the neuron doctrine to neural networks. Nature Re-
views Neuroscience. 2015;16(8):487–497. Available from: http://www.
nature.com/doifinder/10.1038/nrn3962.
2. Braitenberg V, Schutz A. Anatomy of the Cortex. No. 18 in Studies of
Brain Function. Springer; 1991. Statistics and Geometry.
3. Izhikevich EM, Edelman GM. Large-scale model of mammalian thala-
mocortical systems. Proceedings of the National Academy of Sciences.
2008;105(9):3593–3598. Available from: http://www.pnas.org/cgi/
doi/10.1073/pnas.0712231105.
4. Callaway EM. Micro-, Meso- and Macro-Connectomics of the Brain.
Springer; 2016. Available from: http://link.springer.com/10.
1007/978-3-319-27777-6.
18/23
5. Freeman WJ. Simulation of chaotic EEG patters with a dynamic model of
the olfactory system. Biological Cybernetics. 1987;56:139–150.
6. Faure P, Korn H.
Is there chaos in the brain? I. Concepts of nonlin-
ear dynamics and methods of investigation. C R Acad Sci III. 2001
sep;324(9):773–93. Available from: http://www.ncbi.nlm.nih.gov/
pubmed/11558325.
7. Shadlen MN, Newsome WT. Noise, neural codes and cortical organization.
Current Opinion in Neurobiology. 1994;4(4):569–579.
8. Faisal AA, Selen LPJ, Wolpert DM. Noise in the nervous system. Na-
ture Reviews Neuroscience. 2008;9(4):292–303. Available from: http:
//www.nature.com/doifinder/10.1038/nrn2258.
9. Schiff SJ, Jerger K, Duong DH, Chang T, Spano ML, Ditto WL. Control-
ling chaos in the brain. Nature. 1994;370(6491):615–620.
10. van Vreeswijk C, Sompolinsky H. Chaos in neuronal networks with
balanced excitatory and inhibitory activity. Science (New York, NY).
1996;274(5293):1724–6. Available from: http://www.ncbi.nlm.nih.
gov/pubmed/8939866.
time series.
short
from:
11. Celletti A, Froeschle C, Tetko IV, Villa AEP.
Deterministic be-
Meccanica. 1999;34(3):147–154.
http://links.isiglobalnet2.com/gateway/
haviour of
Available
Gateway.cgi?GWVersion=2{&}SrcAuth=mekentosj{&}SrcApp=
Papers{&}DestLinkType=FullRecord{&}DestApp=WOS{&}KeyUT=
000083554700001{%}5Cnpapers3://publication/uuid/
9EE31316-A232-4916-8227-D1E9D35B4355.
12. Stam CJ. Nonlinear dynamical analysis of EEG and MEG: review of an
emerging field. Clin Neurophysiol. 2005 oct;116(10):2266–301. Available
from: http://www.ncbi.nlm.nih.gov/pubmed/16115797.
13. Buzs´aki G, Anastassiou Ca, Koch C. The origin of extracellular fields
and currents–EEG, ECoG, LFP and spikes. Nature reviews Neuroscience.
2012;13(6):407–20. Available from: http://www.ncbi.nlm.nih.gov/
pubmed/22595786.
14. Wright JJ, Liley DTJ. Dynamics of the Brain at Global and Microscopic
Scales: Neural Networks and the EEG. Behavioral and Brain Sciences.
1996;19(2):285.
15. Rabinovich MI, Varona P, Selverston AI, Abarbanel HDI. Dynamical prin-
ciples in neuroscience. Reviews of Modern Physics. 2006;78(4).
al.
et
Science. 2012;338(6111):1202–1205.
16. Eliasmith C, Stewart TC, Choo X, Bekolay T, DeWolf T,
the Functioning
Available from:
Tang C,
Brain.
http://www.sciencemag.org/cgi/doi/10.1126/science.
1225266{%}5Cnhttp://www.ncbi.nlm.nih.gov/pubmed/
23197532.
A Large-Scale Model of
17. Deco G, Ponce-Alvarez A, Hagmann P, Romani G, Mantini D, Corbetta
M. How local excitation-inhibition ratio impacts the whole brain dynam-
ics. The Journal of neuroscience. 2014;34(23):7886–98. Available from:
19/23
http://www.pubmedcentral.nih.gov/articlerender.fcgi?
artid=4044249{&}tool=pmcentrez{&}rendertype=abstract.
18. David O, Friston KJ. A neural mass model for MEG/EEG: Coupling and
neuronal dynamics. NeuroImage. 2003;20(3):1743–1755.
19. Grimbert F, Faugeras O. Bifurcation analysis of Jansen's neural mass
model. Neural computation. 2006;18(12):3052–3068.
20. Cona F, Zavaglia M, Massimini M, Rosanova M, Ursino M. A neural mass
model of interconnected regions simulates rhythm propagation observed
via TMS-EEG. NeuroImage. 2011;57(3):1045–1058.
21. Coombes S. Large-scale neural dynamics: Simple and complex. NeuroIm-
age. 2010;52(3):731–739.
22. Babajani A, Soltanian-Zadeh H. Integrated MEG/EEG and fMRI model
based on neural masses. IEEE Transactions on Biomedical Engineering.
2006;53(9):1794–1801.
23. Babiloni F, Cincotti F, Babiloni C, Carducci F, Mattia D, Astolfi L, et al.
Estimation of the cortical functional connectivity with the multimodal in-
tegration of high-resolution EEG and fMRI data by directed transfer func-
tion. NeuroImage. 2005;24(1):118–131.
24. Bojak I, Oostendorp TF, Reid AT, Kotter R. Connecting mean field
models of neural activity to EEG and fMRI data. Brain Topography.
2010;23(2):139–149.
25. Jansen BH, Rit VG. Electroencephalogram and visual evoked potential
generation in a mathematical model of coupled cortical columns. Biologi-
cal Cybernetics. 1995;73(4):357–366.
26. David O, Harrison L, Friston KJ. Modelling event-related responses in the
brain. NeuroImage. 2005;25(3):756–770.
27. Spiegler A, Knosche TR, Schwab K, Haueisen J, Atay FM. Modeling brain
resonance phenomena using a neural mass model. PLoS Computational
Biology. 2011;7(12).
28. Wright JJ, Robinson PA, Rennie CJ, Gordon E, Bourke PD, Chapman CL,
et al. Toward an integrated continuum model of cerebral dynamics: The
cerebral rhythms, synchronous oscillation and cortical stability. BioSys-
tems. 2001;63(1-3):71–88.
29. Schiff SJ, Sauer T. Kalman filter control of a model of spatiotemporal
cortical dynamics. Journal of neural engineering. 2008;5(1):1–8.
30. Kiebel SJ, Garrido MI, Moran RJ, Friston KJ. Dynamic causal modelling
for EEG and MEG. Cognitive Neurodynamics. 2008;2(2):121–136.
31. Shine JM, Koyejo O, Bell PT, Gorgolewski KJ, Gilat M, Poldrack RA. Es-
timation of dynamic functional connectivity using Multiplication of Tem-
poral Derivatives. NeuroImage. 2015;122:399–407.
32. Ma X, Chou CA, Sayama H, Chaovalitwongse WA. Brain response
pattern identification of fMRI data using a particle swarm optimization-
based approach. Brain Informatics. 2016;3(3):181–192. Available from:
http://link.springer.com/10.1007/s40708-016-0049-z.
20/23
33. Hamilton F, Cressman J, Peixoto N, Sauer T.
Reconstruct-
ing neural dynamics using data assimilation with multiple mod-
els.
Available
EPL (Europhysics Letters). 2014;107(6):68005.
from:
http://www.scopus.com/inward/record.url?eid=2-s2.
0-84907292009{&}partnerID=tZOtx3y1.
34. Freestone DR, Kuhlmann L, Chong MS, Nesic D, Grayden DB,
Patient-specific neural mass modeling - stochas-
Recent Advances in Predicting
Available
https://hal.archives-ouvertes.fr/hal-00876475/
Aram P, et al.
tic and deterministic methods.
and Preventing Epileptic Seizures. 2013;p. 63–82.
from:
{%}5Cnhttp://www.worldscientific.com/doi/abs/10.1142/
9789814525350{_}0005.
35. Freestone DR, Karoly PJ, Nesi´c D, Aram P, Cook MJ, Grayden DB.
Estimation of effective connectivity via data-driven neural modeling.
Frontiers in neuroscience. 2014;8(November):383.
Available from:
http://www.pubmedcentral.nih.gov/articlerender.fcgi?
artid=4246673{&}tool=pmcentrez{&}rendertype=abstract.
36. Cao Y, Ren K, Su F, Deng B, Wei X, Wang J. Suppression of seizures
based on the multi-coupled neural mass model. Chaos. 2015;25(10).
37. Bojak I, Liley D. Modeling the effects of anesthesia on the electroen-
cephalogram. Physical Review E. 2005;71(4):041902.
38. Kuhlmann L, Freestone DR, Manton JH, Heyse B, Vereecke HEM, Lip-
ping T, et al. Neural mass model-based tracking of anesthetic brain states.
NeuroImage. 2016;133:438–456.
39. Zhang Z. A fast method to compute surface potentials generated by dipoles
within multilayer anisotropic spheres. Physics in medicine and biology.
1995;40(3):335–349.
40. Mosher JC, Leahy RM, Lewis PS. EEG and MEG: forward solutions
IEEE Transactions on Biomedical Engineering.
for inverse methods.
1999;46(3):245–259.
41. Haufe S, Tomioka R, Dickhaus T, Sannelli C, Blankertz B, Nolte G, et al.
Large-scale EEG/MEG source localization with spatial flexibility. Neu-
roImage. 2011;54(2):851–859.
42. Verhellen E, Boon P. EEG source localization of the epileptogenic focus in
patients with refractory temporal lobe epilepsy, dipole modelling revisited.
Acta neurologica Belgica. 2007;107(3):71–77.
43. Gotman J. Noninvasive methods for evaluating the localization and prop-
agation of epileptic activity. Epilepsia. 2003;44 Suppl 1:21–29.
44. Jansen BH, Zouridakis G, Brandt ME. A neurophysiologically-based
mathematical model of flash visual evoked potentials. Biological Cyber-
netics. 1993;68:275–283.
45. Merwe RVD, Wan EA. The square-root unscented Kalman filter for state
and parameter-estimation.
In: 2001 IEEE International Conference on
Acoustics, Speech, and Signal Processing. Proceedings. (ICASSP '01).
vol. 6; 2001. p. 3461–3464.
21/23
46. Julier SJ, Uhlmann JK. Unscented Filtering and Nonlinear Estimation.
Proceedings of the IEEE. 2004;92(3):401–422.
47. Rudolf Emil Kalman. A New Approach to Linear Filtering and Predic-
tion Problems. Transactions of the ASME–Journal of Basic Engineering.
1960;82(Series D):35–45.
48. Silva FHLD, Hoeks A, Smits H, Zetterberg LH. Model of brain rhythmic
activity. Kybernetic. 1974;15(1):27–37.
49. Faugeras O, Touboul J, Cessac B. A constructive mean-field analysis
of multi-population neural networks with random synaptic weights and
stochastic inputs. Frontiers in Computational Neuroscience. 2009;3(1).
50. Silva FLD.
In: EEG: Origin and Measurement. World Scien-
tific Publishing Co.; 2011. p. 63–82. Available from: http://www.
worldscientific.com/doi/abs/10.1142/9789814525350_0005.
5.
51. Pons AJ, Cantero JL, Atienza M, Garcia-Ojalvo J. Relating structural and
functional anomalous connectivity in the aging brain via neural mass mod-
eling. NeuroImage. 2010;52(3):848–861.
52. Ary JP, Klein SA, Fender DH. Location of Sources of Evoked Scalp Po-
tentials: Corrections for Skull and Scalp Thicknesses. IEEE Transactions
on Biomedical Engineering. 1981;BME-28(6):447–452.
53. Jurcak V, Tsuzuki D, Dan I. 10/20, 10/10, and 10/5 systems revisited:
Their validity as relative head-surface-based positioning systems. Neu-
roImage. 2007;34(4):1600–1611.
54. Berg P, Scherg M. A fast method for forward computation of multiple-shell
spherical head models. Electroencephalography and Clinical Neurophysi-
ology. 1994;90:58–64.
55. Merwe RVD, Wan EA. The Unscented Kalman Filter for Nonlinear Es-
timation. In: Adaptive Systems for Signal Processing, Communications,
and Control Symposium 2000. AS-SPCC. The IEEE 2000; 2000. p. 153–
158.
56. Solonen A, Hakkarainen J, Ilin A, Abbas M, Bibov A. Estimating model
error covariance matrix parameters in extended Kalman filtering. Nonlin-
ear Processes in Geophysics. 2014;21(5):919–927.
57. Richards JE. Recovering dipole sources from scalp-recorded event-
related-potentials using component analysis: principal component anal-
ysis and independent component analysis.
International Journal of Psy-
chophysiology. 2004;54:201–220.
58. Toral R, Colet P. Stochastic numerical methods: an introduction for stu-
dents and scientists. John Wiley & Sons; 2014.
59. Liu X, Gao Q. Parameter estimation and control for a neural mass model
based on the unscented Kalman filter. Phys Rev E. 2013 Oct;88:042905.
Available from: http://link.aps.org/doi/10.1103/PhysRevE.88.
042905.
60. Grimbert F, Faugeras O. Bifurcation analysis of Jansen's neural mass
model. Neural Computation. 2006;18(12):3052–3068.
22/23
61. Freestone DR, Karoly PJ, Nesi´c D, Aram P, Cook MJ, Grayden DB. Es-
timation of effective connectivity via data-driven neural modeling. Fron-
tiers in Neuroscience. 2014;8:383. Available from: http://journal.
frontiersin.org/article/10.3389/fnins.2014.00383.
62. Schiff SJ. Neural Control Engineering: The Emerging Intersection Be-
tween Control Theory and Neuroscience. Computational neuroscience.
MIT Press; 2012. Available from: https://books.google.es/books?
id=P9UvTQtnqKwC.
63. Campisi P, Rocca DL. Brain waves for automatic biometric-based user
IEEE Transactions on Information Forensics and Security.
recognition.
2014;9(5):782–800.
64. Anokhin AP, Birbaumer N, Lutzenberger W, Nikolaev A, Vogel F. Age
increases brain complexity. Electroencephalography and Clinical Neuro-
physiology. 1996;99(1):63–68.
65. Yang S, Deng B, Wang J, Li H, Liu C, Fietkiewicz C, et al. Efficient im-
plementation of a real-time estimation system for thalamocortical hidden
Parkinsonian properties. Scientific Reports. 2017;.
66. Soekadar SR, Birbaumer N, Slutzky MW, Cohen LG. Brain-machine
interfaces in neurorehabilitation of stroke. Neurobiology of Disease.
2015;83:172–179.
67. Zich C, Debener S, Kranczioch C, Bleichner MG, Gutberlet I, De Vos
M. Real-time EEG feedback during simultaneous EEG-fMRI identifies
the cortical signature of motor imagery. NeuroImage. 2015;114:438–447.
68. Phillips JM, Vinck M, Everling S, Womelsdorf T. A long-range fronto-
parietal 5- to 10-Hz network predicts "top-down" controlled guidance in a
task-switch paradigm. Cerebral Cortex. 2014;24(8):1996–2008.
69. Shan B, Wang J, Deng B, Wei X, Yu H, Zhang Z, et al. Particle swarm
optimization algorithm based parameters estimation and control of epilep-
tiform spikes in a neural mass model. Chaos. 2016;26.
70. Del R Mill´an J, Ferrez PW, Gal´an F, Lew E, Chavarriaga R. Non-Invasive
Brain-Machine Interaction. International Journal of Pattern Recognition
and Artificial Intelligence. 2008;22(05):959–972.
71. Krause B, M´arquez-Ruiz J, Cohen Kadosh R. The effect of transcranial
direct current stimulation: a role for cortical excitation/inhibition balance?
Frontiers in human neuroscience. 2013;7(September):602. Available from:
http://www.pubmedcentral.nih.gov/articlerender.fcgi?
artid=3781319{&}tool=pmcentrez{&}rendertype=abstract.
72. Aram P, Freestone DR, Cook MJ, Kadirkamanathan V, Grayden
DB. Model-based estimation of intra-cortical connectivity using elec-
trophysiological data.
Avail-
able from:
http://www.sciencedirect.com/science/article/
pii/S1053811915005534.
NeuroImage. 2015;118:563 – 575.
73. Stephan KE, Schlagenhauf F, Huys QJM, Raman S, Aponte EA, Broder-
sen KH, et al. Computational neuroimaging strategies for single patient
predictions. NeuroImage. 2015;145:180–199.
23/23
|
1302.3261 | 1 | 1302 | 2013-02-13T22:18:49 | Pavlov's dog associative learning demonstrated on synaptic-like organic transistors | [
"q-bio.NC",
"cond-mat.dis-nn",
"cs.ET",
"cs.NE"
] | In this letter, we present an original demonstration of an associative learning neural network inspired by the famous Pavlov's dogs experiment. A single nanoparticle organic memory field effect transistor (NOMFET) is used to implement each synapse. We show how the physical properties of this dynamic memristive device can be used to perform low power write operations for the learning and implement short-term association using temporal coding and spike timing dependent plasticity based learning. An electronic circuit was built to validate the proposed learning scheme with packaged devices, with good reproducibility despite the complex synaptic-like dynamic of the NOMFET in pulse regime. | q-bio.NC | q-bio | Pavlov’s Dog Associative Learning
Demonstrated on Synaptic-like Organic
Transistors
O. Bichler1 , W. Zhao1 , F. Alibart2 , S. Pleutin2 , S. Lenfant2 , D.
Vuillaume2 , C. Gamrat1
1CEA, LIST, Embedded Computing Laboratory, 91191 Gif-sur-Yvette Cedex, France.
2 Institute for Electronics, Microelectronics and Nanotechnology (IEMN), CNRS,
University of Lille, BP60069, Avenue Poincaré, 59652, Villeneuve d’Ascq, France.
Abstract
In this paper, we present an original demonstration of an associative learning neural
network, inspired by the famous Pavlov’s dogs experiment. A single Nano-particle
Organic Memory Field Effect Transistor (NOMFET) is used to implement each
synapse. We show how the physical properties of this dynamic memristive device can be
used to perform low power write operations for the learning and implement short-term
association using temporal coding and Spike-Timing-Dependent Plasticity (STDP)-
based learning. An electronic circuit was build to validate the proposed learning scheme
with packaged devices, with good reproducibility despite the complex synaptic-like
dynamic of the NOMFET in pulse regime.
1 Introduction
In classical conditioning, associative learning involves repeatedly pairing an
unconditioned stimulus, which always triggers a reflexive response, with a neutral
stimulus, which normally triggers no response. After conditioning, a response can be
triggered for both the unconditioned stimulus and the neutral stimulus, the later one
becoming a conditioned stimulus. This concept goes back to Pavlov’s experiments in
the early 1900s. In his famous experiments, he showed how a neutral stimulus - like
the ring of a bell - could be associated to the sight of food and trigger the salivation
of his dogs (Pavlov, 1927). Associative memory is now a key concept in the learning
and adaptability processes of the brain. If this ability is ubiquitous in everyday life, it
is because it is likely a direct consequence of the structure and plasticity of the brain,
down to the synaptic plasticity level. This is the basis of the Hebbian theory, which can
be summarized as cells that fire together, wire together. It is therefore not surprising if
associative learning is extensively studied in artificial neural networks (Hassoun, 1993;
Norman, 2007).
However, the lack of an efficient implementation of artificial synapses for
associative learning neural networks has greatly impeded the use of associative memory
as a general-purpose type of memory or learning tool. The need of several transistors
to implement dynamical synapses (Hynna and Boahen, 2006) is indeed not efficient
enough compared to re-programmable digital logic, considering that at least several
thousands of synapses are generally needed to perform any large scale information
processing functionality such as associative memory on “real-life” data - the processing
of visual or auditory stimuli for example (Bichler et al., 2012) - with limited reuse for
other tasks. Synapses implemented with the most current CMOS technology would
therefore still be several orders of magnitude behind their biological counterparts in
terms of area and power consumption, which can be roughly estimated to ∼ 1012
synapses/cm3 and 1-10 fJ/spike if one considers an average firing rate of 1-10 Hz, given
a power consumption of the human brain on the order of 10 W (Kandel et al., 2000).
This is the reason why many new nano-devices currently actively researched to replace
flash memory are receiving considerable attention from the neuromorphic engineering
community.
2
Among them, Phase-Change Memory (PCM) devices have been proposed to realize
artificial synapses implementing Spike-Timing-Dependent Plasticity (STDP), by using
gradual crystallization and amorphization of the phase-change material, corresponding
to gradual increase and decrease of conductance, for Long-Term Potentiation (LTP)
and Long-Term Depression (LTD), respectively (Kuzum et al., 2012). In Conductive-
Bridging RAM (CBRAM), the conductance of the device can be modulated by
controlling the growth of the conductive filament in the material (Kund et al., 2005).
If PCM crystallization has been show to be a truly cumulative process, meaning
that successive identical programming pulses lead to gradual increase of conductance
(Suri et al., 2012), it is not clear how this can be achieved with PCM amorphization and
CBRAM, which require more complex programming schemes for synaptic applications
(Bichler et al., 2012; Yu and Wong, 2010). While these two technologies are currently
reaching industrial stage, multi-device synaptic applications are yet to be demonstrated
in practice. Resistive RAM (RRAM or OXRAM), like Univ. Michigan nanoscale
synapses (Jo et al., 2010) or HP TiO2 memristors (Strukov et al., 2008), also present
interesting characteristics for synaptic applications, with a cumulative effect for the
conductance programming (Querlioz et al., 2011), as some carbon nanotube devices
(Agnus et al., 2010; Zhao et al., 2010). All these devices fall in the class of memristive
devices. Their high scalability potential and their programming scheme make them
serious candidates to implement efficient artificial synapses.
Beside the above-mentioned technologies, volatile and/or organic synaptic memory
devices have received comparatively little attention from the community. Bi-
stable organic memories based on PMMA:ZnO (Ramana et al., 2012) or PMMA:C60
(Frolet et al., 2012) nanocomposites have been recently shown. Non-volatile multilevel
conductance has also been demonstrated, for example in organic/Si nanowire
transistor, by controlling the concentration of ions in the thin film layer through
the gate voltage (Lai et al., 2008). Although the physical processes modulating
the conductance might well be cumulative, the practicability of synaptic LTP
or LTD implementation by gradual conductance change remains to be assessed.
This leads us to the Nano-particle Organic Memory Field Effect Transistor
(NOMFET) (Alibart et al., 2010; Bichler et al., 2010). It was shown in previous work
(Alibart et al., 2012) that the NOMFET could be seen as a memristive device, by
3
modulating the conductivity of its channel through the charging of nano-particles
embedded into the organic semi-conducting channel. It is volatile and has a retention
time of typically 10 to 1000 s. With these physical properties, the NOMFET can exhibit
many behaviors of a dynamic synapse (Abbott et al., 1997; Tsodyks et al., 1998), when
used in pulse regime, with a clear cumulative effect.
An elementary associative memory with memristive devices was first proposed by
Pershin and Di Ventra, effectively reproducing the Pavlov’s dogs experiment (Pershin
and Di Ventra, 2010). The memristive devices were emulated with a micro-controller
however, thus mitigating many difficulties that could arise from the interfacing with
physical nano-devices, which are hard to predict in simulation with behavioral models.
Another approach was proposed in (Ziegler et al., 2012) to realize the same elementary
associative memory using a single Pt/Ge0.3Se0.7 /SiO2 /Cu memristive device. In
this paper, we propose an original scheme using NOMFETs to implement dynamic
associative learning and demonstrate it by interfacing NOMFETs to a CMOS discrete
circuit, thoroughly described in the “Methodology” section. We show how the unique
synaptic properties of the NOMFET can be used directly at the device level to
implement what we call a dynamic associative memory, where the association is only
retained as long as there is a minimal activity at its input. In the final “Experiments and
Results” section, the learning dynamic is viewed in relation to the NOMFET volatility,
the learning scheme is interpreted in terms of STDP and the impact of variability is
briefly discussed.
2 The NOMFET
The NOMFET is a three-terminal device, as the conventional MOSFET (see figure
1). As an organic transistor, it features the classical p-type transistor behaviors in
accumulation regime, but with added memristivity. For a fixed gate voltage, its
conductivity can indeed be modulated by charging or discharging the gold nano-
particles (NPs) embedded in its channel, with a negative or positive gate voltage
respectively. When a negative gate voltage is applied, the NPs are positively charged
and the repulsive electrostatic interaction between the holes trapped in the NPs and the
ones in the pentacene p-type channel therefore reduces the channel conductivity. The
4
Gold electrode
Pentacene
Au Nanoparticles
Drain
Source
p+ silicon
Gate
SiO2
200nm
Figure 1: Physical structure of the NOMFET transistor (left). It is composed of a
p+ doped bottom-gate covered with silicon oxide (200 nm thick). Source and drain
electrodes are made of gold and Au NPs (20 nm in diameter) are deposed on the inter-
electrode gap (5 µm), before the pentacene deposition. Scanning electron microscope
image of the NP array between the source and drain electrodes (right).
charge retention time is typically of 10 to 1000 s. The programming of the conductivity
of the channel can be done with no current flowing from the source to the drain, by
keeping VDS = 0.
Figure 2 shows the charging/discharging dynamic characteristic of the NOMFET.
The current is measured before and after a 10 s programming pulse is applied on VG .
The ∆I /I − V curve shows the change in conductivity in function of the pulse voltage,
with a non-linearity between 0 and 15 V. This characteristic remains the same when
no current flows through the channel, thus making the NOMFET an ideal memristive
device (Alibart et al., 2012) in the sense that on average, no power is dissipated for
the programming of its conductivity. The evolution of the conductivity is essentially
controlled by the time integral of the gate voltage, contrary to the memistor (not to
be confused with the memristor), which was programmed by the time integral of the
current flowing through its third terminal (Widrow et al., 1961).
While we have reported synaptic-like behavior in short (200 nm gap, 5 nm NPs)
NOMFET working at a bias of -3 V (Alibart et al., 2010), here we use, for the sake
of demonstration, larger (5 µm gap, 20 nm NPs) NOMFET because they showed the
largest plasticity amplitude (i.e. the largest modulation of the NOMFET conductivity
under the application of programming pulses). With an ON-state current ranging
from 50 to 500 nA and a drain-to-source voltage of 15 V, the equivalent resistance
5
)
V
(
G
V
L
A
I
T
I
N
I
I
/
)
L
A
I
T
I
N
I
I
-
R
E
T
F
A
I
(
Pulse voltage VP (V)
Figure 2: Dynamic characteristic: relative conductivity change as a function of the
pulse voltage (dashed line is a guide for eyes). The current is measured just before
and after the application of a 10 s pulse with different amplitudes. The relative change
in the current is related to the amount of charges in the NPs. For negative pulses,
the NOMFET is in accumulation, NPs are positively charged by holes, thus reducing
the current by coulombic effect. For positive bias, holes are detrapped, leading to
an increase in the current. The curve is non-linear with a zone between 0 V and 15
V where charges in the NPs are not affected by the applied bias. For more details, see
ref. (Alibart et al., 2012).
of the synapse ranges from 30 MΩ to 300 MΩ. While this is relatively high
compared to other memristive technologies (Bichler et al., 2012; Kund et al., 2005),
this is accomplished without sacrificing the OFF/ON resistance ratio thanks to the
transistor field effect (OFF-state current below 0.1 nA, measurement limited). This
makes the NOMFET low power, yet still usable in device arrays for large-scale
networks (Liang and Wong, 2010).
3 Methodology
Our associative memory is constituted of two input neurons, two synapses and one
output neuron (figure 3). A synapse is implemented using a single NOMFET and
the neurons are built with discrete CMOS devices on a custom electronic board. The
association is realized relying solely on the plasticity behavior of the NOMFET: its
conductivity is changed in an unsupervised way by the interaction of the pre-synaptic
6
Input #1
“Food”
Input #2
“Bell”
S
D I1
G
I2
Output
“Salivation”
Feedback
CLKREAD
Figure 3: Equivalent electronic circuit for the associative memory. There are 3
neurons (Input #1, Input #2 and Output) and 2 synapses/NOMFETs. A global
clock signal (CLKREAD ) enables synchronization of the input and output neurons
during the read operations. I1 and I2 are the measured drain-source currents.
pulses (input signal) and the post-synaptic pulses (feedback), such that the association
is made when there is a temporal correlation between the pre- and post-synaptic pulses
for a long time (see how this relates to STDP in the “Experiements and Results” part).
The operations on the synapses are done in two steps, a “read” step and a “write”
step. Separating these two steps allows a greater control on the dynamic of the system,
effectively decoupling the physical plasticity of the NOMFETs from the associative
learning dynamic. More generally, even in non-volatile memristive devices based neural
networks, separating these two steps can have several advantages: because the read can
be done with a lower voltage and/or a shorter pulse, the conductivity change of the
device induced by the read can be minimized, as does the power consumption. Another
interest of this approach is that current does not need to flow through the device during
the write step. Indeed, contrary to resistive-based memristive devices, the nano-particles
in the NOMFET channel can be charged or discharged without the transistor being in
the ON state, thus modulating its conductivity with no current flowing from the source
to the drain. We effectively use this principle in our experimental setup, where the
pre-synaptic pulses are applied to the gate of the NOMFETs. The output neuron is
connected to the source and drain terminals of all the devices and can apply separate
voltages to the source and to the drain.
7
3.1 Read and Output Neuron Activation
A read step occurs periodically and triggers the activation of the output neuron if an
input is present on a highly conductive synapse (meaning that the neuron is currently
“listening” to that synapse). During a read step, the conductive state of the devices
is measured. A feedback is generated if there is an active input signal at one of the
NOMFETs during the read and that it has a conductivity higher than a fixed threshold.
The read step implies that both input and output neurons are synchronized, as the active
input neurons should turn on the NOMFETs by applying a negative voltage to the gate.
The output neuron should apply a negative voltage pulse to the source of the devices
and measure the drain-source current at the drain terminal. This process can be done
sequentially by turning on only one NOMFET at a time so that all the drain and all the
source terminals can be linked together.
3.2 Write (Input and Feedback Interaction)
The conductivity of the NOMFETs is modulated during the write steps, through the
interaction of the pre-synaptic pulses (gate) and the post-synaptic pulses (source and
drain). In order to achieve an associative memory, the conductivity of a NOMFET
should increase only in the occurrence of simultaneous pre- and post-synaptic pulses,
meaning that there is a correlation between the events leading to the activation of the
output neuron and the input events coming to the synapse. This is implemented by
ensuring that only the interaction of pre- and post-synaptic pulses leads to a significant
increase in the conductivity of the NOMFET, by applying a higher voltage across the
gate and source-drain terminals of the device than with a pre- or a post-synaptic pulse
alone. Since the relation between the conductivity change and the applied voltage
amplitude is non-linear (see figure 2), it is possible to effectively maximize the effect of
the two interacting pulses and minimize the effect of a single pulse to the conductivity.
The shape of the pulses is given in figure 4:
• When no feedback is present, the input only has limited effect on the
conductivity of the synapse: no input means a constant voltage of -15 V
is applied to the gate, which tends to decrease the conductivity over time
(loss of memory). This is mitigated by the time when the input is being
active, where the conductivity
8
Equivalent VGS
voltage across the
NOMFET
/ write pulse
Rest (no input) (G)
0V
-15V
30V
0V
Input (G)
No feedback (D, S)
0V
0V
-15V
30V
0V
Feedback (D, S)
0V
-30V
15V
0V
60V
0V
Figure 4: Input and feedback pulse shapes to implement the associative learning.
Input signals are applied to the gate and feedback to the source and drain
terminals. The inputs are negatively biased, meaning that a -15 V voltage is
constantly applied to the inputs when no signal is present. Only the interaction of an
input signal and a feedback signal creates a programming voltage between the gate
and the drain-source terminals large enough to significantly increase the conductivity
of the NOMFET, thus leading to the association when several interactions occurs in a
short time.
9
slightly increases. The overall effect is that the conductivity of the synapse is
maintained as long as there is a sustained activity at its input.
• When a feedback is present, there is no change in conductivity in the absence of
input, whereas a major increase in conductivity occurs upon the interaction of the
input present pulse and the feedback pulse. The -15 V bias was chosen in order to
avoid maintaining an association with only the feedback (thus the voltage across
the NOMFET is 15 V with the feedback, instead of 30 V with an input).
During write operations, the source and drain terminals are connected together,
which implies that the only electric potential difference in the device is across the gate,
regardless of the simultaneity of the input and feedback pulses. In this configuration,
the device act as a capacitor, with the charging and discharging of the embedded NPs
in the channel modulating the conductivity of the transistor, for a given gate voltage
during read operations. Like a capacitor, because no current can flow through the gate,
the change in conductance does not dissipate any power on average. This is unlike
non charge-based, non-volatile, resistive memory devices like PCM or CBRAM, where
conductance change requires Joule heating to induce phase-change or ions migration in
the material. In these technologies, the programming current typically falls in the mA
or µA range (Bichler et al., 2012; Kund et al., 2005).
3.3 Calibration
In a dynamical associative learning network, a calibration step is necessary to initialize
the state of the synapses, in order to enable system responsiveness for an initial set of
unconditioned stimuli. An optimal threshold can also be automatically computed for
this initial configuration. Because the synaptic weights are programmed in parallel, this
initial set-up does not depend on the number of synapses.
In our demonstration, this calibration process is automated in the micro-controller
implementing the neurons (see next sections). It sets up the threshold before the
associative memory learning is run. In addition, a constant gain is automatically applied
to the reading of one of the two NOMFETs in order to compensate for the mismatch
between the two devices. The calibration sequence is the following: (1) Measure the
conductivity state of each device for the four possible programming states (with or
10
without pre-synaptic spike and with or without post-synaptic spike). (2) Compute the
gain to be applied to the second device to compensate for the mismatch. (3) Compute
the optimal threshold. (4) Initialize the conductivity state of both devices.
Although steps (1)-(2) may be unpractical in larger systems, mismatch
compensation for the static IDS characteristic may be eliminated for matched devices
on the same die fabricated in controlled industrial process. Indeed, we found that most
of the mismatch is introduced by the wire bounding during device packaging.
3.4 Experimental board
A custom electronic board was build to interface with the NOMFETs. The NOMFETs
were packaged in a standard through-hole TO (Transistor Outline) metal can package.
The design of electronic boards with off the shelf components to interface with
experimental devices often proves to be challenging, not to mention specific constrains
on temperature or atmosphere control for the wire bounding and packaging depending
on the materials used. For the NOMFET, one has to be capable of measuring current in
the nA range with voltage as high as ±30 V for the first generation of NOMFETs that
we used for this setup.
Despite these possible difficulties, the presented NOMFETs are working in
atmospheric condition without any special encapsulation and have shown a stability
in time over more than 6 months. And we have previously reported that the synaptic
properties of the NOMFET were maintained by decreasing the gate thickness down
to 10 nm, reducing the gap size down to 200 nm and the NP diameter to 5 nm (Alibart
et al., 2010), thus reducing the maximum programming voltage down to 3 V. The bottom
up approach and the low temperature deposition condition for the fabrication of the
NOMFETs is compatible with CMOS and we think that interfacing of experimental
devices with standard electronic is a valuable step towards larger scale prototypes and
co-integration with CMOS.
As the proposed associative learning scheme only requires four different voltages
to generate the read and write pulses, the pulses at each terminal are generated with an
analog multiplexer, the MAX14752, which can switch voltages up to ±36 V. Because
we use the dynamic of the NOMFET in the ms-s range, the switching speed is not
11
Figure 5: Photography of the experimental board, with the NOMFETs in a TO
(Transistor Outline) case at the center. The computer interface, including the FPGA
(Field-Programmable Gate Array)-based control board implementing the control
logic for the neurons, are not shown.
critical.
For the read step, the conductivity is obtained by measuring the current flowing
through the drain terminal when a positive voltage pulse is applied to the source,
which is equivalent to applying a negative pulse to the gate and the drain. The
current measurement is done with a current-to-voltage converter in a feedback-ammeter
topology. In this topology, the operational amplifier delivers a voltage proportional to
the current flowing through the feedback resistor, which is equal to the drain-source
current of the NOMFET. It is particularly adapted for the measurement of small currents
(below the µA). The OPA445 is used for this task, which is one of the few operational
amplifiers to support a large supply voltage of up to ±45 V and to have a low input
bias current (of the order of 10 pA), required to measure current down to the nA. It
is followed by a differential amplifier AD629B to allow measurement of the current
when a feedback pulse is applied to the NOMFET, by suppressing the common mode
voltage of the feedback pulse. It supports a high common mode voltage and has a low
input voltage offset. Finally, the LTC1856 analog-to-digital converter (ADC) is used
to obtain a digital value of the current. It can measure bipolar voltages up to ±10 V
directly applied at its input and can withstand input voltages as high as ±30 V without
damage. It also has eight multiplexed measurement channels with a resolution of 16
12
VREST
VINPUT
Input signal
Input neuron
CLKREAD
VREAD
-
+
+
-
VFB
ADC
Output neuron
Software
implemented
state
machine
Figure 6: Simplified schematic of the experimental setup for one NOMFET. The
blue/thin parts are implemented in software on the Microblaze synthesized on the
FPGA. In read mode, the source terminal is connected to VREAD while the other two
terminals are grounded. When the output neuron activates, the source and drain
terminals are connected to VFB and the device’s conductivity is modulated depending
on the gate voltage VREST (no input) or VINPUT (input).
bits, making it easy to measure several devices with a single component.
A photography of the board is shown in figure 5 and a simplified schematic of the
circuit is shown in figure 6. No special care was needed to measure current down to the
nA, other than careful layout with proper ground plane and decoupling capacitors.
3.5 Neuron emulation
While the experimental board provides the digital-to-analog (pulses generation) and
analog-to-digital (current measurement) interface to the NOMFETs, the neurons
themselves are implemented in digital logic on a FPGA (Field-Programmable Gate
Array)-based board, which does the initial calibration and supervises the learning
process. Instead of implementing the control logic directly as a state machine in a
hardware description language, a programmatic approach was preferred for flexibility.
A Xilinx Microblaze (Xilinx, 2012) micro-controller core was therefore synthesized on
the FPGA and programmed in the C language.
13
4 Experiments and results
To demonstrate the working order of our associative memory, we reproduced the
Palvov’s dog experiment. Figure 7 shows the association process: the two inputs of
the memory models the sight of food and the hearing of a bell respectively, whereas
the output neuron activity models the salivation of the dog. At first, the conductivity of
the devices was programmed so that only the sight of food (unconditioned stimulus)
triggers the activation of the output neuron (salivation). Before the association is
made, the hearing of the bell (neutral stimulus) does not trigger the salivation, as
the conductivity of the corresponding synapse is below the threshold of the output
neuron. When both inputs are active simultaneously, the conductivity of the second
synapse increases until it reaches the threshold (conditioning): the association is made.
From this point, the hearing of the bell alone triggers the activation of output neuron
(conditioned stimulus). In figure 8 we can verify that without feedback from the output
neuron, the conductivity change implied by the pre-synaptic pulses alone is not enough
to create the association.
Our associative learning circuit is symmetrical, which means that there is no
difference between the input receiving the conditioning stimulus (sight of food) and the
input receiving the neutral stimulus (bell sound). The conditioning input is the one that
is initially programmed to have a synaptic weight above the output neuron threshold.
This is in contrast with the asymmetrical scheme introduced in (Ziegler et al., 2012),
where the two inputs cannot be interchanged.
4.1 NOMFET Volatility and Learning Dynamic
Because the conductivity of the NOMFET is volatile, what we really implemented is
a short-term associative memory. Indeed, when no sustained activity is present at the
inputs, the association is lost (figure 9). This functionality is directly provided by the
physics of the device. There is no need for additional circuitry that would periodically
decrease the conductivity of the synapses. We argue that this functionality is essential
for any practical use of associative memory, because otherwise, the synaptic weights
in such a system would eventually saturate due to the random short-time correlations
between inputs.
14
Figure 7: Dynamic associative learning: association is maintained as long as there is
a minimum activity at the inputs. The duration of a cycle is 200 ms. After each
cycle, the current is measured and compared to the neuron’s threshold. In this
experiment, the minimal duration of an input pulse is 5 cycles (1 s). The total duration
of the sequence is 40 s.
15
Figure 8: Same sequence, with disabled feedback: no association is made and the
first synapse loses the association as well. As in the other experiments, the best
threshold is automatically computed at the beginning of the sequence using the same
calibration procedure, in which the feedback was also disabled.
Figure 9: Dynamic associative learning: if no sustained activity is present at the
inputs, the association is lost.
16
The NOMFET exhibits both a short-term and a long-term dynamic, that are clearly
identifiable in figure 7. The short-term dynamic corresponds to the fast increase and
decrease of conductance upon activation and deactivation, respectively, of the input
neuron (typical time dynamic ∼ 1 s). The average conductance or current evolution
during the learning follows a longer term dynamic (typical time dynamic ∼ 10 to
100 s): the average current is indeed higher after the association than before. A
long potentiating pulse (conditioning step) is required to induce this more durable
change. Such dynamical synapse in time is likely to be impractical to emulate with
non-volatile memristive devices, like PCM, as each device would require some external
time reference to implement its time dynamic.
4.2 Relationship with STDP
Spike-Timing-Dependent Plasticity is currently one of the most widely studied
neuromorphic learning algorithm on memristive devices (Yu and Wong, 2010; Kuzum
et al., 2012; Jo et al., 2010; Bichler et al., 2012). There is not a single STDP rule
however: a broad family of synaptic update characteristics in function of the pre-
post synaptic time difference were recorded (Wittenberg and Wang, 2006). Most
notably, asymmetric STDP was observed in hippocampal glutamatergic synapses
(Bi and Poo, 1998) and symmetric STDP was observed in CA1 region of the
hippocampus (Nishiyama et al., 2000) and GABAergic synapses (Woodin et al.,
2003).
Asymmetric STDP with NOMFET was reported in (Alibart et al., 2012). The
programming scheme realized in our associative memory however emulates a form of
symmetric STDP learning rule, as shown qualitatively in figure 10. Indeed, a significant
increase in conductivity is only induced when there is a simultaneous occurrence of pre-
and post-synaptic events (input and feedback pulses), leading to a 60 V equivalent pulse
across the NOMFET.
4.3 Impact of variability
Although there is a factor 10 in the mean conductivity between the two NOMFETs used
in this experiment (before applying a correcting gain), the variability on the dynamical
17
60
40
20
80
)
%
(
W
(cid:0)
e
g
n
a
h
c
t
h
g
i
e
w
e
v
-40
i
t
a
l
R
e
-100
-20
0
W ood in et a l. data
Gaussian fit
-50
(cid:0)T = t
post
0
50
- t
pre (ms)
100
e
g
n
a
l
h
a
i
t
c
i
n
i
t
I
n
/
)
e
l
a
r
i
t
i
r
n
u
i
I
c
-
r
e
e
t
f
a
v
I
i
(
t
a
l
e
R
Pulse across
the NOMFET
Pre-post
interaction
ΔT = t
post - tpre
(ms)
Figure 10: Left: relative synaptic weight change in symmetric STDP, observed in
GABAergic synapses. Experimental data from (Woodin et al., 2003). Right: relative
conductance change realized
in our associative
learning
for a NOMFET
synapse, depending on the pre-post programming pulses interaction (qualitative
diagram).
behavior of the devices is low in comparison (Bichler et al., 2010; Alibart et al., 2012).
Because our associative memory is based on temporal coding, only the relative variation
of the conductivity obtained through programming pulses and natural charge relaxation
of the NPs imposes the dynamic of the learning. In this regard, STDP-based learning
systems were shown to be tolerant to variations from 20% up to 100% in synaptic
weight update steps (Querlioz et al., 2011; Bichler et al., 2012), which was confirmed
by the high reproducibility of our experiment. Incidentally, the current variations for
the two NOMFETs are not necessarily matched during the calibration process, which
is apparent in figures 7-9. Instead, the effect of the calibration process is to match the
temporal dynamic of the conductivity variations. This dynamic remains valid for longer
time scale, with cycle duration up to one order of magnitude higher.
Discussion and Conclusion
Associative memory is a fundamental computing block in the brain and is implemented
extremely efficiently in biological neural networks. An efficient and scalable
implementation of associative memory would certainly benefit many applications,
especially in the area of natural data processing. One could imagine, using the same
principle as the Pavlov’s dog learning, association of visual and auditory features, but
directly below artificial retina and cochlea event-based sensors.
18
We demonstrated experimentally an elementary associative memory, which uses
only one NOMFET memristive nano-device per synapse. Although it may not be able to
ultimately achieve the scalability of biological synapses, it exhibits dynamical behaviors
closer to biology than any other known memristive device. Furthermore, the volatility
of the NOMFET that leads to the dynamic learning scheme introduced in this paper is
not necessarily irrelevant when considering biological synchronous neural models for
short-term memory (Ward, 2003).
Acknowledgments
We thank the following colleagues for help in the device fabrication: K. Lmimouni
(pentacene deposition), D. Guerin for surface chemistry (grafting of NPs), J. Dubois
(earlier electronic circuit prototyping), and D. Querlioz for helpful discussions.
Financial support: European Union through the FP7 Project NABAB (Contract FP7-
216777).
References
Abbott, L., J. Varela, K. Sen, and S. Nelson (1997). Synaptic depression and cortical
gain control. Science 275, 220–224.
Agnus, G., W. Zhao, V. Derycke, A. Filoramo, Y. Lhuillier, S. Lenfant, D. Vuillaume,
C. Gamrat, and J.-P. Bourgoin (2010). 2-terminal carbon nanotube programmable
devices for adaptive architectures. Advanced Material 22(6), 702–706.
Alibart, F., S. Pleutin, O. Bichler, C. Gamrat, T. Serrano-Gotarredona, B. Linares-
Barranco, and D. Vuillaume (2012). A memristive nanoparticle/organic hybrid
synapstor for neuro-inspired computing. Advanced Functional Materials 22(3), 609–
616.
Alibart, F., S. Pleutin, D. Guérin, C. Novembre, S. Lenfant, K. Lmimouni, C. Gamrat,
and D. Vuillaume (2010). An organic nanoparticle transistor behaving as a biological
spiking synapse. Advanced Functional Materials 20(2), 330–337.
19
Bi, G.-Q. and M.-M. Poo (1998). Synaptic Modifications in Cultured Hippocampal
Neurons: Dependence on Spike Timing, Synaptic Strength, and Postsynaptic Cell
Type. Journal of Neuroscience 18, 10464–10472.
Bichler, O., D. Querlioz, S. Thorpe, J. Bourgoin, and C. Gamrat (2012). Extraction
of temporally correlated features from dynamic vision sensors with spike-timing-
dependent plasticity. Neural Networks 32, 339–348.
Bichler, O., M. Suri, D. Querlioz, D. Vuillaume, B. DeSalvo, and C. Gamrat (2012).
Visual Pattern Extraction Using Energy-Efficient ’2-PCM Synapse’ Neuromorphic
Architecture. Electron Devices, IEEE Transactions on. In press.
Bichler, O., W. Zhao, C. Gamrat, F. Alibart, S. Pleutin, and D. Vuillaume (2010).
Functional model of a nanoparticle-organic memory transistor for use as a spiking
synapse. Electron Devices, IEEE Transactions on 57 (11), 3115–3122.
Frolet, N., M. Charbonneau, R. Tiron, J. Buckley, D. Mariolle, D. Boutry, R. Coppard,
and B. De Salvo (2012). Development of organic resistive memory for flexible
electronics. In MRS Proceedings, Volume 1430, pp. mrss12–1430–e05–06.
Hassoun, M. (1993). Associative Neural Memories: Theory and Implementation.
Oxford Press.
Hynna, K. and K. Boahen (2006). Neuronal ion-channel dynamics in silicon. In Circuits
and Systems, 2006. ISCAS 2006. Proceedings. 2006 IEEE International Symposium
on, pp. 3614–17.
Jo, S. H., T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumder, and W. Lu (2010).
Nano
Nanoscale Memristor Device as Synapse in Neuromorphic Systems.
Letters 10(4), 1297–1301.
Kandel, E. R., J. H. Schwartz, and T. M. Jessell (2000). Principles of Neural Science
(4th ed.). McGraw-Hill Medical.
Kund, M., G. Beitel, C.-U. Pinnow, T. Rohr, J. Schumann, R. Symanczyk, K.-D. Ufert,
and G. Muller (2005). Conductive bridging RAM (CBRAM): an emerging non-
20
volatile memory technology scalable to sub 20nm. In Electron Devices Meeting,
2005. IEDM Technical Digest. IEEE International, pp. 754–757.
Kuzum, D., R. G. D. Jeyasingh, B. Lee, and H.-S. P. Wong (2012). Nanoelectronic
Programmable Synapses Based on Phase Change Materials for Brain-Inspired
Computing. Nano Letters 12(5), 2179–2186.
Lai, Q., Z. Li, L. Zhang, X. Li, W. F. Stickle, Z. Zhu, Z. Gu, T. I. Kamins, R. S.
Williams, and Y. Chen (2008). An Organic/Si Nanowire Hybrid Field Configurable
Transistor. Nano Letters 8(3), 876–880.
Liang, J. and H.-S. Wong (2010). Cross-point memory array without cell selectors -
device characteristics and data storage pattern dependencies. Electron Devices, IEEE
Transactions on 57 (10), 2531–2538.
Nishiyama, M., K. Hong, K. Mikoshiba, M.-m. Poo, and K. Kato (2000). Calcium
stores regulate the polarity and input specificity of synaptic modification. Nature 408,
584–588.
Norman, D. (2007). The Brain That Changes Itself. United States: Viking Press.
Pavlov, I. (1927). Conditioned reflexes: An investigation of the physiological activity of
the cerebral cortex. translated and edited by g. v. anrep. London: Oxford University
Press.
Pershin, Y. and M. Di Ventra (2010). Experimental demonstration of associative
memory with memristive neural networks. Neural Networks 23(7), 881–886.
Querlioz, D., O. Bichler, and C. Gamrat (2011). Simulation of a memristor-based
spiking neural network immune to device variations. In Neural Networks (IJCNN),
The 2011 International Joint Conference on, pp. 1775–1781. IEEE.
Querlioz, D., P. Dollfus, O. Bichler, and C. Gamrat (2011). Learning with memristive
devices: How should we model their behavior? In Nanoscale Architectures
(NANOARCH), 2011 IEEE/ACM International Symposium on, pp. 150–156.
21
Ramana, C., M. Moodely, V. Kannan, A. Maity, J. Jayaramudu, and W. Clarke (2012).
Fabrication of stable low voltage organic bistable memory device. Sensors and
Actuators B: Chemical 161(1), 684 – 688.
Strukov, D. B., G. S. Snider, D. R. Stewart, and S. R. Williams (2008). The missing
memristor found. Nature 453, 80–83.
Suri, M., O. Bichler, D. Querlioz, B. Traoré, O. Cueto, L. Perniola, V. Sousa,
D. Vuillaume, C. Gamrat, and B. DeSalvo (2012). Physical Aspects of Low Power
Synapses based on Phase Change Memory Devices. Journal of Applied Physics. In
press.
Tsodyks, M., K. Pawelzik, and H. Markram (1998). Neural networks with dynamic
synapses. Neural Computation 10, 821–835.
Ward, L. M. (2003). Synchronous neural oscillations and cognitive processes. Trends
in Cognitive Sciences 7 (12), 553–559.
Widrow, B., W. H. Pierce, and J. Angell (1961). Birth, life, and death in microelectronic
systems. Military Electronics, IRE Transactions on MIL-5(3), 191–201.
Wittenberg, G. M. and S. S.-H. Wang (2006). Malleability of Spike-Timing-Dependent
Plasticity at the CA3-CA1 Synapse. The Journal of Neuroscience 26(24), 6610–
6617.
Woodin, M. A., K. Ganguly, and M. ming Poo (2003). Coincident Pre- and Postsynaptic
Activity Modifies GABAergic Synapses by Postsynaptic Changes in Cl− Transporter
Activity. Neuron 39(5), 807 – 820.
Xilinx (2012). Xilinx DS865 LogiCORE IP MicroBlaze Micro Controller System.
Xilinx.
Yu, S. and H.-S. Wong (2010). Modeling the switching dynamics of programmable-
metallization-cell (PMC) memory and its application as synapse device for a
neuromorphic computation system. In Electron Devices Meeting (IEDM), 2010 IEEE
International, pp. 22.1.1–22.1.4.
22
Zhao, W., G. Agnus, V. Derycke, A. Filoramo, C. Gamrat, and J. Bourgoin (2010).
Nanotube devices based crossbar architecture: Toward neuromorphic computing.
Nanotechnology 21, 175–202.
Ziegler, M., R. Soni, T. Patelczyk, M. Ignatov, T. Bartsch, P. Meuffels, and H. Kohlstedt
(2012). An Electronic Version of Pavlov’s Dog. Advanced Functional Materials. In
press.
23
|
1510.03507 | 1 | 1510 | 2015-10-13T01:57:12 | The intrinsic value of HFO features as a biomarker of epileptic activity | [
"q-bio.NC",
"cs.LG",
"stat.ML"
] | High frequency oscillations (HFOs) are a promising biomarker of epileptic brain tissue and activity. HFOs additionally serve as a prototypical example of challenges in the analysis of discrete events in high-temporal resolution, intracranial EEG data. Two primary challenges are 1) dimensionality reduction, and 2) assessing feasibility of classification. Dimensionality reduction assumes that the data lie on a manifold with dimension less than that of the feature space. However, previous HFO analyses have assumed a linear manifold, global across time, space (i.e. recording electrode/channel), and individual patients. Instead, we assess both a) whether linear methods are appropriate and b) the consistency of the manifold across time, space, and patients. We also estimate bounds on the Bayes classification error to quantify the distinction between two classes of HFOs (those occurring during seizures and those occurring due to other processes). This analysis provides the foundation for future clinical use of HFO features and buides the analysis for other discrete events, such as individual action potentials or multi-unit activity. | q-bio.NC | q-bio | THE INTRINSIC VALUE OF HFO FEATURES AS A
BIOMARKER OF EPILEPTIC ACTIVITY.
Stephen V. Gliske, William C. Stacey∗
Kevin R. Moon, Alfred O. Hero III†
5
1
0
2
t
c
O
3
1
]
.
C
N
o
i
b
-
q
[
1
v
7
0
5
3
0
.
0
1
5
1
:
v
i
X
r
a
University of Michigan
Department of Neurology
Ann Arbor, Michigan, U.S.A.
ABSTRACT
High frequency oscillations (HFOs) are a promising biomarker of
epileptic brain tissue and activity. HFOs additionally serve as a pro-
totypical example of challenges in the analysis of discrete events in
high-temporal resolution, intracranial EEG data. Two primary chal-
lenges are 1) dimensionality reduction, and 2) assessing feasibility
of classification. Dimensionality reduction assumes that the data lie
on a manifold with dimension less than that of the features space.
However, previous HFO analysis have assumed a linear manifold,
global across time, space (i.e. recording electrode/channel), and in-
dividual patients. Instead, we assess both a) whether linear methods
are appropriate and b) the consistency of the manifold across time,
space, and patients. We also estimate bounds on the Bayes classifi-
cation error to quantify the distinction between two classes of HFOs
(those occurring during seizures and those occurring due to other
processes). This analysis provides the foundation for future clinical
use of HFO features and guides the analysis for other discrete events,
such as individual action potentials or multi-unit activity.
Index Terms -- high-frequency oscillation, intrinsic dimension,
dimensionality reduction, classification error, divergence
1. INTRODUCTION
About one third of epilepsy patients fail to obtain seizure control
with available pharmaceuticals. One of the few options for these re-
fractory patients is resective surgery -- removing the portion of the
brain thought to be causing the seizures. This region is denoted the
seizure onset zone (SOZ). In some cases, determining the SOZ in-
volves a highly invasive surgery to place electrodes on the brain's
surface, followed by one to two weeks of recording and monitoring.
A second invasive surgery is performed if the SOZ can be identified
and safely resected. A schematic relating the implanted electrodes
with the recorded data is shown in Fig 1.
A proposed biomarker to improve the localization of the SOZ
are high frequency oscillations (HFOs) [1, 2]. HFOs are high fre-
quency (about 80 -- 300 Hz), short (< 50 ms), rare events occurring
in intracranial EEG recorded at sampling rates of several kHz. Ex-
ample HFO detections and a recorded seizure are shown in Fig 2.
∗These authors were partially supported by the Doris Duke Foundation
and the US National Institute of Health (NIH) grant 5-K08-NS069783-05.
SG was additionally supported by an NIH Big Data to Knowledge Mentored
Training Grant, 1-K01-ES026839-01A1.
†These authors were partially supported by the US National Science
Foundation (NSF) under grant CCF-1217880 and a NSF Graduate Research
Fellowship to KM under Grant No. F031543.
University of Michigan
Electrical Engineering and Computer Science
Ann Arbor, Michigan, U.S.A.
Fig. 1. Diagram relating the recorded data with the implanted elec-
trodes. A 5×7 grid of electrodes placed over a region of cortex. Each
channel produces a separate time series of data, with some channels
being identified by clinicians as seizure onset zone (SOZ). HFO de-
tections are marked by solid magenta lines under the EEG trace.
Much of the published research on HFOs uses human identified
HFOs in short (10 to 20 minute) recordings [3, 4]. These results have
shown that a high HFO occurrence rate is correlated with the SOZ.
However, recent work is moving towards automated identification
and analysis of HFOs in long term, high resolution data, which re-
quires advanced computational and statistical techniques [5]. Qual-
ity recordings may span 7-14 days, with over 100,000 HFO detec-
tions in several terabytes of data. Thus, the next advances are ex-
pected to come through big data analysis of HFO features, utilizing
millions of recorded HFOs across as many patients as possible [6].
Relatively few research groups have analyzed HFO features in
detail. The most advanced analysis computed six features of about
300,000 HFOs in nine patients and two controls, and utilized a global
PCA across all channels followed by k-means clustering [7, 8, 9].
The authors implicitly assumed that the distribution of these HFO
features lies on a linear manifold in feature space, and that the man-
ifold is consistent across time and space (i.e., recording electrode).
The other most advanced analysis compared HFOs produced in
the motor cortex via movement versus HFOs occurring in the SOZ,
utilizing three features and a support vector machine (SVM) classi-
fier [10]. Some differences were noted, but a more general analysis
that addresses the degree to which HFOs produced by pathologi-
cal activity or networks (denoted pathological HFOs or pHFOs) and
HFOs produced by normal, physiological activity (denoted normal
HFOs or nHFOs) are observably different has not been performed.
The goals of this work are to test the implicit assumptions pre-
viously used in HFO feature analysis. Specifically, the goals are
to 1) assess the type of manifold on which HFO features lie, and
2) assess how discernible pHFOs are from nHFOs, based on their
25 mV
20 ms
25 mV
20 ms
2 V
5 min
Fig. 2. Example HFO detections within 45 min of one channel of
intracranial EEG data. A seizure occurs at about 35 minutes. HFO
detections (72 interictal and preictal, 32 ictal) are shown as small
yellow dots. Two HFOs are also shown using a much smaller scale.
feature-space distributions.
The general outline is as follows. To assess the linearity of the
HFO-feature manifold, a non-linear, local estimate of intrinsic di-
mension [11, 12] is compared with a linear global estimate (PCA)
applied to local subsets. This approach is similar to that in [13]. The
corresponding reduced dimension subspaces are individually com-
pared across time and space using a modified Grassmann distance,
building off the work in [14]. We additionally use a greedy Fisher
LDA algorithm (similar to the greedy LDA in [15]) to identify a
basis that maximally separates pHFOs from nHFOs. Unsupervised
clustering of the subspaces is then compared with channel groups
based on clinical markings of the SOZ and physical groupings of the
electrodes. Lastly, we assess the discernibility of nHFOs and pHFOs
by estimating bounds on the Bayes Error using the Henze-Penrose
divergence (HPD) [16, 17].
2. PATIENT POPULATION AND DATA
EEG data from adult patients with refractory epilepsy who under-
went intracranial EEG monitoring were selected from the IEEG Por-
tal [18] and from the University of Michigan. All patient data was
included which met the following criteria: sampling rate of at least
5 kHz, recording time greater than one hour, data recorded with
traditional intracranial electrodes, and available meta-data regard-
ing seizure times and the resected volume or SOZ. This yielded 17
patients, (nine IEEG portal, eight U. of M.). All data were acquired
with approval of local institutional review boards (IRB), and all pa-
tients consented to share their de-identified data. Of these 17 pa-
tients, 13 had recorded seizures, nine were known to have resection
and obtained seizure-freedom (ILAE class I), with eight patients in
common between these two categories. Because these surgeries are
relatively rare, this patient population size is moderately large for
analysis of intracranial EEG data.
HFOs were detected using the qHFO algorithm [5], resulting in
over 1.6 million HFOs in nearly 100,000 channel-hours of 5 kHz
data. Each HFO was band pass filtered between 80 to 500 Hz using
an elliptical filter and 33 features were computed, including dura-
tion, peak power, mean of the Teager-Kaiser energy [19], and various
spectral properties. Ictal is defined as during seizures, with seizures
assumed to be five minutes long if the length was not specified in
available meta-data. Interictal is defined as at least 30 minutes from
the start or end of a seizure, based on [9].
3. DIMENSIONALITY REDUCTION
3.1. Consistency of local, non-linear intrinsic dimension
To assess both the linearity and local versus global nature of the HFO
feature manifold, the non-linear intrinsic dimension was computed
via the k-nearest neighbor (NN) based estimator in [11]. This esti-
mator estimates the intrinsic dimension by exploiting its relationship
to the total edge length of the k-NN graph. This nonlinear estimator
provides a local estimate of intrinsic dimension which enables us to
identify local variations in data manifolds. The result is an estimate
of the intrinsic dimension for each given HFO, which are then aver-
aged to obtain the mean intrinsic dimension for a given partition of
the data. The consistency of the manifold is measured by comparing
the distribution of intrinsic dimension across time, space and pa-
tients. The specific comparisons are: a) interictal versus ictal times,
per channel, b) time variation within interictal periods, per channel,
and c) comparisons between channels, integrated over time.
A variety of methods could be employed to compare the intrin-
sic dimension for two disjoint sets of HFOs. However, the final di-
mension selected will be an integer. Thus, small differences in the
intrinsic dimension between two sets, no matter how statistically sig-
nificant, are not meaningfully different if they mean value for each
set round to the same integer.
To compare two sets of intrinsic dimension, we define a distance
measure, θI, between collections of integers. Let the two sets of
integers be A and B, and let ni be the fraction of elements in A
equal to i, and mi be the fraction of elements in B equal to i. The
T
n, and the
probability that two elements in set A are equal is n
T
probability that an element of A is equal to an element of B is n
m.
A measure of distance between A and B is how likely an element of
A is equal to an element of B, normalized by the likelihoods of
elements being equal another element in the same group:
θI = arccos(cid:18)
√n
T
T
n
m
n√m
m(cid:19) ,
T
(1)
The value θI is the angle between n and m and provides an easy
interpretation as to the consistency of two different collections of
local intrinsic dimensional. This quantity is also know as the angu-
lar distance or angular dissimilarity, and is the inverse cosine of the
Ochiai-Barkman coefficient [20, 21].
We compute the θI-distance for three different comparisons of
HFOs: 1) a comparison of each pair of 30 minute time windows on
a given channel for a given patient, 2) a comparison of ictal versus
interictal periods, again on a given channel for a given patient, and
3) a comparison of different channels in a given patient during in-
terictal periods. Interictal-ictal comparisons where one set of events
is less than 50 HFOs are ignored. Histograms of the distribution
of θI for each type of comparison are shown in Fig. 3. This fig-
ure involves over 5 million interictal time bin comparisons, 163 ictal
versus interictal comparisons, over 36 thousand interictal channel-
channel comparisons and almost 10 thousand ictal channel-channel
comparisons.
The intrinsic dimension is quite consistent across different time
segments during interictal times (strong peak near 0◦), but has some
variance between ictal and interictal times (still peaked near 0◦ but
the peak is wider). The dimension is less consistent across channels,
with comparisons during ictal times showing small peaks at both
0◦ and 90◦, and interictal having the largest peak at 90◦, showing
maximal difference. Thus we see that the intrinsic dimension varies
significantly across channels, especially for interictal HFOs.
s
t
n
u
o
c
d
e
z
i
l
a
m
r
o
N
0.4
0.3
0.2
0.1
0
0°
Interical Time Bins
Ictal vs Interictal
Interictal Channels
Ictal Channels
30°
60°
θ [degrees]
90°
Fig. 3. Comparisons of the consistency of the intrinsic dimension (θI
from Eq. 1) for four comparisons, as described in the text. Note, 0◦
implies no difference and 90◦ implies maximal difference between
the two collections of intrinsic dimension being compared.
3.2. Comparison with Global Linear Intrinsic Dimension
Next we compare the k-NN intrinsic dimension estimate of Sec-
tion 3.1 with a global, linear method to assess how linear and/or local
the feature manifold is. The most common global, linear method of
intrinsic dimension estimation and dimensionality reduction is prin-
ciple component analysis (PCA), which we perform by first center-
ing the data and then using singular value decomposition (SVD).
PCA is performed multiple times for the same divisions of the HFO
data as done for Fig. 3. Subsets of HFOs with less than 50 events are
ignored, as these are deemed insufficient to estimate the PCA vec-
tors in 33D. This results in PCA vectors being computed for 606 of
the 1318 channels (12 patients) for interictal HFOs, 171 channels for
ictal HFOs (8 patients), and 163 channels comparing ictal versus in-
terictal (8 patients). In patients that had multiple recording sessions,
channels are counted once per each session.
To compare the k-NN (non-linear) and PCA intrinsic dimension
per channel, we 1) select the number of principle components equal
to the non-linear intrinsic dimension estimate, and then 2) report the
fraction of the variance accounted for by that number of principle
components. This is repeated for all 606 channels (interictal) and
171 channels (ictal). When using either ictal or interictal HFOs, the
median fraction of variance was 99.8%, with 99%-tile of the chan-
nels being above 89.5% (interictal) and 97.1% (ictal). It is likely that
that noise in the data could account for up to 10% of the variance,
and thus we conclude that the feature manifolds are approximately
linear over the locality of a given channel and ictal state.
3.3. Comparison between Local Manifolds
In addition to the earlier comparison of the subspace dimensionality
across channels, the next step is to directly compare the subspaces
selected by the dimensionality reduction. We use a generalization of
distance in the Grassmann space, which allows comparison of affine
subspaces with unequal dimension [14]. The method augments the
principle angles (defined in [22]) with enough additional angles (all
equal to π/2) to increase the number of angles to the dimensionality
of the larger space. An additional "direction vector" is also added to
account for the affine offset.
We apply this generalization [14] to a new modification of the
sin2 θi!
k
Xi=1
1/2
,
(2)
chordal distance, defined as
θC = arcsin
1
k
for k principle angles {θi}k
i=1. The two modifications are 1) dividing
by k, which allows the distance to be independent on the dimension-
ality of the spaces being compared, and 2) converting the distance
measure back to an angle, which is more intuitive.
We then compare the subspaces obtained by PCA dimensional-
ity reduction, for the same divisions of the data as used for Fig. 3.
However, we now ignore any subsets of less than 50 HFOs, as these
are deemed unreliable for computing the PCA in 33D. Note this fig-
ure still involves over 200 thousand time bin comparisons, the same
163 ictal versus interictal comparisons, and nearly 3,500 of each type
of channel-channel comparisons.
Results for these comparisons are shown in Fig. 4. We observe
that the PCA subspaces are quite consistent across different time
bins, with the distributions for PCA subspaces all peaking less than
15◦ and not extending much past 20◦.
s
t
n
u
o
c
d
e
z
i
l
a
m
r
o
N
0.6
0.4
0.2
0
0°
Interical Time Bins (PCA)
Ictal vs Interictal (PCA)
Interictal Channels (PCA)
Ictal Channels (PCA)
Channels (LDA)
15°
30°
45°
60°
θ [degrees]
Fig. 4. Comparison of the subspaces from the PCA and greedy
Fisher LDA dimensionality reduction, using θC (Eq. 2), for the same
types of comparisons as Fig. 3. Again, 0◦ implies maximally similar
and 90◦ implies maximally different.
4. PATHOLOGICAL VERSUS NORMAL HFOS
4.1. Dimensionality Reduction: Greedy Fisher LDA
While the subspaces obtained via PCA represent the variance well,
they are not necessarily the optimal directions for separating the fea-
ture distributions of pHFOs and nHFOs. An alternate dimensional-
ity reduction method is used: greedy Fisher's LDA. In this method,
Fisher's LDA is applied, resulting in a single basis direction. The
projection of the data in this direction is then subtracted from the
data, and the process is repeated.
Recall that Fisher's LDA uses the sum of the covariance of each
group, rather than the covariance of the pooled groups [23]. Letting
the mean and covariance of the two groups be denoted µA, µB and
ΣA, ΣB, respectively, the specific direction is given by
w ∝ (ΣA + ΣB)−1 (µA − µB) .
(3)
Note, the rank of the sum of the covariances will be reduced by one
in each step. Thus, to invert the matrix, the eigenvalue method is
used, with the inverse of the eigenvalues corresponding to the re-
moved projections being set to zero.
We compute this basis, comparing ictal versus interictal times,
for the 13 patients with recorded seizures. We select the number of
basis vectors equal to the mean intrinsic dimension. We again com-
pare the basis vectors using Eq. 2, with the results shown in Fig. 4.
The LDA subspaces vary much more across channels than the PCA
subspaces, with the θC distribution for LDA being almost fully local-
ized between 30◦ and 50◦. Note that 45◦ implies that the subspaces
overlap by half. Thus, the Fisher LDA subspaces show that some
differences in ictal versus interictal HFO features are consistent be-
tween channels, while other differences are not conserved.
0.8
0.6
0.4
0.2
s
t
n
u
o
c
d
e
z
i
l
a
m
r
o
N
0
0
In SOZ (ILAE Class I)
Not in SOZ (ILEA Class I)
Other
0.1
0.4
Upper Bound on Bayes Error
0.2
0.3
0.6
r
o
r
r
E
r
a
e
n
L
i
0.4
0.2
0
0.5
0
Lower Bound on Bayes Error
0.4
0.2
0.6
4.2. Bayes Error Estimates of pHFOs versus nHFOs
Next we quantify how distinct the feature distributions of pHFOs and
nHFOs are, per channel. This quantification serves as a guide for fu-
ture work. We utilize the Henze-Penrose divergence (HPD) [16, 17]
to compute bounds on the Bayes Error. Note that small upper bounds
imply highly separable classes, whereas upper bounds near 0.5 imply
inseparable classes. We estimate the HPD bound using the nonpara-
metric ensemble estimator derived in [24, 25], which achieves the
parametric convergence rate.
The left panel of Fig. 5 shows the Bayes Error bound estimates
for the 163 channels (eight patients) with at least 50 HFOs in each
of the ictal and interictal states. Channels are separated between
SOZ (27 channels) and non-SOZ (124 channels) for patients with
ILEA Class I (the best) surgery outcome, and an "other" category
(12 channels), including channels from patients with either worse
surgery outcomes, no surgery, or missing meta-data.
Patients with Both SOZ and non-SOZ channels in ILAE class
I patients (best surgery outcome) have the Bayes Error rate bound
to relatively small values. However, channels in other patients have
a more diffuse distribution of upper bounds, with the most proba-
ble upper bound about 0.25. It is expected that ictal HFOs are al-
most entirely pHFOs (on any channel), that interictal HFOs on non-
SOZ channels are predominately nHFOs, and that interictal HFOs
on SOZ channels are a mixture of both pHFOs and nHFOs. Thus,
it is expected that the Bayes Error bounds would be lower for non-
SOZ channels. Overall, these bounds suggest that there is sufficient
separation between pHFO and nHFO feature distributions to allow
classification of pHFOs and nHFOs in most channels.
The right panel of Fig. 5 displays the lower bound versus a lin-
ear error estimate. The linear estimate was computed with 10-fold
cross validation in the Fisher LDA space by using a "box" classifica-
tion boundary, with the threshold in each dimension being the value
for which the receiver operator curve has largest transverse distance
from the diagonal. Linear regression was also performed, resulting
in an offset of 0.06 (0.04 -- 0.08 at 95% C.L.) and a slope of 1.05
(0.82 -- 1.28 at 95% C.L.). Thus, in aggregate this "box" classifier
is already relatively close to the bound on the Bayes Error, though
many individual channels are still quite far from the bound.
5. CLUSTERING CHANNELS BASED ON SUBSPACES
Given the observed variations in greedy Fisher LDA subspaces
across channels, we seek to compare the natural clustering of these
subspaces with known groupings of channels. The most relevant
groupings of channels are the groups based on the physical con-
figuration of the recording electrodes (several grids or strips of
electrodes are implanted for each patient), as well as the clinically
determined SOZ and resected volume for ILAE Class I patients.
Fig. 5. Distribution of upper Bayes Error bound based on Henze-
Penrose divergences (HPD) between ictal and interictal HFOs per
channel. Left panel: upper bounds stratified by channel type.
Bounds near zero imply high distinction between pHFOs and
nHFOs, and bounds near 0.5 imply no distinction. Right panel: com-
parison between lower bound estimate and a linear error estimate
(see the text), including the linear regression fit.
The unsupervised clustering is obtained by first converting the
matrix of modified chordal distances (Eq. 2) between channels per
each patient to a metric using the method of dual-rooted-trees fol-
lowed by spectral clustering [26]. This method adapts to the natural
geometry of the data and is competitive with other algorithms [26].
We select either two groups (as the SOZ and resected volume clus-
tering is binary) or the number of groups equal to the number of
strips and grids. The unsupervised clustering results are compared
with these labels using the adjusted Rand index (ARI) [27]. Unfor-
tunately, the requirement that there be at least 50 ictal HFOs reduces
the number of channels per patient significantly, resulting in only a
few (4 -- 5) patients having enough channels to cluster. However, the
ARI never exceeded 0.02 for any of these comparisons in any pa-
tients, suggesting that the primary distinction between channels is
not the pathology or the grid/strip placing, but other effects.
6. DISCUSSION AND CONCLUSION
Overall, we observe that the HFO features tend to cluster on linear
manifolds. Both the subspaces of these manifolds and the local in-
trinsic dimension tend to be consistent within interictal periods, but
may change between interictal and ictal periods. We especially note
significant differences in both the intrinsic dimension and feature
manifolds between different channels within the same patient. Thus,
dimensionality reduction and feature analysis must account for vari-
ations between channels. The dominant cause of this inter-channel
variation does not appear to be tissue pathology or grid/strip groups
in the recording. We also observe that pHFOs and nHFOs are indeed
distinct on a large number of channels, suggesting a strong potential
for classifying individual HFOs.
This analysis also demonstrates methods applicable to other dis-
crete events, including using the θI statistic for comparing local,
intrinsic dimension for collections of events, an affine Grassmann
distance θC for comparing consistency of subspaces, and estimating
bounds on the Bayes Error to assess the feasibility of low-error clas-
sification. Future work will extend this analysis to a larger patient
population, and classify HFOs and/or recording channels based on
HFO features.
[15] J Wang, Y Xu, D Zhang, and J You, "An efficient method for
computing orthogonal discriminant vectors," Neurocomputing,
vol. 73, pp. 2168 -- 76, 2010.
[16] KR Moon, V Delouille, and AO Hero III, "Meta learning of
bounds on the bayes classifier error," in IEEE Signal Process-
ing and Signal Processing Education Workshop. IEEE, 2015.
[17] V Berisha, A Wisler, A Hero III, and A Spanias, "Empirically
estimable classification bounds based on a nonparametric di-
vergence measure," Signal Processing, IEEE Transactions on,
2015.
[18] JB Wagenaar, GA Worrell, Z Ives, D Matthias, B Litt, and
A Schulze-Bonhage,
"Collaborating and sharing data in
epilepsy research," J. Clin Neurophysiol., vol. 32, pp. 235 -- 9,
2015.
[19] JF Kaiser, "On a simple algorithm to calculate the energy of
a signal," in IEEE Int. Conf. Acoustic Speech Signal Process.
IEEE, 1990.
[20] A Ochiai, "Zoogeographical studies on the soleoid fishes found
japan and its neighboring regions. ii," Bull. Jap. Soc. Sci. Fish,
vol. 22, no. 9, pp. 526 -- 530, 1957.
[21] JJ Barkman, "Phytosociology and ecology of cryptogamic epi-
phytes, including a taxonomic survey and description of their
vegetation units in europe," Assen. Van Gorcum, 1958.
[22] H Hotelling,
"Relations between two sets of variates,"
Biometrika, vol. 28, pp. 321 -- 377, 1936.
[23] RA Fisher, "The use of multiple measurements in taxonomical
problems," Annals of Eugenics, vol. 7, no. 2, pp. 179 -- 188,
1936.
[24] KR Moon and AO Hero, "Ensemble estimation of multivari-
ate f-divergence," in Information Theory (ISIT), 2014 IEEE
International Symposium on. IEEE, 2014, pp. 356 -- 360.
[25] K Moon and A Hero, "Multivariate f-divergence estimation
with confidence," in Advances in Neural Information Process-
ing Systems, 2014, pp. 2420 -- 2428.
[26] L Galluccio, O Michel, P Comon, M Kliger, and AO Hero III,
"Clustering with a new distance measure based on a dual-
rooted tree," Information Sciences, vol. 251, pp. 96 -- 113, 2013.
[27] WM Rand, "Objective criteria for the evaluation of clustering
methods," Journal of the American Statistical Association, vol.
66, no. 336, pp. 846 -- 850, 1971.
7. REFERENCES
[1] SC Park, SK Lee, H Che, and Chung CK, "Ictal high-gamma
oscillation (60-99 hz) in intracranial electroencephalography
and postoperative seizure outcome in neocortical epilepsy,"
Clin. Neurophysiol., vol. 123, no. 6, pp. 1100 -- 10, 2012.
[2] JR Cho, DL Koo, EY Joo, DW Seo, SC Hong, P Jiruska, and
SB Hong, "Resection of individually identified high-rate high-
frequency oscillations region is associated with favorable out-
come in neocortical epilepsy," Epilepsia, vol. 55, pp. 1872 -- 83,
2014.
[3] C Haegelen, P Perucca, CE Chtillon, L Andrade-Valena, R Zel-
mann, J Jacobs, DL Collins, F Dubeau, A Olivier, and J Got-
man, "Highfrequency oscillations, extent of surgical resection,
and surgical outcome in drugresistant focal epilepsy," Epilep-
sia, vol. 54, pp. 848 -- 57, 2013.
[4] K Kerber, M Dmpelmann, B Schelter, P Le Van, R Korinthen-
berg, A Schulze-Bonhage, and J Jacobs, "Differentiation of
specific ripple patterns helps to identify epileptogenic areas for
surgical procedures," Clin. Neurophysiol., vol. 125, pp. 1339 --
1345, 2014.
[5] S Gliske, Z Irwin, C Chestek, and WC Stacey, "Automated
identification of seizure onset zone with a universal detector
of high frequency oscillations," Clin. Neurophysiol., vol. IN-
SERT, 2015.
[6] GA Worrell, K Jerbi, K Kobayashi, JM Lina, R Zelmann, and
M Le Van Quyen,
"Recording and analysis techniques for
high-frequency oscillations," Prog. Neurobiol., vol. 98, pp.
265 -- 278, 2012.
[7] JA Blanco, M Stead, A Krieger, J Viventi, WR Marsh, KH Lee,
GA Worrell, and B Litt, "Unsupervised classification of high-
frequency oscillations in human neocortical epilepsy and con-
trol patients," J Neurophysiol, vol. 104, pp. 2900 -- 2912, 2010.
[8] JA Blanco, M Stead, A Krieger, W Stacey, D Maus, E Marsh,
J Viventi, KH Lee, R Marsh, B Litt, and GA Worrell, "Data
mining neocortical high-frequency oscillations in epilepsy and
controls," Brain, vol. 134, pp. 2948 -- 2959, 2011.
[9] A Pearce, D Wulsin, JA Blanco, A Krieger, B Litt, and
WC Stacey, "Temporal changes of neocortical high-frequency
oscillations in epilepsy," J. Neurophysiol, vol. 110, pp. 1167 --
1179, 2013.
[10] A Matsumoto, BH Brinkmann, SM Stead, J Matsumoto,
MT Kucewicz, WR Marsh, F Meyer, and G Worrell, "Patho-
logical and physiological high-frequency oscillations in focal
human epilepsy,"
J. Neurophysiol., vol. 110, pp. 1958 -- 64,
2013.
[11] KM Carter, R Raich, and AO Hero, "On local intrinsic dimen-
sion estimation and its applications," Signal Processing, IEEE
Transactions on, vol. 58, no. 2, pp. 650 -- 663, 2010.
[12] JA Costa and AO Hero III, "Determining intrinsic dimension
and entropy of high-dimensional shape spaces," in Statistics
and Analysis of Shapes, pp. 231 -- 252. Springer, 2006.
[13] KR Moon, JJ Li, V Delouille, R De Visscher, F Watson, and
AO Hero III, "Image patch analysis of sunspots and active
regions. i. intrinsic dimension and correlation analysis," arXiv
preprint arXiv:1503.04127, 2015.
[14] K Ye and L-H Lim, "Distance between subspaces of different
dimensions," Preprint, 2014. arXiv:1407.0900.
|
1709.05650 | 1 | 1709 | 2017-09-17T12:25:33 | Closing the loop between neural network simulators and the OpenAI Gym | [
"q-bio.NC"
] | Since the enormous breakthroughs in machine learning over the last decade, functional neural network models are of growing interest for many researchers in the field of computational neuroscience. One major branch of research is concerned with biologically plausible implementations of reinforcement learning, with a variety of different models developed over the recent years. However, most studies in this area are conducted with custom simulation scripts and manually implemented tasks. This makes it hard for other researchers to reproduce and build upon previous work and nearly impossible to compare the performance of different learning architectures. In this work, we present a novel approach to solve this problem, connecting benchmark tools from the field of machine learning and state-of-the-art neural network simulators from computational neuroscience. This toolchain enables researchers in both fields to make use of well-tested high-performance simulation software supporting biologically plausible neuron, synapse and network models and allows them to evaluate and compare their approach on the basis of standardized environments of varying complexity. We demonstrate the functionality of the toolchain by implementing a neuronal actor-critic architecture for reinforcement learning in the NEST simulator and successfully training it on two different environments from the OpenAI Gym. | q-bio.NC | q-bio |
Closing the loop between neural network simulators and the
Jakob Jordan1,2,3∗†
OpenAI Gym
Philipp Weidel2,1,3∗
Abigail Morrison2,1,3
1 Institute of Neuroscience and Medicine (INM-6), Research Centre Jülich, Jülich, Germany
2 Institute for Advanced Simulation (IAS-6), Research Centre Jülich, Jülich, Germany
3 JARA-Institute Brain Structure Function Relationship (JBI 1 / INM-10), Research Centre Jülich, Jülich, Germany
Abstract
Since the enormous breakthroughs in machine learning over the last decade, functional neural network
models are of growing interest for many researchers in the field of computational neuroscience. One major
branch of research is concerned with biologically plausible implementations of reinforcement learning,
with a variety of different models developed over the recent years. However, most studies in this area are
conducted with custom simulation scripts and manually implemented tasks. This makes it hard for other
researchers to reproduce and build upon previous work and nearly impossible to compare the performance of
different learning architectures. In this work, we present a novel approach to solve this problem, connecting
benchmark tools from the field of machine learning and state-of-the-art neural network simulators from
computational neuroscience. This toolchain enables researchers in both fields to make use of well-tested
high-performance simulation software supporting biologically plausible neuron, synapse and network
models and allows them to evaluate and compare their approach on the basis of standardized environments
of varying complexity. We demonstrate the functionality of the toolchain by implementing a neuronal
actor-critic architecture for reinforcement learning in the NEST simulator and successfully training it on
two different environments from the OpenAI Gym.
1 Introduction
The last decade has witnessed major progress in the field of machine learning, moving from small-scale toy
problems to large-scale real-world applications including image [1] and speech recognition [2], complex
motor-control tasks [3] and playing (video) games at super-human performance [4, 5]. This progress has been
driven mainly by an increase in computing power, especially by training deep networks on graphics processing
units [6], and conceptual breakthroughs like layer-wise pretraining [7, 8] or dropout [9]. Even so, this rate
of progress would not have been possible without the wide availability of high-performance ready-to-use
tools, e.g., Torch [10], Theano [11], Caffe [12], TensorFlow [13] and standardized benchmarks and learning
environments, such as the MNIST [14], CIFAR [15] and ImageNET [16] datasets, and the MuJoCo [17], ALE
[18] and OpenAI Gym [19] toolkits. While ready-to-use tools allow researchers to focus on important aspects
rather than basic implementation details, standardized benchmarks can guide the community as a whole
∗These authors contributed equally to this study.
†Corresponding author: [email protected]
1
towards promising approaches, as for example in the case of convolutional networks through the ImageNET
competition [20].
Similarly, researchers in the field of computational neuroscience have benefited from the increase of
computational power and achieved many conceptual breakthroughs over the last decade, with a plethora
of new neuron, synapse and network models being developed. Thanks to a variety of software projects,
researchers have access to simulators for all scales of neural systems from molecular simulations [21] over
complex neuron [22, 23] and network models [24 -- 26] to whole brain simulations using neural fields [27].
However, in computational neuroscience no generally accepted set of benchmarks exist so far (but see [28]).
While it is desirable to compare different neural network models with respect to biological plausibility,
explanatory power and functional performance, only the latter lends itself to the definition of quantitative
benchmarks.
One particular area in which the lack of standardized benchmarks is apparent is research into reinforcement
learning (RL) in neurobiological substrates. Inspired by behavioural experiments, RL is concerned with
the ability of organisms to learn from previous experiences to optimize their behavior in order to maximize
reward and avoid punishment (see, e.g., [29]). RL has a long tradition in the field of machine learning which
has led to several powerful algorithms, such as SARSA and Q-learning [30]. Similarly, a large variety of
neurobiological models have been proposed in recent years [31 -- 44]. However, only a small proportion of
these rely on publicly available simulators and all of them employ custom built environments. Even for fairly
simple environments, this has led to a situation where different network models are difficult to compare
and reproduce, thus creating a fragmentation of research efforts. Instead of building upon and extending
existing models, researchers are forced to spend too much time on recreating basic methods for custom
implementations.
This issue has led to the Human Brain Project (HBP) [45] dedicating significant resources of a subproject
(SP10, Neurorobotics) to the development of the necessary infrastructure that allows users to conduct robotic
experiments in virtual environments and connect these to their neural network implementations with an
easy to use web interface. This approach however, specifically addresses the need of researchers developing
neuronal controllers for robotic applications.
In contrast, the OpenAI Gym [19] provides a rich and generic collection of standardized RL environ-
ments developed to support the machine learning community in evaluating and comparing algorithms. All
environments are accessible via a simple, unified interface, that requires an agent to supply an action and
returns an observation and reward for its current state. The toolkit includes a range of different environments
with varying levels of complexity ranging from low-dimensional fully discrete (e.g., FrozenLake) to high-
dimensional fully continuous tasks (e.g., Humanoid). Consistency of the OpenAI Gym environments across
different releases supports researchers in reproduction and extension of previous work and allows systematic
benchmarking and comparison of learning algorithms and their implementations. The easy accessibility
of different tasks fosters progress by allowing researchers to focus on learning algorithms instead of basic
implementation details of particular environments and provokes researchers to evaluate the performance of
their algorithms on many different tasks.
To make this comprehensive resource available to the computational neuroscience community, we
developed a toolchain to interface neural network simulators with the OpenAI Gym. Using this toolchain,
researchers can rely on well-tested, high-performance simulation engines to power their models and evaluate
them against a curated set of standardized environments, allowing more time to focus on neurobiological
questions.
In the next section we introduce additional pre-existing components on which our toolchain relies, and
afterwards discuss how it links the different tools. We demonstrate its functionality by implementing a neural
actor-critic in NEST ([24], NEural Simulation Tool) and successfully training it on two different environments
2
Figure 1: Interfacing RL toolkits with neural network simulators. The RL toolkit (left) is responsible for
emulating an environment that provides observations and rewards which are communicated via ZeroMQ
sockets and MUSIC adapters (middle) to a neural network simulator (right). Concurrently, the activity of the
simulated neural network is transformed to an action and fed back to the RL toolkit.
from the OpenAI Gym.
2 Pre-existing components
NEST is a neural simulator designed for the efficient simulation of large-scale networks of simple neuron
models with biophysically realistic connectivity. The simulation kernel scales from small simulations on
a laptops to super computers, with the largest simulation containing about 109 neurons and 1013 synapses,
corresponding to about 10% of the human cortex at the resolution of individual cells and connections [46].
NEST is actively developed and maintained by the NEST initiative in collaboration with the community and
freely available under the GPLv2 and is supported by the HBP with the explicit aim of widespread long-term
availability and maintainability.
While the network implementation that we present in the results section relies on the NEST simulator,
the toolchain can also be used with other simulators that support the MUSIC library, for example NEURON
[22]. The MUlti-SImulation Coordinator is a multi-purpose middleware for neural network simulators built
on top of MPI (Message Passing Interface) that enables online interaction of different simulation engines
[47]. MUSIC provides named MPI channels, referred to as MUSIC ports, which allow the user to set up
communication streams between several processes. While originally intented to distribute a single neural
network model across different simulators, the MUSIC library can also be used to connect neural simulators
with robotic applications.
For this purpose the ROS-MUSIC Toolchain (RMT) [48] was recently developed, providing an interface
from MUSIC to the Robotic Operating System (ROS) [49]. ROS is the most popular middleware in the
robotic community which can interact with many robotic simulators and hardware platforms. The RMT
allows exchange of well-defined messages between ROS and MUSIC via stand-alone executables, so called
adapters, that were designed with a focus on modularity. The toolchain contains several different adapters
each performing a rather simple operation on streams of inputs (e.g., filtering). By concatinating several
adapters, the overall transformation of the original data can become more complex, for example converting
high-dimensional continuous data (e.g., sensory data) to low-dimensional discrete data (e.g., action potentials)
3
ObservationRLToolkitNeural-networksimulatorZMQMUSICActionReward or vice-versa. More information and introductory examples can be found on GitHub.1
3 Results
To enable the online interaction of neural network simulators and the OpenAI Gym we rely on two different
libraries, MUSIC, to interface with the neural simulator, and ZeroMQ [50] to exchange messages with the
environment simulated in the OpenAI Gym. In the following we describe these two parts of the toolchain
and demonstrate their functionality by interfacing a neural network simulation in NEST with two different
environments.
3.1 Extending the ROS - MUSIC toolchain
We extended the RMT by adding adapters that support communication via ZeroMQ following a publish-
subscribe pattern. ZeroMQ is a messaging library that allows applications to exchange messages at runtime
via sockets. Continuously developed by a large community, it offers bindings for a variety of languages
including C++ and Python, and supports most operating systems. A single communication adapter of the RMT
sends (receives) data via a ZeroMQ socket and receives (sends) data via a MUSIC port. While the adapters
can handle arbitrary data, we defined a set of specialized messages in JSON format (see supplementary
material) specifically designed to communicate observations, rewards and actions as discrete or continuous
real-valued variables of arbitrary dimensions, as used in the OpenAI Gym. We chose the JSON format due to
its simplicity, easy serialization and broad platform support.
In addition to the ZeroMQ adapters dedicated for communication with MUSIC, we developed several
further adapters that can perform specific transformations of the data. As discussed above, environments can
be defined in continuous or discrete spaces with arbitrary dimensionality. To generate the required closed-loop
functionality, the observations provided by the environment must be consistently transformed to a format that
can be fed into neural network simulations. Conversely, the activity of the neural network must be interpreted
and transformed into valid actions which can be executed in the environment.
A standard way to address the first issue is to introduce so called place cells. Each of these cells is tuned
to a preferred (multidimensional) observation, i.e., is highly active for a specific input and less active for other
inputs [37]. The dependence of the activity of a single place cell on observations is described by its tuning
curve, often chosen as a multidimensional Gaussian. To perform the transformation of observations to activity
of place cells, we implemented a discretize adapter that allows users to specify the position and width of the
tuning curves of an arbitrary number of place cells. While having a certain biological plausibility [51], one
disadvantage of this approach is that the number of place cells required to cover the whole observation space
evenly scales exponentially in the number of dimensions of the observation. For observations with a small
number of dimensions, however, this approach is very suitable.
To perform action selection, we added several adapters that can, respectively, select the most active
neuron (argmax adapter), threshold the activity across neurons to create a binary vector (threshold adapter) or
linearly combine the activity of neurons across many input channels (linear decoder). Depending on the type
of action required (discrete/continuous) by the environment, the user can select a single one or a combination
of these. See the documentation of the RMT for detailed specifications of the adapters.
In general, we followed the design principle behind the RMT and developed modular adapters. This
makes each individual adapter easy to understand and enables users to quickly extend the toolchain with their
1https://github.com/incf-music/ros-music-adapters
4
own adapters. By combining several adapters, the RMT allows arbitrarily complex transformations of the
data and can hence be applied to many use-cases.
3.2 ZeroMQ wrapper for the OpenAI Gym
The second part of the toolchain is a Python wrapper around the OpenAI Gym that exposes ZeroMQ sockets
for communicating actions, observations and rewards (see figure 1). An environment in the OpenAI Gym is
updated in steps. In each step, an agent needs to provide an action and receives an observation and reward for
its current state. The wrapper consists of four different threads that coordinate: (i) performing steps in an
environment, (ii) receiving actions via a ZeroMQ SUB socket, (iii) publishing observations via a ZeroMQ
PUB socket and (iv) publishing rewards via a ZeroMQ PUB socket. Before spawning the threads, the wrapper
starts a user-specified environment and creates the necessary communication buffers. The thread coordinating
the environment reads actions from the corresponding buffer, performs single steps in the environment and
updates the observation and reward buffers based on the return values of the environment. Upon detecting
that a single episode has ended, e.g., by an agent reaching a certain goal position, it resets the environment
and allows a user-specified break before starting another episode. The communication threads continuously
send(receive) messages via ZeroMQ and read from(write to) the corresponding buffers. All threads can be
run with different update intervals, for example, to slow down movement of the agent by performing steps on
a coarse time grid whilst continuously receiving action choices from the neural network simulation running
on a fine time grid. The user can specify a variety of parameters via a configuration file in JSON format (see
supplementary material). See the documentation for detailed specifications of the wrapper.
3.3 Applications
To demonstrate the functionality of the toolchain, we implemented a neural network model with actor-critic
architecture in NEST and trained it on two different environments simulated in the OpenAI Gym. In the first
task, the agent needs to learn to perform a sequence of actions in order to reach the top of a hill in a continous
environment. The second task is a classical grid-world in which an agent needs to learn to navigate to a
goal position in a two-dimensional discrete environment with obstacles. We first describe the neural network
architecture and learning rule and afterwards discuss the network's performance on the two tasks.
3.3.1 Neural network implementation
We consider a temporal-difference learning algorithm [29] implemented as an actor-critic architecture,
originally using populations of spiking neurons [37]. We translated the spike-based implementation to rate
neurons, mainly to simplify the implementation by avoiding issues arising from noise introduced by spiking
neuron models [52, 37]. The neuron dynamics we considered here are determined by the following stochastic
differential equation:
dzi(t)
τ
dt
= −zi(t) + µi + f (hi(t) − θi) + ξi(t) ,
σξ. The input field hi(t) is determined by the activity of other neurons according to hi(t) =(cid:80)
(1)
where τ is some positive time constant, µi a baseline activity level, f (·) some (arbitrary) activation function,
hi(t) a time dependent input field, θi an input threshold and ξi(t) white noise with a certain standard deviation
j wijzj(t),
with wij denoting the strength of the connection (weight) from neuron j to neuron i. Here we will exclusively
consider activation functions of the form f (x) = x (linear case), and f (x) = Θ(x)x (threshold-linear case,
5
Figure 2: Actor-critic architecture for RL with rate neurons. Observations are communicated via a
MUSIC input port to a population of place cells. These project on the one hand to a critic unit and on the
other hand to actor units arranged in a winner-take-all circuit. The critic and an additional MUSIC input port
project to a unit representing the reward-prediction error that modulates the plasticity between place cells and
critic and actors, respectively. The actor units project to a MUSIC output port encoding the selected action.
"relu"). These models have recently been added to the NEST simulator. Their dynamics are solved on a
fixed time-grid by a stochastic-exponential-Euler method with a step size determined by the resolution of the
simulation. For more details on the model implementation see [53].
The input layer is a population of threshold-linear rate neurons which receive inputs through MUSIC
and encode observations from the environment (see figure 2). Via plastic connections these place cells
project to a single neuron representing the value that the network assigns to the current state (the "critic").
An additional neuron calculates the reward-prediction error by combining the reward received from the
environment with input from the critic. Plasticity of the projections from inputs to the critic is modulated by
this reward-prediction error (see below).
In addition, neurons in the input layer project to a population of neurons representing the available actions
(the "actor"). To enforce selection of a specific action, the actor units are arranged in a winner-take-all (WTA)
circuit. This is implemented by recurrent connections between actor units that correspond to short-range
excitation and long-range inhibition, depending on the similarity of the action that actor units encode. The
activity of actor units is transformed to a valid action and communicated to the environment via the RMT.
To derive a learning rule for the critic, we followed similar steps as in [37] applied to rate models (equation
(cid:90) ∞
t(cid:48)
(1)). The critic activity should approximate a continous-time value function defined by [54]
V π(s(t)) :=
r(sπ(t(cid:48)))e
− t(cid:48)−t
τr dt(cid:48) .
(2)
Here s(t) denotes the state of the agent at time t, r(sπ(t)) denotes the reward obtained in state s(t), τr a
discounting factor for future rewards and π the agent's policy. To achieve this, we define the following
objective function which should be minimized by gradient descent on the weights from inputs to the critic:
E(t) =
1
2
(V π(t) − z(t))2 ,
6
(3)
ObservationRewardPlace cellsActorMMCriticPrediction errordActionMMUSIC portthreshold-linear unitlinear unitexcitatory connectioninhibitory connectionplastic connectionMUSIC communicationMneuromodulation where z(t) represents the activity of the critic unit. By performing gradient descent on equation (3), using
a self-consistency equation for V π(t) from the derivative of equation (2) and bootstrapping on the current
prediction for the value (see supplementary material and [54, 37]), we obtain the following local Hebbian
three-factor learning rule that approximately minimizes the objective function (equation (3)):
∆wj = ηδ(t)xj(t)Θ (z(t) − θpost) ,
(4)
where η is a learning rate, xj(t) represents the activity of the jth place cell, Θ(·) the Heaviside function and
θpost a parameter that accounts for noise on the postsynaptic unit (see supplementary material for details). The
term δ(t) = v(t) + r(t) − 1
v(t) corresponds to the activity of the reward-prediction error unit, acting as a
neuromodulatory signal for the Hebbian plasticity between the presynaptic (xj) and postsynaptic (z) units.
To avoid explicit calculation of the derivative, we rewrite δ(t) as:
τr
(cid:18) 1
d
δ(t) ≈
(cid:19)
− 1
τr
v(t) − 1
d
v(t − d) + r(t) .
(5)
To approximate the derivative we hence implement two connections from the critic to the reward-prediction
error unit: one instantaneous, and one with delay d > 0.
To learn an optimal policy, we exploit that the actor units follow the same dynamics as the critic. Similar
to [37], we hence apply the same learning rule to the connections between the inputs and the actor units. In
order to assure that at least one actor unit is active, thus preventing a deadlock, we introduce a minimal weight
for each connection between input and output units and add input noise to the actor units.
3.3.2 Mountain Car
As an example of an environment with continous states, we consider the MountainCar environment. The task
is to steer a toy vehicle that starts at a valley between two hills to the top of the right one (figure 3A, inset).
To make the task more challenging, the car's engine is not strong enough to reach the top in one go, so the
agent needs to learn to gain momentum by swinging back and forth between the two hills. A single episode
in this environment starts when the agent is placed in the valley and ends when it reaches the final position on
the top of the right hill. The state of the agent is described by two continous variables: the x-position x(t)
and the x-velocity x(t). The agent can choose from three different discrete actions that affect the velocity of
the vehicle (accelerate left, no acceleration, accelerate right). It receives punishment from the environment
in every step; the goal is to minimize the total punishment collected over the whole episode. Since this is
difficult to implement in an actor-critic architecture [52], we provide additional reward when the agent reaches
the final position.
To translate the agent's current state into neuronal activity, we distribute 25 place cells evenly across the
two dimensional plane of possible positions and velocities using the discretize adapter of the RMT. The actor
is implemented by a WTA circuit of three units as described in section 3.3.1. The activity of these units is
transformed into an action via the argmax adapter (section 3.1).
Initially, the agent explores the environment by selecting random actions. Due to the WTA circuit
dynamics, a single action stays active over an extended period of time. The constant punishment gradually
decreases the weights from the place cells to the corresponding actor unit, eventually leading to another actor
unit becoming active (figure 3B, left). After a while, the agent reaches the goal by performing actions that
have not been significantly punished. For this task the stable nature of the WTA is advantageous, causing
the agent to perform the same action repeatedly allowing efficient exploration of the state space. After the
agent has found the goal once, the number of steps spent on exploring actions in the following episodes is
much smaller. From the sixth episode on, the performance of the agent is already close to optimal (figure 3A).
7
Figure 3: Network performance on an environment with continuous states and discrete actions. A:
Reward obtained by the agent per episode averaged over 10 simulations with different seeds (solid orange
line). Orange band indicates ± one standard deviation. Light gray line marks average reward per episode
for which the environment is considered solved. Inset: screenshot of the environment with agent (stylized
vehicle), environment with valley and two hills (black solid line) and goal position (yellow flag). The agent
is close to the starting position at the trough. B: Activity traces of place cells (bottom), actor units (second
from bottom), critic unit (second from top) and reward-prediction-error unit (top). Shown are neural activities
during 6.5 s early (left) and late (right) in the simulation.
After learning for about 10 episodes, the agent's performance has converged. The value of the final state is
successfully propagated backwards over different states, leading to a ramping of activity of the critic unit
from the start of an episode to the end (figure 3B, right).
Since the OpenAI Gym offers a variety of environments, we trained the same network model on an
additional task with different requirements.
3.3.3 Frozen Lake
As a second application, we chose the FrozenLake environment consisting of a discrete set of 16 states
arranged in a four-by-four grid (figure 4A, inset). Each state is either a start state (S), a goal state (G), a hole
(H) or a frozen state (F). From the start position, the agent has to reach the rewarded state by navigating
over the frozen states without falling into holes which would reset the agent to the starting position. In each
step, the agent can choose from four different actions: move west, move north, move east and move south.
Usually, the tiles are "slippery", i.e., there is a chance that a random action is executed irrespective of the
action chosen by the agent. However, to simplify learning for demonstration purposes, we turned this feature
off. Upon reaching the goal, the agent receives a reward of magnitude one. Since the optimal path involves
six steps from start to goal, the theoretical optimal reward per step is ∼ 0.16. To encourage exploration, the
agent receives a small punishment in each state, and additionally, to speed up learning, the agent is punished
for falling into holes.
Unlike in the continuous MountainCar environment, the tuning curves of place cells do not overlap in the
discrete case, leading to sharp transitions in the network activity. This leads to severe issues for associating
values and actions with the respective states. To address this problem we introduced a simple eligibility trace
by evaluating the activity of the pre- and post synaptic units in the learning rule with a small delay δt (see
supplementary material). With this addition, the network model is able to find the optimal solution for this
8
AB Figure 4: Network performance on a grid-world environment. A: Average reward collected by the agent
over the next 500 steps (orange solid line) averaged over 5 simulations. Orange band indicates ± one standard
deviation. Gray line: theoretical optimum. Inset: screenshot of the environment with start state (S), frozen
states (F), holes (H) and goal state (G). The position of the agent is indicated in pink. B: The learned policy
and value map of the environment. Red colors indicate positive, blue colors negative values. Arrows indicate
the preferred direction of movement.
task within roughly 2000 steps (figure 4A). It also learns to associate holes with punishment and frozen states
with reward if they are on the path to the goal (figure 4B). Although there are two possible paths to the goal,
the agent prefers the path with less corners, since it is easier to navigate using a WTA circuit.
4 Conclusion
In this manuscript, we have presented a toolchain that closes the loop between the OpenAI Gym and neural
network simulators. We demonstrated the functionality of the toolchain by implementing an actor-critic
architecture in NEST and evaluating its performance on two different environments. The performance of the
network quickly reached near-optimal performance on these two tasks.
Combining neural network simulators with RL toolkits responds to the growing need of researchers to
provide neural network models with rich, dynamic input. Compared to creating customized environments to
this end, using readily available tools is easier, often computationally more efficient, and most importantly,
supports reproducible science. In addition, having the OpenAI Gym environments as common benchmarks in
both fields encourages comparison between traditional machine learning and biologically plausible implemen-
tations. In contrast to models presented in previous studies, our toolchain makes it easy for other researchers
to extend our implementation of an actor-critic architecture to other environments, replace neurons models or
explore alternative learning rules.
While the toolchain currently only supports the OpenAI Gym, the extension to other toolkits is simple
due to a modular design of the wrapper. The RMT can be found on GitHub and is available under the GPLv3.
The OpenAI Gym ZeroMQ wrapper is also available via GitHub under the MIT license. A complementary
development to the work presented here is provided by SPORE, a framework for reward-based learning with
spiking neurons in the NEST simulator.2 It provides support for synapse models with time-driven updates,
2https://github.com/IGITUGraz/spore-nest-module
9
AB additional support for recording and evaluating traces of neuronal state variables and introduces MUSIC ports
for communicating rewards to a running simulation.
With the work presented here we enable researchers to build more easily upon previous studies and
evaluate novel models. We hope this boosts the progress in computational neuroscience in uncovering the
biophysical mechanisms involved in learning complex tasks from delayed rewards.
Acknowledgments
We acknowledge partial support by the German Federal Ministry of Education through our German-Japanese
Computational Neuroscience Project (BMBF Grant 01GQ1343), EuroSPIN, the Helmholtz Alliance through
the Initiative and Networking Fund of the Helmholtz Association and the Helmholtz Portfolio theme "Super-
computing and Modeling for the Human Brain" and the European Union Seventh Framework Programme
(FP7/2007-2013) under grant agreement no. 604102 (HBP). All network simulations carried out with NEST
(http://www.nest-simulator.org).
References
[1] Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural
networks. In Advances in neural information processing systems, pp. 1097 -- 1105.
[2] Hinton, G., Deng, L., Yu, D., Dahl, G. E., Mohamed, A.-r., Jaitly, N., Senior, A., Vanhoucke, V., Nguyen, P., Sainath,
T. N., et al. (2012a). Deep neural networks for acoustic modeling in speech recognition: The shared views of four
research groups. IEEE Signal Processing Magazine 29(6), 82 -- 97.
[3] Mnih, V., Badia, A. P., Mirza, M., Graves, A., Lillicrap, T., Harley, T., Silver, D., & Kavukcuoglu, K. (2016).
Asynchronous methods for deep reinforcement learning. In International Conference on Machine Learning, pp.
1928 -- 1937.
[4] Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., Graves, A., Riedmiller, M., Fidje-
land, A. K., Ostrovski, G., et al. (2015). Human-level control through deep reinforcement learning. Nature 518(7540),
529 -- 533.
[5] Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I.,
Panneershelvam, V., Lanctot, M., et al. (2016). Mastering the game of go with deep neural networks and tree search.
Nature 529(7587), 484 -- 489.
[6] Raina, R., Madhavan, A., & Ng, A. Y. (2009). Large-scale deep unsupervised learning using graphics processors. In
Proceedings of the 26th annual international conference on machine learning, pp. 873 -- 880. ACM.
[7] Hinton, G. E., & Salakhutdinov, R. R. (2006). Reducing the dimensionality of data with neural networks. Sci-
ence 313(5786), 504 -- 507.
[8] Bengio, Y., Lamblin, P., Popovici, D., Larochelle, H., et al. (2007). Greedy layer-wise training of deep networks.
Advances in neural information processing systems 19, 153.
[9] Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I., & Salakhutdinov, R. R. (2012b). Improving neural
networks by preventing co-adaptation of feature detectors. arXiv preprint arXiv:1207.0580.
[10] Collobert, R., Bengio, S., & Mariéthoz, J. (2002). Torch: a modular machine learning software library. Technical
report, Idiap.
10
[11] James, B., Olivier, B., Frédéric, B., Pascal, L., & Razvan, P. (2010). Theano: a cpu and gpu math expression
compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy).
[12] Jia, Y., Shelhamer, E., Donahue, J., Karayev, S., Long, J., Girshick, R., Guadarrama, S., & Darrell, T. (2014). Caffe:
Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093.
[13] Abadi, M., Agarwal, A., Barham, P., Brevdo, E., Chen, Z., Citro, C., Corrado, G. S., Davis, A., Dean, J., Devin,
M., et al. (2016). Tensorflow: Large-scale machine learning on heterogeneous distributed systems. arXiv preprint
arXiv:1603.04467.
[14] LeCun, Y., Cortes, C., & Burges, C. J. (1998). The mnist database of handwritten digits.
[15] Krizhevsky, A., & Hinton, G. (2009). Learning multiple layers of features from tiny images.
[16] Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). ImageNet: A Large-Scale Hierarchical Image
Database. In CVPR09.
[17] Todorov, E., Erez, T., & Tassa, Y. (2012). Mujoco: A physics engine for model-based control. In Intelligent Robots
and Systems (IROS), 2012 IEEE/RSJ International Conference on, pp. 5026 -- 5033. IEEE.
[18] Bellemare, M., Naddaf, Y., Veness, J., & Bowling, M. (2015). The arcade learning environment: An evaluation
platform for general agents. In Twenty-Fourth International Joint Conference on Artificial Intelligence.
[19] Brockman, G., Cheung, V., Pettersson, L., Schneider, J., Schulman, J., Tang, J., & Zaremba, W. (2016). OpenAI
Gym. ArXiv e-prints.
[20] Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein,
M., Berg, A. C., & Fei-Fei, L. (2015). ImageNet Large Scale Visual Recognition Challenge. International Journal of
Computer Vision (IJCV) 115(3), 211 -- 252.
[21] Wils, S., & De Schutter, E. (2009). STEPS: modeling and simulating complex reaction-diffusion systems with
Python. Front. Neuroinformatics 3(15).
[22] Carnevale, T., & Hines, M. (2006). The NEURON Book. Cambridge: Cambridge University Press.
[23] Bower, J. M., & Beeman, D. (2007). GENESIS (simulation environment). Scholarpedia 2(3), 1383.
[24] Gewaltig, M.-O., & Diesmann, M. (2007). NEST (NEural Simulation Tool). Scholarpedia 2(4), 1430.
[25] Goodman, D. F. M., & Brette, R. (2009). The Brian simulator. Front. Neurosci. 3(2), 192 -- 197.
doi:
10.3389/neuro.01.026.2009.
[26] Bekolay, T., Bergstra, J., Hunsberger, E., DeWolf, T., Stewart, T. C., Rasmussen, D., Choo, X., Voelker, A. R.,
& Eliasmith, C. (2013). Nengo: a python tool for building large-scale functional brain models. Frontiers in
neuroinformatics 7.
[27] Sanz Leon, P., Knock, S., Woodman, M., Domide, L., Mersmann, J., McIntosh, A., & Jirsa, V. (2013). The virtual
brain: a simulator of primate brain network dynamics. Front. Neuroinform. 7, 10.
[28] Gerstner, W., & Naud, R. (2009). How good are neuron models? Science 326(5951), 379 -- 380.
[29] Sutton, R. S., & Barto, A. G. (1998). Reinforcement Learning: An Introduction. Adaptive Computation and Machine
Learning. The MIT Press.
[30] Watkins, C. J. C. H. (1989). Learning from delayed rewards. Ph. D. thesis, University of Cambridge England.
11
[31] Vasilaki, E., Frémaux, N., Urbanczik, R., Senn, W., & Gerstner, W. (2009). Spike-based reinforcement learning
in continuous state and action space: When policy gradient methods fail. PLoS Comput. Biol. 5(12), e1000586.
doi:10.1371/journal.pcbi.1000586.
[32] Izhikevich, E. M. (2007). Solving the distal reward problem through linkage of STDP and dopamine signaling.
Cereb. Cortex 17(10), 2443 -- 2452.
[33] Urbanczik, R., & Senn, W. (2009). Reinforcement learning in populations of spiking neurons. Nature neuro-
science 12(3), 250.
[34] Potjans, W., Morrison, A., & Diesmann, M. (2009). A spiking neural network model of an actor-critic learning
agent. Neural Comput. 21, 301 -- 339.
[35] Frémaux, N., Sprekeler, H., & Gerstner, W. (2010). Functional requirements for reward-modulated spike-timing-
dependent plasticity. J. Neurosci. 30(40), 13326 -- 13337.
[36] Jitsev, J., Morrison, A., & Tittgemeyer, M. (2012). Learning from positive and negative rewards in a spiking neural
network model of basal ganglia. In Neural Networks (IJCNN), The 2012 International Joint Conference on, pp. 1 -- 8.
IEEE.
[37] Frémaux, N., Sprekeler, H., & Gerstner, W. (2013). Reinforcement learning using a continuous time actor-critic
framework with spiking neurons. PLoS Comput Biol 9(4), e1003024.
[38] Rasmussen, D., & Eliasmith, C. (2014). A neural model of hierarchical reinforcement learning. In CogSci.
[39] Friedrich, J., Urbanczik, R., & Senn, W. (2014). Code-specific learning rules improve action selection by populations
of spiking neurons. International Journal of Neural Systems 24(05), 1450002.
[40] Rombouts, J. O., Bohte, S. M., & Roelfsema, P. R. (2015). How attention can create synaptic tags for the learning
of working memories in sequential tasks. PLoS Comput Biol 11(3), e1004060.
[41] Baladron, J., & Hamker, F. H. (2015). A spiking neural network based on the basal ganglia functional anatomy.
Neural Networks 67, 1 -- 13.
[42] Aswolinskiy, W., & Pipa, G. (2015). Rm-sorn: a reward-modulated self-organizing recurrent neural network.
Frontiers in computational neuroscience 9, 36.
[43] Friedrich, J., & Lengyel, M. (2016). Goal-directed decision making with spiking neurons. Journal of Neuro-
science 36(5), 1529 -- 1546.
[44] Rueckert, E., Kappel, D., Tanneberg, D., Pecevski, D., & Peters, J. (2016). Recurrent spiking networks solve
planning tasks. Scientific reports 6.
[45] Human Brain Project (2014). Project website. Available at: http://www.humanbrainproject.eu.
[46] Kunkel, S., Schmidt, M., Eppler, J. M., Masumoto, G., Igarashi, J., Ishii, S., Fukai, T., Morrison, A., Diesmann, M.,
& Helias, M. (2014). Spiking network simulation code for petascale computers. Front. Neuroinformatics 8, 78.
[47] Djurfeldt, M., Hjorth, J., Eppler, J. M., Dudani, N., Helias, M., Potjans, T. C., Bhalla, U. S., Diesmann, M., Kotaleski,
J. H., & Ekeberg, O. (2010). Run-time interoperability between neuronal simulators based on the music framework.
Neuroinformatics 8. doi:10.1007/s12021-010-9064-z.
[48] Weidel, P., Djurfeldt, M., Duarte, R. C., & Morrison, A. (2016). Closed loop interactions between spiking neural
network and robotic simulators based on music and ros. Frontiers in neuroinformatics 10.
12
[49] Quigley, M., Conley, K., Gerkey, B., Faust, J., Foote, T., Leibs, J., Wheeler, R., & Ng, A. Y. (2009). Ros: an
open-source robot operating system. In ICRA workshop on open source software, Volume 3, pp. 5. Kobe.
[50] Hintjens, P. (2013). ZeroMQ: Messaging for Many Applications. "O'Reilly Media, Inc.".
[51] Moser, E. I., Kropff, E., & Moser, M.-B. (2008). Place cells, grid cells, and the brain's spatial representation system.
Annu. Rev. Neurosci. 31, 69 -- 89.
[52] Potjans, W., Diesmann, M., & Morrison, A. (2011). An imperfect dopaminergic error signal can drive temporal-
difference learning. PLoS Comput. Biol. 7(5), e1001133.
[53] Hahne, J., Dahmen, D., Schuecker, J., Frommer, A., Bolten, M., Helias, M., & Diesmann, M. (2016). Integration of
continuous-time dynamics in a spiking neural network simulator. arXiv. 1610.09990 [q -- bio.NC].
[54] Doya, K. (2000). Reinforcement learning in continuous time and space. Neural Comput. 12(1), 219 -- 245.
[55] Nordlie, E., Gewaltig, M.-O., & Plesser, H. E. (2009). Towards reproducible descriptions of neuronal network
models. PLoS Comput. Biol. 5(8), e1000456.
Supplementary material
A Derivation of the learning rule
(cid:90) ∞
We consider continuous time, continuous states and continuous actions and follow similar steps as [54, 37]. Starting from
the continuous-time value function
V π(s(t)) :=
(cid:48)
r(sπ(t
))e
− t(cid:48)−t
(cid:48)
τr dt
,
(6)
we take the derivative with respect to t to arrive at a self-consistency equation for V :
t
V π(t) :=
dV π(t)
dt
= −r(t) +
1
τr
V π(t) .
(7)
To implement temporal-difference learning in a neural network architecture, we would like to approximately represent
the true value V π(t) of the state at time t by the rate zi(t) of a critic neuron. This activity will depend on the activity of
input units, i.e., place cells, and the weights between inputs and critic. With initially random weights the self-consistency
criterion will not be fulfilled, and will have a finite error δ(t):
δ(t) = zi(t) + r(t) − 1
τr
zi(t) .
We now define an objective function that should be minimized by gradient descent on the weights:
We take derivative with respect to wij and use the self-consistency equation (7):
E(t) =
1
2
(V π(t) − zi(t))2 .
∂E(t)
∂wij
= −(V π(t) − zi(t))
∂zi(t)
∂wij
= −(τr V π(t) + τrr(t) − zi(t))
≈ − (τr z(t) + τrr(t) − zi(t))
∂zi(t)
∂wij
∂zi(t)
∂wij
(cid:125)
(cid:124)
(cid:123)(cid:122)
τr δ(t)
13
(8)
(9)
(10)
Here we have replaced V π(t) with zi(t) in the last line. For a discussion of the validity of this approximation and the
convergence of the learning rule, see [37]. To perform gradient descent on the objective function, we hence need to
change the weights according to
∆wij := − η
(cid:48) ∂E(t)
∂wij
∂zi(t)
∂wij
=ηδ(t)
,
(11)
where we introduced η = η(cid:48)τr. We hence need to determine the derivative of the critic activity with respect to the weights
between inputs and critic ∂zi(t)
∂wij
We start from the differential equation describing the dynamics of a threshold-linear rate neuron, and assume the
noise to be small, i.e., we drop the term ξi(t). Without loss of generality, we assume µi = 0, θi = 0. The dynamics are
then given by the solution to
.
τ
dzi(t)
dt
= −zi(t) + g Θ
wijxj(t)
wijxj(t)
.
(cid:32)(cid:88)
(cid:32)(cid:88)
j
j
(cid:33)(cid:32)(cid:88)
(cid:33)(cid:33)
j
(cid:33)
(cid:33)
(t) ,
Variation of constants yields the general solution as a convolution equation:
zi(t) = g
Θ (. . . )
wijxj(·)
∗ κ(·)
(cid:32)(cid:32)
(cid:40)
(12)
(13)
(14)
(15)
else
if(cid:80)
if(cid:80)
else
derivative at(cid:80)
where we have introduced the filter kernel κ(t) = 1
τ Θ(t). We now take the derivative with respect to wij. While the
j wijxj(t) = 0 is technically not defined, we follow the usual convention and set it to zero. This yields
τ e− t
(cid:40)
∂zi(t)
∂wij
=
g (xj ∗ κ) (t)
0
j wijxj(t) > 0
By combining equation (11) with equation (14), we obtain the following learning rule:
∆wij =
ηδ(t)g(xj ∗ κ)(t)
0
j wijxj(t) > 0
postsynaptic neuron by observing that zi(t) > 0 iff(cid:80)
By choosing the time constant of critic and actors small, we effectively remove the filtering of the presynaptic activity
(limτ→0 κ(t) = δ(t)) and hence ignore it. To simplify this equation further, we rewrite it as a condition on the rate of the
j wijxj(t) > 0. To implement exploration for similar inputs to all
output units we add noise to the activity of the actor units. We only consider a postsynaptic neuron active, if its activity is
larger than some threshold θpost. This leads to the following form for the learning rule:
∆wij = ηδ(t)gxj(t)Θ(zi(t) − θpost) ,
(16)
where we Θ(·) denotes the Heaviside step function defined as:
(cid:40)
Θ(x) =
1 x > 0
0 else
To implement a simple type of eligibility trace, we introduce an additional parameter δt that can delay the activity of
the pre- and post-synaptic units in the learning rule:
∆wij = ηδ(t)gxj(t − δt)Θ(zi(t − δt) − θpost) ,
(17)
14
A
Populations
Topology
Connectivity
Neuron model
Channel models
Synapse model
Plasticity
External input
External output
Measurements
B
Name
Observation
Reward
Action
Place cells
Critic
Actor
Prediction error
C
Source
Observation
Reward
Actor
Place cells
Place cells
Critic
Critic
Actor
D
Type
Dynamics
Type
Dynamics
Type
Plasticity
E
Type
Observation
Reward
F
Type
Action
Model summary
Seven
None
Population specific
Linear & threshold-linear rate
None
Instantaneous & delayed continuous coupling
Three-factor Hebbian
Continuous MUSIC ports
Continuous MUSIC ports
Rates of all neurons
Elements
MUSIC in port
MUSIC in port
MUSIC out port
Threshold-linear
Threshold-linear
Threshold-linear
Linear
Target
Place cells
Prediction error
Action
Critic
Actor
Prediction error
Prediction error
Actor
Populations
Size
1
1
1
16(25)
1
4(3)
1
Connectivity
Pattern
One-to-one (by MUSIC channel), instantaneous, static, weight wo
One-to-one (by MUSIC channel), instantaneous, static, weight wr
One-to-one (by MUSIC channel), instantaneous, static, weight wa
All-to-all, instantaneous, plastic, initial weight wpc
All-to-all, instantaneous, plastic, initial weight wpa
One-to-one, instantaneous, static, weight 1/d − 1/τr
One-to-one, delay d, static, weight −1/d
All-to-all, instantaneous, static, weight α exp(−∆a/σ) + β
Neuron and synapse model
dt = −z(t) + µ + (h(t) − θ) + ξ(t)
dt = −z(t) + µ + Θ (h(t) − θ) (h(t) − θ) + ξ(t)
Linear rate neuron
τ dz(t)
Threshold-linear rate neuron
τ dz(t)
Three-factor Hebbian synapse
∆wij = ηδ(t)gxj(t − δt)Θ(zi(t − δt) − θpost)
Input
Description
Rate r ∈ [−1, 1] according to tuning of place cell (using discretize adapter)
Rate r ∈ [−1, 1] according to reward provided by the environment
Description
Rates ri ∈ [0,∞) according to activities of the actor units
Output
Table 1: Description of the network model (according to [55]).
15
B
Name
τ
g
µ
σξ
θ
B
Name
τ
g
µ
σξ
θ
B
Name
τ
g
µ
σξ
θ
B
Name
τ
g
µ
σξ
θ
C
Name
wo
wr
wa
wpc
wmin
pc
wmax
pc
θpc
post
wpa
wmin
pa
wmax
pa
θpa
post
d
τr
α
β
σξ
ηcritic
ηactor
δt
E
Name
σx
σy
Populations: place cells
Populations: critic
Populations: reward
Populations: actor
Connectivity
16
Input: discretize adapter
Values
5.0 (1.0)
1.0
0.0
0.0
-0.5
Values
0.1
1.0
-1.0
0.0
-1.0
Values
1.0
1.0
0.0
0.0
0.001 (-0.0999)
Values
0.1
1.0
0.0
0.2 (0.05)
0.0
Values
0.5
0.1
1.0
0.0
−1.0
1.0
−1.0
0.9(0.3)
0.1(0.05)
1.0
0.5(0.1)
1.0
20000.0
1.2
−0.55
0.1
0.01(0.125)
0.2(0.250)
19.0(0.0)
Values
0.01 (0.2)
-- (0.2)
Table 2: Table of the network parameters used for both tasks (according to [55]). Values in brackets are used
for the MountainCar environment.
B Network description
The tables 1 and 2 summarize the network architecture and parameters.
C JSON Message types
Listing 1 show the standard message types used for communication between the OpenAI Gym and the RMT. All messages
are serialized using JSON and communicated via ZeroMQ.
Listing 1: Message types used for communication.
BasicMsg
{
f l o a t min
f l o a t max
f l o a t v a l u e
f l o a t
t i m e s t a m p
}
ObservationMsg
{
BasicMsg [ ] o b s e r v a t i o n s
# one b a s i c msg p e r d i m e n s i o n
}
RewardMsg
{
BasicMsg [ ]
reward
# reward i s
always one d i m e n s i o n a l
}
ActionMsg
{
BasicMsg [ ]
a c t i o n s
}
f or d i s c r e t e
# one d i m e n s i o n a l
# o r one d i m e n s i o n p e r p o s s i b l e
a c t i o n s
a c t i o n
D Example wrapper configuration file
Listing 2 shows an example configuration file for running the mountain car environment.
Listing 2: Example configuration file for the wrapper to run the "MountainCar-v0" environment.
" A l l " :
{
" s e e d " : 12345 ,
" t i m e _ s t a m p _ t o l e r a n c e " : 0 . 0 1 ,
" p r e f i x " : null ,
" w r i t e _ r e p o r t " :
" r e p o r t _ f i l e " :
" o v e r w r i t e _ f i l e s " :
" f l u s h _ r e p o r t _ i n t e r v a l " : n u l l
true ,
" . / r e p o r t . j s o n " ,
f a l s e ,
17
} ,
" Env " :
{
" MountainCar−v0 " ,
" env " :
" i n i t i a l _ r e w a r d " : null ,
" f i n a l _ r e w a r d " : null ,
" min_reward " : −1.0 ,
" max_reward " : 1 . 0 ,
" r e n d e r " :
" m o n i t o r " :
" m o n i t o r _ d i r " :
" m o n i t o r _ a r g s " :
{
true ,
f a l s e ,
" w r i t e _ u p o n _ r e s e t " :
" v i d e o _ c a l l a b l e " :
f a l s e
" . / e x p e r i m e n t −0/ " ,
true ,
}
} ,
" EnvRunner " :
{
" u p d a t e _ i n t e r v a l " : 0 . 0 1 ,
" i n t e r _ t r i a l _ d u r a t i o n " : 0 . 4
} ,
" CommandReceiver " :
{
" s o c k e t " : 5555 ,
" t i m e _ s t a m p _ t o l e r a n c e " : 0 . 0 1
} ,
" O b s e r v a t i o n S e n d e r " :
{
" s o c k e t " : 5556 ,
" u p d a t e _ i n t e r v a l " : 0 . 0 1
} ,
" RewardSender " :
{
" s o c k e t " : 5557 ,
" u p d a t e _ i n t e r v a l " : 0 . 0 1
}
E Example MUSIC configuration file
Listing 3 shows an example MUSIC configuration file to run the MountainCar environment. It shows the different
processes with parameters which are spawned by MUSIC including RMT adapters and NEST.
Listing 3: Example MUSIC configuration file to run the MountainCar environment.
s t o p t i m e =150.
r t f = 1 .
[ reward ]
b i n a r y = z m q _ i n _ a d a p t e r
a r g s =
18
np=1
m u s i c _ t i m e s t e p =0.001
m e s s a g e _ t y p e =GymObservation
zmq_topic =
zmq_addr= t c p : / / l o c a l h o s t :5557
[ s e n s o r ]
b i n a r y = z m q _ i n _ a d a p t e r
a r g s =
np=1
m u s i c _ t i m e s t e p =0.001
m e s s a g e _ t y p e =GymObservation
zmq_topic =
zmq_addr= t c p : / / l o c a l h o s t :5556
[ d i s c r e t i z e ]
b i n a r y = d i s c r e t i z e _ a d a p t e r
a r g s =
np=1
m u s i c _ t i m e s t e p =0.001
g r i d _ p o s i t i o n s _ f i l e n a m e = g r i d _ p o s . j s o n
[ n e s t ]
b i n a r y = . . / a c t o r _ c r i t i c _ n e t w o r k / network . py
a r g s=−t 1 5 0 . −n 25 −m 3 −p network_params . j s o n
np=1
[ argmax ]
b i n a r y = a r g m a x _ a d a p t e r
a r g s =
np=1
m u s i c _ t i m e s t e p =0.001
[ command ]
b i n a r y = z m q _ o u t _ a d a p t e r
a r g s =
np=1
m u s i c _ t i m e s t e p =0.01
m e s s a g e _ t y p e =GymCommand
zmq_topic =
zmq_addr= t c p : / / ∗:5555
s e n s o r . out−> d i s c r e t i z e . i n [ 2 ]
d i s c r e t i z e . out−>n e s t . i n [ 2 5 ]
reward . out−>n e s t . r e w a r d _ i n [ 1 ]
n e s t . out−>argmax . i n [ 3 ]
argmax . out−>command . i n [ 1 ]
F Environments
Table F shows parameters for the OpenAI Gym environments and the ZeroMQ wrapper.
19
Name
Version
Name
Version
Max episode steps
Initial reward*
Final reward*
Inter-trial duration*
Update interval (env runner)*
Name
Version
Max episode steps
Slippery
Final reward null*
Inter-trial duration*
Update interval (env runner)*
OpenAI Gym
MountainCar
FrozenLake
Values
0.8.1
Values
0
None
-1.0
-0.4
0.4
0.02
Values
0
None
False
-0.1
0.1
0.1
Table 3: Table of the environment parameters. Values marked with * indicate values for the ZeroMQ wrapper.
20
|
1812.02870 | 3 | 1812 | 2019-07-05T21:18:11 | The Role of Engagement, Honing, and Mindfulness in Creativity | [
"q-bio.NC"
] | As both our external world and inner worlds become more complex, we are faced with more novel challenges, hardships, and duress. Creative thinking is needed to provide fresh perspectives and solve new problems.Because creativity can be conducive to accessing and reliving traumatic memories, emotional scars may be exacerbated by creative practices before these are transformed and released. Therefore, in preparing our youth to thrive in an increasingly unpredictable world, it could be helpful to cultivate in them an understanding of the creative process and its relationship to hardship, as well as tools and techniques for fostering not just creativity but self-awareness and mindfulness. This chapter is a review of theories of creativity through the lens of their capacity to account for the relationship between creativity and hardship, as well as the therapeutic effects of creativity. We also review theories and research on aspects of mindfulness attending to potential therapeutic effects of creativity. Drawing upon the creativity and mindfulness literatures, we sketch out what an introductory 'creativity and mindfulness' module might look like as part of an educational curriculum designed to address the unique challenges of the 21st Century. | q-bio.NC | q-bio | Gabora, L., & Unrau, M. (2019). The role of engagement, honing, and mindfulness in creativity. In Mullen, C.
(Ed.), Creativity Theory and Action in Education: Vol. 3. Creativity under duress in education? Resistive
theories, practices, and actions (pp. 137-154). doi:10.1007/978-3-319-90272-2_8
The Role of Engagement, Honing, and Mindfulness in Creativity
Liane Gabora and Mike Unrau
University of British Columbia, Canada
Abstract As both our external world and inner worlds become more complex, we are faced with
more novel challenges, hardships, and duress. Creative thinking is needed to provide fresh
perspectives and solve new problems. Because creativity can be conducive to accessing and
reliving traumatic memories, emotional scars may be exacerbated by creative practices before
these are transformed and released. Therefore, in preparing our youth to thrive in an increasingly
unpredictable world, it could be helpful to cultivate in them an understanding of the creative
process and its relationship to hardship, as well as tools and techniques for fostering not just
creativity but self-awareness and mindfulness. This chapter is a review of theories of creativity
through the lens of their capacity to account for the relationship between creativity and hardship,
as well as the therapeutic effects of creativity. We also review theories and research on aspects of
mindfulness attending to potential therapeutic effects of creativity. Drawing upon the creativity
and mindfulness literatures, we sketch out what an introductory 'creativity and mindfulness'
module might look like as part of an educational curriculum designed to address the unique
challenges of the 21st Century.
Keywords: classroom, creativity, dual process, duress, education, hardship, honing,
mindfulness, teaching, trauma
Please direct correspondence regarding the manuscript to
Liane Gabora
Email: liane [dot] gabora [at] ubc [dot] ca
Dept. of Psychology, University of British Columbia, Fipke Centre for Innovative Research,
3247 University Way, Kelowna BC V1V 1V7, CANADA
Liane Gabora, PhD, is a Professor in the Psychology Department at the Okanagan Campus of
the University of British Columbia, Canada. Her research focuses on the mechanisms underlying
creativity, and how creative ideas -- and culture more generally -- evolve, using both
computational modeling and empirical studies with human participants. She has almost 200
articles published in scholarly books, journals, and conference proceedings, has procured over
one million dollars in research funding, supervised numerous graduate and undergraduate
students, and given talks worldwide on creativity and related topics. Her research on creativity is
informed by her own experiences creating literature. She has a short story published in Fiction,
another forthcoming in Fiddlehead, and has a novel underway titled Quilandria that merges her
scholarly and creative writing interests. Her paintings have been exhibited in the United States at
Anastasia's Asylum Coffee House (Los Angeles), Ten Women Gallery (Los Angeles), and Java
Joe's Coffee House (Santa Fe). Her animated short film titled "Self-referential Face" was shown
at the Artificial Life VI Conference, University of California, Los Angeles. Her electronic music
composition Stream Not Gone Dry was performed at Royce Hall, University of California, Los
Angeles (a piano version can be found at https://people.ok.ubc.ca/lgabora/artistic_files/Gabora-
Liane-Stream.wav).
Mike Unrau, MFA, is a PhD student in Interdisciplinary Graduate Studies at the University of
British Columbia (Okanagan Campus), Canada. He is studying creativity and social innovation,
focusing on how creative mindfulness impacts collective trauma towards societal change. He is
currently adjunct faculty with the University of Calgary and Mount Royal University, working in
field education and simulated educational experiences, as well as social-based theatre and
creativity. He has held international fellowships, given lectures and conference presentations,
conducted workshops and led research projects in different parts of the world, including a pre-
social lab in India. These days Mike's passion is Somativity, a physical movement mindfulness
approach he developed after living in a Buddhist monastery (Thailand and Canada) and co-
founding a physical theatre company (Calgary, AB). He has published findings on somatic
awareness as well as creativity, is a published poet, and a competition finalist in song writing and
internationally in photography. He is trained in expressive arts, a mindfulness approach called
Living Inquiries, and is a certified Transformational Arts facilitator.
1.1 Introduction
With intelligence increasing across generations (a phenomenon referred to as the Flynn Effect),
our networks of thoughts and ideas are taking on more diverse, complex structures. As one
effect, our minds are traveling down less-trodden, newer paths. Some thought-paths lead to
revolutionary innovations, heart-wrenching tunes, and riveting movies. There are other thought-
paths that lead to ruminations about the past, and fears about the future, all of which may play a
role in anxiety or depression. Others lead to ever-subtler ways of manipulating each other,
probing and bringing to light repressed and offensive parts of ourselves.
2
Thus, as both our external world and inner worlds become more complex, we are faced
with novel challenges, hardships, and duress. Creative thinking is needed to provide fresh
perspectives and solve new problems, but since creativity can be conducive to accessing and
reliving traumatic memories, emotional scars may be exacerbated by creative practices before
they are transformed and released. This suggests that in order to prepare our youth to thrive in an
increasingly unpredictable world, it could be helpful to cultivate in them an understanding of the
creative process and its relationship to hardship, as well as tools and approaches for fostering,
not just creativity, but self-awareness and mindfulness. As a working definition, mindfulness is
the awareness of what is happening presently, by paying attention to our experience and the
novel distinctions of it we actively draw upon, without judgement (Kabat-Zinn, 2003; Haller,
Bosma, Kapur, Zafonte, Langer, 2016). Mindfulness is essentially creative, in that as we
experience life mindfully, what we notice is new to us as a fresh perspective (Langer,
Moldoveanu, 2000).
This chapter begins with a brief overview of a few theories of creativity. Our focus is on
how, and to what extent, they address the relationship between creativity and hardship, as well as
the well-documented therapeutic impact of creative engagement. Next, we investigate some
theoretical aspects of mindfulness, again attending to its relationship to hardship and well-being.
Finally, we sketch out the basics of what a 'creativity and mindfulness' module might look like.
1.2 Hardship and the Therapeutic Effects of Creativity
It is widely believed that creativity is fostered by a warm, supportive, nurturing, and trustworthy
environment conducive to self-actualization (Maslow, 1971; Rogers, 1959). However, there is a
negative correlation between creativity and parental warmth (Siegelman, 1973) and a high
incidence of early parental loss in eminent creators (Eisenstadt, 1978). More generally,
childhood adversity is believed to be a developmental antecedent of eminent creativity
(MacKinnon, 1962; Rhue & Lynn, 1987; Simonton, 1994), perhaps in part because of a
relationship between adversity and diversifying experiences (Damian & Simonton, 2015). The
relationship between creativity and hardship is not restricted to childhood adversity; for example,
stories written by adults in response to mildly threatening stimuli were rated as more creative
than stories responding to non-threatening stimuli (Riley & Gabora, 2012). Thus, a
Nevertheless, creativity is believed to have therapeutic effects (Barron, 1963; Forgeard,
2013). It can be intrinsically rewarding (Gruber, 1995; Kounios & Beeman, 2014; Martindale,
1984). The creative process can at times be frustrating and draining and involve working through
negative material. However, there is evidence that high levels of creativity are correlated with
positive affect (Hennessey & Amabile, 2010) and the ability to manage intense feelings (Moon,
1999). Clinical practitioners of art therapy note that imagery and creative engagement can
deepen communication between client and therapist (Moon, 2009). Art therapy can also enhance
self-understanding, in addition to facilitating the process of finding healthier ways of handling
situations and interacting with others (Dunn-Snow & Joy-Smellie, 2000; Riley, 1999). By
providing access to issues that are difficult to verbalize, art therapy can bring these to the surface
3
in a nonverbal form or provide a springboard for discussion (Malchiodi, 2007). There is also
evidence that creativity can enhance one's sense of self (Garailordobil & Berrueco, 2007;
MacKinnon, 1962).
Therapeutic effects of creativity may also stem from the capacity of therapy to enhance
feelings of connection to, and appreciation by, others. In the verification of a creative work, the
creator generates an internal context for the idea that encompasses a typical individual who will
encounter the work. For an inventor this (what?) might involve developing a working prototype.
For an artist it might involve arranging artworks for showing at a gallery. By finding a form for
the idea that is palatable (e.g., comprehensible or intriguing) to others, one's worldview merges
with and expands those of others. To the extent that a creative product responds to universal
features of worldviews, it may have a healing effect on others. Creative products are felt to be a
highly personal form of self-disclosure; self-disclosure has therapeutic value (Pennebaker, 1997)
and even beneficial effects on the immune system (Pennebaker, Kiecolt-Glaser, & Glaser, 1998).
Since creative people often feel disconnected from others because they defy convention
(Sternberg & Lubart, 1995; Sulloway, 1996), the benefits of creative self-disclosure may be
mediated by an enhanced sense of belonging.
1.3 Theories of Creativity
To what extent do theories of creativity incorporate and account for (1) the relationship between
creativity and hardship, and (2) the transformative and sometimes therapeutic effects of
creativity? In this section we address these questions.
A starting point for much research into creativity is Wallas' (1926) classification of the
creative process into a series of stages. The first stage is preparation, which involves obtaining
background knowledge relevant to the problem, and its history, such as any past attempts, or
preconceptions regarding how to solve it. It also involves conscious, focused work on the
problem. The second stage is incubation -- unconscious processing of the problem that continues
while one is engaged in other tasks. The first and second stages may be interleaved. Wallas
proposed that after sufficient preparation and incubation, the creative process is often marked by
a sudden moment of illumination or insight during which the creator glimpses a way of going
about the task, which may require substantial work to bring to completion. The final phase is
referred to as verification. This involves not just fine-tuning the work and making certain it is
works not just in theory but in in practice, but putting it in a form that can be understood and
appreciated by others.
While early research supported Wallas' classic four-stage theory of creativity, subsequent
studies, in particular those examining the need for incubation, did not support Wallas' theory
(Eindhoven & Vinacke, 1952). One criticism of Wallas' theory is that it is merely descriptive
and thus fails to explain how or why the stages occur. More importantly with respect to our
purposes, it does not address the relationship between creativity and hardship or the therapeutic
impact of creativity. However, variants of Wallas' theory have continued to serve as a platform
for theoretical and empirical research on creativity.
4
1.4 Heuristic Search
Inspired by the metaphor of the mind as a computer (or computer program), early research on
creativity focused on the notion of heuristic search. In heuristic search, [rules of thumb guide the
inspection of different states within a particular state space (i.e., a set of possible solutions) until
a satisfactory solution is found (Eysenck, 1993; Newell, Shaw, & Simon 1957; Newell & Simon,
1972). In heuristic search, the relevant variables are defined up front; thus, the state space is
generally fixed. Examples of heuristics include breaking a complex problem into sub-problems,
hill climbing (reiteratively modifying the current state to look more like the goal state), and
working backward from the goal state to the initial state. Heuristic search may include the
restructuring of mental representations. This restructuring may be accomplished, for example, by
(1) re-encoding the problem such that new elements are perceived to be relevant, or (2) relaxing
goal constraints (Weisberg, 1995).
The idea that creativity could be construed as a heuristically guided search gave hope to
those who sought a scientific understanding of creativity, since search is formally tractable.
However, in many creative tasks, and particularly artistic forms of creativity, the goal state is
unspecified, and some elements of the eventual solution may not be present when the problem
presents itself. It has been suggested that creativity involves heuristics that guide the search for a
new state space itself, not just a possibility within a given state space (Boden, 1990; Kaplan &
Simon, 1990, Ohlsson, 1992). However, search based approaches to creativity start with pre-
existing state spaces and do not address how a new state space comes into existence.
Furthermore, like Wallas' four-stage theory, the heuristic search approach to creativity neither
addresses the relationship between creativity and hardship nor the therapeutic impact of
creativity.
1.5 Dual Process Theories
Several proposals for two forms of cognitive processing come from largely disconnected
literatures in cognitive and social psychology (Evans, 2008; Sowden, Pringle, & Gabora, 2014).
There is interaction between implicit and explicit processes, including synergy effects. Dual
processing theories are diverse; while some are concerned with parallel competing processes
involving explicit and implicit this repeats from above knowledge systems, others are concerned
with the contextualizing and shaping of deliberative reasoning and decision-making by
preconscious processes (source?). Nevertheless, these proposals both concern the basic need to
distinguish between cognitive processes that are fast, automatic, and unconscious and those that
are slow, deliberative, and conscious.
Dating back to Freud's (1949) distinction between primary process and secondary
process thinking, most creativity researchers espouse some variant of a dual process theory (e.g.,
Barron, 1963; Eysenck, 1995; Feist, 1999; Finke, Ward, & Smith, 1992; Fodor, 1995; Gabora,
2003; Martindale, 1995; Richards et al., 1988; Russ, 1993; Simonton, 1999). Psychological
theories of creativity typically involve (1) a divergent stage that predominates during idea
5
generation, and (2) a convergent stage that predominates during the refinement, implementation,
and testing of an idea (for a review see Runco, 2010). Divergent thought is characterized as
intuitive and reflective; also, it involves the generation of multiple discrete, often unconventional
possibilities. Divergent thinking ability is sometimes measured in terms of fluency: the number
of ideas generated. Convergent thought, characterized as critical and evaluative, involves the
selection or tweaking of the most promising possibilities. Neural models of the mechanisms
underlying these two modes of thought have been proposed (Gabora, 2002, 2010, 2018; Gabora
& Ranjan, 2013).
One well-known dual process theory of creativity is the Geneplore model (Finke et
al.,1992). This theory posits that the creative process consists of two stages: generate and
explore. (Indeed the name 'Geneplore' is a condensation of "generate" and "explore.") The
generation stage involves coming up with crudely formed ideas referred to as pre-inventive
structures that contain the kernel of an idea as opposed to an idea in its entirety. The exploration
stage involves fleshing out these pre-inventive structures through elaboration and testing.
Use of the term exploration to refer to the second phase of the creative process can be
misleading. Explore is often used to refer to surveying the space of possibilities as generally
occurs during the first phase of the creative process, as opposed to refining a single possibility as
generally occurs during the second phase. However, the notion of a pre-inventive structure does
capture the intuition that early on in the creative process one is working with cognitive structures
that are different in kind from those being worked with later in the creative process. The
Geneplore model does not attempt to formalize how a pre-inventive structure differs from a full-
fledged ideaor what differentiates a promising pre-inventive structure from a mundane one.
Another theory of creativity that could be considered a dual process theory emphasizes
ideation-evaluation cycles (Basadur, 1995). Creative thinking is said to involve three major
stages -- problem finding, problem solving, and solution implementation.Each of these involves
alternating cycles ideation and evaluation to varying degrees, depending on the domain. Domains
that emphasize problem finding have a higher ratio of ideation to evaluation, whereas domains
that emphasize implementation show the opposite.
A dual process theory of analogy is structure mapping (Gentner, 1983). In brief, analogy
generation occurs in two steps: first, searching memory in a "structurally blind" manner
(Gentner, 2010, p. 753) for an appropriate source and aligning it with the target. Second,
mapping the correct one-to-one correspondences between the source and the target.
Yet another well-known dual process theory of creativity is the Darwinian theory of
creativity (Campbell 1960; Simonton, 2011). Like biological species, creative ideas exhibit the
kind of cumulative complexity and adaptation over time as an evolutionary process, not just
when they are expressed to others but in the mind of the idea's creator (Gabora, 1996; Terrell,
Hunt, & Gosden, 1997; Thagard, 1980; Tomasello 1996). Thus, it has been proposed that in
creativity, as in natural selection, there is a phase conducive to generating variety and another
conducive to pruning out inferior variants. According to the Darwinian theory, we generate new
ideas through essentially a trial-and-error process involving blind generation of ideational
6
variants followed by selective retention of the fittest variants for development into a finished
product. Thus, the Darwinian theory is sometimes referred to as Blind Variation Selective
Retention (BVSR). The variants are said to be 'blind' in the sense that the creator has no
subjective certainty about whether they are a step in the direction of the final creative product.
In addition to serious theoretical flaws with BVSR (e.g., Gabora, 2007), although the
relationship among creativity, hardship, and well-being is at times mentioned in BVSR and other
dual process accounts, it does not play a central role in these theories. If we were to find out
suddenly that we were wrong about the research relating creativity to hardship and the
therapeutic benefits of creativity, these theories would not require substantial revision as a result.
1.6 Honing Theory (HT)
This relationship among creativity, hardship, and well-being plays a fundamental role in another
theory of creativity known as the honing theory (Gabora, 2017). While the central aim of the
above-mentioned theories of creativity is to account for the existence of creative products -- i.e.,
products that are new and useful to society -- the central aim of the honing theory of creativity is
to account for the cumulative nature of cultural evolution. This is explained not as creative
outputs but as the minds that generate them. Thus, HT focuses not just on restructuring as it
pertains to the conception of the task, but also as it pertains to the global structure of the mind,
what we call the worldview.
A worldview is a mind experienced subjectively, from the inside. It is a way of seeing and
being in the world that emerges as a result of the structure of one's web of understandings,
beliefs, and attitudes. A worldview reveals itself through behavioural regularities in how it is
expressed and responds to situations (Gabora, 2017). The creative process reflects the natural
tendency of a worldview to self-organize to achieve a state of dynamical equilibrium through
interactions amongst its components, whether they be ideas, attitudes, or bits of knowledge. Most
people are familiar with the experience of "catching themselves" in internal dialogue. Internal
dialogue is evidence of the self-organization in action of one's worldview.
An evolutionary theory, HT developed from the premise that creativity is the novelty-
generating component in cultural evolution. This is an evolutionary process, that is, a process of
cumulative, open-ended, adaptive change over time. However, it is not a Darwinian or
selectionist evolutionary theory (Gabora, 2011). Although selection as the term is used in the
layperson sense may play a role (i.e., people may be selective about which aspects of their
worldviews they express or which paintings they show at a gallery), the process does not involve
selection in its technical sense. (This involves change over generations due to the effect of
differential selection on the distribution of heritable variation across a population.) As in any
kind of evolutionary process, novelty generation must be balanced by novelty preservation. In
biological evolution, novelty is generated through genetic mutation and recombination, and the
novelty is preserved through the survival and reproduction of "fit" variants. In cultural evolution,
novelty's generation is through creativity, and novelty's preservation through imitation and other
forms of social learning.
7
HT posits that the creative process begins with being alert to psychological entropy,
arenas of one's worldview that, on the spectrum from orderly to chaotic, are relatively chaotic
and in need of creative restructuring (Gabora, 2017; Hirsh, Mar, & Peterson, 2012). The process
can be "jogged along" by stimuli that capture attention or pique interest; creativity often involves
a seed incident that gets the creative juices flowing (Doyle, 1998).
Honing an idea involves looking at it from the different angles proffered by one's
particular worldview: "putting one's own spin on it," making sense of it in one's own terms, and
expressing it outwardly (Gabora, 2017). HT posits that creativity involves viewing the task from
a new context, which may restructure the internal conception of it. This restructuring may be
amenable to external expression. Thus, honing enables the creator's understanding of the
problem or task to shift, and in so doing a form may be found that fits better with the worldview
as a whole. In this way, not only does the task get completed (or worked on and put aside) but
also the worldview transforms, becoming more robust as it evolves.
The transformative impact of immersion in the creative process extends far beyond the
"problem domain." It can bring about sweeping changes to that second (psychological) level of
complex, adaptive structure that alters one's self-concept and view of the world. Creative acts
and products render such cognitive transformation culturally transmissible. This is why HT
posits that what evolves through culture are not creative contributions but worldviews. Cultural
contributions give hints about the worldviews that generate them. When faced with a creatively
demanding task, not only does one's worldview offers perspectives that alters the conception of a
task, but, likewise, immersion in the task subtly or profoundly alters one's worldview. The
above-mentioned finding that childhood adversity is a developmental antecedent of creativity is
consistent with viewing creativity as the honing and expressing of a unique worldview, since
adversity and isolation generate the need and mental space to figure things out for oneself. It is
through the creative honing of networks of understandings that worldviews self-organize, and it
through the communal expression of honed ideas that culture evolves.
Midway through the creative process, one may have made associations between the
current task and previous experiences, but not disambiguated which aspects of those previous
experiences are relevant to the current task. At this point, the idea may feel "half-baked." It can
be said to be in a potentiality state because how it will actualize depends on the particular
perspectives from which it is considered. These perspectives may be internally generated
(imagining what would happen if …) or externally generated (e.g., building a prototype and
trying it out). Thus, the recursive process described in which an external change suggests a new
context from which to think about the creative task, and so forth recursively until the task is
complete is sometimes referred to as context-driven actualization of potential. Each time the idea
is looked at from a new context it undergoes a change-of-state such that some of its potential
becomes more readily actualized. When the task is complete, the conception of it is said to be in
an eigenstate, because one's worldview is no longer spontaneously generating new contexts from
which to consider it. The creator may express this state as a product, which can cause someone
8
else's worldview to be in a potentiality state. This is when it is the other's turn to adapt it to their
own needs or tastes. Through this process culture evolves in new directions.
A worldview not only self-organizes in response to perturbations but it is imperfectly
reconstituted and passed down through culture. This is because it is not just self-organizing but
self-regenerating: people share experiences, ideas, and attitudes, thereby influencing the process
by which others' worldviews form and transform. Children expose elements of what was
originally an adult's worldview to different experiences and bodily constraints, and thereby forge
unique internal models of the relationship between self and world. Thus, worldviews evolve by
interleaving (1) internal interactions amongst their parts, and (2) external interactions with
others. Through these social interactions, novelty accumulates and culture evolves. Elements of
culture create niches for one another. One creative ideas begets another and modifications build
on each other, a phenomenon sometimes referred to as the ratchet effect.
1.7 Theory and Research on Mindfulness
At the beginning of the chapter, we defined mindfulness as the awareness of what is happening
presently, by paying attention to our experience from a novel perspective, without judgement.
Like creativity, the practice of mindfulness appears to reduce dissonance, enhance feelings of
connection, and facilitate the movement of repressed emotion. This correlates with indicators of
well-being (the reduction of stress, anxiety, and depression) and it has also been related to
satisfaction of life, vitality, a sense of flourishing, and self-actualization (Beitel, Ferrer. &
Cecero, 2004; Brown & Codon, 2016; Brown & Ryan, 2003; Cardaciotto, Herbert, Forman,
Moitra, & Farrow, 2008; Carlson & Brown, 2005; Feldman, Hayes, Kumar, Greeson, &
Laurenceau, 2007; Lawlor, Schonert-Reichl, Gadermann, & Zumbo, 2014; Walach, Buchheld,
Buttenmuller, Kleinknecht, & Schmidt, 2006).
If creativity can reduce the dissonance in a worldview, could it then lead to mindfulness,
which reduces disharmony in well-being? If a person knows nothing about what mindfulness is,
can he or she in the heart of the creativity become more (unknowingly or knowingly) mindful? If
so, does the individual afterward feel less gripped by internally aroused or externally provoked
stress, and therefore more connected to others? If this is the case, does this person feel more
revitalized, open to possibility (an increase in potentiality), and able to influence social networks
towards a creative shift? Perhaps a feedback loop of creativity > mindfulness > creativity might
transformatively impact not only an individual's creative process but also her or his
psychological structure to adaptably alter self-concept and thus move towards self-actualization.
The impact of mindfulness on creativity is being studied (Baas, Nevicka, Ten Velden,
2014; Greenberg, Reiner & Meiran, 2012; Ostafin & Kassman, 2012; Ren, Luo, Wei, Ying, Ding
et al., 2011). Such research connects mindfulness practice, such as focused attention meditation
and open monitoring meditation, to divergent and convergent creative thinking. Research also
finds a positive correlation with mindfulness meditation on positive insight problem solving,
reduction in habitual responses, an increase in cognitive flexibility, fluency and originality (Baas
et al., 2014; Capurso, Fabbro, Crescentini, 2014; Chambers, Gullone, & Allen, 2009; Colzato,
9
Ozturk, & Hommel, 2012; Greenberg, et al., 2012), as well as attentional focus (Davidson &
Lutz, 2008; Valentine & Sweet, 1999) and the inhibition of automatic responding (Schmertz,
Anderson, & Robins, 2009). With these considerations, mindfulness can be said to lead to
creativity.
Our question, however, is can creativity lead to mindfulness? For our purposes here,
creativity that is enhanced or engaged in due to a pre or concurrent state of mindfulness we will
call mindful creativity, which is in common use (Haller, 2015). However, mindfulness that is
enhanced or engaged in due to a pre or concurrent process of creativity we will refer to as
creative mindfulness, which redirects how this term is often used to focus on creativity leading to
mindfulness. In brief, a person who is being more mindful might become more creative, and
someone being more creative might become more mindful.
Some work that could help answer the question of the extent to which creativity can
cultivate mindfulness is the research on flow. Csikszentmihalyi (2014) defines flow as a "holistic
sensation present when we act with total involvement," (p. 136) that is experienced in creativity,
but not always or is limited to it. He distinguishes flow from mindfulness in the sense that in
flow one is aware of one's actions, but not aware of being aware. For example, one may be
aware of painting an artistic scene on a canvas, in part due a narrowing of attention on the
canvas, but not consciously aware that one is in a room with a cat in the periphery, focused
directly on painting a painting. However, he argues that flow fuels the motivational drivers that
lead to creativity (Csikszentmihalyi (2015).
Mindfulness has been described as "the awareness that emerges through paying attention
on purpose, in the present moment, and nonjudgmentally to the unfolding of experience moment
by moment" (Kabat-Zinn, 2003, p. 145). It also includes "simultaneously drawing novel
distinctions in the present moment" (Haller, et al., 2016, p. 894). We posit here that mindfulness
has three components (Shapiro & Carlson, 2009; Young, 2016):
1) Awareness of what is happening presently, from a subjective (self-referential) and
concurrently objective (non-self-referential) point of view (Vázquez Campos & Gutiérrez,
2015).
2) Attention, through the novel distinctions in observation compared to the habitual. One
might ask "what is new in what I'm observing?" This includes focused attention on a
particular object (both externally and internally) and an open attention, which defocuses
from a particular object or, rather, is a distributed attentional focus where one is attentive to
experience and the interpretation of it
3) Equanimity, including the non-judgmental acceptance of what is being experienced and
attended to.
Mindfulness can result in observation, non-evaluative description,, acceptance (Baas et al.,
2014), as well as a "greater sensitivity to one's environment, more openness to new information,
10
the creation of new categories for structuring perception, and enhanced awareness of multiple
perspectives in problem solving" (Langer & Moldoveanu, 2000).
Mindfulness can be contextualized within four modes of perspective. These are subjective
(thoughts, emotions, perceptions and sensations); objective (a time/space material observation of
body and externalities); inter-subjective (values, relationships, and meaning amongst social
connections), and inter-objective (systems, networks, and environments), as per integral theory
(Forbes, 2016). Integral mindfulness practices create the context of a mindfulness experience
within not only an internal (subjective) and external (objective) awareness, but also one that is
collective (inter-subjective) and systemic (inter-objective). From the perspective of its systemic
mode, mindfulness programs, for example, have been used to engage in social actions around the
Occupy movement, and anti-racist, climate change, and social justice initiatives in schools and
communities (Magee, 2015; Rowe, 2015a, 2015b; Forbes, 2016). In this way, the systemic mode
of mindfulness can lead to social inclusivity, equitable citizenry, communal well-being
(eudemonia), and a "shared meaning of the common good" (Forbes, 2016, p. 1267; also, Giroux
2014; Healey 2015a, 2015b).
Although mindfulness has distinct components such as attention, observation, and
equanimity, it has been shown that the "ability to carefully observe, notice, or attend to a variety
of internal and external phenomena consistently predict[s]enhanced creativity" (Baas et al. 2014,
p. 1103). Still, observation has been considered one of the stronger indicators of creativity; it is
associated with being open to experience and assisting in adaptive responses to uncertain or
complex situations (Siegel & Siegel, 2014). While a mindfulness induced observation leads to
improved creativity, it has not been shown conclusively that a creativity induced observation
leads to improved mindfulness.
Should a study be conducted in this manner, creativity could be seen to initiate
mindfulness, and beyond this possibly a deeper sense of personal awareness and social
awareness. A heightened awareness could then lead to a positive shift towards creating
transformative environments for individuals and communities to deal with complex problems
innovatively. Creative mindfulness, meaning-making, and social capital could result as
demonstrated by enhanced trust, community engagement, and meaningful individual and social
action (Ponder, 2012). Bridging outward social action with inner transformation in a collective
way can enable social and self-actualization. While such cause and effect may not be so direct,
such possibilities act as starting points for the conscious honing of the networks of
understandings by which worldviews self-organize, which is in part how culture evolves. As
Mouchiroud and Bernoussi (2008) conclude: "It may be that only socially creative individuals
will be able to act efficiently on these global issues and invent viable social solutions" (p. 379).
1.8 Duress and Well-being in Theory on Mindfulness and Creativity
We all want well-being of some kind in our lives. We seek it for ourselves and others in many
ways: financial security, support of family members and friends, and in our own personal sense
of self. Self-agency, emotional regulation, and higher self-esteem are correlated with
11
mindfulness (Siegel, Siegel, & Parker, 2016), which may help to facilitate our ability to meet
basic needs we have for autonomy, competence, and relatedness (Brown & Ryan, 2003).
Research has been done to determine if the Mindfulness-Based Stress Reduction (MBSR)
program and the Eight Point Program (EPP), a concentration-based meditation program,
decrease duress and enhance well-being for those with medical and psychiatric conditions, as
measured by the Mindful Attention and Awareness Scale (MAAS; Brown & Ryan, 2003;
Easwaran, 1991). In this study, mindfulness was defined by the non-judgemental awareness of
moment-to-moment experiences. Findings suggested significant increases in well-being
(Shapiro, Oman, Thoresen, Plante, & Flinders, 2008), with well-being defined by improved
effects in perceived stress, rumination, and forgiveness. Another study showed that when well-
being was measured for self-esteem, neuroticism, positive or negative affect, self-actualization,
autonomy, competence, and physical health, MAAS-measured mindfulness had wide-ranging
and inclusive correlations to enhanced well-being (Brown & Ryan, 2003). There is evidence that
daily creative activity in everyday environments has a positive impact on emotional experience
(Conner, DeYoung & Silva, 2018). This supports the growing perspective that everyday
creativity can cultivate positive psychological functioning and thus a sense of flourishing and
well-being (Richards, 2010).
Well-being has been shown to be correlated with social and emotional competence (SEC;
Jennings, 2016), which refers to the ability to manage stress, emotional reactivity, and related
cognition and behaviour to optimize effectiveness in daily activity such as the classroom. SEC's
five competencies are relationship skills, social awareness, responsible decision-making, self-
awareness, and self-regulation. Our external abilities toward social competency are related to our
internal abilities toward personal competency. Interpersonal connections, trust, compassion,
internal kind-heartedness, and self-awareness all result from mindfulness (Siegel et al, 2016).
Although the SEC has been applied to teaching, students may also benefit.
When a contemplative capacity such as mindfulness along with social competence is
relatively high, an individual is more likely to be creative (Zajonc, 2014). The type of creativity
that instigates or results from social awareness could be called social creativity (Watson, 2007),
although this term also applies to multiple conditions. Social mindfulness has been described as a
"benevolent focus on the needs and interests of others. General mindfulness starts with paying
attention to the little things available to individual awareness; social mindfulness starts at a
similar basic level" (van Doesum, van Lange & van Lange, 2013).
For the sake of this chapter, we propose that social mindfulness is the attentiveness of
greater individual awareness as applied to the needs and interests of others. Social mindfulness is
founded in both inter-subjective and inter-objective modes of mindfulness. Similarly, social
creativity has been related to a higher sense of meaning-making, such as self-actualization
(Maslow, 1971; Serotkin, 2010). Social creativity has been used in social awareness and social
justice programs, anti-oppression pedagogy, democratic reform, and social advocacy (Beyerbach
& Ramalho, 2011; Boal, 2000, 2005; Dewhurst, 2014; Hick & Furlotte, 2009; Lampert, 2013)
and it is at the heart of social innovation programs in that such programs offer novel approaches
12
to social problems (Mumford & Moertl, 2003). Both social mindfulness and social creativity
increase an individual's awareness of the needs of others, connections within groups and
communities, which deepen a sense of meaning and increase personal and social well-being
while augmenting social change.
1.9 Educational Applications of Creativity and Mindfulness Research
The pace of cultural change is accelerating more quickly than ever. In biological systems,
environmental change often induces a sudden increase in the mutation rate. This makes adaptive
sense, although most mutations are detrimental while others are beneficial and enhance their
survival in changing environments. Similarly, in times of accelerated cultural change it is
adaptive to increase the rate of creativity and generate innovative solutions to any unforeseen
problems. This is particularly important now; in our high-stimulation environment, children
spend so much time processing new stimuli that there is less time to go deep with the stimuli
they have encountered. They also do not have much time to think about ideas and situations from
different perspectives, such that their ideas become more interconnected and their mental models
of understanding more integrated. It is this kind of deep processing and the resulting integrated
webs of understanding that make the crucial connections leading to important advances and
innovations.
Therefore, given the fast pace of change in society today we are led to ask, how can
creativity and mindfulness be cultivated in the classroom? Starting with creativity, one way is by
focusing less on the reproduction of information, and more on problem solving and critical
thinking. Another way of cultivating creativity in this context is by posing questions and
challenges, followed by experiencing opportunities for solitude and reflection or group
discussion in an effort to foster the honing of new approaches and ideas.
In addition, to incorporate teaching methods that encourage creativity, we suggest that
classroom curricula might include a regular Creativity and Mindfulness Module that) includes
activities and approaches inspired by the creativity and mindfulness literatures. Creativity-
enhancing activities can be, for example, assignments that transcend traditional disciplinary
boundaries. Examples include painting murals that depict historical events, acting out plays
about animals that share a biological ecosystem, and writing poetry about black holes. After all,
the world does not come carved up into different subject areas. It is only through enculturation
that we come to believe these disciplinary boundaries are real -- our thinking becomes trapped in
and by them.
There are also ways in which mindfulness can be cultivated in the classroom both on a
day-to-day basis and as part of a possible Creativity and Mindfulness Module. If formal
mindfulness programs are not in place or practical, a creative mindfulness approach would offer
students the benefits of creative engagement and possibly heightened well-being. This can be
accomplished with students through emphasis (general or specific) on being creative in present-
moment awareness, by providing the physical and psychological space for full engagement in the
creative act (e.g., in art, science, physical education, etc.) via awareness, attention, and
13
equanimity. For example, a teacher could guide students into focusing their attention on the
details of what a student is doing by not only paying attention to their task, but also their process
and environment (being aware of one's surroundings as well as the study at hand in a balanced
way), and not judging themselves or others' experience or outcome. By helping students 'get
lost' in their creative experience, to the point of them being engrossed with the very act of
creating itself, teachers may find their students become more involved as whole beings, more
able to pay attention with greater awareness, and overall more mindful as a result.
Creative mindfulness, when directed to bring the group into the heart of the creative
process in any domain, could lead to self-awareness and then to social awareness of the group
dynamic. If focused, this could bring about peak experiences such as flow that are rich with a
sense of vitality and total involvement. As William James (1923), acclaimed father of
psychology, claimed, "the faculty of voluntarily bringing back a wandering attention, over and
over again, is the very root of judgment, character, and will. … An education which should
improve this faculty would be the education par excellence'' (p. 424, italics in original). With
these considerations, youth can be prepared to thrive with uncertainty, open their minds to the
novelty found in observing themselves and their studies with awareness, while nurturing
personal and social well-being.
1.10 Conclusions
Creativity is much-needed in today's fast-paced, ever-changing world. In the long term, it can be
transformative, therapeutic, and even self-actualizing. In the short term, however, creativity can
re-open painful memories, causing duress. Yet, there is reason to believe that -- paired with the
potentially life-affirmating effects of mindfulness towards well-being -- it may be possible to
delve into painful experiences and tendentious material but nevertheless feel held and accepted
by one's higher self throughout the process. With this type of self-regulation, creativity may help
one improve a generalized sense of self-awareness and even social mindfulness, as the needs of
others and connections to community, meaning-making, and social well-being develop. Such
changes could add to a possible collective shift in social-actualization and a potential change in
worldview. Drawing upon the literatures in the psychology of creativity and mindfulness, we
have loosely sketched out a creativity and mindfulness module that introduces a creative
mindfulness approach, for prospective application in a classroom setting. This is just an initial
outline that requires development and refinement. Nevertheless, we suggest that a move in this
direction could play a part in an educational curriculum designed to address unique challenges
of the 21st Century.
Acknowledgements
This research was supported in part by a grant (62R06523) from the Natural Sciences and
Engineering Research Council of Canada.
References
14
Baas, M., Nevicka, B., & ten Velden, F. S. (2014). Specific mindfulness skills differentially
predict creative performance. Personality and Social Psychology Bulletin, 40(9), 1092-1106.
doi: 10.1177/0146167214535813
Barron, F. (1963). Creativity and psychological health. Princeton, NJ: Van Nostrand.
Barron, F. (1969). Creative person and creative process. Oxford, England: Holt, Rinehart, &
Winston.
Basadur, M. (1995). Optimal ideation-evaluation ratios. Creativity Research Journal, 8, 63 -- 75.
Beitel, M., Ferrer, E., & Cecero, J. J. (2004). Psychological mindedness and awareness of self
and others. Journal of Clinical Psychology, 61, 739 -- 750.
Beyerbach, B., & Ramalho, T. (2011). Chapter fifteen: Activist art in social justice pedagogy.
Counterpoints, 403, 202-217.
Boal, A. (2000). Theater of the Oppressed. Pluto Press.
Boal, A. (2005). Legislative theatre: Using performance to make politics. New York, NY:
Routledge.
Brown, K. W., & Cordon, S. (2016). Toward a phenomenology of mindfulness: Subjective
experience and emotional correlates. In Schonert-Reichl, K. A. & Roeser, R. W.
(Eds.), Handbook of mindfulness in education: Integrating theory and research into practice
(pp. 59−81). New York, NY: Springer.
Brown, K. W., & Ryan, R. M. (2003). The benefits of being present: Mindfulness and its role in
psychological well-being. Journal of Personality and Social Psychology, 84, 822 -- 848.
Boden, M. (1990/2004). The creative mind: Myths and mechanisms. London, England:
Weidenfeld & Nicolson.
Campbell, D. T. (1960). Blind variation and selective retention in creative thought as in other
knowledge processes. Psychological Review, 67, 380−400.
Capurso, V., Fabbro, F., & Crescentini, C. (2014). Mindful creativity: The influence of
mindfulness meditation on creative thinking. Frontiers in Psychology, 4, Article 1020. doi:
10.3389/fpsyg.2013.01020
Cardaciotto, L., Herbert, J. D., Forman, E. M., Moitra, E., & Farrow, V. (2008). The assessment
of present-moment awareness and acceptance: The Philadelphia Mindfulness Scale.
Assessment, 15, 204 -- 223.
Carlson, L. E., & Brown, K. W. (2005). Validation of the Mindful Attention Awareness Scale in
a cancer population. Journal of Psychosomatic Research, 58, 29 -- 33.
Chambers, R., Gullone, E., & Allen, N. B. (2009). Mindful emotion regulation: An integrative
review. Clinical Psychology Review, 29, 560-572.
Colzato, L. S., Ozturk, A., & Hommel, B. (2012). Meditate to create: The impact of focused-
attention and open-monitoring training on convergent and divergent thinking. Frontiers in
Psychology, 3, Article 116.
Conner, T. S., DeYoung, C. G., & Silvia, P. J. (2018) Everyday creative activity as a path to
flourishing. The Journal of Positive Psychology, 13(2), 181-189. doi:
10.1080/17439760.2016.
15
Csikszentmihalyi, M. (2014). Flow and the foundations of positive psychology: The collected
works of Mihaly Csikszentmihalyi. Dordrecht, Netherlands: Springer.
Csikszentmihalyi, M. (2015). The systems model of creativity: The collected works of Mihaly
Csikszentmihalyi. Dordrecht, Netherlands: Springer.
Damian, R. I., & Simonton, D. K. (2015). Psychopathology, adversity, and creativity:
Diversifying experiences in the development of eminent African Americans. Journal of
Personality and Social Psychology, 108(4), 623.
Davidson, R. J., & Lutz, A. (2008). Buddha's brain: Neuroplasticity and meditation. IEEE Signal
Processing Magazine, 25, 172-176.
Dewhurst, M. (2014). Social justice art: A framework for activist art pedagogy. Cambridge,
MA:Harvard Education Press.
Dunn-Snow, P., & Joy-Smellie, S. (2000). Teaching art therapy techniques: Mask-making, a case
in point. Art Therapy, 17, 125 -- 131.
Doyle, J. J. (1998). The writer tells: The creative process in the writing of literary fiction.
Creativity Research Journal, 11, 29-37.
Eisenstadt, J. M. (1978). Parental loss and genius. American Psychologist, 33, 211 -- 223.
Easwaran, E. (1991). Meditation: A simple eight-point program for translating spiritual ideals
into daily life (2nd ed.). Tomales, CA: Nilgiri Press. (Original work published 1978).
Retrieved from http://www.easwaran.org
Eindhoven, J. E., & Vinacke, W. E. (1952). Creative processes in painting. Journal of General
Psychology, 47(2), 139-164.
Evans, J. (2008). Dual-processing accounts of reasoning, judgment, and social cognition. Annual
Review of Psychology, 59, 255 -- 278.
Eysenck, H. J. (1993). Creativity and personality: Suggestions for a theory. Psychological
Inquiry, 4, 147 -- 178.
Eysenck, H. J. (1995). Genius: The natural history of creativity. Cambridge, England:
Cambridge University Press.
Feist, G. (1999). The influence of personality on artistic and scientific creativity. In R. J.
Sternberg (Ed.), Handbook of creativity. Cambridge, UK: Cambridge University Press.
Feldman, G., Hayes, A., Kumar, S., Greeson, J., & Laurenceau, J. (2007). Mindfulness and
emotion regulation: The development and initial validation of the Cognitive and Affective
Mindfulness Scale-Revised (CAMS-R). Journal of Psychopathology and Behavioral
Assessment, 29, 177 -- 190.
Fodor, J. D. (1998). Learning to parse. Journal of Psycholinguistic Research, 27, 285 -- 319.
Forbes, D. (2016). Modes of mindfulness: Prophetic critique and integral emergence.
Mindfulness, 7(6), 1256-1270. doi: 10.1007/s12671-016-0552-6
Forgeard, M. (2013). Perceiving benefits after adversity: The relationship between self-reported
posttraumatic growth and creativity. Psychology of Aesthetics, Creativity, Arts 7, 245 -- 264.
Freud, S. (1949). An outline of psychoanalysis. New York, NY: Norton.
16
Finke, R. A., Ward, T. B., & Smith, S. M. (1992). Creative cognition: Theory, research, and
applications. Cambridge, MA: MIT Press.
Gabora, L. (1996). A day in the life of a meme. Philosophica, 57, 901 -- 938.
Gabora, L. (2000). Toward a theory of creative inklings. In (R. Ascott, Ed.) Art, Technology, and
Consciousness (pp. 159-164). Intellect Press, Bristol, UK.
Gabora, L. (2003). Contextual focus: A cognitive explanation for the cultural transition of the
Middle/Upper Paleolithic. In R. Alterman & D. Kirsh (Eds.), Proceedings of the 25th annual
meeting of the Cognitive Science Society. Hillsdale, NJ: Lawrence Erlbaum Associates.
Gabora, L. (2007). Why the creative process is not Darwinian. Creativity Research Journal,
19(4), 361-365.
Gabora, L. (2010). Revenge of the 'neurds': Characterizing creative thought in terms of the
structure and dynamics of human memory. Creativity Research Journal, 22, 1 -- 13.
Gabora, L. (2011). An analysis of the Blind Variation and Selective Retention (BVSR) theory of
creativity. Creativity Research Journal, 23(2), 155-165.
Gabora, L. (2017). Honing theory: A complex systems framework for creativity. Nonlinear
Dynamics, Psychology, and Life Sciences, 21(1), 35-88.
Gabora, L. (2018). The neural basis and evolution of divergent and convergent thought. In O.
Vartanian & R. Jung (Eds.), The Cambridge Handbook of the Neuroscience of Creativity
(pp. 58-70). Cambridge MA: Cambridge University Press.
Gabora, L., & Ranjan, A. (2013). How insight emerges in distributed, content-addressable
memory. In A. Bristol, O. Vartanian, & J. Kaufman (Eds.), The neuroscience of creativity
(pp. 19-43). Cambridge, MA: MIT Press.
Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science,
7, 155 -- 170.
Giroux, H. A. (2014). Barbarians at the gates: Authoritarianism and the assault on public
education. Truthout. Retrieved from http://www.truth-out.org/news/item/28272-barbarians-
at-the-gatesauthoritarianism-and-the-assault-on-public-education
Greenberg, J., Reiner, K., & Meiran, N. (2012). Mind the trap: Mindfulness practice reduces
cognitive rigidity. PLoS ONE, 7(5), e36206. doi: 10.1371/journal.pone.0036206
Gruber, H. E. (1995). Insight and affect in the history of science. In R. J. Sternberg & J. E.
Davidson (Eds.), The nature of insight (pp. 397-431). Cambridge, MA: The MIT Press.
Guilford, J. P. (1950). Creativity. American Psychologist, 5, 444−454.
Haller, C. S. (2015). Mindful creativity scale (MCS): Validation of a German version of the
Langer mindfulness scale with patients with severe TBI and controls. Brain Injury, 29(4),
517-526. doi: 10.3109/02699052.2014.989906
Haller, C. S., Bosma, C. M., Kapur, K., Zafonte, R., & Langer, E. J. (2016). Mindful creativity
matters: Trajectories of reported functioning after severe traumatic brain injury as a function
of mindful creativity in patients' relatives: A multilevel analysis. Quality of Life
Research, 26(4), 893-902. doi: 10.1007/s11136-016-1416-1
17
Healey, K. (2015a). Disrupting Wisdom 2.0: The quest for 'mindfulness' in Silicon Valley and
beyond. Journal of Religion, Media and Digital Culture, 4(1), 67 -- 95.
Healey, K. (2015b). Contemplative media studies: Religions. Next steps in religion and popular
media, 6(3), 948 -- 968.
Hennessey, B. A., & Amabile, T. (2010). Creativity. Annual Review of Psychology, 61, 569 -- 598.
Hick, S., & Furlotte, C. R. (2009). Mindfulness and social justice approaches: Bridging the mind
and society in social work practice. Canadian Social Work Review, 26(1), 5-24.
Hirsh, J. B., Mar, R. A., & Peterson, J. B. (2012). Psychological entropy: A framework for
understanding uncertainty-related anxiety. Psychological Review, 119, 304-320.
Jennings, P. A. (2016). CARE for Teachers: A mindfulness-based approach to promoting
teachers' social and emotional competence and well-being. In K. A. Schonert-Reichl & R.
W. Roeser (Eds.), Handbook of mindfulness in education: Integrating theory and research
into practice (pp. 133-148). New York, NY: Springer.
James, W. (1923). The principles of psychology. New York, NY: Holt.
Kabat-Zinn, J. (2003). Mindfulness-based interventions in context: Past, present, and future.
Clinical Psychology: Science and Practice, 10, 144 -- 156.
Kaplan, C. A., & Simon, H. A. (1990). In search of insight. Cognitive Psychology, 22, 374−419.
Kounios, J., & Beeman, M. (2014). The cognitive neuroscience of insight. Annual Review of
Psychology, 65, 71 -- 93.
Lampert, N. (2013). A people's art history of the United States: 250 years of activist art and
artists working in social justice movements. New York, NY: The New Press.
Langer, E. J. (1989). Mindfulness. Longman: Addison-Wesley.
Lawlor, M. S., Schonert-Reichl, K. A., Gadermann, M., & Zumbo, B. D. (2013). A validation
study of the Mindful Attention Awareness Scale Adapted for children. Mindfulness, 5(6),
730 -- 741. doi http://dx.doi.org/10.1007/s12671-013-0228-4 .
MacKinnon, D. W. (1962). The nature and nurture of creative talent. American Psychologist, 17,
484 -- 495.
Magee, R. (2015, April). Breathing together through "I can't breathe": The ethics and efficacy
of mindfulness in working toward justice for all. [Keynote address, spring conference],
University of Massachusetts Center for Mindfulness in Medicine, Health Care, and Society.
Shrewsbury, MA. Retrieved from
http://www.fleetwoodonsite.com/ppSD2/catalog.php?id=18
Malchiodi, C. (2007). The art therapy sourcebook. New York, NY: McGraw-Hill.
Martindale, C. (1984). The pleasures of thought: A theory of cognitive hedonics. Journal of
Mind & Behavior, 5, 49 -- 80.
Martindale, C. (1995). Creativity and connectionism. In S. M. Smith, T. B. Ward, & R. A. Finke
(Eds.), The creative cognition approach (pp. 249−268). Cambridge MA: MIT Press.
Maslow, A. H. (1971). Creativity in self-actualizing people. In A. Rothenburg & C. R. Hausman
(Eds.), The creative question (pp. 86-92). Durham, NC: Duke University Press.
18
Langer, E., & Moldoveanu, M. (2000). The construct of mindfulness. Journal of Social Issues,
56(1), 1 -- 9.
Moon, B. (1999). The tears make me paint: The role of responsive art making in adolescent art
therapy. Art Therapy: Journal of the American Art Therapy Association, 16, 78 -- 82.
Moon, B. (2009). Existential art therapy: The canvas mirror. Springfield, IL: Charles Thomas
Pub.complete
Mouchiroud, C., & Bernoussi, A. (2008). An empirical study of the construct validity of social
creativity. Learning and Individual Differences, 18(4), 372-380.
doi: 10.1016/j.lindif.2007.11.008
Mumford, M. D., & Moertl, P. (2003). Cases of social innovation: Lessons from two innovations
in the 20th century. Creativity Research Journal, 15, 261-266.
Newell, A., Shaw, C., & Simon, H. (1957). The process of creative thinking. In H. E. Gruber, G.
Terrell, & M. Wertheimer (Eds.), Contemporary approaches to creative thinking (pp. 153 --
189). New York, NY: Pergamon.
Newell, A., & Simon, H. (1972). Human problem solving. Edgewood Cliffs NJ: Prentice-Hall.
Ohlsson, S. (1992). Information-processing explanations of insight and related phenomena. In M.
T. Keane & K. J. Gilhooly (Eds.), Advances in the psychology of thinking, 1 (pp. 1 -- 44). New
York, NY: Harvester Wheatsheaf.
Ostafin, B. D., & Kassman, K. T. (2012). Stepping out of history: Mindfulness improves insight
problem solving. Consciousness and Cognition, 21, 1031-1036.
Pennebaker, J. W. (1997). Writing about emotional experiences as a therapeutic process.
Psychological Science, 8(3), 162-166.
Pennebaker, J. W., Kiecolt-Glaser, J. K., & Glaser, R. (1998). Disclosure of traumas and immune
function: Health implications for psychotherapy. Journal of Consulting and Clinical
Psychology, 56, 239-245.
Ponder, L. M. (2012). An exploratory study of the potential impacts of yoga on self and
community: Creating mindfulness, self-actualization and social capital (Doctoral
dissertation, Clemson University). Retrieved from
http://ezproxy.library.ubc.ca/login?url=https://search-proquest-
com.ezproxy.library.ubc.ca/docview/1039269015?accountid=14656
Ren, J., Huang, Z., Luo, J., Wei, G., Ying, X., Ding, Z., ... & Luo, F. (2011). Meditation
promotes insightful problem-solving by keeping people in a mindful and alert conscious
state. Science China Life Sciences, 54(10), 961-965.
Richards, R. (2010). Everyday creativity: Process and way of life -- Four key issues. In J. C.
Kaufman & R. J. Sternberg (Eds.), The Cambridge handbook of creativity (pp. 189 -- 215).
New York, NY: Cambridge University Press.
Richards, R., Kinney, D., Lunde, I., Benet, M., & Merzel, A. (1988). Creativity in manic
depressives, cyclothymes, their normal relatives, and control subjects. Journal of Abnormal
Psychology, 97, 281−289.
19
Riley, S. (1999). Brief therapy: An adolescent intervention. Art Therapy: Journal of the
American Art Therapy Association, 16, 83 -- 86.
Riley, S., & Gabora, L. (2012). Evidence that threatening situations enhance creativity.
Proceedings of the 34th annual meeting of the Cognitive Science Society (pp. 2234 -- 2239).
Houston TX: Cognitive Science Society.
Rogers, C. (1959). Toward a theory of creativity. In H. Anderson (Ed.), Creativity and its
cultivation. New York, NY: Harper & Row.
Rowe, J. K. (2015a). Learning to love us-versus-them thinking. Open democracy. Retrieved from
https://www.opendemocracy.net/transformation/james-krowe/learning-to-love-us-versus-
them-thinking#
Rowe, J. K. (2015b). Zen and the art of social movement maintenance. Waging nonviolence.
http://wagingnonviolence.org/feature/mindfulness-and-the-art-of-social-
movementmaintenance.
Rhue, J. W., & Lynn, S. J. (1987). Fantasy proneness: Developmental antecedents. Journal of
Personality, 55, 121 -- 137.
Russ, S. W. (1993). Affect and creativity. Hillsdale, NJ: Erlbaum.
Runco, M. A. (2010). Divergent thinking, creativity, and ideation. In J. Kaufman & R. Sternberg,
(Eds.), The Cambridge handbook of creativity (pp. 414 -- 446). Cambridge, UK: Cambridge
University Press.
Schmertz, S. K., Anderson, P. L., & Robins, D. L. (2009). The relation between self-report
mindfulness and performance on tasks of sustained attention. Journal of Psychopathology
and Behavioral Assessment, 31, 60-66.
Serotkin, S. V. (2010). The relationship between self-actualization and creativity as a self-
growth practice. (California Institute of Integral Studies.) Retrieved from
http://ezproxy.library.ubc.ca/login?url=https://search-proquest-
com.ezproxy.library.ubc.ca/docview/816337923?accountid=14656
Shapiro, S. L., & Carlson, L. E. (2009). The art and science of mindfulness: Integrating
mindfulness into psychology and the helping professions. Washington, DC: American
Psychological Association.
Shapiro, S. L., Oman, D., Thoresen, C. E., Plante, T. G., & Flinders, T. (2008). Cultivating
mindfulness: Effects on well‐being. Journal of Clinical Psychology, 64(7), 840-862. doi:
10.1002/jclp.20491
Siegel, D. & Siegel, M. (2014). Thriving with uncertainty: Opening the mind and cultivating
inner well-being through contemplative and creative mindfulness. In A. Ie, C. T. Ngnoumen,
E. J., & Langer (Eds.), The Wiley Blackwell handbook of mindfulness (pp. 21-47).
Chichester, UK: John Wiley & Sons.
Siegel, D., Siegel, M., & Parker, S. C. (2016). Internal education and the roots of resilience:
Relationships and reflection as the new R's of education. In K. A. Schonert-Reichl & R. W.
Roeser (Eds.), Handbook of mindfulness in education: Integrating theory and research into
practice (pp. 47-63). New York, NY: Springer.
20
Siegelman, M. (1973). Parent behavior correlates of personality traits related to creativity in sons
and daughters. Journal of Consulting and Clinical Psychology, 40, 43 -- 47.
Simonton, D. K. (1994). Greatness: Who makes history and why. New York, NY: Guilford
Press.
Simonton, D. K. (2011). Creativity and discovery as blind variation: Campbell's (1960) BVSR
model after the half-century mark. Review of General Psychology, 15, 158-174.
Sowden, P., Pringle, A., & Gabora, L. (2014). The shifting sands of creative thinking:
Connections to dual process theory. Thinking & Reasoning.
doi: 10.1080/13546783.2014.885464
Sternberg, R. J., & Lubart, T. I. (1995). Defying the crowd: Cultivating creativity in a culture of
conformity. New York, NY: Free Press.
Sulloway, F. J. (1996). Born to rebel: Birth order, family dynamics, and creative lives. New
York, NY: Pantheon.
Terrell, J. E., Hunt, T. L., & Gosden, C. (1997). The dimensions of social life in the Pacific:
Human diversity and the myth of the primitive isolate. Current Anthropology, 38, 155 -- 195.
Thagard, P. (1980). Against evolutionary epistemology. In P. Asquith & R. Giere (Eds.), PSA
1980. East Lansing, MI: Philosophy of Science Association.
Tomasello, M. (1996). Do apes ape? In C. M. Heyes & B. G. Galef, Jr. (Eds.), Social learning in
animals. The roots of culture (pp. 319 -- 343). London, England: Academic Press.
Valentine, E. R., & Sweet, P. L. G. (1999). Meditation and attention: A comparison of the effects
of concentrative and mindfulness meditation on sustained attention. Mental Health, Religion
& Culture, 2, 59-70.
van Doesum, N. J., van Lange, D. A. W., & van Lange, P. A. M. (2013). Social mindfulness:
Skill and will to navigate the social world. Journal of Personality and Social
Psychology, 105(1), 86-103. doi: 10.1037/a0032540
Vázquez Campos, M. & Liz Gutiérrez, A.M. (2015). Subjective and objective points of view. In
M. Vázquez Campos & A.M. Liz Gutiérrez (Eds.), Temporal points of view: Studies in
applied philosophy, epistemology and rational ethics (pp. 59-104). Switzerland: Springer.
doi: 10.1007/978-3-319-19815-6_2
Walach, H., Buchheld, N., Buttenmuller, V., Kleinknecht, N., & Schmidt, S. (2006). Measuring
mindfulness: The Freiburg Mindfulness Inventory (FMI). Personality and Individual
Differences, 40, 1543 -- 1555.
Wallas, G. (1926). The art of thought. London: Cape.
Watson, E. (2007). Who or what creates? A conceptual framework for social creativity. Human
Resource Development Review, 6(4), 419-441. doi: 10.1177/1534484307308255
Weisberg, R. W. (1995). Prolegomena to theories of insight in problem solving: Definition of
terms and a taxonomy of problems. In R. J. Sternberg & J. E. Davidson (Eds.) The nature of
insight (pp. 157−196). Cambridge MA: MIT Press.
21
Young, S. (2016). What is mindfulness? A contemplative perspective. In K. A. Schonert-Reichl
& R. W. Roeser (Eds.), Handbook of mindfulness in education: Integrating theory and
research into practice (pp. 29-45). New York, NY: Springer.
Zajonc, A. (2016). Contemplative education. In K. A. Schonert-Reichl & R. W. Roeser (Eds.),
Handbook of mindfulness in education: Integrating theory and research into practice (pp.
17-28). New York, NY: Springer.
22
|
1801.01651 | 1 | 1801 | 2018-01-05T07:27:47 | Estimating the impact of structural directionality: How reliable are undirected connectomes? | [
"q-bio.NC"
] | Directionality is a fundamental feature of network connections. Most structural brain networks are intrinsically directed because of the nature of chemical synapses, which comprise most neuronal connections. Due to limitations of non-invasive imaging techniques, the directionality of connections between structurally connected regions of the human brain cannot be confirmed. Hence, connections are represented as undirected, and it is still unknown how this lack of directionality affects brain network topology. Using six directed brain networks from different species and parcellations (cat, mouse, C. elegans, and three macaque networks), we estimate the inaccuracies in network measures (degree, betweenness, clustering coefficient, path length, global efficiency, participation index, and small worldness) associated with the removal of the directionality of connections. We employ three different methods to render directed brain networks undirected: (i) remove uni-directional connections, (ii) add reciprocal connections, and (iii) combine equal numbers of removed and added uni-directional connections. We quantify the extent of inaccuracy in network measures introduced through neglecting connection directionality for individual nodes and across the network. We find that the coarse division between core and peripheral nodes remains accurate for undirected networks. However, hub nodes differ considerably when directionality is neglected. Comparing the different methods to generate undirected networks from directed ones, we generally find that the addition of reciprocal connections (false positives) causes larger errors in graph-theoretic measures than the removal of the same number of directed connections (false negatives). These findings suggest that directionality plays an essential role in shaping brain networks and highlight some limitations of undirected connectomes. | q-bio.NC | q-bio | Estimating the impact of structural directionality: How reliable are undirected
connectomes?
Penelope Kale1, 2, Andrew Zalesky3, Leonardo L. Gollo1, 2
1. QIMR Berghofer Medical Research Institute, Australia
2. The University of Queensland, Australia
3. Melbourne Neuropsychiatry Centre and Department of Biomedical Engineering, The University of Melbourne, Australia
Keywords: directionality, connectome, structural connectivity, graph theory, hubs, false positives
Abstract
Directionality is a fundamental feature of network connections. Most structural brain networks are intrinsically
directed because of the nature of chemical synapses, which comprise most neuronal connections. Due to limitations of
non-invasive imaging techniques, the directionality of connections between structurally connected regions of the
human brain cannot be confirmed. Hence, connections are represented as undirected, and it is still unknown how this
lack of directionality affects brain network topology. Using six directed brain networks from different species and
parcellations (cat, mouse, C. elegans, and three macaque networks), we estimate the inaccuracies in network measures
(degree, betweenness, clustering coefficient, path length, global efficiency, participation index, and small worldness)
associated with the removal of the directionality of connections. We employ three different methods to render directed
brain networks undirected: (i) remove uni-directional connections, (ii) add reciprocal connections, and (iii) combine
equal numbers of removed and added uni-directional connections. We quantify the extent of inaccuracy in network
measures introduced through neglecting connection directionality for individual nodes and across the network. We find
that the coarse division between core and peripheral nodes remains accurate for undirected networks. However, hub
nodes differ considerably when directionality is neglected. Comparing the different methods to generate undirected
networks from directed ones, we generally find that the addition of reciprocal connections (false positives) causes
larger errors in graph-theoretic measures than the removal of the same number of directed connections (false
negatives). These findings suggest that directionality plays an essential role in shaping brain networks and highlight
some limitations of undirected connectomes.
Introduction
Connectomes provide a comprehensive network description of structural brain connectivity (Sporns et al., 2005).
Large-scale connectomes mapped in humans are typically represented and analyzed as undirected networks, due to the
inability of non-invasive connectome mapping techniques to resolve the directionality (afferent or efferent) of white
matter fibers. Reducing an inherently directed network such as the connectome to an undirected network is a
simplification that may introduce inaccuracies in graph-theoretic analyses. For example, the flow of action potentials
along an axon is mostly only ever in one direction, and thus analyses of information flow are critically dependent on
connection directionality. This study aims to systematically and comprehensively characterize the impact of
representing and analyzing connectomes as undirected networks.
At the neuronal level, the connections between nodes (neurons) are given by synapses, and the great majority of them
are chemical, which have distinctive pre- and post-synaptic terminals determining the direction of neurotransmitter
flux (Kandel et al., 2000). This structural feature of chemical synapses emphasizes the importance of directionality for
the connections, and therefore for the whole network. Invasive techniques to map connectomes such as tract tracing
(Scannell et al., 1999, Kötter, 2004, Sporns et al., 2007, Oh et al., 2014) or electron microscopy (White et al., 1986,
Achacoso and Yamamoto, 1992) can detect the directionality of the connections. Conversely, human connectomes are
currently mapped with non-invasive tractography methods performed on diffusion-weighted magnetic resonance
imaging data (Assaf and Basser, 2005, Hagmann et al., 2008, Tournier et al., 2012). While methods for improving the
quality of diffusion-based connectomes have advanced in recent years, and numerous tractography algorithms have
been developed to reconstruct axonal fiber bundles, they cannot provide any information about the directionality of the
connections. Therefore, analyses of the human connectome, as well as modeling studies that use the human
connectivity matrix, are compromised by the lack of information regarding directionality, which is one of the most
fundamental features of complex networks.
In the absence of directionality, networks are considered undirected and therefore the connections only represent the
existence of a relationship between nodes. This is the case for scientific co-authorship networks (Newman, 2004), film
actor networks (Watts and Strogatz, 1998), and functional networks defined by symmetric functions such as the
Pearson correlation (Biswal et al., 1995) or the phase locking value (Aydore et al., 2013). Among others, studies of
tractography-derived human brain networks have revealed a variety of important features such as hub regions (van
den Heuvel and Sporns, 2013), modularity and clustering (Sporns, 2011, Sporns and Betzel, 2016), small worldness
(Bassett and Bullmore, 2006, Medaglia and Bassett, 2017), core-periphery structure (Hagmann et al., 2008) and the
existence of a rich club (van den Heuvel and Sporns, 2011). These topological properties are not specific to the human
brain. Comparisons across many species have recapitulated these features (Harriger et al., 2012, Towlson et al., 2013,
Betzel and Bassett, 2016, van den Heuvel et al., 2016). However, the topological characteristics of connectomes, as well
as many other graph-theoretic measures, are affected by the directionality of connections (Rubinov and Sporns, 2010).
When directionality cannot be identified, undirected representations of connectomes are incomplete. Undirected
networks inform the presence of a relationship between two brain regions. But these networks lack information about
the asymmetry of this relationship. For example, if a directed network is represented as an undirected network, uni-
directional connections are either present, which can be interpreted as a spurious addition of a reciprocal connection
(false positives), or overlooked (false negatives). More specifically, if a uni-directional connection exists from node u to v,
but not from v to u, then the undirected representation of this connection is either: (i) an undirected connection
between u and v, which can be construed as admitting a false positive from node v to u; or, (ii) absence of an undirected
connection between u and v, which can be construed as a false negative from node u to v. In either case, a potential
error (false positive or false negative) is introduced to the undirected network.
Beyond the effect of directionality, connectomes also contain errors in the balance between overlooked and spurious
connections owing to imprecisions in currently available mapping techniques (Calabrese et al., 2015, Donahue et al.,
2016). Although both error types impact the network topology, spurious (false positive) connections introduce
inaccuracies in a few graph-theoretic measures (network clustering, efficiency and modularity) in different
connectomes that are at least twice as large as those found with the same number of overlooked (false negative)
connections (Zalesky et al., 2016). This finding indicates that the importance of specificity is much greater than
sensitivity for general connectivity in which false positives could be any absent connection and false negatives, any
present connection. However, the impact of representing a directed connection as undirected, which, for practical
purposes, is typically indistinguishable from a bidirectional connection, is currently unknown. Therefore, when
directed networks are mapped with techniques that cannot infer directionality, it is important to establish what
undirected representation is the most detrimental with respect to directionality: admitting spurious reciprocal
connections (false positives) or overlooking uni-directional connections (false negatives).
Moreover, the effect of directionality on the identification of network hubs may also be important as hubs play an
important role for normal brain function (van den Heuvel et al., 2012, Mišić et al., 2015) as well as in neuropsychiatric
disorders (Bassett et al., 2008, Crossley et al., 2014, Fornito et al., 2015). But how are these highly connected regions
affected by directionality? Does the classification of nodes into hubs still hold if directionality is taken into account?
Furthermore, to what extent do graph-theoretic measures at the node level remain valid? The characterization of the
human brain as an undirected network is often overlooked and requires investigation.
The aim of this study is to understand the limitations of analyzing inherently directed connectomes as undirected
networks. Beginning with directed connectomes of the macaque, cat, mouse, and Caenorhabditis elegans (C. elegans),
we study how seven graph-theoretic measures are affected as we progressively modify uni-directional connections,
either deleting them or making them undirected. More specifically, we consider three schemes to progressively
eliminate directionality information: removing uni-directional connections (creating false negatives), adding reciprocal
connections to existing uni-directional connections (creating false positives), and removing one uni-directional
connection for each reciprocal connection added, thus preserving the density and mean degree of the original network.
We show how essential network features, such as the identification and classification of hubs, are affected by
perturbations in directionality. Moreover, we quantify how graph-theoretic measures are affected at both the node and
network level and determine whether false positive or false negative uni-directional connections are more detrimental
to the characterization of graph-theoretic measures.
Materials and Methods
Connectivity Data
Following a comparative connectomics approach (van den Heuvel et al., 2016), we analyzed structural connectivity
data from several species and various parcellations including three macaque connectomes, a cat and mouse
connectome, and a C. elegans nervous system connectome (Fig. 1). Each network possesses a different number of
nodes, proportion of uni-directional connections, modularity, and network density (see Supplementary Table 1).
Crucially, these networks include information on the directionality of connections (all networks are directed) obtained
through invasive techniques that have different proportions of connection reciprocity (Garlaschelli and Loffredo, 2004).
Among the meso- and macro-scale connectomes, nodes represent cortical regions and the directed connections
represent axons or white matter fibers linking these regions via chemical synapses. In the case of the micro-scale C.
elegans connectome, nodes represent neurons, the directed connections represent chemical synapses, and the electrical
synapses (or gap junctions) are bidirectional connections.
To accommodate the analysis of such a wide range of directed connectomes, the strength of connections was
disregarded (for the cat and mouse connectomes) to make each network binary. This procedure allowed us to
characterize all connectomes using the same methods for binary and directed networks as a first step to understand the
role of directionality in structural brain networks. Other high-quality weighted connectomes can be used in future
Page 2 of 29
studies (Bezgin et al., 2012, Markov et al., 2012, Shih et al., 2015, Ypma and Bullmore, 2016, Gămănuţ et al., 2017). As
recently reported, the combination of both directionality and weight can be crucial to uncover relationships between
structural connectivity and univariate brain dynamics (Sethi et al., 2017).
Macaque networks: The first macaque network, used in a study by Honey et al. (2007), (with number of nodes N = 47
and connections E = 505, Fig. 1A) follows the parcellation scheme of Felleman and Van Essen (1991) including the
visual and sensorimotor cortex, and motor cortical regions. Relevant data was collated in the CoCoMac database
(Modha and Singh, 2010) following the procedures of Kötter (2004) and Stephan et al. (2001), and translated to the
brain map using coordinate independent mapping (Stephan et al., 2000, Kötter and Wanke, 2005).
The second macaque connectome (N = 71 and E = 746, Fig. 1B) was derived from a whole cortex model generated by
Young (1993) with regions of the hippocampus and amygdala eliminated. The parcellation was based mostly on the
scheme by Felleman and Van Essen (1991), except for the fields of the superior temporal cortex (Yeterian and Pandya,
1985). Yeterian and Pandya (1985) utilized an autoradiographic technique (radioactively labeled amino acids) to
establish the existence and trajectory of fibers.
The final macaque connectome (N = 242 and E = 4090, Fig. 1C) was generated by Harriger et al. (2012). This network
comprises anatomical data from over 400 tract tracing studies collated in the CoCoMac database (Modha and Singh,
2010) following the procedures of Kötter (2004) and Stephan et al. (2001), focusing on the right hemisphere with all
subcortical regions removed as well as regions without at least one incoming and one outgoing connection.
The data collated for the CoCoMac database used a range of tracer substances (with anterograde, retrograde or
bidirectional transport properties) and methods (as discussed in Stephan et al. (2001)). Each contributing study must
discern a source and target for the connection. If the reciprocal direction had not been tested for, the connection was
assumed to be uni-directional. Some connections have been confirmed to be uni-directional, for example, the
connection from V2 to FST, see Boussaoud et al. (1990). Regarding macaque connectomes Felleman and Van Essen
(1991) have also suggested that the reciprocity of connections may vary between individuals.
Cat network: The cat matrix is a connectome reconstructed by Scannell et al. (1999) and curated from a database of
thalamo-cortico-cortical connections from a large number of published studies in the adult cat. The parcellation was
based on a previous scheme by Reinoso-Suarez (1984) and adapted by Scannell et al. (1995). Areas ALG, SSF, SVA, DP,
Amyg and 5m were discarded (and some regions grouped) to create a weighted network (N = 52 and E = 818, Fig. 1D).
This connectome was generated from the available data across numerous studies. It is noted that each study used a
different type of anterograde and/or retrograde tracer, methodology and parcellations. Some connections lacked data
on the existence of a reciprocal direction between brain regions (these were left as uni-directional), and all connections
between the cortex and thalamus were assumed to be reciprocal.
Mouse network: We obtained the mouse connectome (N = 213 and E = 2105, Fig. 1E) from the Allen Mouse Brain
Connectivity Atlas generated by Oh et al. (2014). The major advantage of this connectome is that the connectivity data,
obtained at a cellular level (axons and synaptic terminals), is generated for the whole mouse brain. Therefore, all 469
individual experiments use the same anterograde tracer and consistent techniques. Each brain is applied to a 3D
template, which itself is averaged across 1231 brain specimens, and the regions are matched against the Allen reference
atlas (Oh et al., 2014). We thresholded this dense and weighted network using the disparity filter (Serrano et al., 2009)
maintaining only connections with a p-value smaller than 0.05. Thresholding was performed such that the resulting
network was binary.
C. elegans network: The C. elegans nervous system matrix (N = 279 and E = 1943, Fig. 1F) was collated by Varshney et
al. (2011), and includes data mapped by White et al. (1986) using electron microscopy, in addition to various other
sources (White et al., 1976, Durbin, 1987, Hall and Russell, 1991). This microscale connectome is comprised of a
directed chemical synapse network and an undirected gap junction network. Although gap junctions may possess
directionality, this has not yet been demonstrated in C. elegans. For the purpose of analysis, the connections from the
gap junction network were treated as bidirectional connections.
Page 3 of 29
Figure 1: The six connectomes analyzed in this study. Brain and connectome for three different parcellations of the macaque cortex (A) nodes
N=47 (Honey et al., 2007), (B) N=71 (Young, 1993), and (C) N=242 (Harriger et al., 2012), as well as three additional species including a (D) cat
(Scannell et al., 1999), (E) mouse (Oh et al., 2014), and (F) C. elegans (White et al., 1986, Varshney et al., 2011). The connectomes represent
connectivity matrices with rows and columns denoting brain regions (or nodes), and the elements within the matrices denoting the presence (filled)
or absence (blank) of a connection between two regions. Uni-directional connections are highlighted in light blue (with the number of uni-directional
connections stated below each connectome) and the nodal regions are arranged into modular communities. The bars below each connectome display
the density of each network (A= 0.234, B= 0.15, C= 0.07, D= 0.308, E= 0.073, F=0.063) and the proportion of uni-directional and bidirectional
connections. The latter is segmented to display the proportion of uni-directional connections between modules (dark green: A= 0.123, B= 0.046, C=
0.238, D= 0.142, E= 0.304, F= 0.165) and within modules (light green: A= 0.117, B= 0.129, C= 0.255, D= 0.117, E= 0.404, F= 0.232) separately, as well
as proportion of bidirectional connections between modules (dark purple: A= 0.214, B= 0.236, C= 0.147, D= 0.21, E= 0.064, F= 0.147) and within
modules (light purple: A= 0.547, B= 0.59, C= 0.359, D= 0.536, E= 0.229, F= 0.457).
Perturbed Networks
To investigate the effects of directionality on the characteristics of the brain, each empirical connectome was altered by
progressively removing connection directionality information, generating a spectrum of perturbed networks. This
spectrum comprised the empirical connectome at one end, and a fully undirected representation of the connectome at
the opposite end. For this purpose, the empirical networks were considered to be approximately the ground-truth
connectomes for a given parcellation. Figure 2 illustrates the three different approaches used to generate perturbed
networks for the macaque (N=47) connectome. The empirical connectome is shown in Fig. 2A, and the uni-directional
connections of this network are shown in Fig. 2B. Perturbed networks (Fig. 2C-E) were generated by altering the
directionality or presence of the uni-directional connections. In this example, we only show the extreme case in which
all information about connection directionality is removed, yielding a fully undirected perturbed network.
For further analyses we present three schemes that were developed to progressively eliminate connection
directionality information from the empirical connectomes, yielding perturbed networks that increasingly resembled
undirected networks.
False negative perturbed networks: The first perturbed network was generated by removing a fixed number of
randomly chosen uni-directional connections, leading to a connectome with false negative uni-directional connections
(FN network, Fig. 2C). The perturbed network was undirected in the extreme case when all uni-directional connections
were removed. This perturbation assumes that uni-directional connections are weaker in strength (weight) relative to
their bidirectional counterparts, and thus uni-directional connections are most vulnerable to elimination with weight-
based thresholding procedures (Rubinov and Sporns, 2010). Such thresholding is commonly used to eliminate weak
connections obtained with tractography, which are often attributed to noise or error (Maier-Hein et al., 2017). As an
example, the majority of the weighted mouse connectome is comprised of uni-directional connections (57%), and they
are also weaker than bidirectional connections. The mean of the strength of uni-directional connections is 0.066,
whereas the mean strength of bidirectional connections is 0.165, which is significantly weaker (P<10-45, Welch's t-test).
False positive perturbed networks: If the weight of a uni-directional connection exceeds the weight-based threshold,
the connection will be represented in the perturbed network as an undirected connection (i.e. a uni-directional
connection from node u to v becomes an undirected connection between nodes u and v). In this case, the undirected
connection is treated as a bidirectional connection, and thus construed as a false positive. To model this case, we
generated perturbed networks by adding reciprocal connections to a fixed number of randomly chosen existing uni-
directional connections, leading to a perturbed network with false positive reciprocal connections (FP network, Fig.
2D). In the extreme case when all reciprocal connections were added, the perturbed network effectively became an
undirected network.
Density-preserving perturbed networks: Finally, to preserve basic properties of the empirical connectome, an
additional perturbed connectome termed the density-preserving network was generated (DP network, Fig. 2E). In this
perturbed connectome, for each reciprocal connection added to a uni-directional connection, another uni-directional
connection is removed (at randomly selected locations). The DP network has an equal number of false negative and
positive connections and also preserves the mean degree of the empirical connectome, but not the degree of each node.
To generate undirected perturbed networks, we progressively applied one of the above three schemes to randomly
chosen uni-directional connections in the empirical connectomes until a desired proportion of connections were
changed. We generated perturbed networks in which 5%, 10%, 20% and 100% of directed connections were altered
(eliminated or the reciprocal connection added). This process was repeated for multiple trials to generate an ensemble
of perturbed networks. Ensemble averages for all graph-theoretic measures were then computed. Each perturbed
network was associated with a rewiring scheme (FN, FP, and DP) and a proportion of changed connections.
Supplementary Table 2 provides the details of the proportion of uni-directional connections altered in the perturbed
networks and other relevant parameters used for each analysis.
The perturbed networks can comprise isolated nodes that are not connected to any other nodes (see Supplementary
Fig. 1). Isolated nodes are more likely to occur in the FN perturbed networks, potentially having a greater impact on
graph-theoretic measures as more connections are changed. Therefore, in cases where only a subset of uni-directional
connections are modified (<100%), the trials that cause nodes to become disconnected are rejected.
Figure 2: Structural connectome for the macaque N=47 cortex and perturbed undirected variants, with an exemplar sub-network. Sub-
network (top) encompassing the PITd region (white node) and neighboring nodes, the adjacency matrix (middle), and the entire network (bottom)
for: (A) Macaque empirical connectome with the community modules outlined in red; (B) Uni-directional connections of the connectome; (C)
Connectome with uni-directional connections removed (false negative network); (D) Connectome with reciprocal connections added to uni-
directional connections (false positive network); (E) Connectome with one randomly selected reciprocal connection added to a uni-directional
connection for each randomly selected uni-directional connection removed (density-preserving network). In each connectome the connections
linking PITd (dorsal posterior inferotemporal) to the rest of the network are colored orange.
Network Measures
Connectome analyses were performed using a range of common graph-theoretic network measures (Costa et al., 2007).
These measures enable the quantitative comparison of connectomes across species and neuroimaging techniques while
remaining computationally inexpensive (Rubinov and Sporns, 2010). Furthermore, the graphical properties of cortical
systems have previously been associated with functional connectivity and evolutionary adaptations in behavior and
cognition (Bullmore and Sporns, 2012, van den Heuvel et al., 2016). For each empirical connectome and associated
perturbed network, we computed several graph-theoretic measures (see Supplementary Table 3), using the Brain
Connectivity Toolbox (Rubinov and Sporns, 2010). Graph-theoretic measures for directed networks were used in all
cases where applicable.
Page 5 of 29
Measures of centrality: The degree of each node was calculated as the sum of the in- and out-degree, or the sum of all
directed connections connecting that node to the rest of the network (Rubinov and Sporns, 2010). Network centrality
identifies nodes that act as important points of information flow between regions. We used a betweenness
centrality measure, defined as the fraction of all the shortest paths between regions that pass through a particular node
(Freeman, 1978). The participation index or coefficient describes the proportion of intra- and inter-modular
connections linking each node (Guimera and Amaral, 2005a). As shown in Supplementary Table 3, we used the out-
participation index with the Louvain algorithm (Blondel et al., 2008) to define network modules (Rubinov and Sporns,
2010). Further details about module delineation are provided below.
Measures of functional segregation: We calculated the clustering coefficient, a measure describing the proportion of a
node's neighbors that are connected to each other (Fagiolo, 2007). In undirected networks it is calculated as the
probability that two connections (linking three nodes) will be closed by a third connection to form a triangle. In
directed networks however, a set of three nodes can generate up to eight different triangles. The function utilized in
this study, clusteringcoef_bd, (Rubinov and Sporns, 2010), takes this into account.
Measures of functional integration: A path is defined as a sequence of nodes and connections that represent potential
routes of information flow between two brain regions. In a directed network, connections comprising a path must be
arranged such that the head of one connection always precedes the tail of the subsequent connection.
The characteristic path length for each network was calculated as the average shortest distance between all pairs of
nodes (Watts and Strogatz, 1998). We also calculated the global efficiency of each network as the average nodal
efficiency, which is the reciprocal of the harmonic mean of the shortest path length between all pairs of nodes (Latora
and Marchiori, 2001).
Small worldness: Lastly, we measured the small-world characteristics of each network (Watts and Strogatz, 1998). For
each node and for the network (see Supplementary Table 3), the small-world index was classified as the clustering
coefficient divided by the characteristic path length of the network, with a comparison to a directed random network,
makerandCIJ_dir, (Rubinov and Sporns, 2010), unless otherwise stated (Humphries and Gurney, 2008). This index
combines local and global topological properties and has been linked to network efficiency (Bassett and Bullmore,
2006).
Community detection and modularity: We generated consensus matrices to describe the community structure of
each empirical connectome (Lancichinetti and Fortunato, 2012). Specifically, 100 runs of the Louvain modularity
algorithm (Blondel et al., 2008) were performed to generate a set of modular decompositions for each empirical
connectome. The different runs did not necessarily yield identical decompositions due to degeneracy of the solution
space and the stochastic nature of the algorithm. A consensus modularity matrix was determined for the 100
decompositions such that each element in the consensus matrix stored the proportion of runs for which a particular
pair of nodes comprised the same module. The consensus modularity matrix was then thresholded (retaining values
>0.4) and 100 runs of the Louvain algorithm were performed on the thresholded consensus matrix. This process was
iterated until the consensus matrix converged and did not change between successive iterations. The macaque N=47
network required a greater number of iterations before a consistent community structure could be achieved (macaque
N=47: 408, macaque N=71: 2, macaque N=242: 5, cat: 4, mouse: 36, C. elegans: 2).
For the perturbed networks with all uni-directional connections altered, a single consensus matrix and consistent
modularity was obtained for the FN and FP networks. For the rank correlation-coefficient analyses, the modularity for
each perturbed network remained the same as that assigned to the associated empirical connectome. These perturbed
networks only had a small percentage of uni-directional connections altered (5%). With these measures we intended to
isolate the effect of directionality on the ranking of nodes by each graph-theoretic measure, and, therefore, used the
empirical consensus modularity for the (participation index) calculations on each type of perturbed network.
For DP networks with 100% of connections altered, a consensus matrix was obtained for each trial (see Supplementary
Table 2 for more details). For other perturbed networks where 5%, 10% and 20% of uni-directional connections are
altered, consensus modularity matrices were obtained for each run (50 runs, see Supplementary Table 2) and for each
type of network (FN, FP, and DP).
Classification of Highly Connected Regions
Core nodes were determined using the core-periphery algorithm, function core_periphery_dir from the Brain
Connectivity Toolbox (Rubinov and Sporns, 2010), with gamma=1, which subdivides all nodes in the network into
either core or periphery groups of similar size. Hubs were defined as regions with a degree at least one standard
deviation above the mean (Sporns et al., 2007), and super hubs were classified as those with a degree of at least 1.5
standard deviations above the mean (see Fig. 4A for an example). Super hubs were defined to evaluate the robustness
of hub nodes to the progressive removal of connection directionality. More specifically, we aimed to assess whether
super hubs would be demoted to hubs or non-hub nodes as directionality information was lost.
Page 6 of 29
We tested the resilience of the classification of nodes belonging to the core of the network, or the set of hubs and super
hubs. For each perturbed network, the accuracy of the classification of nodes into each of these three groups (core,
hubs, and super hubs) was compared with the empirical connectomes. For each group, the accuracy, or matching index,
A was computed taking into account the number of nodes with common classification and the number of mismatched
nodes that had a different classification between the empirical and the perturbed networks. More precisely, A was given
by the simple matching index:
𝐴 =
(1)
,
𝐶
𝐶+ 𝑁𝑒 − 𝐶 + (𝑁𝑏 − 𝐶)
Where C was the number of overlapping nodes within the same group between the empirical and perturbed networks;
Ne was the number of nodes within this group for the empirical connectome; and Nb was the number of nodes within
this group for the perturbed network. This measure of accuracy attained a minimum of 0 when there was no overlap
between the connectomes and a maximum of 1 for a perfect overlap.
The participation index can be used to classify nodes, and has been applied to hubs (Guimera and Amaral, 2005b). Hubs
with large participation index connect areas from different modules. Supplementary Table 4 lists the regions classified
as hubs for each empirical network, as either connector (with a participation index Y > 0.35) or provincial (Y ≤ 0.35)
hubs. Consistent with other studies (Sporns et al., 2007), node degree (as the sum of the in- and out- degree) was used
to define the set of hubs based on their topological role within the network.
Quantifying Changes in Network Measures
D
N
i=1
(2)
RSI = Ei − Bi
,
To investigate changes in node-specific features between the empirical connectomes and corresponding perturbed
networks, we developed a measure to quantify the change in the ranking of nodes. Nodes can be ranked with any of a
number of graph-theoretic measures. The Rank Shift Index (RSI) represents the sum of the absolute value of the
difference between the ranking of the empirical (E) and perturbed (B) matrices for each node, divided by the maximum
possible difference (D) in which the ranks of the network are reversed:
A RSI of zero indicates no change, and an index of one indicates a complete inversion in the rank order (see Fig. 5).
Node-level changes were also measured by the Spearman rank correlation (Spearman, 1904) and Kendall coefficient
(Kendall, 1938).
Results
To understand the effects of neglecting connection directionality on the structural properties of connectomes, we
compared several directed brain networks across multiple species, including three macaque connectomes (with
different parcellation schemes), a cat, a mouse, and a C. elegans connectome. The characteristics of each of these
networks were analyzed using a range of network measures: degree, betweenness centrality, clustering coefficient,
characteristic path length, global efficiency, participation index, and small world index.
We altered uni-directional connections according to one of three schemes (see Methods) to progressively eliminate
information about connection directionality. We then quantified the inaccuracies in graph-theoretic measures admitted
through this loss of directionality information. We begin with the degree-preserving (DP) scheme and consider the
extreme case in which all uni-directional connections are eliminated, resulting in an undirected network. In particular,
we compare the network characteristics of selected regions of interest (ROIs) across the empirical connectomes and
single-trial DP counterparts (Fig. 3). These ROIs (shown as the red matrix entries in Fig. 3A) occupy peripheral
locations in the network topology, have low degree, and the sub-network of the local neighborhood surrounding each
ROI can be clearly represented (Fig. 3B). From the empirical to the DP sub-networks, uni-directional connections are
eliminated and made bidirectional, resulting in changes to graph-theoretic measures characterizing these regions.
Figure 3C illustrates the relative graph-theoretic metrics at these exemplar regions for the empirical and DP sub-
networks. Although the mean degree of the DP network is preserved, at the node level, the degree may increase or
decrease depending on whether the uni-directional connections surrounding the node of interest received more false
positive or false negative alterations. Likewise, clustering and small worldness also exhibit trial-dependent changes
based on how the neighbors of these exemplar regions and the whole network topology are affected.
Page 7 of 29
Figure 3: Graph-theoretic measures for a specific region of interest from each empirical and density-preserving connectome. (A) Empirical
(blue) and density-preserving (red, an illustrative single trial with 100% of uni-directional connections altered) connectomes. Nodal regions are
arranged into modular communities and the connections connecting the region of interest to the rest of the network in the empirical connectome are
colored red. (B) Labels for each region of interest (top), and sub-networks of the local neighborhood around each region of interest (white node). (C)
Graph theoretic measures at the selected brain region for the empirical and density-preserving networks. Graph-theoretic measures are as follows:
→). *Normalized by the maximum value of that measure across all nodes in their
K=Degree, C=Clustering coefficient and S=Small-world index (𝑆𝑖
respective network. PITd: dorsal posterior inferotemporal, A32: anterior cingulate area 32, 28m: medial entorhinal cortex, AAF: anterior auditory
field, MOB: main olfactory bulb, VC05: ventral cord neuron 5.
Highly Connected Regions
Connectivity across brain regions and connections is heterogeneously distributed. Hub nodes are identified as the most
connected neural regions, and have enhanced importance in information integration for cognitive functions (van den
Heuvel and Sporns, 2013). Hub nodes can be further classified based on their participation index as either provincial or
connector hubs, depending on their level of intra- vs. inter-module connectivity (Guimera and Amaral, 2005b, Sporns et
al., 2007). Provincial hubs, with a high intra-module degree and low participation index, are thought to facilitate
modular segregation. Conversely, connector hubs, with a higher participation index, are thought to assist with
intermodular integration (Rubinov and Sporns, 2010). When hub regions are more densely connected among
themselves than to other nodes they form a 'rich club', consisting of a central but costly backbone of pathways that
serves an important role in global brain communication (Colizza et al., 2006, van den Heuvel et al., 2012, Aerts et al.,
2016). Hence, alterations to directionality at hub nodes influence the network activity observed in functional
connectivity. But how is the identification and characteristics of these highly significant hub regions affected when
directionality is modified?
Inaccuracies may be introduced to node-specific graph-theoretic measures as connection directionality information is
lost. By comparing the empirical connectomes to corresponding perturbed networks with all uni-directional
connections eliminated according to the DP scheme, we see that peripheral, core and hub nodes are all impacted (Fig.
4). Even the degree, a fundamental network characteristic, is affected in these perturbed networks, as shown in Fig. 4A
for each cortical area in the macaque N=47 connectome. In particular, the degree of some hub and super-hub nodes
falls below the threshold used for their classification in the empirical connectome. This implies that hub nodes
identified based on degree can be inaccurate when directionality within the network is neglected or unknown. To
further investigate this, we redefined core, hub, and super-hub nodes for each perturbed network, and calculated their
accuracy according to the empirical connectome. Fig. 4B shows the percentage of nodes that retain the same
classification for core, hub and super-hub nodes across all perturbed networks. We find that the estimation of core
nodes from the perturbed networks were the most accurate compared to the empirical connectomes (mean=86.7%).
However, the estimation of hubs and super hubs is less precise (mean=79% and 68.2% respectively). The accuracy of
nodes belonging to core, hub, and super-hub was tested with paired sample t-tests and found to be significantly
different. Core (including results from all connectomes and each type of perturbed network) vs. hubs P=0.0027, core vs.
super hubs P=0.00001, and hubs vs. super hubs P=0.003. In Supplementary Fig. 2 these results are shown for each type
of perturbed network and connectome separately.
A recent study in the mouse brain (Sethi et al., 2017) showed a strong correlation between the in-degree characteristics
of a brain region and its resting state functional-MRI dynamics. We therefore sought to investigate in- and out-degree
separately. Supplementary Fig. 3A and B display the in- and out-degree of all cortical regions in the macaque N=47
empirical connectome and perturbed networks. In this case, the delineation of hubs and super-hub nodes depends on
the directed degree, and therefore a different set are identified in Figures 2A and B. However, due to the methodology
for generating the perturbed networks, the resulting in- and out-degree of each node becomes equal. This is because
(when 100% of uni-directional connections are altered) the only remaining connections in each case (FN, FP or DP) are
represented as bidirectional and therefore, each region has the same number of incoming connections as it has
outgoing connections. Previous studies in the cat connectome have found that high in-degree nodes also show (on
average) a high out-degree as well. In this connectome, 66% of rich-club nodes (defined by the summed degree) had a
higher in-degree than out-degree (de Reus and van den Heuvel, 2013). A comparison across the connectomes analyzed
in this study (Supplementary Fig. 3C) showed that four out of six sets of hub regions had a higher mean in-degree than
out-degree. The mouse connectome however, was an interesting case for which all hub regions had a much larger out-
degree.
Page 9 of 29
Figure 4: Identification of hubs, changes in graph-theoretic measures at the node level, and provincial/connecter hub classification. (A)
Cortical areas of the macaque N=47 connectome sorted by degree for the empirical and each perturbed network. Hubs are defined as nodes that have
a total degree (in-degree plus out-degree) one standard deviation above the mean, and super hubs are defined as nodes that have a degree 1.5
standard deviations above the mean. The density-preserving results are from an illustrative single-trial and show the standard deviation in degree for
each node (over 1000 trials). (B) Percentage of core, hub, and super-hub nodes across the perturbed networks of all six connectomes that retain
correct classification according to their empirical connectome (as the mean over 1000 trials). (C) Change in the participation index of each brain
region from the empirical macaque N=47 connectome to an illustrative case of the density-preserving network. (D) Identification and classification of
hub nodes for the empirical (blue) macaque N=47 connectome and an illustrative case of the density-preserving (red) network. The dotted line
represents the hub definition based on the degree and the dashed line represents the sub-classification of hubs as either connector (Y > 0.35) or
provincial (Y ≤ 0.35), based on the participation index. (E) Mean probability (across all connectomes over 1000 trials) that hub nodes will cross over
either, or both of the threshold lines following density-preserving alterations in directionality, resulting in a classification that is inconsistent with the
empirical connectomes. (A-E) Each perturbed network has 100% of uni-directional connections altered. Hub nodes are defined in the empirical
network and remain the same in the perturbed networks.
Page 10 of 29
Next, we investigate the classification of hubs based on the participation index. In comparison to peripheral regions, the
participation index of hub nodes is more resilient as illustrated in Fig. 4C as the change for each region from the
empirical macaque N=47 connectome to a (typical) DP example network. Because peripheral nodes have a low degree,
the alterations in directionality may affect a larger proportion of these connections. Therefore, peripheral regions often
show greater change in the participation index than both core and hub nodes. As illustrated in Supplementary Fig. 4,
this also occurs for other graph-theoretic measures.
The relationship between participation index and degree for the set of hub nodes (defined in the empirical
connectome) are displayed in Fig. 4D for the empirical macaque N=47 connectome and an illustrative DP network.
Directionality alterations to the network cause changes in these measures, both of which were used to define and
classify the set of hubs in the empirical connectome. As such, some of these regions in the DP network exceed the
degree and participation index thresholds (degree K=1 SD above the mean and Y=0.35) resulting in misclassifications
according to the empirical network. Across all connectomes, hub nodes are more likely to lose their classification based
on degree, indicating that the definition of hubs based on the degree is on average 3.5 times more vulnerable to changes
in directionality in comparison to the misclassification of hubs based on the participation index (Fig 4E and
Supplementary Fig. 5). Supplementary Fig. 6 displays the number of core, hub, and super hubs across the connectomes
(A: mean, B: individually), as defined in the empirical and each perturbed network.
Quantifying the Errors in Node Rank when Directionality is Lost
All the results presented thus far have pertained to perturbed networks in which all uni-directional connections are
altered, yielding perturbed networks that are effectively undirected. Next, we investigate the impact of losing only a
small proportion of connection directionality information. To this end, we generate perturbed networks in which the
proportion of uni-directional connections altered is 5%. Changes in node-specific network measures were quantified
using the rank-shift index (RSI, see Methods). This measure calculates the change in the ranking of nodes by a specific
graph-theoretic measure from the empirical to the perturbed networks (see Fig. 5A). We first focus on the set of hub
nodes for each connectome, finding that differences in the RSI can be seen across perturbed networks and graph-
theoretic measures (Fig. 5B, super-hub results were similar). Figure 5C directly compares the effects of the FN and FP
connections (perturbations) on the graph-theoretic measures, first across all nodes in the network, and then for the set
of hub nodes. It can be seen that the FP connections consistently have a greater effect on the betweenness centrality
and participation index, whereas the clustering coefficient and small worldness are more affected by the FN
connections. For hub nodes, the RSI shows that the degree is also more affected by FP connections.
The RSI calculation is similar to the Spearman rank correlation coefficient (Spearman, 1904) and Kendall rank
coefficient (Kendall, 1938) at the network level. Supplementary Fig. 7 pertains to analyses repeated with these similar,
yet alternative measures and should be compared with Figs. 5B and C. Regardless of the measure used, the overall
trends in the data between Fig. 5 B-C and Supplementary Fig. 7 are consistent.
Directly comparing each of the methods for altering directionality (Fig. 5D), we find that the DP networks showed the
greatest RSI across almost all measures. Across connectomes the summed RSI for all graph-theoretic measures were
quite similar (Fig. 5E). In particular, the mouse connectome, which has the largest proportion of uni-directional
connections (see Fig. 1 and Supplementary Table 1), showed larger differences for the same percentage of altered
connections.
Page 11 of 29
Figure 5: Nodal changes measured by the rank-shift index. (A) The rank-shift index quantifies the change in the rank of nodes from the empirical
connectome to the perturbed network when they are ordered by a particular graph-theoretic measure. More specifically, it calculates the sum of the
difference between graph-theoretic values for each node in the empirical and perturbed matrices, divided by the maximum potential difference that
could exist between these two networks (where a value of 0 indicates no change, and a value of 1 indicates the maximum change). See methods for
further explanation. (B) Rank-shift index of hub nodes across all perturbed networks, for each graph-theoretic measure. (C) Difference in the rank-
shift index between the false-negative and false-positive networks for all nodes (left), and hub nodes (right). A positive value indicates that the false
negative connections cause greater changes in the ranking of nodes, whereas a negative value indicates the same for false positive connections. (D)
Rank-shift index for each graph-theoretic measure summed across all connectomes. (E) Rank-shift index values summed across all graph-theoretic
measures for each density-preserving connectome. (B-E) Results correspond to the mean over 50 trials for which 5% of randomly selected uni-
directional connections are modified in each perturbed network (error bars show the standard error of the mean). Graph-theoretic measures are as
→). M47: the macaque
follows: K=Degree, B=Betweenness centrality, C=Clustering coefficient, Y=Participation index and S=Small-world index (𝑆𝑖
connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Quantifying the Importance of Directed Connections in the Whole Network
We next considered the mean changes in graph-theoretic measures in the whole network caused by the loss of
directionality. We focus our analysis on perturbed networks with alterations to a small percentage of the uni-
directional connections (5%, see Fig. 6). In the initial two perturbed connectomes, false negative and false positive
alterations have opposite effects on network measures (Fig. 6A). The changes in betweenness (B), characteristic path
length (L) and global efficiency (G) are directly dependent on the degree (K), as these connections facilitate a shorter
route between nodes. The effects pertaining to clustering (C), participation index (Y), and small-world index (S) are
more complex because they depend on whether the changes increase or decrease the inter-neighbor or the inter-
Page 12 of 29
modular connectivity. Aside from the mean degree (which is preserved in the DP networks), the effects on graph-
theoretic measures were mostly similar across the FP and DP perturbed networks. To better understand the role of uni-
directional connections, we next compare how false positive and false negative modifications affect the mean graph-
theoretic measures of networks (Fig 6B). When it is not possible to distinguish the directionality of the connections, is it
better to assume that they are bidirectional or to disregard uni-directional connections?
In the case where a subset of connections is altered, for most graph-theoretic measures the false positive uni-
directional connections were more detrimental. It can be seen in Supplementary Fig. 8 that this trend remains robust as
the proportion of uni-directional connections is increased (to 10% and 20%). However, the error present in each
graph-theoretic measure is predictably increased. With the exception of the small worldness and degree, the FP
perturbed networks consistently show the greatest changes in the mean graph-theoretic measures (Fig. 6C and
Supplementary Fig. 8C, F). The participation index is the only measure directly affected by the modularity of the
networks.
Figure 6: Relative changes in mean graph-theoretic measures for perturbed networks. (A) Changes in mean graph-theoretic measures across all
connectomes and each type of perturbed network. (B) Difference between the changes in mean graph-theoretic measures for the false negative and
false positive networks. (C) Mean changes in graph-theoretic measures for each of the perturbed networks, summed across all connectomes. Two
separate modularity inputs are used the participation index calculations for the perturbed networks: the consensus modularity of the empirical
networks (light colors) and the new modularity assignments for each generated perturbed network (dark colors). (A-C) All results correspond to
perturbed networks with 5% of randomly selected uni-directional connections modified. The results represent the mean of these networks over 50
trials, and describe the change in the mean graph-theoretic measure (from the empirical to perturbed network) normalized by the mean of the
empirical network (error bars show the standard error of the mean). Graph-theoretic measures are as follows: K=Degree, B=Betweenness centrality,
C=Clustering coefficient, L=Characteristic path length, G=Global efficiency, Y=Participation index and S=Small-world index (𝑆→, changes in this
measure are presented as the mean over 1000 trials). M47: the macaque connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242,
C52: cat, M213: mouse, C279: C. elegans.
The changes in mean graph-theoretic measures are emphasized across connectomes in Supplementary Fig. 9. In the FN
and FP networks, the changes for each graph-theoretic measure depend on the degree and proportion of uni-
directional connections. Once again, the degree is correlated with the global efficiency and inversely correlated with
the characteristic path length and betweenness. Moreover, the clustering coefficient is also correlated with the changes
in degree but this is caused by the elimination of triangles from false negatives and addition of triangles from false
positives.
Discussion
Over ten years ago, Sporns et al. (2005) proposed an influential coordinated research strategy to map the
human connectome, which motivated and guided many researchers. A lot of progress has been made towards this goal
with the development of diffusion-weighted imaging and tractography methods, enabling the reconstruction of several
descriptions of the human connectome (Assaf and Basser, 2005, Goulas et al., 2014). However, much more research is
needed to achieve an accurate, reliable and standardized representation of connectivity in the human brain. It must also
be acknowledged that the methods of collation and reconstruction for these large datasets, including diffusion imaging
and tract tracing, can give rise to errors and inconsistencies in the data, as discussed elsewhere (Calabrese et al., 2015,
Donahue et al., 2016, Gămănuţ et al., 2017). Beyond this, several parcellation schemes have been proposed for the
human connectome (Cloutman and Ralph, 2012, de Reus and van den Heuvel, 2013, Honnorat et al., 2015, Glasser et al.,
2016), which can each have different effects on the characterization of the network (Zalesky et al., 2010). Furthermore,
the inability to resolve connection directionality noninvasively, which was originally classified as a crucial task (Sporns
Page 13 of 29
et al., 2005), has remained surprisingly overlooked. Without improvements in neuroimaging techniques, directionality
can only be indirectly estimated for the human connectome, for example, investigating effective connectivity (Stephan
et al., 2009, Friston, 2011). With current macro-scale connectome mapping techniques, connection directionality
cannot be explicitly resolved.
Here, we quantified the impact of disregarding directionality in connectome analysis. Specifically, we estimated the
inaccuracies in brain networks quantified by graph-theoretic measures following modifications to the uni-directional
connections in connectomes of different species and parcellations.
Our analyses indicate that several network measures are susceptible to error when directionality is lost. Graph-
theoretic measures are affected at both the individual-node, and network level, as is the definition of hubs. Across all
networks analyzed those with a larger proportion of uni-directional connections were more extensively affected by the
loss of connection directionality. This proportion is closely related to the parcellation, as finer parcellations tend to
have a larger proportion of uni-directional connections. We have also compared three different schemes to generate
undirected networks, which showed that the addition of reciprocal connections to a subset of existing connections
(false positives) is more detrimental to graph-theoretic measures than the removal of uni-directional connections (false
negatives).
Error in the Classification of Hub Nodes
Heterogeneity in cortical regions plays an important role in structural brain networks: Highly connected hub regions
support integration of functionally and structurally segregated brain regions (van den Heuvel et al., 2012, Mišić et al.,
2015, van den Heuvel et al., 2016). At these regions, neuronal dendrites have larger spine density (van den Heuvel and
Sporns, 2013, Scholtens et al., 2014) and increased transcription of metabolic genes (Fulcher and Fornito, 2016).
Moreover, hub nodes have high wiring cost and demand for metabolic resources, meaning their connections are more
likely to become structurally damaged and symptomatic in a wide range of neuropsychiatric disorders (Crossley et al.,
2014, Fornito et al., 2015, Fulcher and Fornito, 2016). For example, the increased vulnerability of hubs in Alzheimer's
disease could be explained by excessive neuronal activity at these regions (Kitsak et al., 2010, de Haan et al., 2012, Raj
et al., 2012). Hence, the correct identification and classification of hub regions is crucial to understanding the effects of
their normal functioning (van den Heuvel and Sporns, 2013) and dysfunction (Fornito et al., 2015) within the brain
network.
Our results indicate that a proportion of hubs and super-hub nodes of the human connectome are vulnerable to
misclassification because the directionality of connections is not available. In particular, the classification of super-hub
nodes was found to have a significant lower accuracy than hub nodes. As a caveat, we need to be aware that this
measure is sensitive to noise because the number of super-hub nodes in some of the connectomes is limited.
Hubs were also classified as either connector or provincial based on their level of intra-module vs. inter-module
connectivity (Guimera and Amaral, 2005b, van den Heuvel and Sporns, 2011). Previous studies have found that
targeted attacks on connector hubs have a widespread effect on network dynamics due to their role in functional
integration, whereas attacks on provincial hubs produce a more localized effect within communities (Honey and
Sporns, 2008). It has been hypothesized that such localized damage would cause specific clinical deficits, whereas
damage to connector hubs would cause complex, distributed dysfunction throughout the network (Fornito et al., 2015).
We found that alterations to uni-directional connections lead to multiple errors in the classification of hub regions.
Hubs were more likely to be defined incorrectly based on degree (losing their classification) rather than the
participation index (changing classification between connector and provincial).
Effect of False Positive and False Negative Connections
Diffusion weighted and diffusion tensor imaging allow detailed reconstructions of the structural human brain network
(Iturria-Medina et al., 2008, Van Essen et al., 2013). Depending on the data and specific tractography algorithms used,
crossing fiber geometries can give rise to two types of errors during network reconstruction: absent connections (false
negatives) and spurious connections (false positives) (Dauguet et al., 2007, Jbabdi and Johansen-Berg, 2011). These
errors cannot be completely eliminated from the reconstructed network; however, when there are multiple subjects, a
group threshold can be used to minimize these errors and achieve a balance between the exclusion of false positives
and false negatives (de Reus and van den Heuvel, 2013, Roberts et al., 2017).
In a recent study, these two types of errors were investigated in undirected connectomes, where false negative
connections were generated by pruning existing connections and false positive connections were generated by
connecting pairs of unconnected nodes (Zalesky et al., 2016). False positive connections were at least twice as
detrimental as false negatives to the estimation of common graph-theoretic measures: clustering coefficient, network
efficiency, and modularity. This has been attributed to the modular topology of the network (Sporns and Betzel, 2016).
Because nodes within the same module are likely to have a higher connection density, false negative connections were
more likely to occur within modules and to be more redundant to network topology. Conversely, false positive
Page 14 of 29
connections were more likely to occur between modules, introducing shortcuts that have a greater impact on the
graph-theoretic metrics of the network. Here we investigated the impact of perturbations to a subset of uni-directional
connections, which were about half intra-modular and half inter-modular. Despite the similarity of this analysis, here
we generated false negative connections by removing existing uni-directional connections and false positive
connections by adding the reciprocal connections and making them bidirectional.
Our results also show that false positive connections were overall more detrimental than false negatives. This occurs
for betweenness, path length, global efficiency and participation index. Notably, the small-world index and the
clustering (for some connectomes) are exceptions, in which false-negative directed connections are more detrimental
than false positives. For these measures, the removal of directed connections reduces the number of closed 3-node
motifs in the network, which may be more detrimental. These findings suggest that graph-theoretic measures are
overall more susceptible to addition of shortcuts introduced by false positive connections. A simple and immediate
recommendation that follows from our results is that connectomes should be thresholded stringently to maximize
specificity at the cost of sensitivity. This recommendation is very straightforward to implement and does not require
the development of any new methodologies. In the mouse as well as other connectomes that have weaker uni-
directional connections, a more stringent thresholding would create more false negative uni-directional connections
and avoid many false positive uni-directional connections that are more detrimental for network measures. Our
findings also suggest that the development of future connectome mapping methodologies should place more
importance on specificity. In this way, our work can inform and guide the development of future tractography
algorithms.
Connectome Mapping and Directionality Estimation
For the reconstruction of the macroscopic human connectome, parcellation schemes range from less than 102 nodes or
regions up to more than 105 [see for example, Tzourio-Mazoyer et al. (2002), Salvador et al. (2005), Aleman-Gomez
(2006), Hagmann et al. (2007), van den Heuvel et al. (2008), Glasser et al. (2016)]. The choice of parcellation can affect
several local and global topological parameters of the network, lowering the reliability of comparisons between
connectomes (Zalesky et al., 2010). The parcellation also affects the proportion of uni-directional connections, as
coarser parcellations correspond to larger brain regions that are more likely to have reciprocal connections. For
example, three of the connectomes can be considered coarse parcellations and have a relatively small proportion of uni-
directional connections (macaque N=47, N=71 and cat connectomes). Nonetheless, even for these connectomes, the
identification of hubs and their graph theoretic measures can result in inaccuracies due to loss of connection
directionality.
We have used connectomes from various species and parcellations that were obtained using different techniques.
These factors make it a complex task to compare and interpret some subtle features of the results across all
connectomes. Nonetheless, the consistency of most results across connectomes suggests that they reflect general
properties of brain networks and are largely independent from the techniques used to obtain these connectomes.
Hence, they are also expected to be valid in other connectomes.
Effect of Connectome Structure on Brain Dynamics
Although the problem of directionality is a recurrent topic in connectomics, with few exceptions (Négyessy et al., 2008,
Rosen and Louzoun, 2014), most work has focused on identifying the directionality of the interactions from the
dynamics of nodes. The directionality of the interactions of nodes in motifs and networks is paramount to shaping the
dynamics of systems (Bargmann and Marder, 2013). The dynamics of small circuits or network motifs can be
substantially altered by subtle differences in connectivity patterns. For example, the presence of a single reciprocal
connection can amplify the synchronization due to resonance (Gollo and Breakspear, 2014, Gollo et al., 2014); the
presence of triangles (loops) can increase metastability (Gollo and Breakspear, 2014) or multistability (Levnajić, 2011)
due to frustration. Moreover, the presence of an inhibitory feedback can cause anticipated synchronization between
neurons (Matias et al., 2016) or cortical regions (Matias et al., 2014). Naturally, this susceptibility of the dynamics to
structural perturbations goes beyond network motifs, affecting the dynamics of the whole network (Eguíluz et al., 2011,
Hu et al., 2012, Gollo et al., 2015, Esfahani et al., 2016).
A basic and influential manner of summarizing the dynamics of brain networks corresponds to functional connectivity
(Biswal et al., 1995), which captures linear correlations between pairs of regions, and are symmetric and undirected
(Friston, 2011). Disambiguating the directionality of connections between pairs of cortical regions has been a priority
in the field (Friston et al., 2003, Friston, 2011) as this directionality can reveal causal interaction between regions, or
how they effectively interact (Friston et al., 2017). Furthermore, a number of methods have been proposed and utilized
to determine the causal interactions between nodes (Friston et al., 2013), or to reconstruct the underlying network
structure from the network dynamics (Stam et al., 2007, Timme, 2007, Napoletani and Sauer, 2008, Vicente et al., 2011,
Friston et al., 2013, Tajima et al., 2015, Deng et al., 2016, Ching and Tam, 2017, López-Madrona et al., 2017, Wei et al.,
2017). A better understanding of the relationship between directionality in network structure and dynamics may aid in
determining causal interactions (Stephan et al., 2009).
Page 15 of 29
At the network level, it is important to distinguish the roles of in- and out-degree in affecting brain dynamics. A recent
study found strong relationships between the structural connectivity of a region and its BOLD (blood oxygen level
dependent) signal dynamics (Sethi et al., 2017). Furthermore, several graph-theoretic measures showed stronger
correlations to the network dynamics (resting state functional MRI) when directionality was taken into account. Brain
regions receiving more input (larger in-degree) required longer integration time to process and combine all these
inputs, which is consistent with the attributed function of rich-club association areas (Heeger, 2017), and also supports
the notion of a hierarchy of timescales recapitulating the anatomical hierarchy of brain structure (Kiebel et al., 2008,
Murray et al., 2014, Chaudhuri et al., 2015, Gollo et al., 2015, Cocchi et al., 2016, Gollo et al., 2016). Overall, these
findings highlight the importance of the directionality of the structural connectivity to understand brain dynamics.
Despite intensive efforts, the structure-function relationship remains far from elucidated, and the issue of inferring
directionality in undirected anatomical connectomes has yet to be addressed. Here we have focused on characterizing
the effect of directionality on brain structure via graph-theoretic measures, and future work will characterize how
perturbations to the directionality of connections influence network dynamics.
Conclusions
Connectomes are inherently directed networks. The majority of non-invasive techniques for mapping connectomes are
unable to resolve connection directionality, thereby yielding undirected approximations in which truly uni-directional
connections are either overlooked or rendered bidirectional. We found that the inability to resolve connection
directionality can introduce substantial error to the estimation of topological descriptors of brain networks,
particularly with respect to the classification and identification of hubs. We analyzed the effect of progressively
eliminating connection directionality information in six directed connectomes that were mapped with invasive
techniques capable of resolving afferent and efferent connections (C. elegans, mouse, cat, and three macaque networks).
We demonstrated that the identification of the most connected hubs is especially affected by the loss of connection
directionality. We also found that the addition of reciprocal uni-directional connections (false positives) is more
detrimental to the estimation of most topological measures than removal of uni-directional connections (false
negatives). Our findings underscore the need for non-invasive connectome mapping techniques that can: (i) provide
estimates of connection directionality; and (ii) yield relatively sparse and highly specific fiber maps that preference
false negatives over false positives. Given that most topological properties have been found to be recapitulated across
directed (macaque) and undirected (human) connectomes, at least qualitatively, resolving the directionality of human
connectomes in the future will most likely not result in a radical re-appraisal of human brain network organization, but
it will enable a more accurate characterization of the human connectome.
Acknowledgements
We would like to sincerely thank Madeleine Flynn, QIMR Berghofer Medical Research Institute for her illustrations
(Figure 1 brain/nervous system images). The authors acknowledge the support of the National Health and Medical
Research Council of Australia (AZ, APP1047648; LLG, APP1110975).
References
Achacoso T, Yamamoto W (1992) AY's Neuroanatomy of C. Elegans for
Computation.
Aerts H, Fias W, Caeyenberghs K, Marinazzo D (2016) Brain networks
under attack: robustness properties and the impact of lesions. Brain
aww194.
Aleman-Gomez Y (2006) IBASPM: toolbox for automatic parcellation of
brain structures. In: 12th Annual Meeting of the Organization for
Human Brain Mapping June 11-15, 2006 Florence, Italy.
Assaf Y, Basser PJ (2005) Composite hindered and restricted model of
diffusion (CHARMED) MR imaging of the human brain. Neuroimage
27:48-58.
Biswal B, Zerrin Yetkin F, Haughton VM, Hyde JS (1995) Functional
connectivity in the motor cortex of resting human brain using
echo‐planar mri. Magnetic resonance in medicine 34:537-541.
Blondel VD, Guillaume J-L, Lambiotte R, Lefebvre E (2008) Fast
unfolding of communities in large networks. Journal of statistical
mechanics: theory and experiment 2008:P10008.
Boussaoud D, Ungerleider LG, Desimone R (1990) Pathways for motion
analysis: cortical connections of the medial superior temporal and
fundus of the superior temporal visual areas in the macaque. Journal
of Comparative Neurology 296:462-495.
Bullmore E, Sporns O (2012) The economy of brain network
Aydore S, Pantazis D, Leahy RM (2013) A note on the phase locking
organization. Nature Reviews Neuroscience 13:336-349.
value and its properties. Neuroimage 74:231-244.
Bargmann CI, Marder E (2013) From the connectome to brain function.
Nature methods 10:483-490.
Bassett DS, Bullmore E (2006) Small-world brain networks. The
neuroscientist 12:512-523.
Bassett DS, Bullmore E, Verchinski BA, Mattay VS, Weinberger DR,
Meyer-Lindenberg A (2008) Hierarchical organization of human
cortical networks
Journal of
Neuroscience 28:9239-9248.
in health and schizophrenia.
Betzel RF, Bassett DS (2016) Multi-scale brain networks. NeuroImage.
Bezgin G, Vakorin VA, van Opstal AJ, McIntosh AR, Bakker R (2012)
Hundreds of brain maps in one atlas: registering coordinate-
independent primate neuro-anatomical data to a standard brain.
Neuroimage 62:67-76.
Calabrese E, Badea A, Cofer G, Qi Y, Johnson GA (2015) A diffusion MRI
tractography connectome of the mouse brain and comparison with
neuronal tracer data. Cerebral Cortex 25:4628-4637.
Chaudhuri R, Knoblauch K, Gariel M-A, Kennedy H, Wang X-J (2015) A
large-scale circuit mechanism for hierarchical dynamical processing
in the primate cortex. Neuron 88:419-431.
Ching ES, Tam H (2017) Reconstructing links in directed networks from
noisy dynamics. Physical Review E 95:010301.
Cloutman LL, Ralph L (2012) Connectivity-based structural and
functional parcellation of the human cortex using diffusion imaging
and tractography. Frontiers in neuroanatomy 6:34.
Cocchi L, Sale MV, Gollo LL, Bell PT, Nguyen VT, Zalesky A, Breakspear M,
Mattingley JB (2016) A hierarchy of timescales explains distinct
effects of local inhibition of primary visual cortex and frontal eye
fields. eLife 5:e15252.
Page 16 of 29
Colizza V, Flammini A, Serrano MA, Vespignani A (2006) Detecting rich-
club ordering in complex networks. Nature physics 2:110-115.
Costa LdF, Rodrigues FA, Travieso G, Villas Boas PR (2007)
Characterization of complex networks: A survey of measurements.
Advances in physics 56:167-242.
Crossley NA, Mechelli A, Scott J, Carletti F, Fox PT, McGuire P, Bullmore
ET (2014) The hubs of the human connectome are generally
implicated in the anatomy of brain disorders. Brain 137:2382-2395.
Dauguet J, Peled S, Berezovskii V, Delzescaux T, Warfield SK, Born R,
Westin C-F (2007) Comparison of fiber tracts derived from in-vivo
DTI
tract
tracer
reconstruction on a macaque brain. Neuroimage 37:530-538.
tractography with 3D histological neural
de Haan W, Mott K, van Straaten EC, Scheltens P, Stam CJ (2012) Activity
dependent degeneration explains hub vulnerability in Alzheimer's
disease. PLoS Comput Biol 8:e1002582.
de Reus MA, van den Heuvel MP (2013) The parcellation-based
connectome: limitations and extensions. Neuroimage 80:397-404.
Deng B, Deng Y, Yu H, Guo X, Wang J (2016) Dependence of inter-
neuronal effective connectivity on synchrony dynamics in neuronal
network motifs. Chaos, Solitons & Fractals 82:48-59.
Donahue CJ, Sotiropoulos SN, Jbabdi S, Hernandez-Fernandez M,
Behrens TE, Dyrby TB, Coalson T, Kennedy H, Knoblauch K, Van
Essen DC (2016) Using diffusion tractography to predict cortical
connection strength and distance: a quantitative comparison with
tracers in the monkey. Journal of Neuroscience 36:6758-6770.
Durbin RM (1987) Studies on the development and organisation of the
nervous system of Caenorhabditis elegans. University of Cambridge
Cambridge.
Eguíluz VM, Pérez T, Borge-Holthoefer J, Arenas A (2011) Structural and
functional networks in complex systems with delay. Physical Review
E 83:056113.
Esfahani ZG, Gollo LL, Valizadeh A (2016) Stimulus-dependent
synchronization in delayed-coupled neuronal networks. Scientific
reports 6.
Fagiolo G (2007) Clustering in complex directed networks. Physical
Review E 76:026107.
Felleman DJ, Van Essen DC (1991) Distributed hierarchical processing in
the primate cerebral cortex. Cerebral cortex 1:1-47.
Fornito A, Zalesky A, Breakspear M (2015) The connectomics of brain
disorders. Nature Reviews Neuroscience 16:159-172.
Freeman LC
(1978) Centrality
in social networks conceptual
clarification. Social networks 1:215-239.
Friston K, Moran R, Seth AK (2013) Analysing connectivity with Granger
in
causality and dynamic causal modelling. Current opinion
neurobiology 23:172-178.
Goulas A, Bastiani M, Bezgin G, Uylings HB, Roebroeck A, Stiers P (2014)
Comparative analysis of the macroscale structural connectivity in
the macaque and human brain. PLoS Comput Biol 10:e1003529.
Guimera R, Amaral LAN (2005a) Cartography of complex networks:
modules and universal roles. Journal of Statistical Mechanics:
Theory and Experiment 2005:P02001.
Guimera R, Amaral LAN (2005b) Functional cartography of complex
metabolic networks. Nature 433:895-900.
Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, Wedeen VJ,
Sporns O (2008) Mapping the structural core of human cerebral
cortex. PLoS Biol 6:e159.
Hagmann P, Kurant M, Gigandet X, Thiran P, Wedeen VJ, Meuli R, Thiran
J-P (2007) Mapping human whole-brain structural networks with
diffusion MRI. PloS one 2:e597.
Hall DH, Russell RL (1991) The posterior nervous system of the
nematode Caenorhabditis elegans: serial reconstruction of identified
neurons and complete pattern of synaptic interactions. Journal of
Neuroscience 11:1-22.
Harriger L, van den Heuvel MP, Sporns O (2012) Rich club organization
of macaque cerebral cortex and its role in network communication.
PloS one 7:e46497.
Heeger DJ (2017) Theory of cortical function. Proceedings of the
National Academy of Sciences 114:1773-1782.
Honey CJ, Kötter R, Breakspear M, Sporns O (2007) Network structure of
cerebral cortex shapes functional connectivity on multiple time
scales. Proceedings of the National Academy of Sciences 104:10240-
10245.
Honey CJ, Sporns O (2008) Dynamical consequences of lesions in cortical
networks. Human brain mapping 29:802-809.
Honnorat N, Eavani H, Satterthwaite TD, Gur RE, Gur RC, Davatzikos C
(2015) GraSP: geodesic graph-based segmentation with shape
priors for the functional parcellation of the cortex. Neuroimage
106:207-221.
Hu Y, Trousdale J, Josić K, Shea-Brown E (2012) Motif statistics and
spike correlations in neuronal networks. BMC Neuroscience 13:P43.
'small-world-ness': a
canonical network
Humphries MD, Gurney K (2008) Network
for determining
quantitative method
equivalence. PloS one 3:e0002051.
Iturria-Medina Y, Sotero RC, Canales-Rodríguez EJ, Alemán-Gómez Y,
Melie-García L (2008) Studying the human brain anatomical
network via diffusion-weighted MRI and Graph Theory. Neuroimage
40:1064-1076.
Jbabdi S, Johansen-Berg H (2011) Tractography: where do we go from
here? Brain connectivity 1:169-183.
Kandel ER, Schwartz JH, Jessell TM, Siegelbaum SA, Hudspeth A (2000)
Friston K, Preller KH, Mathys C, Cagnan H, Heinzle J, Razi A, Zeidman P
Principles of neural science: McGraw-hill New York.
(2017) Dynamic causal modelling revisited. NeuroImage.
Kendall MG (1938) A new measure of rank correlation. Biometrika
Friston KJ (2011) Functional and effective connectivity: a review. Brain
30:81-93.
connectivity 1:13-36.
Kiebel SJ, Daunizeau J, Friston KJ (2008) A hierarchy of time-scales and
Friston KJ, Harrison L, Penny W (2003) Dynamic causal modelling.
the brain. PLoS Comput Biol 4:e1000209.
Neuroimage 19:1273-1302.
Fulcher BD, Fornito A (2016) A transcriptional signature of hub
connectivity in the mouse connectome. Proceedings of the National
Academy of Sciences 113:1435-1440.
Gămănuţ R, Kennedy H, Toroczkai Z, Van Essen D, Knoblauch K,
Burkhalter A (2017) The Mouse Cortical Interareal Network Reveals
Well Defined Connectivity Profiles and an Ultra Dense Cortical
Graph. bioRxiv 156976.
Garlaschelli D, Loffredo MI (2004) Patterns of link reciprocity in
directed networks. Physical review letters 93:268701.
Glasser MF, Coalson TS, Robinson EC, Hacker CD, Harwell J, Yacoub E,
Ugurbil K, Andersson J, Beckmann CF, Jenkinson M (2016) A multi-
modal parcellation of human cerebral cortex. Nature.
Gollo LL, Breakspear M (2014) The frustrated brain: from dynamics on
motifs to communities and networks. Phil Trans R Soc B
369:20130532.
Gollo LL, Mirasso C, Sporns O, Breakspear M (2014) Mechanisms of zero-
in cortical motifs. PLoS Comput Biol
lag synchronization
10:e1003548.
Gollo LL, Roberts JA, Cocchi L (2016) Mapping how local perturbations
preprint
systems-level
dynamics.
influence
arXiv:160900491.
brain
arXiv
Gollo LL, Zalesky A, Hutchison RM, van den Heuvel M, Breakspear M
(2015) Dwelling quietly
the rich club: brain network
determinants of slow cortical fluctuations. Phil Trans R Soc B
370:20140165.
in
Kitsak M, Gallos LK, Havlin S, Liljeros F, Muchnik L, Stanley HE, Makse
HA (2010) Identification of influential spreaders in complex
networks. Nature physics 6:888-893.
Kötter R (2004) Online retrieval, processing, and visualization of
the CoCoMac database.
connectivity data
from
primate
Neuroinformatics 2:127-144.
Kötter R, Wanke E (2005) Mapping brains without coordinates.
Philosophical Transactions of the Royal Society of London B:
Biological Sciences 360:751-766.
Lancichinetti A, Fortunato S (2012) Consensus clustering in complex
networks. Scientific reports 2:336.
Latora V, Marchiori M (2001) Efficient behavior of small-world
networks. Physical review letters 87:198701.
Levnajić Z (2011) Emergent multistability and frustration in phase-
repulsive networks of oscillators. Physical Review E 84:016231.
López-Madrona VJ, Matias FS, Pereda E, Canals S, Mirasso CR (2017) On
the role of the entorhinal cortex in the effective connectivity of the
hippocampal formation. Chaos: An Interdisciplinary Journal of
Nonlinear Science 27:047401.
Maier-Hein KH, Neher PF, Houde J-C, Côté M-A, Garyfallidis E, Zhong J,
Chamberland M, Yeh F-C, Lin Y-C, Ji Q (2017) The challenge of
mapping the human connectome based on diffusion tractography.
Nature Communications 8:1349.
Markov NT, Ercsey-Ravasz M, Ribeiro Gomes A, Lamy C, Magrou L,
Vezoli J, Misery P, Falchier A, Quilodran R, Gariel M (2012) A
weighted and directed interareal connectivity matrix for macaque
cerebral cortex. Cerebral cortex 24:17-36.
Page 17 of 29
Matias FS, Gollo LL, Carelli PV, Bressler SL, Copelli M, Mirasso CR (2014)
Modeling positive Granger causality and negative phase lag between
cortical areas. NeuroImage 99:411-418.
Matias FS, Gollo LL, Carelli PV, Mirasso CR, Copelli M (2016) Inhibitory
loop robustly induces anticipated synchronization in neuronal
microcircuits. Physical Review E 94:042411.
Medaglia JD, Bassett DS (2017) Network Analyses and Nervous System
Disorders. arXiv preprint arXiv:170101101.
Mišić B, Betzel RF, Nematzadeh A, Goñi J, Griffa A, Hagmann P, Flammini
A, Ahn Y-Y, Sporns O (2015) Cooperative and competitive spreading
dynamics on the human connectome. Neuron 86:1518-1529.
Modha DS, Singh R (2010) Network architecture of the long-distance
pathways in the macaque brain. Proceedings of the National
Academy of Sciences 107:13485-13490.
Murray JD, Bernacchia A, Freedman DJ, Romo R, Wallis JD, Cai X, Padoa-
Schioppa C, Pasternak T, Seo H, Lee D (2014) A hierarchy of intrinsic
timescales across primate cortex. Nature neuroscience 17:1661-
1663.
Napoletani D, Sauer TD (2008) Reconstructing the topology of sparsely
connected dynamical networks. Physical Review E 77:026103.
Négyessy L, Nepusz T, Zalányi L, Bazsó F (2008) Convergence and
divergence are mostly reciprocated properties of the connections in
the network of cortical areas. Proceedings of the Royal Society of
London B: Biological Sciences 275:2403-2410.
Newman ME (2004) Coauthorship networks and patterns of scientific
collaboration. Proceedings of the national academy of sciences
101:5200-5205.
Oh SW, Harris JA, Ng L, Winslow B, Cain N, Mihalas S, Wang Q, Lau C,
Kuan L, Henry AM (2014) A mesoscale connectome of the mouse
brain. Nature 508:207-214.
Raj A, Kuceyeski A, Weiner M (2012) A network diffusion model of
disease progression in dementia. Neuron 73:1204-1215.
Reinoso-Suarez
F
(1984)
in
parietotemporooccipital association cortex of the feline cerebral
cortex. Cortical
integration: Basic archicortical and cortical
association levels of neural integration IBRO monograph series Val
11:255-278.
Connectional
patterns
Roberts JA, Perry A, Roberts G, Mitchell PB, Breakspear M (2017)
the human connectome.
thresholding of
Consistency-based
Neuroimage 145:118-129.
Rosen Y, Louzoun Y (2014) Directionality of real world networks as
predicted by path length in directed and undirected graphs. Physica
A: Statistical Mechanics and its Applications 401:118-129.
Rubinov M, Sporns O (2010) Complex network measures of brain
connectivity: uses and interpretations. Neuroimage 52:1059-1069.
Salvador R, Suckling J, Schwarzbauer C, Bullmore E (2005) Undirected
graphs of frequency-dependent functional connectivity in whole
brain networks. Philosophical Transactions of the Royal Society of
London B: Biological Sciences 360:937-946.
Scannell J, Burns G, Hilgetag C, O'Neil M, Young MP (1999) The
connectional organization of the cortico-thalamic system of the cat.
Cerebral Cortex 9:277-299.
Scannell JW, Blakemore C, Young MP (1995) Analysis of connectivity in
the cat cerebral cortex. Journal of Neuroscience 15:1463-1483.
Scholtens LH, Schmidt R, de Reus MA, van den Heuvel MP (2014) Linking
to microscale
Journal of
in the macaque connectome.
macroscale
graph
neuroarchitectonics
Neuroscience 34:12192-12205.
organization
analytical
Serrano MÁ, Boguná M, Vespignani A (2009) Extracting the multiscale
backbone of complex weighted networks. Proceedings of the
national academy of sciences 106:6483-6488.
Sethi SS, Zerbi V, Wenderoth N, Fornito A, Fulcher BD (2017) Structural
connectome topology relates to regional BOLD signal dynamics in
the mouse brain. Chaos: An Interdisciplinary Journal of Nonlinear
Science 27:047405.
Shih C-T, Sporns O, Yuan S-L, Su T-S, Lin Y-J, Chuang C-C, Wang T-Y, Lo C-
C, Greenspan RJ, Chiang A-S (2015) Connectomics-based analysis of
information flow in the Drosophila brain. Current Biology 25:1249-
1258.
Spearman C (1904) The proof and measurement of association between
two things. The American journal of psychology 15:72-101.
Sporns O (2011) The human connectome: a complex network. Annals of
the New York Academy of Sciences 1224:109-125.
Sporns O, Betzel RF (2016) Modular brain networks. Annual review of
psychology 67:613-640.
Sporns O, Honey CJ, Kötter R (2007) Identification and classification of
hubs in brain networks. PloS one 2:e1049.
Sporns O, Tononi G, Kötter R (2005) The human connectome: a
structural description of the human brain. PLoS Comput Biol 1:e42.
Stam CJ, Nolte G, Daffertshofer A (2007) Phase lag index: assessment of
functional connectivity from multi channel EEG and MEG with
diminished bias from common sources. Human brain mapping
28:1178-1193.
Stephan KE, Kamper L, Bozkurt A, Burns GA, Young MP, Kötter R (2001)
Advanced database methodology for the Collation of Connectivity
data on the Macaque brain (CoCoMac). Philosophical Transactions of
the Royal Society of London B: Biological Sciences 356:1159-1186.
Stephan KE, Tittgemeyer M, Knösche TR, Moran RJ, Friston KJ (2009)
Tractography-based priors for dynamic causal models. Neuroimage
47:1628-1638.
Stephan KE, Zilles K, Kötter R (2000) Coordinate–independent mapping
of structural and
functional data by objective relational
transformation (ORT). Philosophical Transactions of the Royal
Society of London B: Biological Sciences 355:37-54.
Tajima S, Yanagawa T, Fujii N, Toyoizumi T (2015) Untangling brain-
wide dynamics in consciousness by cross-embedding. PLoS Comput
Biol 11:e1004537.
Timme M (2007) Revealing network connectivity from response
dynamics. Physical review letters 98:224101.
Tournier
J, Calamante F, Connelly A (2012) MRtrix: diffusion
tractography in crossing fiber regions. International Journal of
Imaging Systems and Technology 22:53-66.
Towlson EK, Vértes PE, Ahnert SE, Schafer WR, Bullmore ET (2013) The
rich club of the C. elegans neuronal connectome. Journal of
Neuroscience 33:6380-6387.
Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O,
Delcroix N, Mazoyer B, Joliot M (2002) Automated anatomical
labeling of activations in SPM using a macroscopic anatomical
parcellation of the MNI MRI single-subject brain. Neuroimage
15:273-289.
van den Heuvel MP, Bullmore ET, Sporns O (2016) Comparative
connectomics. Trends in cognitive sciences 20:345-361.
van den Heuvel MP, Kahn RS, Goñi J, Sporns O (2012) High-cost, high-
capacity backbone for global brain communication. Proceedings of
the National Academy of Sciences 109:11372-11377.
van den Heuvel MP, Sporns O (2011) Rich-club organization of the
human connectome. The Journal of neuroscience 31:15775-15786.
van den Heuvel MP, Sporns O (2013) Network hubs in the human brain.
Trends in cognitive sciences 17:683-696.
van den Heuvel MP, Stam CJ, Boersma M, Pol HH (2008) Small-world and
scale-free organization of voxel-based resting-state functional
connectivity in the human brain. Neuroimage 43:528-539.
Van Essen DC, Smith SM, Barch DM, Behrens TE, Yacoub E, Ugurbil K,
Consortium W-MH (2013) The WU-Minn human connectome
project: an overview. Neuroimage 80:62-79.
Varshney LR, Chen BL, Paniagua E, Hall DH, Chklovskii DB (2011)
Structural properties of the Caenorhabditis elegans neuronal
network. PLoS Comput Biol 7:e1001066.
Vicente R, Wibral M, Lindner M, Pipa G (2011) Transfer entropy-a
model-free measure of effective connectivity for the neurosciences.
Journal of computational neuroscience 30:45-67.
Watts DJ, Strogatz SH (1998) Collective dynamics of
'small-
world'networks. nature 393:440-442.
Wei Y, Liao X, Yan C, He Y, Xia M (2017) Identifying topological motif
patterns of human brain functional networks. Human Brain
Mapping 38:2734-2750.
White JG, Southgate E, Thomson JN, Brenner S (1976) The structure of
the ventral nerve cord of Caenorhabditis elegans. Philosophical
Transactions of the Royal Society of London B: Biological Sciences
275:327-348.
White JG, Southgate E, Thomson JN, Brenner S (1986) The structure of
the nervous system of the nematode Caenorhabditis elegans. Philos
Trans R Soc Lond B Biol Sci 314:1-340.
Yeterian EH, Pandya DN (1985) Corticothalamic connections of the
in the rhesus monkey. Journal of
posterior parietal cortex
Comparative Neurology 237:408-426.
Young MP (1993) The organization of neural systems in the primate
cerebral cortex. Proceedings of the Royal Society of London B:
Biological Sciences 252:13-18.
Ypma RJ, Bullmore ET (2016) Statistical analysis of tract-tracing
experiments demonstrates a dense, complex cortical network in the
mouse. PLoS computational biology 12:e1005104.
Page 18 of 29
Zalesky A, Fornito A, Cocchi L, Gollo LL, van den Heuvel MP, Breakspear
M (2016) Connectome sensitivity or specificity: which is more
important? NeuroImage.
Zalesky A, Fornito A, Harding IH, Cocchi L, Yücel M, Pantelis C, Bullmore
ET (2010) Whole-brain anatomical networks: does the choice of
nodes matter? Neuroimage 50:970-983.
SUPPLEMENTARY MATERIAL
Supplementary Figure 1: Probability that a node will disconnect from the density preserving and false negative networks. Probability
that a node will disconnect from the network as uni-directional connections are removed in the false negative perturbed networks (lower x-
axis). The probability of nodal disconnection for the density preserving networks (top x-axis prior to the red dashed line) as randomly selected
uni-directional connections are removed, with a reciprocal connection added to another uni-directional connection with each modification.
These results describe the mean over 1000 trials. M47: the macaque connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242,
C52: cat, M213: mouse, C279: C. elegans.
Supplementary Figure 2: Percentage of high degree nodes classified correctly as defined by the empirical network. These figures display
the percentage of core, hub, and super-hub nodes across all connectomes that are accurate (see methods) according to those in the empirical
networks when these nodes are redefined for each of the perturbed connectomes (with 100% of uni-directional connections altered; the
density-preserving results show a mean over 1000 trials with the standard deviation). M47: the macaque connectome with 47 nodes, M71:
macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Page 19 of 29
Supplementary Figure 3: Directionality effects on the in- and out- degree. (A-B) Cortical areas of the macaque N=47 connectome sorted by in-degree (A) or out-degree (B) for the empirical and each perturbed
network. Hubs are defined as nodes that have a total in-degree (A) or out-degree (B) one standard deviation above the mean, and super-hubs are defined as nodes that have an in-/out-degree 1.5 standard deviations
above the mean (the density-preserving results are from an illustrative single trial). (C) The mean of the in-degree minus the out-degree for the set of hub nodes in each connectome. M47: the macaque connectome
with 47 nodes, M71: macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Page 20 of 29
Supplementary Figure 4: Nodal variation in graph-theoretic measures for each region in the macaque N=47connectome from the
empirical to the density preserving network. (A) Change in the degree of each region (left) from the empirical to the density preserving
network, and the relative change (right). (B) Change in the betweenness centrality of each node (left) and the relative change (right). (C) Change
→). These results were generated from the
in the clustering coefficient of each node. (D) The change in the small-world index of each node (𝑆𝑖
mean change across 50 trial density preserving networks, and were relative to the maximum graph-theoretic value (for each measure and
region) across the macaque empirical network.
Page 21 of 29
Supplementary Figure 5: Relationship between the participation index and degree for hub nodes in each empirical and density
preserving connectome. Plots (participation vs. degree) displaying the hub nodes of the empirical (blue) network, and the same regions in the
density preserving (red) networks for each connectome (taken from a single illustrative trial). The dotted line represents the hub definition
based on the degree (K > one standard deviation above the mean), and the dashed line represents the further classification of hubs based on the
participation index (connector hubs Y > 0.35 and provincial hubs Y ≤ 0.35). The bar figures below show the probability that a hub node will cross
over either, or both of the threshold lines following density-preserving alterations in directionality (over 1000 trials), resulting in a classification
that is inconsistent with the empirical connectome.
Supplementary Figure 6: Number of core, hub and super-hub nodes identified in each empirical and perturbed network across all
connectomes. (A) Mean number of each type of highly connected region across all connectomes, for the empirical and each perturbed network
(100% of uni-directional connections altered). (B) Number of each type of hub region across all connectomes, for the empirical and each
perturbed network. Results for the density preserving network are the mean of 50 trials and show the standard deviation. M47: the macaque
connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Page 22 of 29
Supplementary Figure 7: Nodal changes measured by the Spearman correlation and Kendall coefficient. (A) Spearman correlation of hub
nodes across all perturbed networks, for each graph-theoretic measure. (B) Difference in the spearman correlation between the false-negative
and false-positive networks for all nodes in the network. A positive value indicates the false-negative connections cause greater changes in the
ranking of nodes, whereas a negative value indicates the same for false-positive connections. (C) Kendall coefficient of hub nodes across all
perturbed networks, for each graph-theoretic measure. (D) Difference in the Kendall coefficient between the false-negative and false-positive
networks for all nodes in the network. (A-D) Results correspond to the mean over 50 trials for which 5% of randomly selected uni-directional
connections are modified in each perturbed network. Graph-theoretic measures are as follows: K=Degree, B=Betweenness centrality,
→). M47: the macaque connectome with 47 nodes, M71: macaque
C=Clustering coefficient, Y=Participation index and S=Small-world index (𝑆𝑖
N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Page 23 of 29
Supplementary Figure 8: Relative change in mean graph-theoretic measures from the empirical connectomes for perturbed networks
with 10% and 20% of uni-directional connections altered. Changes in mean graph-theoretic measures across all connectomes and each type
of perturbed network with 10% of uni-directional connections altered (A) and 20% of uni-directional connections altered (D). Difference
between the changes in mean graph-theoretic measures for the 10% (B) and 20% (E) false-negative and false-positive networks. Mean changes
in graph-theoretic measures for each of the perturbed networks with 10% of connections altered (C) or 20% of connections altered (F) summed
across all connectomes. (A-F) The results represent the mean of these networks over 50 trials, and describe the change in the mean graph-
theoretic measure (from the empirical to perturbed network) normalized by the mean of the empirical network (error bars show the standard
error of the mean). Graph-theoretic measures are as follows: K=Degree, B=Betweenness centrality, C=Clustering coefficient, L=Characteristic
path length, G=Global efficiency, Y=Participation index and S=Small-world index (𝑆→, this measure is the mean over 1000 trials). M47: the
macaque connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
Page 24 of 29
Supplementary Figure 9: The change in mean graph-theoretic measures across connectomes sorted by the proportion of uni-directional connections (M71 = 17%, M47 = 24%, C52 = 26%, C279 = 40%,
M242 = 49%, M213 = 71%). (A) Relative change in mean graph theoretic measures (see legend) for each type of perturbed network with 5% of uni-directional connections altered. (B) Results for perturbed networks
with 20% of uni-directional connections altered. The results represent the mean of these networks over 50 trials, and describe the change in the mean graph-theoretic measure (from the empirical to perturbed
network) normalized by the mean of the empirical network. The legend shows each graph theoretic measure as follows: K=Degree, B=Betweenness centrality, C=Clustering coefficient, L=Characteristic path length,
G=Global efficiency, Y=Participation index and S=Small-world index (𝑆→, this measure is the mean over 1000 trials). M47: the macaque connectome with 47 nodes, M71: macaque N=71, M242: macaque N=242, C52:
cat, M213: mouse, C279: C. elegans
.
Page 25 of 29
Connectome
Number of
Proportion of
Density of
Number of
connections
Proportion of
uni-directional
connections
Density of
Proportion of
uni-directional
connections
Mean
Network
measures
Nodes
Modules
Connections
Uni-directional
connections
All connections
Uni-directional
connections
Inter-/intra-
modular ratio
Core
Feeder
Peripheral
Core
Feeder
Peripheral
Hub-hub
connections
Hub-peripheral
connections
Hub-hub
connections
Hub-peripheral
connections
Degree
Betweenness
centrality
Clustering
coefficient
Macaque Macaque Macaque
242
5
4090
47
4
505
71
4
746
Cat Mouse
52
213
3
5
3301
818
C. elegans
279
4
4903
0.24
0.23
0.06
0.51
263
194
48
0.19
0.29
0.29
0.69
0.17
0.15
0.03
0.39
337
339
70
0.12
0.21
0.29
0.57
0.49
0.26
0.07
0.31
0.03
0.08
0.71
0.07
0.05
0.63
0.53
0.58
2080
1705
305
0.43
0.56
0.59
362
346
110
0.15
0.39
0.20
1025
1729
547
0.63
0.76
0.70
0.43
0.85
0.27
1.64
1.55
0.75
1.62
1.37
0.16
0.05
0.30
0.05
0.54
0.20
21.5
48
0.19
21.0
92
0.43
0.26
0.84
33.8
31.5
31.0
368
41
335
0.40
0.06
0.03
0.45
1763
2117
1023
0.37
0.43
0.37
0.35
0.66
0.31
0.36
35.1
417
0.58
0.47
0.37
0.59
0.28
0.29
Participation index
0.378
0.288
0.380 0.381
0.325
0.332
Small-world index
2.23
2.78
4.71
1.79
3.31
4.20
Characteristic path
length
2.05
2.33
2.53
1.81
2.63
2.50
Global efficiency
0.57
0.50
0.44
0.64
0.43
0.44
Supplementary Table 1: Network characteristics of each empirical connectome. Details of modularity determination are presented in
Methods.
Page 26 of 29
Proportion of
randomly selected
connections
Modularity in
perturbed
networks
Number of
trials
Connectomes
Definition of hubs in
perturbed networks
Fig. 1
Fig. 2
Fig. 3
-
100%
100%
-
Redefined
-
1
Redefined
A: 1, B: 1
All
M47
All
-
-
-
Fig. 4
100%
Redefined
A, C&D: 1,
B&E: 1000
A, C&D: M47,
A&C-E: Empirical,
B&E: All
B: Redefined
Empirical
50
50, except small
world index:
1000
10, 000
1
1000
50
Plots: 1,
Bars: 1000
50
50
All
All
All
Empirical
-
-
A, B: M47,
C: All
A, B: As per in-/
out- degree of
Empirical network,
C: Empirical
All
M47
All
All
All
All
All
Redefined
Empirical
Empirical
Redefined
Empirical
-
-
Fig. 5
Fig. 6
5%
5%
S.Fig. 1
0-100%
S.Fig. 2
100%
S.Fig. 3
100%
S.Fig. 4
100%
Empirical &
Redefined
-
-
-
-
S.Fig. 5
100%
Redefined
S.Fig. 6
100%
Both Displayed
S.Fig. 7
5%
Empirical
S.Fig. 8
10% and
20%
50, except small
Redefined
world index:
1000
50, except small
S.Fig. 9 5% and 20%
Redefined
world index:
1000
Supplementary Table 2: Methodological details for analyses presented in each figure and in the supplementary material. M47: the
macaque connectome with 47 nodes.
Page 27 of 29
Graph-theoretic Measure Formula
Degree
(Rubinov and Sporns,
2010)
𝐾𝑖= degree of node i(sum of directed
in- and out-degree)
𝐾𝑖
(both) 𝐾𝑖 =
𝑗 ∈𝑁
𝑎𝑖𝑗
𝑖𝑛 = 𝑎𝑗𝑖
𝑗 ∈𝑁
𝐾𝑖
𝑜𝑢𝑡 = 𝑎𝑖𝑗
𝑗 ∈𝑁
Betweenness
Centrality
(Freeman, 1978)
Number of Triangles
(Rubinov and Sporns,
2010)
Clustering Coefficient
(Fagiolo, 2007)
Shortest
Path Length
(Rubinov and Sporns,
2010)
Characteristic
Path Length
(Watts and Strogatz,
1998)
Global Efficiency
(Latora and Marchiori,
2001)
Participation Index
(Guimera and Amaral,
2005)
Small Worldness
(Humphries and Gurney,
2008)
𝐵𝑖 = betweenness centrality of node i,
𝐵ℎ𝑗 (𝑖) = number of shortest paths
between h &j passing through i,
𝐵ℎ𝑗 = number of shortest paths
between h &j
→= number of triangles around node i
𝑡𝑖
𝐵𝑖 =
1
𝑛 − 1 (𝑛 − 2)
𝐵ℎ𝑗 (𝑖)
𝐵ℎ𝑗
ℎ,𝑗 ∈𝑁
ℎ≠𝑗 ,ℎ≠𝑖,𝑗 ≠𝑖
→ =
𝑡𝑖
1
2
𝑎𝑖𝑗 + 𝑎𝑗𝑖 𝑎𝑖ℎ + 𝑎ℎ𝑖 (𝑎𝑗 ℎ + 𝑎ℎ𝑗 )
𝑗 ,ℎ∈𝑁
→= clustering coefficient of node i
𝐶𝑖
→ =
𝐶𝑖
1
𝑛
𝑖∈𝑁
𝐾𝑖
𝑜𝑢𝑡 + 𝐾𝑖
𝑖𝑛 𝐾𝑖
𝑖𝑛 − 1 − 2
𝑖∈𝑁
𝑎𝑖𝑗 𝑎𝑗𝑖
→
𝑡𝑖
𝑜𝑢𝑡 + 𝐾𝑖
𝐶→= mean clustering coefficient of the network
𝑁
𝐶→ =
𝑖=1
→
𝐶𝑖
𝑁
→ = shortest path length between nodes i & j,
𝑑𝑖𝑗
where 𝑔𝑖→𝑗 is the shortest path between i & j
→ =
𝑑𝑖𝑗
𝑎𝑖𝑗
𝑎𝑖𝑗 ∈𝑔𝑖→𝑗
𝐿→ = average distance between all nodes
𝐿→ =
1
𝑛
𝑖∈𝑁
𝑗 ∈𝑁,𝑗 ≠𝑖
𝑛 − 1
→
𝑑𝑖𝑗
𝐺→ = global efficiency of the network
𝑜𝑢𝑡 =out-participation index, 𝑀= set of modules,
𝑜𝑢𝑡 (𝑚) = number of out- connections between i
𝑌𝑖
𝐾𝑖
& all nodes in module m
→ = small worldness of node i
𝑆𝑖
𝑆→ = small world index of network
Ci
→
→
𝐿 𝑟𝑎𝑛𝑑
rand = clustering of a random network,
= path length of a random network
→ =
𝑆𝑖
𝑆→ =
𝐺→ =
1
𝑛
𝑖∈𝑁
→)−1
𝑗 ∈𝑁,𝑗 ≠𝑖
(𝑑𝑖𝑗
𝑛 − 1
𝑜𝑢𝑡 = 1 −
𝑌𝑖
2
→
𝐾𝑖
𝑜𝑢𝑡 (𝑚)
𝑜𝑢𝑡
𝐾𝑖
𝑚∈𝑀
→/𝐶𝑖 𝑟𝑎𝑛𝑑
𝐶𝑖
𝐿→/𝐿 𝑟𝑎𝑛𝑑
𝐶→/𝐶𝑟𝑎𝑛𝑑
𝐿→/𝐿 𝑟𝑎𝑛𝑑
→
→
→
Notations: N = all nodes in the network, n = number of nodes, L = all connections, l = number of connections, (i, j) =
connection between nodes i & j, (i, j∈N), aij = connection status between i & j: = 1 when a connection from node i to j
exists, otherwise = 0, l =
(each bidirectional connection is counted twice, as aij & as aji), rand = random
network, → indicates formulae that consider directionality
𝒂𝒊𝒋
𝒊,𝒋∈𝑵
Supplementary Table 3: Reference, description and formula for each graph-theoretic measure used in this study. Table adapted from
Rubinov and Sporns (2010).
Page 28 of 29
M242
C52
M213
C279
Connector
Provincial
M47
V4
FEF
46
7a
TF
5
FST
MT
7b
M71
A46
TF
TPT
A7a
V4
FEF
TS3
ER
TH
A7b LIP
S2
PS
7b
25
23
8A
6M
46
24
46v
LIP 23c 24c 24b
32 PIT
TF 10
13a TE
12o PGm 8B
9
13
TH 36
11
12l
ENT
14
F5
F7
lai
CGp EPp
35
AES
36
Ia
Ig
7
6m
5Al
20a
PA
NOT
PTLp
SPFm
LGv MH
VM MPT
GPi
STN
CLA
RM
CLI MM
PP
PPN MGd BLA
PERI RCH VISam
PT
ACAd SUBd
SPFp LGd
LGd
PAA
ILA
PVR
RIBL DVC
AVJR RIMR
AVAR PVPL AIBL RMGL
AVAL AVDL RIAL AVG
AVBR PVT
AVBL AVHR RIBR HSNR
DVA
PVCR PVNR RMDL RIH
PVCL AIBR AVKL AVL
AVER PVPR AVJL
AVDR AVHL RIML
AVEL RIAR RIGL
RIML RIH
Supplementary Table 4: Connector and provincial hubs identified for each connectome. The hub definition based on the degree is K greater than one standard deviation above the mean and further
classification of hubs based on the participation index for connector hubs is Y > 0.35 and provincial hubs is Y ≤ 0.35. Each section (and column) is sorted by the highest degree. M47: the macaque connectome with
47 nodes, M71: macaque N=71, M242: macaque N=242, C52: cat, M213: mouse, C279: C. elegans.
|
1703.03065 | 1 | 1703 | 2017-03-08T22:55:45 | Discrete modes of social information processing predict individual behavior of fish in a group | [
"q-bio.NC",
"physics.bio-ph",
"q-bio.PE"
] | Individual computations and social interactions underlying collective behavior in groups of animals are of great ethological, behavioral, and theoretical interest. While complex individual behaviors have successfully been parsed into small dictionaries of stereotyped behavioral modes, studies of collective behavior largely ignored these findings; instead, their focus was on inferring single, mode-independent social interaction rules that reproduced macroscopic and often qualitative features of group behavior. Here we bring these two approaches together to predict individual swimming patterns of adult zebrafish in a group. We show that fish alternate between an active mode in which they are sensitive to the swimming patterns of conspecifics, and a passive mode where they ignore them. Using a model that accounts for these two modes explicitly, we predict behaviors of individual fish with high accuracy, outperforming previous approaches that assumed a single continuous computation by individuals and simple metric or topological weighing of neighbors behavior. At the group level, switching between active and passive modes is uncorrelated among fish, yet correlated directional swimming behavior still emerges. Our quantitative approach for studying complex, multi-modal individual behavior jointly with emergent group behavior is readily extensible to additional behavioral modes and their neural correlates, as well as to other species. | q-bio.NC | q-bio | Discrete modes of social information processing
predict individual behavior of fish in a group
Roy Harpaz1, Gašper Tkačik2, and Elad Schneidman1
1Department of Neurobiology, Weizmann Institute of Science, Rehovot, Israel
2Institute of Science and Technology, Klosterneuburg, Austria
Individual computations and social interactions underlying collective behavior
in groups of animals are of great ethological, behavioral, and theoretical
interest. While complex individual behaviors have successfully been parsed
into small dictionaries of stereotyped behavioral modes, studies of collective
behavior largely ignored these findings; instead, their focus was on inferring
single, mode-independent social interaction rules that reproduced macroscopic
and often qualitative features of group behavior. Here we bring these two
approaches together to predict individual swimming patterns of adult zebrafish
in a group. We show that fish alternate between an 'active' mode in which they
are sensitive to the swimming patterns of conspecifics, and a 'passive' mode
where they ignore them. Using a model that accounts for these two modes
explicitly, we predict behaviors of individual fish with high accuracy,
outperforming previous approaches that assumed a single continuous
computation by individuals and simple metric or topological weighing of
neighbors' behavior. At the group level, switching between active and passive
modes is uncorrelated among fish, yet correlated directional swimming
behavior still emerges. Our quantitative approach for studying complex, multi-
modal individual behavior jointly with emergent group behavior is readily
extensible to additional behavioral modes and their neural correlates, as well
as to other species.
1
Introduction
Group behavior has been studied in a wide range of species – bacteria (1),
slime mold (2), insects (3–5), fish (6–13), birds (14–16), and mammals (17–20)
– seeking the design principles of collective information processing, decision
making, and movement. Theoretical models have suggested possible classes
of computations and interaction rules that generate complex collective
behavior, qualitatively replicate macroscopic features of behavior observed in
real animal groups (21–25), and also have algorithmic, behavioral, and
economic implications (26). The ability to record the movement patterns of
animals in a group with high temporal and spatial precision for long periods (10,
14, 16, 17, 27, 28), allows for direct exploration of individual traits and
interactions between group members. Such attempts have considered
topological vs. metric relations between conspecifics (16), effective social
"forces" depending on the distance between individuals (6, 7), inference of
functional interactions based on maximum entropy models of observed
directional correlations (15), hierarchical spatial ordering (14, 29, 30), and
active signaling (3, 31, 32).
Because individual behavior is complex, previous studies have mostly focused
on modeling various group level statistics, e.g., polarization or moments of the
distribution of inter-individual distances (6–8, 16, 33, 34); see also (35)). These
approaches, however, do not necessarily yield a unique solution for the
underlying interactions between individuals (35). Furthermore, the resulting
models were often non-physiological in terms of response times or temporal
causality, ignored physical constraints such as momentum and friction, or
omitted the role of non-social sensory information. Somewhat surprisingly, most
models of individual behavior in a group commonly assume that animals
continuously update their movement based on the location or velocity of their
neighbors (22, 24, 25). In contrast, characterization of movement patterns of
individual zebrafish larva, C. elegans, and Drosophila, for example, suggest
that a distinct and relatively small set of stereotyped modes underlies complex
individual behavior (36–38). Here, we ask how discrete behavioral modes at
the level of the individual affect sensory and social information processing
underlying group-level motion decisions.
We studied individual behavior in groups of adult zebrafish in a large arena,
using high spatio-temporal tracking of fish under different behavioral contexts.
The adult zebrafish live in nature in groups of 4-20 fish either in still waters or
in running rivers (39), exhibit social behaviors and shoaling tendencies both in
the wild and in the laboratory (9, 39) (unlike the transparent larvae that allow
for imaging neuronal circuits underlying sensory-motor processing (40–43) but
2
exhibit a limited behavioral repertoire (44)). We analyze the behavior of
individuals in the group and identify distinct behavioral modes, which are used
to build a highly accurate mathematical model of swimming behavior of
individual fish in a group. The model is based on the sensory and social
information that is available to each animal and takes into account spatial and
temporal biophysical constraints. Importantly, we evaluate the models in terms
of their power to predict individual fish trajectories, rather than statistical
averages over the whole group.
3
the
into
Results
To study individual computations and interactions underlying group behavior in
zebrafish, we tracked individuals in groups of 2, 3, and 6 adult fish for up to an
hour at a time, in a large circular arena with shallow waters constituting an
effective 2D environment (Fig. 1A, and S1A, Movie M1, SI Methods). We
sampled the trajectory of the center of mass of each fish 𝑖 in the group, denoted
as 𝑥#(𝑡), with high spatial and temporal resolution (see SI Methods).
fish, 𝑣#𝑡 ,
instantaneous swimming speed, 𝑠#𝑡 = 𝑣#𝑡 , and instantaneous direction
𝑑#(𝑡)= ,-(.)
,-(.) , revealed a clear segmentation of the trajectories into acceleration
time-dependent velocity of each
Decomposing
and deceleration epochs (Fig. 1B-C, Movie M2). Acceleration epochs of the fish
were very accurately described by a family of sigmoid functions that differed by
their slope and duration (Fig. 1C-D). Decelerations were very accurately
described by a single exponential, corresponding to a simple drag force (Fig.
1C-D), where the inferred friction coefficient showed very little variance within
and between fish (Fig. S1B-D). The durations of successive epochs of
acceleration (~200±104 𝑚𝑠) and deceleration (~250±160𝑚𝑠) were very
weakly correlated, and the rate of switching between them was strongly related
to the speed of the fish (Fig. S1E-G). We further found that fish made turns
mostly during acceleration epochs (Fig. 1E, S1H,I and movie M2). We note that
the continuous motion of the adult fish makes these kinematic states very
different than the distinct stop-and-go bouts of zebrafish larvae (45).
The segmentation of fish kinematics into clear epochs that have simple
functional forms suggests that fish may not be using a universal ongoing
computation to determine their behavior at every time instant, as has been
suggested previously (8, 24, 25). Furthermore, we find clear anisotropies in
group structure, implying that simple distance-based or topology-based models
of social interactions, common in the literature, may fall short in explaining
individual zebrafish trajectories (22, 24, 25). Specifically, we find that the adult
zebrafish prefer to be on the side of other fish (Fig. 1F) within ~1.5 body lengths,
and that they typically demonstrated aligned swimming directions when they
are directly in front, behind, or on the side of another fish (Fig. 1G).
4
Fig 1: Kinematic states of individual fish and group structure. A. Snapshot of the
tracks of 6 freely swimming fish in a pseudo 2D circular arena. B. A short segment of
5
the swimming pattern of a single fish from the group, down-sampled to 50Hz for
visualization (dots). Dot color indicate if a fish is accelerating (red) or decelerating
(blue) C. Speed profile of the trajectory in B. D. Functional fits to the acceleration and
deceleration epochs in C (see SI Methods). E. Heading direction vs. time for the
segment shown in B. Directional changes occur predominantly during acceleration
epochs (see Fig S1H). F. Density map of neighboring fish relative to a focal fish
situated at [0,0] pointing north. G. Density map of directional alignment of neighboring
fish relative to the direction of motion of the focal fish – each point shows the mean
alignment value of fish in that bin, with 0 representing perfect alignment (see SI
Methods).
We therefore modeled the behavior of individual fish in a group using two
modes of information processing: a 'passive' mode where inertia and friction
control the movement of the fish, with no sensory or social influence, and an
'active' mode where an additional sensory term, described by a spatio-temporal
receptive field (RF) model of sensory and social processing, contributes to the
change in velocity. In detail, we discretize time into bins of size 𝛥𝑡 and denote
the measured instantaneous change in velocity of fish 𝑖 in the group as 𝛥𝑣#(𝑡).
We model the change in velocity in the passive mode as 'gliding' where water
friction slows down the fish (Fig. 2A):
𝛥𝑣#;<==#,>𝑡 =−𝜂𝑣#(𝑡−𝜏#B>C)
(1)
50 ms).
In the active mode, we assume that sensory information and social interactions
where 𝜂 is the friction coefficient, estimated from fitting deceleration epochs
(Fig. 1D and SI Methods), and 𝜏#B>C is a short time-constant (chosen here to be
are taken into account by the fish, and the change in velocity of fish 𝑖 at time 𝑡
The interaction term 𝛥𝑣#FG𝑡 is given by a spatio-temporal receptive field (RF)
𝛥𝑣#FG𝑡 =
𝛥𝑣#<D.#,>𝑡 =𝛥𝑣#;<==#,>𝑡 + 𝛥𝑣#FG𝑡
𝛽I𝑘 ⋅𝑣I𝑡−𝑘𝛥𝑡 +
𝛽L𝑘 ⋅𝑑L𝑡−𝑘𝛥𝑡
model (Fig. 2B):
is given by
.
(3)
(2)
I,M
L,M
The first term is a social interaction term, summing over the past swimming
6
velocities of neighboring fish, where the weights of spatial bin 𝑗 at time 𝑡−𝑘𝛥𝑡
are given by 𝛽I𝑘 , and 𝑣I𝑡−𝑘𝛥𝑡 is the velocity of the fish in that bin. The
𝑑L𝑡−𝑘𝛥𝑡 is a vector tangent to the wall closest to the fish, and 𝛽L𝑘 are the
second term is the contribution of non-social sensory information, where
weights associated with that bin. Models were fit on labeled training data, taken
from acceleration epochs (see SI Methods); the number of spatial bins and the
extent of the temporal history (that together determine the number of
parameters) were chosen to maximize model performance using penalized
regularization.
The 'passive' and 'active' models give very different predictions for 𝑣#𝑡 at
different times along the trajectory of a fish swimming in a group. Fig. 2C shows
examples of the different predictions of the two models, on top of a segment of
a complex swimming pattern of one fish in a group of three (neighboring fish
not shown). Along most of the trajectory, the two models alternate in terms of
their accuracy in predicting behavior. Fig. 2D shows the models' prediction
errors as a function of time on a short segment of held-out test data, suggesting
that the passive model makes smaller errors mostly during decelerations,
whereas the active model makes smaller errors mostly during accelerations.
This observation was further supported by analyzing complete fish trajectories
and multiple groups of the same size recorded independently (N=6-7 for the
different sizes): the active model significantly outperformed the passive model
in acceleration epochs, while the passive model outperformed the active model
in deceleration epochs (Fig. 2E, P<0.0005 for all group sizes, t-test for
dependent samples). Learning a separate RF-model for the deceleration
epochs did not result in a significant improvement over the passive model (Fig.
S2A,B), reasserting that fish show very weak social responses during
decelerations. These results indicate that individual fish alternate between two
distinct modes of social information processing, which roughly correspond to
acceleration and decelerations epochs; in other words, we hypothesize that the
kinematic states of the fish are a good indicator for the mode of social
information processing.
7
Fig 2: Modeling fish behavior using active and passive models of social
information processing. A. In the passive model (left), the change in velocity, 𝛥𝑣#𝑡,
is given by inertia and friction. In the active model (right), 𝛥𝑣#(𝑡) is the sum of the
passive component and the contribution of a sensory and social component. B. The
sensory and social component of the active model is given by a computation based on
a spatio-temporal 'receptive field' (RF), where the behavior of conspecifics in spatial
'bins' is weighed with time-dependent parameters (see inset at right and text). C.
Example of a trajectory of one fish in a group of three, with a comparison of model
predictions (passive model in blue, active model in red) and the measured velocity
(black). Insets show zoom-in on three representative segments of the trajectory,
reflecting the different prediction accuracy of the models at different times. D. Top:
prediction errors, 𝐸PQR>L=𝑣C><L−𝑣PQR>L, as a function of time for one fish in group
of three, using the active mode (red) and the passive mode (blue). Background color
designates whether the fish was accelerating (pink) or decelerating (white). Bottom:
the difference between the errors of the active and passive models; typically, each of
the two models is much more accurate in one of the kinematic states. E. Average
values of the difference between errors shown in D for groups of 2 fish (N=6 groups),
3 fish (N=7), and 6 fish (N=7); error bars represent SEM.
Since we do not have access to the actual information processing state of the
fish, we asked how well can we explain fish behavior if we were to pick the best
8
model for each time point (the one that gives the lowest error when compared
to the real velocity). This combined model (see Movie M3) gives an excellent fit
to the data both in terms of the speed (Fig. 3A top), and the heading direction
of swimming (Fig. 3A bottom). Over all groups, the correlation between the real
and the estimated trajectory of the fish was ~0.97 for direction and ~0.94 for
speed on test data (Fig. 3B). To further illustrate the importance of the two
interleaved modes for describing individual behavior, we compared the
accumulated effect of the errors in predicting the instantaneous velocity vectors
that each of the models make. Figure 3C shows the 'reconstructed' swimming
trajectory of a fish in a group that would result from summing over the
instantaneous velocity predictions of each model to obtain a complete trajectory
segment (see SI Methods and also Fig. S3C). Repeating this analysis for 5000
3s long segments of a group of 3 fish, we found that combining between the
active model of information processing and the passive model (again by
choosing the best model at each time point) gave much more accurate
reconstructions than either model alone (Fig. 3D left). The reconstruction errors
over many trajectory segments for all groups of 3 fish were lower by 37±5%
compared to the passive model alone, and 19±11% compared to the active
model alone; these improvements were similar for groups of 2 and 6 fish (Fig.
3D right, P<0.005 for all group sizes and for both comparisons, t-test for
matched samples). Even though the combined model used here is an upper
bound for the performance of any mix of the 'active' and 'passive' models, the
majority of its performance gains are retained in a model where the kinematic
state of the fish is used directly as an indicator of its information processing
state (Fig. S3A-E), as hypothesized above.
9
Fig 3: Accurate prediction of velocity and fish trajectory reconstruction. A.
Examples of the measured speed (top) and heading direction (bottom) of a single fish
in a group of 3 (black) and the prediction obtained using the combination of the active
and passive models, picking the better model at each time point (green). B. Average
Pearson's correlation between measured speed and heading direction and model
predictions for all fish in groups of the same size (error bars = SEM, N=6,7,7 groups)
C. A short segment (3s) of the trajectory of a fish in a group of 3 (black) and the
reconstructed trajectory obtained by summing the velocity predictions of the active
model (red), the passive model (blue), and the optimal combination of the two (green).
D. Left: Distribution of the ratio between the reconstruction error of the combined model
and the reconstruction error of either the active model (red) or the passive model (blue)
alone, for 5000 segments similar to the one shown in C. Values below 1 (black dashed
line) represent advantage to the combined model. Colored dashed lines represent
distribution medians. Right: Median values for the reduction in prediction error of the
combined model vs. the active model or the passive model alone (as in the left panel),
for groups of different sizes (error bars = STD).
A significant part of the high correlation between model predictions and the data
(shown in Fig. 3) originates from the auto-correlation of individual swimming
behavior. This is especially true in deceleration epochs, where the correlation
between the measured 𝑣#𝑡 and prediction based on 𝛥𝑣#;<==#,>𝑡 , was 0.986±
0.002 (see also low prediction error values in Fig. S2A). We therefore focused
component, which we denote as 𝛥𝑢#𝑡 = 𝛥𝑣#(𝑡) −𝛥𝑣#;<==#,>𝑡 . Clearly, in the
to explain. In the acceleration epochs, the correlation between 𝛥𝑢#𝑡 and the
prediction of 𝛥𝑣#FG𝑡 was ~0.5. When we examined the social or sensory
on the change in velocity that is not explained by autocorrelation and friction.
Figure 4A shows the change in velocity that is not explained by the passive
deceleration epochs removing the passive component leaves very little change
contributions to the RF model in isolation, the resulting model's prediction
performance was significantly lower than when both information types were
included (Fig. 4B, P<0.0005 for all group sizes, t-test for matched samples),
with small differences between group sizes (Fig. S4A). The non-additivity of
social information and non-social sensory information reflects the redundancy
between them. In the current setup, it is impossible to discern whether fish
'read' sensory information about the environment from their own senses, or
from the behavior of other fish. We note that the relation between the 𝛥𝑣# and
the predictions of the models did not indicate a need for a non-linear extension
of the (active) RF model (Fig. S4C, cf. LN models in neuroscience; (46)).
Predicting the entire acceleration epoch using a similar RF model, from the
sensory and social information at the beginning of the epoch, performed
significantly worse (see SI Methods and Fig. S7).
10
Our RF model significantly outperformed common models of collective
movement in predicting 𝛥𝑣#, even when the parameters of these competing
models were optimized to our data (see SI Methods): we predicted 𝛥𝑢#𝑡 by
~8±2.5% better on average than a 'zonal model' (24) and ~11.3±4% better
than a 'topological model' (16) (Fig. 4C, P<0.05 for all group sizes and both
model comparisons, t-test for matched samples). The advantage of the RF
model is even more pronounced when comparing the accuracy of prediction
using only social information, as sensory information which is similar across all
models obscures part of these differences (Fig. S4B), indicating that the
assumptions of the RF-model better match the behavior of swimming zebrafish.
Fig 4: Active movement changes are accurately predicted by the RF model using
both social and sensory information. A. Top: An example of 𝛥𝑢𝑖𝑡, the magnitude
measured 𝛥𝑢#𝑡 and the prediction obtained using the RF model (red arrows) in the
acceleration epochs. B. Prediction accuracy of 𝛥𝑢#𝑡 by RF models that use only the
of the measured change in velocity after subtracting the passive component (see text)
of a single fish in a group of 3 over 4s; background colors mark accelerations (pink)
and decelerations (white). Bottom: the corresponding comparison between the
social information component in Eq. 3, only the sensory information component in Eq.
11
Improvement in predicting 𝛥𝑢#𝑡 in acceleration epochs by our RF model relative to
3, or both, for different group sizes (error bars = SEM, N=6,7,7 groups). C.
its zonal or topological versions (shown in insets; see Methods for details).
Improvement values are averaged over all groups of different sizes (error bars
represent SEM).
To characterize the spatio-temporal effects of social and sensory information
on the movement decisions of a focal fish, we compared the weight maps of
the RF models under two different behavioral contexts – fish swimming freely
in the arena as described above, and fish who were trained to seek food that is
randomly scattered in the arena (see SI Methods). Inhomogeneity in the
receptive field map reflect the effects of the relative distance and relative angle
of neighbors on the focal fish (Fig. 5A): social effects are strongest in front of
the fish and weaker behind it. The weights of the non-social information show
the opposite structure, with walls directly to the side of the fish having the
strongest effect on its behavior. In general, responses to neighbors are weaker
for longer temporal delays, but keep their positive sign. In contrast, the effect of
the wall decreases faster with time (see middle weight map in Fig. 5A) and
ultimately switches sign. Interestingly, the way fish integrate information from
their surroundings changes with the behavioral context (Fig. 5B): effects of
arena walls are weaker in food-searching fish, and the effects of fish positioned
directly behind the focal fish are positive and stronger (Fig. S5A-B for
statistically significant differences between weight maps).
A.
wall
-150ms
-350ms
-550ms
i
g
n
m
m
w
s
i
e
e
r
F
6 body lengths
B.
i
g
n
k
e
e
s
d
o
o
F
0.2
0.1
0
-0.1
-0.2
12
Fig 5: Receptive field maps show distinct spatial, temporal, and behavioral state
dependencies. A. Receptive fields for a fish in group of 3 fish (N=7), freely swimming
in the arena, for the model illustrated in Fig 2B (shown are average 𝛽 values from Eq.
3). Outer circles represent weights associated with the walls of the arena. B. Receptive
fields obtained for a fish in group of 3 fish (N=7), trained to seek food in the arena (no
food is present during the analyzed session).
What does switching between the two modes of information processing at the
individual level imply for the behavior of the group? Figures 6A-B show an
example of the swimming velocities of 3 fish, decomposed into the speed 𝑠#𝑡,
(Fig. 6A) and the direction of swimming 𝜃(𝑑#) (Fig. 6B). We asked what are the
correlation (𝜏P<[) did not show any structure and was indistinguishable from
temporal relations between kinematic states in pairs of fish in the group, by
seeking the time lag that would maximize the correlation for short movement
segments (1s long) for each fish pair (keeping the identity of the fish throughout
the analysis, (47) see Methods). The distribution of the time of maximal
the expectation of fish changing states independently (Fig. 6C); the correlation
values also did not differ from what was expected by chance (Fig. S6A). This
suggests that the transitions between the two behavioral modes of individual
fish are independent. Such organization could give the group a way to sample
the sensory space in a distributed and interleaved manner, with no temporal
processing gaps, without the need for scheduling. In contrast, analogous
analyses identified significant correlations between swimming directions in
pairs of fish (Fig. S6A) and a corresponding significant peak in the distribution
of temporal lags, suggesting causal relationships (Fig 6D). The absence of
statistical dependencies between kinematic states of fish in a group and the
presence of dependencies for swimming direction was corroborated by
estimating the probability of synchronized states among the fish in the group:
i.e. the probability to find 𝑘 out of the 𝑁 fish in the group to be accelerating
synchronously (Fig. 6E), and the probability of 𝑘 fish to swim in a similar
direction (Fig. 6F). For synchronous accelerations, the probability distribution
was symmetric and matched closely the expected distribution if fish were
switching states independently of one another (Fig. 6E). The distribution of
number of fish swimming synchronously in the same direction had a clear
structure and was very different from the expectation for independent fish (Fig.
6F). Thus, independent switching between modes of information processing in
individual fish on a time scale of several seconds is consistent with the
emergence of correlated directional behavior with clear temporal ordering at
the group level.
13
Fig 6: Asynchronous switching between information processing modes among
individual fish in a group and synchronous heading directions of group
members. A. Example of the simultaneous acceleration (pink)/deceleration (white)
profiles of 3 fish in a group. B. Same as in A, but for the heading direction of the fish.
pairs of fish over 1s long windows, in a group of 3 (colored lines) and shuffled controls
(light gray - see SI Methods). Right: Peak correlation value in units of standard
deviation of the shuffled data averaged for all pairs of fish within all groups of 3 fish
(N=6, P=0.75, t-test for matched samples). D. Same as in C, but for the cross-
correlation of direction of motion of pairs of fish. Here directional correlation show clear
structure and peak times (N=6, P<0.0005, t-test for matched samples). E. Average
C. Left: Distributions of delay time (𝜏P<[) that gave the maximum correlations between
empirical probability distribution 𝑃(𝑎_,𝑎',…,𝑎B) of the kinematic states of the fish
(acceleration or deceleration), where 𝑎# represent the state of fish 𝑖 (solid green), and
the prediction of a model assuming independence between fish 𝑝𝑎_𝑝𝑎' …𝑝𝑎B
distribution of heading direction 𝑃(𝑑_,𝑑',…,𝑑B), where 𝑑# is the direction of fish 𝑖
obtained under the assumption of independence 𝑝𝑑_𝑝𝑑' …𝑝𝑑B .
(dashed grey line); light shadings represent SEM. F. Same as in E, only for the
discretized into 6 even sized bins (see Methods) (solid green), and the distribution
14
Discussion
Predicting individual behavior of fish in a group, by combining active and
passive models of sensory and social information processing, proved to be
highly accurate, outperforming commonly used models that assume a universal
ongoing computation by individuals. Specifically, spatio-temporal receptive
fields captured the computation that a fish performs, surpassing current models
that assume simple topological or metric based computation. Moreover, a
comparison between food seeking vs. free swimming behavior revealed that
the computation employed by the fish depends strongly on context. At the group
level, the behavioral modes of individuals seem temporally independent among
fish, yet signatures of collective behavior still arise. The approach we presented
here merges two distinct lines of inquiry of animal behavior: studies of single-
animal behavior that have shown 'discrete behavioral modes' (36–38), and
group behavior models that have focused on qualitatively capturing complex
collective behavior (8, 24, 25, 33) emerging in groups of simple interacting
individuals described by a single behavioral mode. Our results show that
individual behavioral modes: (i) have clear kinematic proxies, (ii) suggest
distinct information processing/computation modes in individuals, and (iii) have
a significant impact on group behavior. Beyond an improved model for
individual behavior in a group, our approach portrays the group as a collection
of diverse individuals whose computations seem temporally discrete and
context-dependent, with interactions that are dynamic in space and time.
The model presented here can be improved in several ways. One possibility is
to further optimize spatiotemporal filters used to describe the visual field of a
fish and to add non-linear components to the prediction model. Improved
accuracy would allow us to explore the limits of the computation of individuals
and study the implications of noise (sensory or motor) on behavior. Another
interesting possibility would be to capture additional aspects of individual and
social computations: First, a finer dissection of individual behavior into multiple
behavioral modes might reveal further intricate processing. Second, one could
define and quantify the transitions between behavioral states in individuals and
their dependence on internal factors, as well as social ones. Third, the
differences between the receptive fields inferred under different behavioral
contexts reflect a dynamic and possibly learned nature of these receptive fields.
Modeling how individual fish use different computations based on 'personal'
tendencies, past experience, or current needs would bring us closer to
dissecting idiosyncratic behavior and understanding its effect at the group level.
The approach we presented here can be readily extended to other animal
groups. Moreover, it could be used for exploring different aspects of fine motor
15
behavior and group traits. For example, for fish this could entail mapping their
exact visual stimuli (13) to tailbeats, which would enable the study of the
mapping of sensory and social information into action, possibly in closed-loop
experimental settings. Finally, combining our approach with recording of neural
activity in members of the group (48), would allow for direct study of social and
sensory integration and processing at behavioral and neuronal levels
simultaneously.
16
Acknowledgements
We thank Ehud Karpas, Oren Forkosh, Iain Couzin, Ofer Feinerman, and Yadin
Dudai for helpful discussions and support. This work was supported by Human
Frontier Science Program, European Research Council grant # 311238, an
Israel Science Foundation grant #1629/12, as well as research support from
Martin Kushner Schnur; and Mr. and Mrs. Lawrence Feis; ES is the Joseph and
Bessie Feinberg professorial
chair
17
References:
1. Shklarsh A, Ariel G, Schneidman E, Ben-Jacob E (2011) Smart Swarms of
Bacteria-Inspired Agents with Performance Adaptable Interactions. PLoS
Comput Biol 7(9):e1002177.
2. Reid CR, et al. (2016) Decision-making without a brain: how an amoeboid
organism solves the two-armed bandit. J R Soc Interface 13(119):20160030.
3. Perna A, et al. (2012) Individual Rules for Trail Pattern Formation in Argentine
Ants (Linepithema humile). PLoS Comput Biol 8(7):e1002592.
4. Gelblum A, et al. (2015) Ant groups optimally amplify the effect of transiently
informed individuals. Nat Commun 6:7729.
5. Greenwald E, Segre E, Feinerman O (2015) Ant trophallactic networks:
simultaneous measurement of interaction patterns and food dissemination. Sci
Rep 5:12496.
6. Herbert-Read JE, et al. (2011) Inferring the rules of interaction of shoaling fish.
Proc Natl Acad Sci 108(46):18726–18731.
7. Katz Y, Tunstrøm K, Ioannou CC, Huepe C, Couzin ID (2011) Inferring the
structure and dynamics of interactions in schooling fish. Proc Natl Acad Sci
108(46):18720–18725.
8. Gautrais J, et al. (2012) Deciphering Interactions in Moving Animal Groups.
PLoS Comput Biol 8(9):e1002678.
9. Miller N, Gerlai R (2012) From Schooling to Shoaling: Patterns of Collective
Motion in Zebrafish (Danio rerio). PLoS ONE 7(11):e48865.
10. Tunstrøm K, et al. (2013) Collective States, Multistability and Transitional
Behavior in Schooling Fish. PLoS Comput Biol 9(2):e1002915.
11. Berdahl A, Torney CJ, Ioannou CC, Faria JJ, Couzin ID (2013) Emergent
Sensing of Complex Environments by Mobile Animal Groups. Science
339(6119):574–576.
12. Arganda S, Pérez-Escudero A, Polavieja GG de (2012) A common rule for
decision making in animal collectives across species. Proc Natl Acad Sci
109(50):20508–20513.
13. Rosenthal SB, Twomey CR, Hartnett AT, Wu HS, Couzin ID (2015) Revealing
the hidden networks of interaction in mobile animal groups allows prediction of
complex behavioral contagion. Proc Natl Acad Sci 112(15):4690–4695.
14. Nagy M, Ákos Z, Biro D, Vicsek T (2010) Hierarchical group dynamics in pigeon
15. Bialek W, et al. (2012) Statistical mechanics for natural flocks of birds. Proc Natl
flocks. Nature 464(7290):890–893.
Acad Sci 109(13):4786–4791.
16. Ballerini M, et al. (2008) Interaction ruling animal collective behavior depends
on topological rather than metric distance: Evidence from a field study. Proc
Natl Acad Sci 105(4):1232–1237.
17. Shemesh Y, et al. (2013) High-order social interactions in groups of mice. eLife
2:e00759.
18
18. Strandburg-Peshkin A, Farine DR, Couzin ID, Crofoot MC (2015) Shared
decision-making drives collective movement in wild baboons. Science
348(6241):1358–1361.
19. Faria JJ, Dyer JRG, Tosh CR, Krause J (2010) Leadership and social
information use in human crowds. Anim Behav 79(4):895–901.
20. Moussaïd M, Helbing D, Theraulaz G (2011) How simple rules determine
pedestrian behavior and crowd disasters. Proc Natl Acad Sci U S A
108(17):6884–6888.
21. Vicsek T, Czirók A, Ben-Jacob E, Cohen I, Shochet O (1995) Novel Type of
Phase Transition in a System of Self-Driven Particles. Phys Rev Lett
75(6):1226.
22. D'Orsogna MR, Chuang YL, Bertozzi AL, Chayes LS (2006) Self-Propelled
Particles with Soft-Core Interactions: Patterns, Stability, and Collapse. Phys
Rev Lett 96(10):104302.
23. Couzin ID, Krause J, Franks NR, Levin SA (2005) Effective leadership and
decision-making in animal groups on the move. Nature 433(7025):513–516.
24. Huth A, Wissel C (1992) The simulation of the movement of fish schools. J
Theor Biol 156(3):365–385.
25. Couzin ID, Krause J, James R, Ruxton GD, Franks NR (2002) Collective
memory and spatial sorting in animal groups. J Theor Biol 218(1):1–11.
26. Bonabeau E, Dorigo M, Theraulaz G (2000) Inspiration for optimization from
social insect behaviour. Nature 406(6791):39–42.
27. Branson K, Robie AA, Bender J, Perona P, Dickinson MH (2009) High-
throughput ethomics in large groups of Drosophila. Nat Meth 6(6):451–457.
28. Nathan R, et al. (2012) Using tri-axial acceleration data to identify behavioral
modes of free-ranging animals: general concepts and tools illustrated for griffon
vultures. J Exp Biol 215(Pt 6):986–996.
29. Ákos Z, Beck R, Nagy M, Vicsek T, Kubinyi E (2014) Leadership and Path
Characteristics during Walks Are Linked to Dominance Order and Individual
Traits in Dogs. PLOS Comput Biol 10(1):e1003446.
30. Nagy M, et al. (2013) Context-dependent hierarchies in pigeons. Proc Natl Acad
Sci 110(32):13049–13054.
31. Khuong A, et al. (2016) Stigmergic construction and topochemical information
shape ant nest architecture. Proc Natl Acad Sci U S A 113(5):1303–1308.
32. Seeley TD, et al. (2012) Stop Signals Provide Cross Inhibition in Collective
Decision-Making by Honeybee Swarms. Science 335(6064):108–111.
33. Huth A, Wissel C (1994) The simulation of fish schools in comparison with
experimental data. Ecol Model 75–76(0):135–146.
34. Zienkiewicz A, Barton DAW, Porfiri M, Bernardo M di (2014) Data-driven
stochastic modelling of zebrafish locomotion. J Math Biol:1–25.
35. Parrish JK, Edelstein-Keshet L (1999) Complexity, Pattern, and Evolutionary
Trade-Offs in Animal Aggregation. Science 284(5411):99–101.
36. Girdhar K, Gruebele M, Chemla YR (2015) The Behavioral Space of Zebrafish
Locomotion and Its Neural Network Analog. PLoS ONE 10(7):e0128668.
37. Stephens GJ, Johnson-Kerner B, Bialek W, Ryu WS (2008) Dimensionality and
Dynamics in the Behavior of C. elegans. PLOS Comput Biol 4(4):e1000028.
19
38. Berman GJ, Choi DM, Bialek W, Shaevitz JW (2014) Mapping the stereotyped
behaviour of freely moving fruit flies. J R Soc Interface 11(99):20140672.
39. Suriyampola PS, et al. (2015) Zebrafish Social Behavior in the Wild. Zebrafish
13(1):1–8.
40. Portugues R, Feierstein CE, Engert F, Orger MB (2014) Whole-Brain Activity
Maps Reveal Stereotyped, Distributed Networks for Visuomotor Behavior.
Neuron 81(6):1328–1343.
41. Aoki T, et al. (2013) Imaging of Neural Ensemble for the Retrieval of a Learned
Behavioral Program. Neuron 78(5):881–894.
42. Bianco IH, Engert F (2015) Visuomotor Transformations Underlying Hunting
Behavior in Zebrafish. Curr Biol 25(7):831–846.
43. Ahrens MB, et al. (2012) Brain-wide neuronal dynamics during motor adaptation
in zebrafish. Nature 485(7399):471–477.
44. Buske C, Gerlai R (2011) Shoaling develops with age in Zebrafish (Danio rerio).
Prog Neuropsychopharmacol Biol Psychiatry 35(6):1409–1415.
45. Müller UK, Leeuwen JL van (2004) Swimming of larval zebrafish: ontogeny of
body waves and implications for locomotory development. J Exp Biol
207(5):853–868.
46. Chichilnisky EJ (2001) A simple white noise analysis of neuronal light
responses. Netw Bristol Engl 12(2):199–213.
47. Pérez-Escudero A, Vicente-Page J, Hinz RC, Arganda S, de Polavieja GG
(2014) idTracker: tracking individuals in a group by automatic identification of
unmarked animals. Nat Methods 11(7):743–748.
48. Vinepinsky E, Donchin O, Segev R (2017) Wireless electrophysiology of the
brain of freely swimming goldfish. J Neurosci Methods 278:76–86.
20
Supplementary information
Methods:
Experimental system.
75 adult zebrafish (Danio Rerio), purchased from Aquazone Israel LTD, at
approximately 1:1 male: female ratio were studied. Fish were housed separately in
their designated groups for at least one month prior to experiments. Environmental
conditions were constant using a re-circulating system and multistage filtration, with
water temperature of 27-28°C, conductivity of 600-700 𝜇𝑠𝑖𝑒𝑚𝑒𝑛𝑠 and PH levels of 7.7–
8. Lighting was kept at 14:10 light/dark cycle with fluorescent lights. Unless otherwise
indicated, fish were fed twice a day a mixture of dry flake food. Experimental arena
consisted of a large 1.2m over 1.1m rectangular aquarium with circular arenas of
different diameters placed in it (see Movie M1). Water levels were kept at a depth of
about 5 cm to constitute a pseudo 2-dimensional environment.
Video recording was done using an industrial recording system with a Vieworks VC-
2MC-M340 digital camera with an 8mm lens, connected to a Karbon-CL frame
grabber. Camera was attached to the ceiling over the test aquarium approximately
150cm above water level to capture the entire arena.
Data extraction.
Videos were analyzed off-line to extract the physical properties of the fish (size,
position, orientation). Position data was used to estimate fish trajectories using a
designated tracker. All image processing and tracking was done using Matlab with
software written in our lab. Briefly, fish were first detected as darker blobs over the
lighter background, and their physical properties calculated. Next, the center of mass
of fish were connected frame by frame to give the estimated track of each fish. When
several fish were close to one another an additional step was taken to estimate the
most likely number of fish in the large blob and their centers. This tracking method
provided accurate detection of fish in >90% of the frames analyzed, but did not ensure
constant identities of the fish. When needed, fish identities were corrected using the
IdTracker software (1). Fish trajectories were smoothed using a Savitzky-Golay filter
(2) spanning 33 frames which constitutes ~1/3 of a second. Fish positions were set
as the coordinates of the fish center: 𝑐#(𝑡)=[𝑥𝑡 #,𝑦(𝑡)#], and fish velocity was
estimated as the difference between two consecutive points: 𝑣𝑡 #=𝑐𝑡 #−𝑐𝑡−1 #.
Direction of motion of the fish was defined as 𝑑𝑡 #= ,.-
,.- , or 𝜃𝑡 # as the angle of
𝑑𝑡 #, and angular velocity was given by 𝜔𝑡 #=𝜃𝑡 #−𝜃𝑡−1 #. As fish tend to
respond strongly to walls of the arena, we did not use data from fish positioned 'close'
to the boundary - all distances smaller than the median of the wall distance distribution
were discarded from further analysis.
Behavioral experiments.
Free-swimming. Prior to behavioral experiments, fish were habituated to the circular
arena (95cm diameter) for short sessions ~10 minutes for 2 days. We then filmed their
free-swimming behavior for 30 or 60 minutes.
Food-seeking. To train fish to seek for food in the arena we conducted a 7-day training
protocol. On each day, fish were transferred from their home tank to the experimental
arena where a constant number of flakes (~4mm in diameter) were scattered randomly
on the water surface. Fish were allowed 5 minutes to consume the flakes and then
netted and returned to their home tanks. To facilitate learning, the effective size of the
arena used by the fish was increased over days (from a small box of 25X25cm to
21
circular arenas of diameter of 57cm, 82cm, 95cm). During training, no other food was
given to the fish, and on days 6-7 fish were deprived of food. On day 8, fish were
transferred again to the test tank and their behavior was recorded, with no food
present.
Kinematic models of fish acceleration and deceleration segments.
Segmentation of speed profiles into acceleration and deceleration epochs was done
by detecting the minima and maxima of the speed profile using numerical
differentiation. To overcome local noise, we constrained two local minima to be
separated by at least 4 frames (or 40 ms) and the single highest extrema between two
such minima points was taken as the local maximum point (and end of acceleration).
Each segment 𝑗 of fish 𝑖 was fitted individually: a single exponential 𝑠#R>D[I]𝑡 =
𝑚𝑎𝑥 (𝑠#R>D[I])⋅𝑒(kl-m⋅.) was used to fit decelerations and a sigmoid 𝑠#<DD[I](𝑡)=
_p>q-m(rsrt) for accelerations. Maximum speeds for both models - 𝑚𝑎𝑥 (𝑠#<DD[I]) and
P<[ (=-noo[m])
𝑚𝑎𝑥 (𝑠#R>D[I]) were obtained form the data as was the midpoint of the sigmoidal function
𝑡u, leaving the 'slope' of both models 𝜂#I and 𝛾#I as free parameters to be fitted (see
Fig. S1B-D). Fitting was done only for segments that were at least 100ms long (i.e.
having at least 10 data points for fitting).
Neighbor maps.
Neighbor position maps were based on 2-dimensional histograms of neighbors'
positions in space, relative to a focal fish whose orientation was defined as "north".
Histograms were smoothed using a rotationally symmetric Gaussian low pass filter
(𝜎'= 0.13 [body lengths]).
To estimate the neighbor alignment maps, the average direction of motion of fish in
each bin was calculated using all frames in which the bin was occupied. The angular
deviation of that average vector from the direction of motion of the focal was expressed
in angles [-180 180] where positive values represent deviation to the right ('east') and
negative values represent deviation to the left ('west').
Model Fitting.
Receptive field models were fit using a Lasso least-squares regression (3) with cross
validation. Briefly, for a given non-negative 𝜆 we calculated
,
β(cid:134)
where 𝑣. is the empirical velocity of fish in time instance 𝑡, 𝑥 is a set of velocities of
main text), 𝛽 is a vector of model parameters or bin weights (of length p), and N is the
total number of observations. Repeating this procedure for different values of 𝜆, we
found the set of parameters 𝛽 that minimized the cross-validation error for held out
neighboring fish in the spatio-temporal receptive field around the fish (see Eq. (3) in
v(cid:127)−x(cid:129)⋅β '
data. This regularization process usually reduces the effective number of parameters
(setting some of the weights to zero) resulting in a sparser model.
Competing models.
We have compared our receptive-field parameterization of space, which depends on
both angle and distance of neighbors from the focal fish (Fig. 2B), to two commonly
used parameterization of space: a zonal model (4–6) where only the distance of
min
12N
(cid:131)
(cid:127)(cid:132)_
+λ
(cid:135)
(cid:134)(cid:132)_
(1)
22
(cid:142),(cid:146)
(2)
neighboring fish is taken into account, and a topological model (7) where neighbors
are weighed according to their topological order (first neighbor, second neighbor, etc)
ignoring their metric distance (see insets in Fig. 4C for model sketch). Mathematically
we get a similar expression as in eq. (3) in the main text
β(cid:134)k ⋅v(cid:134)t−kΔt +
β(cid:142)k ⋅d(cid:142)t−kΔt
Δv(cid:137)(cid:138)(cid:139)(cid:140)(cid:141)(cid:142)t =
(cid:134),(cid:146)
where the first term is the social interaction term, but spatial weights are assigned to
bins according to their distance from the focal fish (zonal model) or to topological
neighbors. Temporal binning was kept similar to that used in the RF model for
comparison.
Model parameters.
Discretization of the receptive field into spatio-temporal bins and temporal bins of the
topological and zonal models were chosen using a non-exhaustive search of
parameter space, whereas zonal model radii, were optimized for the cases of either 2
or 3 rings. Thus, the RF parameters used here give only a lower bound on the accuracy
of this model since an exhaustive search of the parameter space can optimize the
obtained results. All parameters used in the models are listed below:
For the RF-model:
Parameter name
# of sectors
Value used in analysis
6 sectors
Meaning
Number of sectors in the
Receptive field
The middle point of the
first sector with respect to
fish motion.
Number of rings in the
receptive field.
𝑅#p_−𝑅# where 𝑅# is the
radius of ring i.
Meaning
Number of rings in the
zonal model
𝑅#p_−𝑅# where 𝑅# is the
radius of ring i.
Number of steps back in
time used for prediction
Time window between fish
motion and neighbor first
neighbor
response
configurations
Time window between first
and
response
second
configurations,
and
between second and third.
0 degrees (north)
6 rings
1 body length
Value used in analysis
2 rings
3 body length
3 steps
150 ms (15 frames)
200 ms (20 frames)
Start angle
# of rings
Ring size
For the Zonal models
Parameter name
# of rings
Ring size
𝛥𝑡_
𝛥𝑡',(cid:148)
Temporal parameters similar for all models:
# of steps back
23
sec (in 10ms increments):
Trajectory reconstruction.
obtained either by the Active model or the Passive model (see main text). The
'combined model' is where for each time step t we use the prediction of the model that
velocities for that time step. We note that this is only a locally 'optimal' choice, that
does not guarantee that the total error between real and predicted trajectories will be
minimal.
Directional and behavioral cross-correlation analysis.
Time windowed directional cross-correlation (8, 9) was estimated using a short window
Predicted positions of fish 𝑖, 𝐶#(𝑡) over time, were obtained using the instantaneous
velocities 𝑣#(𝑡) predicted by the models where 𝐶#(𝑡)=𝐶#(𝑡−1)+ 𝑣#(𝑡); 𝑣#𝑡 were
gave the lower error 𝐸PQR>L=𝑣C><L−𝑣PQR>L between the real and predicted
of the response of fish 𝑖 (𝐿= 1𝑠), and the responses of fish 𝑗 over a range −4≤𝜏≤4
where 𝑣# and 𝑣I are fish velocities. The time delay 𝜏 that gives the maximal correlation
then calculate 𝐶𝑡 for all time points and for all pairs of fish in a given group and find
the maximal correlation for each segment 𝐶#IP<[ (Fig. S6A), and the temporal delay
corresponding to this maximal correlation 𝜏P<[(D). Histograms of temporal relations
(Fig 6D and Fig. S6B bottom) are then constructed using these 𝜏P<[(D) values, keeping
only temporal delays of high correlations 𝐶#IP<[>0.9, to focus only on instances where
below). Importantly, choosing different thresholds for 𝐶#IP<[ did not qualitatively change
use 𝑥# and 𝑥I which are binary variables denoting acceleration (1) or decelerations (0)
𝜏P<[(D) values corresponding to 𝐶#IP<[≥0.8 for the behavioral state correlation, as to
the results obtained (Fig. S6B, bottom).
Behavioral state correlation was estimated in a similar manner, but in this case, we
the fish are actually responding to one another and to ensure that more that 10% of all
correlations are retained both for the actual data and for the shuffled analysis (see
(Fig S6A Left). When constructing the temporal delay distributions (Fig. 6C) we kept
for this segment is then considered to be the temporal relation between animals. We
𝐶#I𝑡 =
(cid:152)
𝑣#𝑡 𝑣I𝑡−𝜏
retain at least 10% of the correlation in the shuffled analysis. Choosing different
threshold values for this analysis as well gave similar results (Fig. S6B top).
To obtain a null distribution of expected correlation values we repeated the same time-
windowed cross-correlation analysis as formulated above for both direction and
behavioral state, but with each fish's trajectories randomly shifted in time. We repeated
this full analysis 1000 times for each pair of fish (see gray lines in Fig 6C,D and in Fig.
S6A,B). To compare the magnitude of the peaks of the time delay distributions
𝑚𝑎𝑥(𝑃(𝜏P<[)) found in real pairs and that of the time-shuffled pairs, we normalized
the maximal values of 𝑃(𝜏P<[) using the mean and standard deviation of the maxima
found in the repetitions of the shuffled analysis
𝑠𝑐𝑜𝑟𝑒=𝑝𝑒𝑎𝑘−𝜇=(cid:157)(cid:158)(cid:159)(cid:159)L>R
𝜎=(cid:157)(cid:158)(cid:159)(cid:159)L>R
24
where 𝜇=(cid:157)(cid:158)(cid:159)(cid:159)L>R and 𝜎=(cid:157)(cid:158)(cid:159)(cid:159)L>R are the mean and standard deviation of the peak in the
shuffled data. This why we are effectively conducting a bootstrap analysis comparing
the maximal probability in our data to that found in the shuffled analysis (Fig 6C,D
right).
Estimating empirical distributions of collective states.
We evaluated
joint probability distributions of synchronous
fish states,
the
occupying the same state. For comparison, we estimate the independent probability
be in each state, taken as the average over the entire session.
Similarly, we estimated the probability distribution of the synchronous swimming
𝑃(𝑎_,𝑎',…,𝑎B), where 𝑎# is the kinematic state of fish 𝑖 (we set 1 for accelerations and
0 for decelerations) and calculate the probability of seeing either 1,2,…,𝑘 out of 𝑛 fish
distribution, 𝑝𝑎_𝑝𝑎' …𝑝𝑎B , where 𝑝(𝑎#) is the independent probability of fish 𝑖 to
direction of fish in a group as 𝑃(𝑑_,𝑑',…,𝑑B), where 𝑑# is the direction of swimming of
fish 𝑖 binned into 6 angular even sized bins. We compared it to the distribution of
directions obtained under a similar independence assumption, 𝑝𝑑_𝑝𝑑' …𝑝𝑑B .
taken to be 𝜔=R¡R. where 𝜃 is the heading direction of the fish. Fitting an RF model to
Predicting full acceleration epochs.
We define the response of a fish over a full acceleration epoch as the integral of the
speed
(see Fig S7A). Angular velocity is
and of the angular velocity 𝜔#𝑑𝑡
.u
predict these quantities was a similar process to the one used for velocity prediction.
For comparison, we also fitted the RF-model for instantaneous velocities again,
reducing the amounts of the data used in the fitting process so it would match the
number of complete acceleration epochs. To this end, we chose a single
representative velocity positioned 150ms after the transition between deceleration and
acceleration. This choice gave very similar results to the RF-model fits on the entire
data set (see Fig S7C).
𝑆#𝑑𝑡
.u
25
SI Figures:
Figure S1: Kinematic states of individual fish in a group. A. A snapshot of the
tracks of 3 fish. B. All acceleration and deceleration epochs for a single fish over 5
minutes of swimming, with time normalized between zeros and 2 (1 marks the
transition from acceleration to deceleration) and speed normalized between zeros and
one. Bold lines are the means over all epochs. C. Distributions of 𝑅' values for all the
segments shown in B, with a median value of 0.99 for accelerations and 0.95 for
decelerations representing high model fit accuracies. D. Fitted parameters for
26
of successive acceleration and deceleration epochs (shown in B) plotted one against
accelerations (left) and decelerations (right). Sigmoid slope 𝛾 shows a distribution
peaked around −30 [ __uP=], and friction coefficient around 𝜂=−2.7[ __uP=]. E. Duration
the other, shows a very week linear relationship between the two (Pearson's 𝐶𝐶 =
0.13, 𝑃=0.0016). F. Distributions of acceleration and deceleration durations for all
fish in all groups. Acceleration epochs had a mean duration of ~200±104 𝑚𝑠, and
were generally shorter than decelerations with a mean duration of ~250±160𝑚𝑠 G.
4.15 BL/s) H. Distributions of the angular velocity 𝜔=R¡R. with 𝜃 being the heading of
distribution centered around zero for decelerations. I. Angular velocity 𝜔 of successive
Probability of switching states as a function of swimming speed. As can be expected,
the probability of switching from deceleration to acceleration is higher for low speed
values with a peak at 0.45 BLs/s and a narrow distribution (middle 95% of the
distribution between 0.05-1.85 BLs/s), while switching back to deceleration is peaked
at 1 BLs/s with a much wider distribution (middle 95% of the distribution between 0.2-
the fish, for the acceleration and deceleration epochs shown in B. Note the narrow
acceleration and deceleration epochs. No clear relationship seems to exist between
successive epochs.
27
Figure S2: Error analysis confirms that fish show little social responses during
decelerations. A. Comparison of the normalized error in prediction 𝑣PQR>L−𝑣C><L/
𝑣C><L during acceleration epochs (left) and deceleration epochs (right), using an
Example distributions of the errors 𝑣PQR>L−𝑣C><L in prediction during acceleration
active model -- a RF model learned separately on each kinematic state (Eq. (3) main
text) and a passive model (Eq. 1 main text). There is no advantage in learning a
separate RF-model to predict active movements in the deceleration epochs. B.
epochs (left) and deceleration epochs (right) for the two models as in A, again, error
distributions in the deceleration epochs are similar for active or passive computation
in these parts.
28
Figure S3: Switching in combined model is comparable to the kinematic state if
the fish. A. The trajectory segment from Fig. 3C (black) and the predicted trajectory
using the optimal combination of the two models (green), overlaid with the predicted
trajectory based on switching between models that is done according to the
acceleration or deceleration state of the fish (purple). B. An example of the optimal
switching between models in the reconstruction shown in Fig. 3C and panel A (green
line), and their high correspondence to the acceleration/deceleration state of the fish
(pink background represents accelerations. C. Left: example distributions of the
reconstruction error: 𝐸𝑟𝑟𝑜𝑟= 𝑐PQR>L−𝑐C><L using the combined model (green)
compared to the active (red) and passive (blue) models. Right: comparison of the
combined model (green) to switching according to the acceleration/deceleration state
of the fish (purple). D. Average reduction in the reconstruction error of the combined
model compared to the active model alone (red), the passive model alone (blue), and
to the prediction obtained by switching according to the acceleration/deceleration state
of the fish (purple). Error bars represent STD. E. Similar to D only for reconstruction of
longer trajectory segments (24 sec), giving highly similar results.
29
Figure S4: Active movement changes are accurately predicted by the RF model
using both social and sensory information. A. Improvement in prediction accuracy
(correlation between data and model prediction) of the RF model using both social and
sensory information, compared to the social information alone (blue) and sensory
information alone (red), for all group sizes (N = 6,7,7, error bars represent SEM). B.
Improvement in prediction accuracy by using the RF-model compared to the zonal and
topological models when using only sensory information for prediction for all group
sizes (N=6,7,7 error bars represent SEM). C. Predicted velocity components 𝛥𝑣[,𝛥𝑣£
of 𝛥𝑣 plotted against their measured values. The linearity and homogeneity of variance
across real values suggest we should expected a limited benefit from adding non-
linarites to our model.
30
Figure S5: Receptive field maps show distinct behavioral state dependencies. A.
Average difference maps of the receptive field for groups of 3 fish during acceleration,
when performing free swimming and food seeking (see above). Bins with weights that
were larger in the free-swimming state are colored red and opposite bins are in blue.
B. Same as A, but showing only bins with differences that pass a significance test (t-
test for matched samples, P<0.05).
31
Figure S6: Asynchronous switching between information processing modes
among individual fish in a group and synchronous heading directions of group
members. A. Distributions of maximal correlation values for a group of 3 fish (from Fig
6), for the state correlation analysis (left) and for the directional correlation analysis
32
(right), and the correlations obtained using shuffled data in gray (see text above). B.
Distributions of delay time (𝜏P<[) of the maximum correlations between pairs of fish in
different correlation thresholds, where only 𝜏P<[ values corresponding to maximal
a group of 3 (colored lines) and shuffled controls (light gray). Different colors represent
correlations above these thresholds are used to construct the distributions. For both
the state correlation (top – green lines) and for directional correlation (bottom – purple
lines), using different thresholds does not change the structure of the result presented
in Fig 6. For comparison, shuffled controls are plotted using correlation threshold of
0.95. C. Group state correlation for groups of 3 fish, similar to the analysis presented
in Fig. 6E-F for 6 fish.
33
Figure S7: Prediction of complete speeding and turning profiles using receptive
field models. A. Examples of speeding (left) and turning profiles (angular velocity -
right), colored by the strength of the response ('low', 'medium', and 'high'), which were
determined by discretizing the integrals of responses (insets) into even-sized bins. B.
Inset: RF maps showing bins with significantly different neighboring fish behavior
before each of the binned responses depicted in A (see methods above). Main:
Average velocity vectors of the fish in the bin marked in red in the inset, for different
speeding strengths (left) and turning strengths (right); strength is denoted by color as
in A. C. Comparison of the accuracy of prediction for the instantaneous model when
and speeding -
profiles (purple and green respectively). Values are average
correlation coefficient (N=6,7,7) and error bars represent SEM. D. Prediction accuracy
predicting 𝛥𝑣#(𝑡) (blue), and the accuracy of predicting the complete turning - 𝜔𝑑𝑡
.u
𝑣𝑑𝑡
.u
34
of complete turning and complete speeding responses using social and sensory
information preceding these events, represented as percentage of the accuracy of
predicting point accelerations (𝛥𝑣#(𝑡)), using similar data. Indeed, the ability to predict
properties of entire acceleration epoch is significantly lower.
35
References:
1. Pérez-Escudero A, Vicente-Page J, Hinz RC, Arganda S, de Polavieja GG
(2014) idTracker: tracking individuals in a group by automatic
identification of unmarked animals. Nat Methods 11(7):743–748.
2. Savitzky A, Golay MJE (1964) Smoothing and Differentiation of Data by
Simplified Least Squares Procedures. Anal Chem 36(8):1627–1639.
3. Tibshirani R (1994) Regression Shrinkage and Selection Via the Lasso. J
R Stat Soc Ser B 58:267--288.
4. Huth A, Wissel C (1992) The simulation of the movement of fish
schools. J Theor Biol 156(3):365–385.
5. Huth A, Wissel C (1994) The simulation of fish schools in comparison
with experimental data. Ecol Model 75–76(0):135–146.
6. Couzin ID, Krause J, James R, Ruxton GD, Franks NR (2002) Collective
memory and spatial sorting in animal groups. J Theor Biol 218(1):1–11.
7. Ballerini M, et al. (2008) Interaction ruling animal collective behavior
depends on topological rather than metric distance: Evidence from a
field study. Proc Natl Acad Sci 105(4):1232–1237.
8. Ákos Z, Beck R, Nagy M, Vicsek T, Kubinyi E (2014) Leadership and Path
Characteristics during Walks Are Linked to Dominance Order and
Individual Traits in Dogs. PLOS Comput Biol 10(1):e1003446.
9. Nagy M, Ákos Z, Biro D, Vicsek T (2010) Hierarchical group dynamics in
pigeon flocks. Nature 464(7290):890–893.
36
|
1003.1200 | 1 | 1003 | 2010-03-05T07:50:49 | Learning as a phenomenon occurring in a critical state | [
"q-bio.NC"
] | Recent physiological measurements have provided clear evidence about scale-free avalanche brain activity and EEG spectra, feeding the classical enigma of how such a chaotic system can ever learn or respond in a controlled and reproducible way. Models for learning, like neural networks or perceptrons, have traditionally avoided strong fluctuations. Conversely, we propose that brain activity having features typical of systems at a critical point, represents a crucial ingredient for learning. We present here a study which provides novel insights toward the understanding of the problem. Our model is able to reproduce quantitatively the experimentally observed critical state of the brain and, at the same time, learns and remembers logical rules including the exclusive OR (XOR), which has posed difficulties to several previous attempts. We implement the model on a network with topological properties close to the functionality network in real brains. Learning occurs via plastic adaptation of synaptic strengths and exhibits universal features. We find that the learning performance and the average time required to learn are controlled by the strength of plastic adaptation, in a way independent of the specific task assigned to the system. Even complex rules can be learned provided that the plastic adaptation is sufficiently slow. | q-bio.NC | q-bio |
Learning as a phenomenon
occurring in a critical state
Lucilla de Arcangelis ∗ † ‡ and Hans J. Herrmann †
†Institute Computational Physics for Engineering Materials, ETH, Schafmattstr.6, 8093 Zurich, CH, and ‡Department of Information Engineering and CNISM, Second University
of Naples, 81031 Aversa (CE), Italy
Submitted to Proceedings of the National Academy of Sciences of the United States of America
Recent physiological measurements have provided clear evidence
about scale-free avalanche brain activity and EEG spectra, feeding
the classical enigma of how such a chaotic system can ever learn
or respond in a controlled and reproducible way. Models for learn-
ing, like neural networks or perceptrons, have traditionally avoided
strong fluctuations. Conversely, we propose that brain activity hav-
ing features typical of systems at a critical point, represents a crucial
ingredient for learning. We present here a study which provides novel
insights toward the understanding of the problem. Our model is able
to reproduce quantitatively the experimentally observed critical state
of the brain and, at the same time, learns and remembers logical
rules including the exclusive OR (XOR), which has posed difficulties
to several previous attempts. We implement the model on a net-
work with topological properties close to the functionality network
in real brains. Learning occurs via plastic adaptation of synaptic
strengths and exhibits universal features. We find that the learning
performance and the average time required to learn are controlled
by the strength of plastic adaptation, in a way independent of the
specific task assigned to the system. Even complex rules can be
learned provided that the plastic adaptation is sufficiently slow.
learning
scale-free network
self-organized criticality
scaling
Spontaneous activity is an important property of the cere-
bral cortex that can have a crucial role in information pro-
cessing and storage. Recently it has been shown that a
novel spatio-temporal form of spontaneous activity is neuronal
avalanches, which can involve from a few to a very large num-
ber of neurons. These bursts of firing neurons have been first
observed [1, 2] in organotypic cultures from coronal slices of
rat cortex, where the size and duration of neuronal avalanches
follow power law distributions with very stable exponents.
The presence of a power law behaviour is the typical feature of
a system acting in a critical state [3], where large fluctuations
are present and the response does not have a characteristic
size. The same critical behaviour, namely the same power law
exponents, has been recently measured also in vivo from su-
perficial layers of cortex in anesthetized rats during early post-
natal development [4], and awake adult rhesus monkeys [5],
using micro-electrode array recordings. Results confirm that
indeed spontaneous cortical activity adjusts in a critical state
where the spatio-temporal organization of avalanches is scale
invariant. Moreover, the investigation on the spontaneous ac-
tivity of dissociated neurons from different networks as rat
hippocampal neurons [6], rat embryos [7] or leech ganglia [6],
has also confirmed the robustness of this scaling behaviour. In
all these cases, the emergence of power law distributions has
been interpreted in terms of self-organized criticality (SOC)
[8]. The term SOC usually refers to a mechanism of slow en-
ergy accumulation and fast energy redistribution driving the
system toward a critical state, where the avalanche extensions
and durations do not have a characteristic size.
The understanding of the fundamental relations between
electro-physiological activity and brain organization with re-
spect to performing even simple tasks, is a long-standing fas-
cinating question. A number of theoretical models [9] have
been proposed for learning, from the simple perceptron [10]
to Attractor Neural Networks (ANNs) [11] of artificial two-
state neurons [12]. In these models the state of the "brain"
is the snapshot of the ensemble of the individual states of all
neurons, which explores phase space following an appropriate
dynamics and eventually recovers memories. The ability of
the brain to self-organize connections in an efficient way is a
crucial ingredient in biologically plausible models. The break-
through of Hebbian plasticity, postulating synapse strengthen-
ing for correlated activity at the pre- and post-synaptic neuron
and synapse weakening for decorrelated activity, triggered the
development of algorithms for neuronal learning and memory,
as, for instance, "reinforcement learning" [13] or error back-
propagation [14], leading for the first time to the XOR rule
learning. Recent results have shown that extremal dynam-
ics, where only the neuron with the largest input fires, and
Fig. 1. One neuronal avalanche in the scale-free network. 40 neurons are con-
nected by directed bonds (direction indicated by the arrow at one edge), representing
the synapses. The size of each neuron is proportional to the number of in-connections,
namely the number of dendrites. The two red neurons are the two input sites, whereas
the black neuron is the output. Connections involved in the avalanche propagation
are shown in red, whereas inactive connections are black.
The authors declare no conflict of interest
∗To whom correspondence should be addressed. Email: ¡[email protected]¿
c(cid:13)2007 by The National Academy of Sciences of the USA
www.pnas.org - -
PNAS
Issue Date
Volume
Issue Number
1–7
uniform negative feedback are sufficient ingredients to learn
the following task: to identify the right connection between
an input and an output node [15, 16]. Similarly, low activ-
ity probabilistic firing rules, where again a single neuron fires
at each step of the iteration, together with a uniform nega-
tive feedback plastic adaptation acting on time scales slower
than the neuron firing time scale, enables learning the XOR
rule without error back-propagation [17]. Both results sug-
gest that the system learns by mistakes, namely depression
rather than enhancement of synaptic strength is the crucial
mechanism for learning. However, in both studies a single
neuron fires at each step of the evolution, not in complete
agreement with recent experimental discoveries. Cooperative
effects leading to self-organization and learning are completely
neglected in the aforementioned approaches.
Operating at a critical level, far from an uncorrelated sub-
critical or a too correlated supercritical regime, may opti-
mize information management and transmission in real brains
[18, 1, 19, 20], as recently confirmed by experiments [21].
Moreover, a recent study of visual perceptual learning has
evidenced that training to a specific task induces dynamic
changes in the functional connectivity able to "sculpt" the
spontaneous activity of the resting human brain and to act
as a form of "system memory" [22].
It is therefore tempt-
ing to investigate the role that critical behaviour plays in the
most important task of neuronal networks, namely learning
and memory. The emergence of a critical state with the same
critical behaviour found experimentally has been recently re-
produced by a neuronal network model based on SOC ideas
[23, 24]. The model implements several physiological proper-
ties of real neurons: a continuous membrane potential, firing
at threshold, synaptic plasticity and pruning. Extensive nu-
merical studies on regular, small world and scale free networks
have shown that indeed the system exhibits a robust critical
behaviour. The distributions of avalanche size and duration
scale with exponents independent of model parameters and
in excellent agreement with experimental data (Fig.2). More
precisely, the distribution of avalanche sizes, measured exper-
imentally either in terms of number of active electrodes or
summed local field potentials in a micro-electrode array [1, 2],
decreases with an exponent −1.5, whereas the distribution
of avalanche temporal durations decreases with an exponent
close to −2.0. A critical avalanche activity has been also found
on fully connected [25] and random networks [26]. Moreover,
the temporal signal for electrical activity and the power spec-
trum of the resulting time series have been compared with
EEG data [23, 24]. The spectrum exhibits a power law be-
haviour, P (f ) ∼ f −0.8, with an exponent in good agreement
with EEG medical data [27] and physiological signal spectra
for other brain controlled activities [28]. This model therefore
seems to capture many of the essential ingredients of sponta-
neous activity, as measured in cortical networks.
Here we study the learning performance of a neuronal net-
work acting in a critical state. The response of the system to
external stimuli is therefore scale-free, i.e. no characteristic
size in the number of firing neurons exists. The approach
reproduces closely the physiological mechanisms of neuronal
behaviour and is implemented on a plausible network having
topological properties similar to the brain functionality net-
work. Neuronal activity is a collective process where all neu-
rons at threshold can fire and self-organize an efficient path
for information transmission. Plastic adaptation is introduced
via a non uniform negative feedback procedure with no error
back propagation.
The model
We consider N neurons positioned at random in a two dimen-
sional space. Each neuron is characterized by the potential vi.
Connections among neurons are established by assigning to
each neuron i a random out-going connectivity degree, kouti .
The distribution of the number of out-connections is then cho-
sen in agreement with the experimentally measured properties
of the functionality network [29] in human adults. Functional
magnetic resonance imaging has indeed shown that this net-
work has universal scale free properties, namely it exhibits a
scaling behaviour n(kout) ∝ k−2
out, independent of the different
tasks performed by the patient. We adopt this distribution
for the number of pre-synaptic terminals of each neuron, over
the range of possible values between kmin
out = 100,
as in experimental data. Two neurons are then connected ac-
cording to a distance dependent probability, p(r) ∝ e−r/r0 ,
where r is their spatial distance [30] and r0 a typical edge
length. To each synaptic connection we then assign an initial
random strength gij, where gij 6= gji, and an excitatory or in-
hibitory character, with a fraction pin of inhibitory synapses.
An example of such a network is shown in Fig.1.
out and kmax
The firing dynamics implies that, whenever at time t the
value of the potential at a site i is above a certain thresh-
old vi ≥ vmax = 6.0, approximately equal to −55mV for real
neurons, the neuron sends action potentials leading to the pro-
duction of an amount of neurotransmitter proportional to vi.
As a consequence, the total charge released by a neuron is pro-
portional to the number of synaptic connections, qi ∝ vikouti .
Each connected neuron receives charge in proportion to the
strength of the synapse gij
vj (t + 1) = vj (t) ±
qi(t)
kinj
gij(t)
Pk gik(t)
[ 1 ]
where kinj is the in-degree of neuron j and the sum is ex-
tended to all out-going connections of i.
In Eq.(1) it is as-
10000
1000
100
10
1
0.1
0.01
1
S
T
10
100
1000
Fig. 2. Demonstration of the critical behaviour of the neuronal network avalanche
activity. The distribution of the sizes of the avalanches of firing neurons, n(S) (cir-
cles), follows a power law behaviour with an exponent 1.5 ± 0.1 (dashed line). The
size is measured as the number of firing neurons. The distribution of avalanche dura-
tions, n(T ) (squares), exhibits a power law behaviour with an exponent 2.2 ± 0.2
(dot-dashed line), followed by an exponential cutoff. Data are obtained for 40 real-
izations of a network of 4000 neurons with pin = 0.05.
2
www.pnas.org - -
Footline Author
sumed that the received charge is distributed over the surface
of the soma of the post-synaptic neuron, proportional to the
number of in-going terminals kinj . The plus or minus sign in
Eq.(1) is for excitatory or inhibitory synapses, respectively.
After firing a neuron is set to a zero resting potential and in
a refractory state lasting tref = 1 time step, during which it
is unable to receive or transmit any charge. We wish to stress
that the unit time step in Eq.(1) does not correspond to a
real time scale, it is simply the time unit for charge propaga-
tion from one neuron to the connected ones. In real system
this time could vary and be as large as 100 ms for longer
firing periods. The synaptic strengths have initially a ran-
dom value gij ∈ [0.5, 1.0], whereas the neuron potentials are
uniformly distributed random numbers between vmax − 1 and
vmax. Moreover, a small random fraction (10%) of neurons is
chosen to be boundary sites, with a potential fixed to zero,
playing the role of sinks for the charge.
In order to start activity we identify input neurons at
which the imposed signal is applied and the output neuron at
which the response is monitored. These nodes are randomly
placed inside the network under the condition that they are
not boundary sites and they are mutually separated on the
network by kd nodes. kd represents the chemical distance on
the network and plays the role of the number of hidden layers
in a perceptron. We test the ability of the network to learn dif-
ferent rules: AND, OR, XOR and a random rule RAN which
associates to all possible combinations of binary states at three
inputs a random binary output. More precisely, the AND, OR
and XOR rules are made of three input-output relations (we
disregards the double zero input which is a trivial test leading
to zero output), whereas the RAN rule with three input sites
implies a sequence of seven input-output relations. A single
learning step requires the application of the entire sequence
of states at the input neurons, monitoring the state of the
output neuron. For each rule the binary value 1 is identified
with the output neuron firing, namely the neuron membrane
potential at a value greater or equal to vmax at some time
during the activity. Conversely, the binary state 0 at the out-
% success
% succ. α0.05
put neuron corresponds to the physiological state of a real
neuron which has been depolarized by incoming ions but fails
to reach the firing threshold membrane potential during the
entire avalanche propagation. Once the input sites are stimu-
lated, their activity may bring to threshold other neurons and
therefore lead to avalanches of firings. We impose no restric-
tion on the number of firing neurons in the propagation and let
the avalanche evolve to its end according to Eq.(1). If at the
end of the avalanche the propagation of charge did not reach
the output neuron, we consider that the state of the system
was unable to respond to the given stimulus, and as a conse-
quence to learn. We therefore increase uniformly the potential
of all neurons by units of a small quantity, β = 0.01, until the
configuration reaches a state where the output neuron is first
perturbed. We then compare the state of the output neuron
with the desired output. Namely we follow the evolution in
phase space of the initial state of the system and verify if the
non-ergodic dynamics has led to an attractor associated with
the right answer.
Plastic adaptation. Plastic adaptation is applied to the sys-
tem according to a non uniform negative feedback algorithm.
Namely, if the output neuron is in the correct state according
to the rule, we keep the value of synaptic strengths. Con-
versely, if the response is wrong we modify the strengths of
those synapses involved in the information propagation by
±α/dk, where dk is the chemical distance of the presynaptic
neuron from the output neuron. Here α represents the ensem-
ble of all possible physiological factors influencing synaptic
plasticity. The sign of the adjustment depends on the mis-
take made by the system: If the output neuron fails to be in a
firing state we increase the used synapses by a small additive
quantity proportional to α. Synaptic strengths are instead
decreased by if the expected output 0 is not fulfilled. Once
the strength of a synapse is below an assigned small value
gt = 10−4, we remove it, i.e. set its strength equal to zero,
which corresponds to the so-called pruning. This ingredient
is very important as since decades the crucial role of selec-
tive weakening and elimination of unneeded connections in
adult learning has been recognized [31, 32]. The synapses
involved in the signal propagation and responsible for the
0.8
0.6
0.4
0.2
α=0.01
α=0.005
α=0.003
α=0.0005
α=0.0001
τ=1/α
0
100 101 102 103 104 105
steps
10-310-2 10-1 100 101 102 103
steps/τ
0.8
0.6
0.4
0.2
0
105
104
103
1
0.8
0.6
0.4
>
e
m
i
t
g
n
i
n
r
a
e
l
<
y
t
f
n
i
c
c
u
s
%
0.2
α-0.05
RAN pin=0.1
XOR pin=0.1
XOR pin=0
XOR
RAN
AND
1/ α
0.001
α
Fig. 3. Demonstration of the learning ability of the neuronal network. (Left) Per-
centage of configurations learning the XOR rule as function of the number of learning
steps for different plastic adaptation strengths α (decreasing from bottom to top).
Data are for 400 realizations of a network with N = 1000 neurons, pin = 0.1,
r0 = 15, kmin = 3 and kd = 5. (Right) Collapse of the curves by rescaling
the number of learning steps by the characteristic learning time τ = 1/α and the
percentage of success by α−0.05.
0.0001
0.01
Fig. 4. Demonstration of the scaling behaviour of learning time and success.
(Top) Scaling of the average learning time for different rules as function of α.
(Bottom) Scaling of the asymptotic percentage of configurations learning different
rules as function of α. Data are obtained for kmin = 3, kd = 5 and pin = 0.1.
Both quantities follow a power law with an exponent independent of the rule and the
percentage of inhibitory synapses.
Footline Author
PNAS
Issue Date
Volume
Issue Number
3
wrong answer, are therefore not adapted uniformly but in-
versely proportional to the chemical distance from the output
site. Namely, synapses directly connected to the output neu-
ron receive the strongest adaptation ±α. This adaptation rule
intends to mimic the feedback to the wrong answer triggered
locally at the output site and propagating backward towards
the input sites. This could be the case, for instance, of some
hormones strongly interfering with learning and memory, as
dopamine suppressing LTD [34] or adrenal hormones enhanc-
ing LTD [35]. Moreover a new class of messenger molecules as
nitric oxide has been found to have an important role in plas-
tic adaptation [36]. For all these agents, released at the output
neuron, the concentration is reduced with the distance from
the origin. In our neuronal network simulation this non uni-
form adaptation has a crucial role since it prevents, in case
of successive wrong positive answers, synapses directly con-
nected to the input sites to decrease excessively, hindering
any further signal transmission. This plastic adaptation is a
non-Hebbian form of plasticity and can be interpreted as a
subtractive form of synaptic scaling [33], where synapses are
changed by an amount independent of their strength. The
procedure mimics the performance of a good critic who does
not tell the system which neurons should have fired or not.
However it tells more than just "right" or "wrong", it ex-
presses an evaluation on the type of error. Finally, we wish to
stress that this model naturally sets the system in a critical
state and therefore the study of the response of the system in
a subcritical or supercritical state requires the introduction of
additional parameters. We can however suppose that in both
cases learning becomes a more difficult task. For instance,
in a subcritical state, being the size of neuronal avalanches
smaller, it would be more complex to generate a firing state
in the output site. Conversely, in a supercritical state it would
be more difficult to generate a non-firing state in the output
site.
Results and Discussion
We analyse the ability of the system to learn the different
rules. Fig. 3 shows the fraction of configurations learning the
XOR rule versus the number of learning steps for different val-
ues of the plastic adaptation strength α. We notice that the
larger the value of α the sooner the system starts to learn the
rule, however the final percentage of learning configurations
is lower. The final rate of success increases as the strength of
plastic adaptation decreases. This result is due to the highly
non linear dynamics of the model, where firing activity is an
all or none event controlled by the threshold. Very slow plastic
adaptation allows then to tune finely the role of the neurons
involved in the propagation and eventually recover the right
answer. Moreover, very slow plastic adaptation also makes
the system more stable with respect to noise, since too strong
synaptic changes may perturb excessively the evolution ham-
pering the recovery of the right answer. The dependence of
the learning success on the plasticity strength is found consis-
tently for different values of the parameters kd, kmin and pin,
where a higher percentage of success is observed in systems
with no inhibitory synapses. Moreover, the dependence on the
plastic adaptation α is a common feature of all tested rules.
Data indicate that the easiest rule to learn is OR, where a
100% percentage of success can be obtained. AND and XOR
present similar difficulties and lead both to a percentage of
final success around 80%, whereas the most difficult rule to
learn is the RAN rule with three inputs where only 50% of
final success is obtained. This different performance is mainly
due to the higher number of inputs, since the system has to
organized a more complex path of connections leading to the
output site.
The most striking result is that all rules give a higher
percentage of success for weaker plastic adaptation. Indeed
this result is in agreement with recent experimental findings
on visual perceptual learning, where better performances are
measured when minimal changes in the functional network oc-
cur as a result of learning [22]. We characterize the learning
ability of a system for different rules by the average learning
time, i.e. the average number of times a rule must be applied
to obtain the right answer, and the asymptotic percentage of
learning configurations. This is determined as the percentage
of learning configurations at the end of the teaching routine,
namely after 106 applications of the rule. Fig. 4 shows that
the average learning time scales as τ ∝ 1/α for all rules and
independently of parameter values. Since some configurations
never learn and do not contribute to the average learning time,
we also evaluate the median learning time which exhibits the
same scaling behaviour as the average learning time. The
asymptotic percentage of success increases by decreasing α as
a very slow power law, ∝ α−0.05. Since this quantity has a
finite upper bound equal to unity, this scaling suggests that in
a finite, even if very long, time any configuration could learn
the rule by applying an extremely slow plastic adaptation. It
is interesting to notice that a larger fraction of systems with
no inhibitory synapses finds the right answer and the aver-
age learning time for these systems is slightly shorter. We
understand this result by considering that for only excita-
tory synapses the system more easily selects a path of strong
enough synapses connecting inputs and output sites and giv-
ing the right answer. Conversely, the presence of inhibitory
synapses may lead to frustration in the system as not all lo-
cal interactions contribute in the optimal way to provide the
right answer and the system has to find alternative paths.
We check this scaling behaviour by appropriately rescaling
the axes in Fig. 3. The curves corresponding to different α
values indeed all collapse onto a unique scaling function. Sim-
ilar collapse is observed for the OR, AND and RAN rules and
for different parameters kd, kmin and pin. In fact, two differ-
ent cases of distributions of inhibitory synapses, one in which
they are chosen randomly among all synapses, the other where
certain randomly chosen neurons have all outgoing synapses
inhibitory, provide equivalent results. The learning dynam-
ics shows therefore universal properties, independent of the
details of the system or the specific task assigned.
The learning behaviour is sensitive to the number of neu-
rons involved in the propagation of the signal, and therefore
depends on the distance between input and output neurons
and the level of connectivity in the system. We then inves-
tigate the effect of the parameters kd and kmin on the per-
formance of the system. Fig. 5 shows the percentage of con-
figurations learning the XOR rule for different minimum val-
ues of the neuron out degree. Systems with larger kmin have
a larger average number of synapses per neuron, producing
a more branched network. The presence of several alterna-
tive paths facilitates information transmission from the in-
4
www.pnas.org - -
Footline Author
puts to the output site. However, the participation of more
branched synaptic paths in the learning process may delay
the time the system first gives the right answer. As expected
the performance of the system improves as the minimum out-
connectivity degree increases, with the asymptotic percentage
of success scaling as ∼ k0.4
min. The dependence of the learning
performance on the level of connectivity is confirmed by the
analysis of systems with different number of neurons N , the
same out-degree distribution and the same set of parameters.
We verify that larger systems exhibit better performances. In
larger systems, in fact, the number of hubs, i.e. highly con-
nected neurons, increases improving the overall level of con-
nectivity. Indeed, the existence of complex patterns of activa-
tion has been recently recognized as very important in linking
together large scale networks in visual perceptual learning[22].
On the other hand, also the chemical distance between the
input and output sites has a very important role, as the num-
ber of hidden layers in a perceptron. Indeed, as kd becomes
larger (Fig. 5), the length of each branch in a path involved in
the learning process increases. As a consequence, the system
needs a higher number of tests to first give the right answer
and a lower fraction of configurations learns the rule after the
same number of steps. The percentage of learning configura-
tions after 106 applications is found, as expected, to decrease
as ∼ k−0.3
and similar behaviour is detected for the OR, AND
and RAN rules.
d
Learning stability and Memory. The existence of systems that
are unable to learn, even after many learning steps, raises in-
triguing questions about the learning dynamics. We question
what happens when a second chance is given to the configu-
rations failing the right answer. We then restart the learning
routine after imposing a small change in the initial configura-
tion of voltages. This small perturbation leads to about 25%
more configurations learning the rule. The initial state of the
system can therefore influence the ability to learn, especially
for complex rules as XOR or RAN. On the other hand, the
analysis of the out-degree distribution in configurations which
did and did not give the right answer indicates that "dumb"
configurations tend to have less highly connected nodes than
the "'smart" ones. Namely, giving repeatedly wrong positive
answers leads to pruning of several synapses. This affects in
particular the highly connected neurons, which have a crucial
role in identifying the right synaptic learning path. Finally we
test the ability of the configurations that do learn to remem-
ber the right answer once the initial configuration is changed.
The memory performance of the system is expected to depend
on the intensity of the variation imposed, namely on the ex-
tension of the basin of the attraction of states leading to the
right answer. The system is able to recover the right answer
in more than 50% of the configurations if a very small per-
turbation (of the order of 10−3) is applied to all neurons or
else a larger one (of the order of 10−2) to 10% of neurons.
The system has a different memory ability depending on the
rule: almost all configurations remember OR, whereas typi-
cally 80% remember AND and at most 70% the XOR rule.
Conclusions
In conclusion, we investigate the learning ability of a model
able to reproduce the critical avalanche activity as observed
for spontaneous activity in in vitro and in vivo cortical net-
works. The ingredients of the model are close to most func-
tional and topological properties of real neuronal networks.
The implemented learning dynamics is a cooperative mecha-
nism where all neurons contribute to select the right answer
and negative feedback is provided in a non-uniform way. De-
spite the complexity of the model and the high number of
degrees of freedom involved at each step of the iteration, the
system can learn successfully even complex rules as XOR or
a random rule with three inputs.
In fact, since the system
acts in a critical state, the response to a given input can be
highly flexible, adapting more easily to different rules. The
analysis of the dependence of the performance of the sys-
tem on the average connectivity confirms that learning is a
truly collective process, where a high number of neurons may
be involved and the system learns more efficiently if more
branched paths are possible. The role of the plastic adap-
tation strength, considered as a constant parameter in most
studies, provides a striking new result: The neuronal network
has a "universal" learning dynamics, even complex rules can
be learned provided that the plastic adaptation is sufficiently
slow. This important requirement for plastic adaptation is
confirmed by recent experimental results [22] showing that the
learning performance, in humans trained to a specific visual
task, improves when minimal changes occur in the function-
ality network. Stronger modifications of the network do not
necessarily lead to better results.
0,8
% success
0,6
0,4
0,2
0
1
10
kmin=2
kmin=3
kmin=4
kmin=5
10000
100
1000
trials
kd=3
kd=4
kd=5
kd=6
10000
1e+05
1e+05
10
100
1000
trials
Fig. 5. Demonstration of the dependence of the learning performance on the min-
imum connectivity degree and the chemical distance between input and output sites.
(Left) Percentage of configurations learning the XOR rule as function of the num-
ber of learning steps, for α = 0.005 in systems with kd = 5 and different values
of kmin (increasing from bottom to top). (Right) Percentage of configurations
learning the XOR rule as function of the number of learning steps, for α = 0.005 in
systems with kmin = 3 and different values of kd (decreasing from bottom to top).
Data are for 400 realizations of a network with N = 1000 neurons, pin = 0.1.
Footline Author
PNAS
Issue Date
Volume
Issue Number
5
1. Beggs JM, Plenz D (2003) Neuronal avalanches in neocortical cir-
20. Eurich CW, Herrmann JM, Ernst UA (2002) Finite- size effects
cuits. J Neurosci 23: 11167-11177.
of avalanche dynamics. Physical Review E 66:0661371-15.
2. Beggs JM, Plenz D (2004) Neuronal avalanches are diverse and
precise activity patterns that are stable for many hours in cortical
slice cultures. J Neurosci 24: 5216-5229.
21. Shew WL, Yang H, Petermann T, Roy R, Plenz D (2009) Neuronal
avalanches imply maximum dynamic range in cortical networks at
criticality. J Neurosci 29: 155595-15600.
3. Stanley HE (1971) Introduction to Phase Transitions and Critical
Phenomena (Oxford Univ Press, New York).
4. Gireesh ED, Plenz D (2008) Neuronal avalanches organize as
nested theta- and beta/gamma - oscillations during development
of cortical layer 2/3. Proc Natl Acad Sci USA 105: 7576-7581.
5. Petermann T, Thiagarajan TC, Lebedev MA, Nicolelis MAL,
Chialvo DR, Plenz D (2009) Spontaneous cortical activity in
awake monkeys composed of neuronal avalanches. Proc Natl Acad
Sci USA 106: 15921-15926.
6. Mazzoni A, Broccard FD, Garcia-Perez E, Bonifazi P, Ru-
aro ME, Torre V (2007) On the dynamics of spontaneous
activity in neuronal networks. PLoS ONE 2(5):
e439, 1-12,
doi:10.1371/journal.pone.0000439.
7. Pasquale V, Massobrio P, Bologna LL, Chiappalone M, Martinoia
S (2008) Self-organization and neuronal avalanches in networks of
dissociated cortical neurons. Neuroscience 153: 1354-1369.
8. Bak P (1996) How nature works. The science of self-organized crit-
icality (Springer, New York).
9. Amit D (1989) Modeling brain function ( Cambridge University
Press, Cambridge).
10. Rosenblatt F (1958) The perceptron: A probabilistic model for
information storage and organization in the brain. Psycological
review 65: 386-408.
11. Hopfield J (1982) Neural networks and physical systems with
emergent collective computational abilities. Proc Natl Acad Sci
USA 79: 2554-2558.
12. McCulloch WS, Pitts W (1943) A logical calculus of the ideas
immanent in nervous activity. Bull Math Biophys 5: 115-133.
13. Barto AG, Sutton RS, Anderson CW (1983) Neuron-like adaptive
elements that can solve difficult learning control problems. IEEE
Trans Syst Man Cybern 15: 835-846.
14. Rumelhart DE, Hinton GE, Williams RJ (1986) Learning repre-
sentations by back-propagating errors. Nature 323: 533-536.
15. Bak P, Chialvo DR (2001) Adaptive learning by extremal dynam-
ics and negative feedback. Physical Review E 63: 0319121-12
16. Chialvo DR, Bak P (1999) Learning from mistakes. Neuroscience
90: 1137-1148.
17. Klemm K, Bornholdt S, Schuster HG (2000) Beyond Hebb:
Exclusive-OR and biological learning. Physical Review Letters 84:
3013-3016.
18. Turing AM (1957) Computing machines and intelligence. Mind
59: 433-460.
19. Kinouchi O, Copelli M (2006) Optimal dynamical range of ex-
citable networks at criticality. Nat Phys 2: 348-351.
22. Lewis CM, Baldassarre A, Committeri G, Romani GL, Corbetta
M (2009) Learning sculpts the spontaneous activity of the resting
human brain. Proc Natl Acad Sci USA 106: 17558-17563.
23. de Arcangelis L, Perrone-Capano C, Herrmann HJ (2006) Self-
organized criticality model for brain plasticity. Physical Review
Letters 96: 0281071-4.
24. Pellegrini GL, de Arcangelis L, Herrmann HJ, Perrone-Capano
C (2007) Activity-dependent neural network model on scale-free
networks. Physical Review E 76: 0161071-9.
25. Levina A, Herrmann JM, Geisel T(2007) Dynamical synapses
causing self-organized criticality in neural networks. Nat Phys 3:
857-860
26. Teramae JN , Fukai T (2007) Local cortical circuit model inferred
from power-law distributed neuronal avalanches. J Comput Neu-
rosci 22: 301-321.
27. Novikov E, Novikov A, Shannahoff-Khalsa D, Schwartz B, Wright
J (1997) Scale-similar activity in the brain. Physical Review E 56:
R2387-R2389.
28. Hausdorff JM, Ashkenazy Y, Peng CK, Ivanov PC, Stanley EH
Goldberger AL (2001) When human walking becomes random
walking:
fractal analysis and modeling of gait rhythm fluctua-
tions. Physica A 302: 138-147.
29. Eguiluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV
(2005) Scale-free brain functional networks. Physical Review Let-
ters 94: 0181021-4.
30. Roerig B, Chen B (2002) Relationships of local inhibitory and
excitatory circuits to orientation preference maps in ferret visual
cortex. Cerebral Cortex 12: 187-198.
31. Young JZ (1964) A model of the Brain (Clarendon, Oxford)
32. Changeux JP (1985) Neuronal Man: The Biology of Mind (Oxford
University Press, New York)
33. Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the
beast. Nature Neuroscience 3: 1178-1183.
34. Otmakhova NA, Lisman JE (1998) D1/D5 Dopamine receptors in-
hibit depotentiation at CA1 synapses via cAMP-dependent mech-
anism. J Neurosci 18: 1270-1279.
35. Coussens CM, Kerr DS, Abraham WC (1997) Glucocorticoid
receptor activation lowers the threshold for NMDA-receptor-
dependent homosynaptic long-term depression in the hippocam-
pus through activation of voltage-dependent Calcium channels. J
Neurophysiol 78: 1-9.
36. Reyes-Harde M, Empson R, Potter BVL, Galione A, Stanton
PK (1999) Evidence of a role for cyclic ADP-ribose in long-term
synaptic depression in hippocampus. Proc Natl Acad Sci 96: 4061-
4066.
6
www.pnas.org - -
Footline Author
|
1904.01862 | 1 | 1904 | 2019-04-03T09:06:27 | Respiratory sympathetic modulation is augmented in chronic kidney disease | [
"q-bio.NC",
"q-bio.TO"
] | Respiratory modulation of sympathetic nerve activity (respSNA) was studied in a hypertensive rodent model of chronic kidney disease (CKD) using Lewis Polycystic Kidney (LPK) rats and Lewis controls. In adult animals under in vivo anaesthetised conditions (n=8-10/strain), respiratory modulation of splanchnic and renal nerve activity was compared under control conditions, and during peripheral (hypoxia), and central, chemoreceptor (hypercapnia) challenge. RespSNA was increased in the LPK vs. Lewis (area under curve (AUC) splanchnic and renal: 8.7$\pm$1.1 vs. 3.5$\pm$0.5 and 10.6$\pm$1.1 vs. 7.1$\pm$0.2 $\mu$V.s, respectively, P<0.05). Hypoxia and hypercapnia increased respSNA in both strains but the magnitude of the response was greater in LPK, particularly in response to hypoxia. In juvenile animals studied using a working heart brainstem preparation (n=7-10/strain), increased respSNA was evident in the LPK (thoracic SNA, AUC: 0.86$\pm$0.1 vs. 0.42$\pm$0.1 $\mu$V.s, P<0.05), and activation of peripheral chemoreceptors (NaCN) again drove a larger increase in respSNA in the LPK with no difference in the response to hypercapnia. Amplified respSNA occurs in CKD and may contribute to the development of hypertension. | q-bio.NC | q-bio | Highlights
Sympathetic nerve activity is modulated by an interaction between central nervous
system respiratory and sympathetic networks that leads to the generation of bursts of
sympathetic nerve activity in phase with the respiratory cycle, termed respiratory-
sympathetic coupling.
In disease states altered respiratory sympathetic coupling may drive the development
and/or maintenance of sympatho-excitation, and in turn hypertension, as is seen in
chronic kidney disease.
Using a rat model of chronic kidney disease, we showed in both young and adult
animals the presence of enhanced respiratory-sympathetic coupling and an
exaggerated sympathoexcitatory response to stimulation of the peripheral
chemoreceptors with sodium cyanide (NaCN) and hypoxia, respectively.
These results indicate that the pathways responsible for respiratory-sympathetic
coupling, may be potential therapeutic targets to reduce high blood pressure in
association with chronic kidney disease.
Respiratory sympathetic modulation is augmented in chronic kidney disease
Manash Saha a b c d, Clement Menuet e f, Qi-Jian Sun a, Peter G.R. Burke g, Cara M.
Hildreth a, Andrew M. Allen e, *Jacqueline K. Phillips a
a Department of Biomedical Sciences, Macquarie University, Australia
b Department of Nephrology, National Institute of Kidney Disease and Urology,
Bangladesh
c Graduate School of Medicine, Wollongong University, Australia
d Department of Medicine, Wollongong Hospital, Australia
e Department of Physiology, University of Melbourne, Australia
f Institut de Neurobiologie de la Méditerranée, INMED UMR1249, INSERM, Aix-
Marseille Université, Marseille, France
g Neuroscience Research Australia, Sydney NSW, Australia
Running title
Respiratory sympathetic coupling in CKD
*Author for correspondence
Professor Jacqueline Phillips
Department of Biomedical Science, Faculty of Medicine and Health Sciences
Macquarie University, Sydney NSW 2109 Australia
EMAIL: [email protected]
PHONE: +61 2 9850 2753
1
Abstract
Respiratory modulation of sympathetic nerve activity (respSNA) was studied in a
hypertensive rodent model of chronic kidney disease (CKD) using Lewis Polycystic
Kidney (LPK) rats and Lewis controls. In adult animals under in vivo anaesthetised
conditions (n=8-10/strain), respiratory modulation of splanchnic and renal nerve activity
was compared under control conditions, and during peripheral (hypoxia), and central,
chemoreceptor (hypercapnia) challenge. RespSNA was increased in the LPK vs. Lewis
(area under curve (AUC) splanchnic and renal: 8.71.1 vs. 3.50.5 and 10.61.1 vs.
7.10.2 µV.s, respectively, P<0.05). Hypoxia and hypercapnia increased respSNA in
both strains but the magnitude of the response was greater in LPK, particularly in
response to hypoxia. In juvenile animals studied using a working heart brainstem
preparation (n=7-10/strain), increased respSNA was evident in the LPK (thoracic SNA,
AUC: 0.86±0.1 vs. 0.42±0.1 µV.s, P<0.05), and activation of peripheral
chemoreceptors (NaCN) again drove a larger increase in respSNA in the LPK with no
difference in the response to hypercapnia. Amplified respSNA occurs in CKD and may
contribute to the development of hypertension.
Keywords
Respiratory sympathetic coupling, peripheral chemoreceptors, hypertension, chronic
kidney disease, working heart brainstem preparation.
2
1. Introduction
Hypertension is a major comorbidity associated with chronic kidney disease (CKD),
arising early in the development of CKD and acting as a significant causal factor for the
development of end-organ damage (Vanholder et al., 2005). Increased sympathetic
nerve activity (SNA) is believed to play an important role in the development and/or
maintenance of hypertension associated with kidney disease, with direct sympathetic
nerve recording and plasma catecholamines levels elevated in individuals with CKD
(Campese et al., 2011; Grassi et al., 2012; Klein IH, 2003; Phillips, 2005; Schlaich et
al., 2009). Treatment with centrally acting sympatholytic agents can ameliorate
hypertension in renal failure patients (Campese and Massry, 1983; Levitan et al., 1984;
Schlaich et al., 2009), however, we still do not understand the mechanisms driving this
increase in SNA.
Modulation of sympathetic nerve discharge occurs during the phases of respiration both
in animals and humans, although the temporal relationship can differ between species
and different nerve beds (Boczek-Funcke et al., 1992; Seals et al., 1993; Simms et al.,
2009; Zoccal et al., 2008). This respiratory modulation of SNA (respSNA) creates
synchronous alterations in blood pressure that allow optimal tissue perfusion. In
experimental animals, where open-chest experiments have been performed, a
component of the respSNA involves coupling between the relevant neuronal circuits
within the brainstem (Machado et al., 2017; Simms et al., 2010). In the spontaneously
hypertensive rat (SHR), a model of essential hypertension, respSNA is augmented and
when compared to normotensive Wistar Kyoto control rats, the phase relationship is
shifted from the post-inspiratory to inspiratory phase (Czyzyk-Krzeska and Trzebski,
1990; Simms et al., 2009). This exaggerated respSNA contributes to the development of
3
hypertension in the SHR (Menuet et al., 2017). In CKD it is not known whether the
phasic pattern of respSNA is altered or whether it contributes to the observed increase
in SNA and the associated hypertensive state (Augustyniak et al., 2002; Salman et al.,
2015).
The respSNA originating from central connections of respiratory and sympathetic
networks (Haselton and Guyenet, 1989; Koshiya and Guyenet, 1996; Spyer, 1993; Sun
et al., 1997) contributes independently to increased SNA in animal models following
chronic intermittent hypoxia (Zoccal et al., 2008), with evidence of input from excitable
central chemoreceptors (Molkov et al., 2011). The peripheral chemoreceptors may also
provide a critical respiratory-related input that contributes to respSNA and, in disease
states, drives the development and/or maintenance of sympatho-excitation, and in turn
hypertension. This hypothesis is supported by a large body of work showing that
exposure to chronic intermittent hypoxia (CIH) is correlated with enhanced respSNA
(Machado et al., 2017). Patients with CKD are more vulnerable to respiratory disorders
such as obstructive sleep apnoea (Hanly, 2004), and those with sleep apnoea have
elevated blood pressure compared to those with CKD alone (Sekizuka et al.). Further,
deactivation of the peripheral chemoreceptors, with hyperoxia, reduces muscle SNA in
individuals with renal failure (Hering et al., 2007). In this context, there is exciting
therapeutic potential in the investigation of the role of central respSNA and its
underlying mechanism in hypertension CKD.
The main purpose of this study, therefore, was to test the hypothesis that respSNA is
amplified in an animal model of CKD. Studies were undertaken in the Lewis polycystic
kidney (LPK) rat, a genetic model of CKD presenting with kidney disease (McCooke et
4
al., 2012) in which we have previously demonstrated hypertension, enhanced tonic SNA
and perturbed reflex responses to both peripheral and central chemoreceptor stimulation
(Salman et al., 2014; Salman et al., 2015; Yao et al., 2015).
2. Methods
All experimental procedures were approved by the Animal Ethics Committees of
Macquarie University, NSW or the University of Melbourne, Victoria, Australia, and
were carried out in accordance with the Australian Code of Practice for the Care and
Use of Animals for Scientific Purposes.
2.1 Study 1: Adult in vivo anaesthetised experiments
Adult (12-13-week-old) male LPK (n = 8) and control Lewis (n = 10) rats were used.
Animals were purchased from the Animal Resources Centre, Murdoch, Western
Australia.
2.1.1 Renal function
A 24 h urine sample was collected from all animals 48 h prior to experimentation and
urine volume, urinary creatinine and protein levels examined using a IDEXX Vetlab
analyser (IDEXX Laboratories Pty Ltd., Rydalmere, NSW, Australia). At the
commencement of the surgical procedure, an arterial blood sample was collected for
determination of plasma urea and creatinine, and creatinine clearance calculated as
described previously (Yao et al., 2015).
2.1.2 Surgical procedures:
Animals were anaesthetised with 10% (w/v) ethyl carbamate (Urethane, Sigma Aldrich,
NSW, Australia) in 0.9% NaCl solution (1.3 g/kg i.p.). Reflex responses to hind-paw
5
pinch were assessed to determine adequate depth of anaesthesia and supplemental doses
of ethyl carbamate given as required (65 mg/kg i.p. or i.v.). Body temperature was
measured and maintained at 37 ± 0.5 C using a thermostatically controlled heating
blanket (Harvard Apparatus, Holliston, MA, USA) and infrared heating lamp. The right
femoral vein and artery were cannulated for administration of fluid (Ringer's lactate, 5
ml/kg/h) and measurement of arterial pressure (AP) and blood collection for
measurement of blood gases, respectively. The AP signal was sampled at 200 Hz and
acquired using a CED 1401 plus and Spike2 software v.7 (Cambridge Electronic
Designs (CED) Ltd, Cambridge, UK, RRID:SCR_000903). A tracheostomy was
performed and an endotracheal tube placed in situ. A bilateral vagotomy was performed
to cut afferent inputs from the stretch receptors of the lung and the animal was
ventilated with oxygen enriched room air (7025 Rodent Ventilator, UgoBasile, Italy)
and paralysed with pancuronium bromide (2 mg/kg iv for induction, 1 mg/kg for
maintenance as required; AstraZeneca, North Ryde, NSW, Australia). The left phrenic,
splanchnic and renal nerves were dissected and the distal end of each nerve was tied and
cut. All nerves were bathed in a liquid paraffin pool and activity was recorded from the
central end using bipolar silver wire recording electrodes. The activity was 10 times
amplified, band-pass filtered between 10-1000Hz by a bio amplifier (CWE Inc.,
Ardmore, PA, USA) and sampled at 5kHz using CED 1401 plus and Spike2 software.
All recordings were made with the same bioamplifier calibrated to a pre-set setting 50
µV.
Following surgical preparation, animals were then stabilised for 30 min. Arterial blood
samples were collected and analysed for pH, pCO2, HCO3
- and SaO2 using a VetStat
Electrolyte and Blood Gas Analyser (IDEXX Laboratories Pty Ltd., Rydalmere, NSW,
6
Australia). After the initial blood gas measurement, arterial blood gases were then
corrected by adjusting the ventilator pump rate and volume and/or slow bolus injections
of 5% sodium bicarbonate as required to maintain parameters under control conditions
within the following range: pH = 7.4 ± 0.5; PCO2 = 40 ± 5 mmHg; HCO3
- = 24 ± 2
mmol/L, SaO2 = 100%; and end tidal (ET)CO2 = 4.5 ± 0.5%. Repeat blood gas analysis
was undertaken as required.
2.1.3 Experimental protocol
After the period of stabilization, the integrity of renal and splanchnic nerve recordings
(rSNA and sSNA respectively) was confirmed by pulse modulation of SNA and
demonstration of a baroreflex response to a bolus injection of phenylephrine (50 µg/kg
iv, Sigma-Aldrich, St. Louis, MO, USA). Baseline parameters were measured under
control conditions for a period of 30 min. Pilot studies indicated that hypoxic
stimulation to stimulate the peripheral chemoreceptors using 10% O2 in N2 (Abbott and
Pilowsky, 2009) could not be used for the 3 min stimulation period without affecting
pH and PCO2 in the LPK, and therefore animals were ventilated with room air, without
oxygen supplementation, for a period of 3 min without any change in ventilator rate or
volume. This induced a stable modest hypoxia - blood gas analysis parameters in the
range of pH = 7.4 ± 0.5; PCO2 = 43 ± 3 mmHg; HCO3
- = 24 ± 2 mmol/L and SaO2 =
82-84%. After a 30 min recovery period, central chemoreceptor stimulation was
performed by ventilating the animals with 5% carbogen (5% CO2 in 95% O2) for 3 min
without any change in the ventilator rate or volume, which was confirmed by blood gas
analysis to produce hypercapnia with parameters in the range of pH = 7.4 ± 0.5; PCO2 =
62 ± 7 mmHg; HCO3
- = 24 ± 2 mmol/L and SaO2 =100%.
7
At the end of experiment, all animals were euthanized with potassium chloride (3M i.v.)
and the electrical noise levels for phrenic nerve activity (PNA), sSNA and rSNA
recorded and later subtracted for data analysis.
2.1.4 Data analysis
All data were analysed offline using Spike 2 software. From the AP signal, mean
(MAP), systolic blood pressure (SBP), diastolic (DBP) and pulse (PP) pressure and
heart rate (HR) were derived. The sSNA and rSNA were rectified and smoothed with a
time constant of 0.1 s and PNA was rectified and smoothed with a time constant of 0.05
s. Baseline data was determined from a 30 s period at the end of the control ventilation
period. Response to respiratory challenge was analysed over a 30 s period once PNA
had stabilised for a period of 1-2 min and expressed as a change (∆) relative to the 30 s
period immediately prior to each stimulus. The PNA amplitude, frequency (number
cycle.min -1), duration (s) and minute activity (MPA=PNA amplitude x PNA frequency)
were measured to quantify the changes. The phrenic cycle was divided into three
phases: inspiratory (I), post-inspiratory (PI) and expiratory (E). For SNA, the mean
level was measured from the period 200 ms prior to the onset of the phrenic burst, and
this was considered the baseline for all other measurements. A phrenic-triggered
average of the rectified and smoothed sSNA and rSNA recordings was performed and
the phases divided into I, PI, and E. From this trace the following parameters were
calculated for sSNA and rSNA: the peak amplitude (PA [V]), the maximum amplitude
at the apex of the peak of the SNA burst coincident with inspiratory/post-inspiratory
phase, the duration (from onset of excitatory activity to return to baseline [s]) and area
under curve (AUC) of respSNA excitatory peak (V.s) determined as the integral of the
8
waveform. AUC of respSNA for I, PI and E phase of phrenic cycle were also calculated
from the integral of the waveform.
2.2. Study 2: Juvenile working heart brainstem preparation
To determine respSNA in early stages of the disease process, another series of
experiments were performed in juvenile rats using the working heart brainstem in-situ
preparation, which while a reduced preparation has the advantage of allowing
assessment of physiological patterns of central respiratory control and cardiovascular
regulation in the absence of the effects of anaesthesia (Wilson et al., 2001). Recordings
were made in 5 week old male Lewis (n = 7) and LPK (n = 10) as described previously
(Menuet et al., 2014). At this age, renal function is only mildly impaired in the LPK
(Phillips et al., 2007). Animals were purchased from the Animal Resources Centre,
Murdoch, Western Australia.
2.2.1 Surgical procedure
Animals were deeply anaesthetized with isoflurane until loss of the pedal withdrawal
reflex and then dissected below the diaphragm, exsanguinated, cooled in Ringer
solution on ice (composition in mM: 125 NaCl, 24 NaHCO3, 5 KCl, 2.5 CaCl2, 1.25
MgSO4, 1.25 KH2PO4 and 10 dextrose, pH 7.3 after saturation with carbogen gas (5%
CO2, 95% O2) and decerebrated precollicularly. Lungs were removed and the
descending aorta isolated and cleaned. Retrograde perfusion of the thorax and head was
achieved via a double-lumen catheter (ø 1.25 mm, DLR-4, Braintree Scientific,
Braintree, MA, USA) inserted into the descending aorta. The perfusate was Ringer
solution containing Ficoll (1.25%) warmed to 31C and gassed with carbogen (95% O2
and 5% CO2; closed loop reperfusion circuit). The second lumen of the cannula was
9
connected to a transducer to monitor perfusion pressure in the aorta. Neuro-muscular
blockade was achieved using vecuronium bromide added to the perfusate (2 -- 4 μg/mL,
Organon Teknika, Cambridge, UK). Simultaneous recordings of PNA, thoracic
sympathetic nerve activity (tSNA, T8-10), vagus nerve activity (VNA) and abdominal
nerve activity (AbNA, T9-T12) were obtained using glass suction electrodes. The
activity was amplified (10 kHz, Neurolog), filtered (50 -- 1500 kHz, Neurolog), digitized
(CED) and recorded using Spike2 (CED).
2.2.2 Experimental protocol
After the period of stabilization, the peripheral chemoreceptors were stimulated using
sodium cyanide (NaCN; 0.05%; 100 µL bolus) injected into the aorta via the perfusion
catheter (Chang et al., 2015; Menuet et al., 2016). Central chemoreceptors were
stimulated using hypercapnic conditions by changing the level of CO2 in the perfusate
source from 95% O2 and 5% CO2 to 90% O2 and 10% CO2 for ~5 min. The electrical
noise levels for tSNA recordings were determined at the end of experiments by
sectioning the sympathetic chain at the proximal paravertebral ganglion level and were
subtracted for data analysis. All chemicals were purchased from Sigma-Aldrich,
Australia.
2.2.3 Data analysis
All data were analysed offline using Spike 2 software. The signals were rectified and
integrated with a 50 ms time constant. The HR was derived using a window
discriminator to trigger from the R-wave of the electrocardiogram recorded
simultaneously with PNA. The PNA, and VNA signals were used to assess respiratory
parameters associated with inspiration, post-inspiration and expiration, respectively.
10
During baseline and hypercapnic conditions, phrenic-triggered (end of inspiratory burst)
averaging of tSNA was used for the analysis of tSNA parameters related to the burst of
respSNA tSNA and tonic non-respiratory modulated tSNA (baseline tSNA). NaCN-
induced respSNA was measured as delta increase in AUC compared to a pre-stimulus
control period with same duration (Menuet et al., 2016), while peak activity of respSNA
of tSNA was calculated as the maximum amplitude of the tSNA burst coincident with
inspiratory/post-inspiratory transition.
2.3 Statistical analysis
All data are presented as mean ± SEM. A Student's t-test was used to determine baseline
differences in cardiorespiratory function and respiratory sympathetic coupling between
the LPK and Lewis. A two-way ANOVA with repeated measures followed by
Bonferroni's post-hoc analysis was used to analyse the effect of peripheral or central
chemoreceptor challenge on respiratory sympathetic coupling with strain and
chemoreceptor challenge as the variables. All analysis was performed using GraphPad
Prism software v6.0 (GraphPad Software Inc., La Jolla, California, USA,
RRID:SCR_002798). Differences were considered statistically significant where
P<0.05.
3. Results
3.1 Adult baseline parameters
Adult male LPK rats showed a phenotypic elevation in blood urea (24.3 ± 2.3 vs. 5.9 ±
0.3 mmol/L) and plasma creatinine (60.6 ± 11.7 vs. 22.3 ± 4.6 µmol/L) and reduction in
creatinine clearance (2.08 ± 0.4 vs. 10.1 ± 2.3 mL/min) (all LPK n=8 vs. Lewis n=10;
P< 0.001) reflective of impaired renal function.
11
During the control period, individual recordings were made of rSNA, sSNA, PNA and
AP from Lewis and LPK rats (Figure 1). Consistent with our previous work (Salman et
al., 2014; Yao et al., 2015), SBP, MAP, DBP, PP, HR and SNA (both sSNA and rSNA)
were elevated in the adult LPK compared with age-matched Lewis controls (Table 1).
Phrenic frequency was also significantly higher in the adult LPK, suggestive of
increased central respiratory drive; however, there was no significant difference in PNA
amplitude or MPA between strains (Table 1).
We examined the temporal relationship between SNA and the respiratory cycle in LPK
and Lewis rats under control conditions. Both the rSNA and sSNA exhibited distinct
respSNA, with a clear burst in SNA in the PI period in both the Lewis and LPK (see
Figures 2 and 4, control panels A -B [renal], E-F [splanchnic]). In the Lewis rat, both
sSNA and rSNA showed a tendency to increase from baseline during I, before the sharp
peak occurred in PI. After the PI peak, activity fell quickly to baseline and was
relatively flat during E. The LPK showed a distinct difference -- with a decrease in
activity below baseline during early I that was most obvious in rSNA (rSNA, I phase
AUC, LPK vs Lewis: -0.07 ± 0.03 vs 0.45 ± 0.2; P ≤ 0.5; sSNA, I phase AUC, LPK vs
Lewis: 0.06 ± 0.05 vs 0.18 ± 0.08 ; P > 0.5). There was a gradual decline in both sSNA
and rSNA during the E period which was not significantly different between the strains
(rSNA, E phase AUC, LPK vs Lewis: 0.07 ± 0.04 vs 0.08 ± 0.1; P > 0.5; sSNA, E
phase AUC, LPK vs Lewis: 0.02 ± 0.03 vs 0.04 ± 0.02; P > 0.5 [Figures 2 and 4,
control panels A -B (renal), E-F (splanchnic)].
Quantitative parameters characterizing respSNA under control conditions for both
nerves are provided in Table 2. PA and AUC for both sSNA and rSNA were
12
significantly greater in the LPK compared with Lewis. The duration of the respSNA for
both sSNA and rSNA was not different between the Lewis and LPK indicating that the
greater AUC observed in the LPK was driven primarily by the greater PA.
3.2 Adult responses to chemoreceptor stimulation
In response to a hypoxic challenge, a greater increase in AP was observed in the LPK
compared with Lewis (Table 3). In the LPK rat this was associated with a slight, but
significant, slowing of PNA frequency, but not amplitude or duration. The hypercapnic
challenge produced comparable increases in AP and central respiratory drive in the two
strains, with only a small difference in the HR response.
In both strains, under hypoxic and hypercapnic conditions, the respSNA curves of
sSNA and rSNA retained a consistent PI peak (see Figures 2 and 4). In response to
hypoxia, respSNA (sSNA and rSNA) was increased in both strains as reflected by an
increase in PA in Lewis and an increase in PA and AUC in LPK (Figure 3) that was
greater in LPK rats (delta AUC, sSNA: Lewis vs. LPK: 0.8 ± 0.6 vs. 5.9 ± 1.9, rSNA:
2.8 ± 0.7 vs. 7.5 ± 2.1 µV.s, both P< 0.05). Hypercapnia induced an increase in the
respSNA of sSNA as reflected by an increase in PA in Lewis and an increase in PA and
AUC in LPK. In the rSNA there was an increase in PA in the Lewis only (Figures 4, 5).
However, the magnitude of the change in the respSNA of both nerves was similar
between strains (Delta AUC, sSNA, Lewis vs. LPK: 1 ± 0.4 vs. 2.5 ± 0.7; rSNA, Lewis
vs. LPK: 3.2 ± 0.5 vs. 2.5 ± 1.4 µV.s, both P > 0.05). During hypoxia, the inspiratory
inhibition in the LPK increased in rSNA (I phase AUC, rSNA, control vs. hypoxia: -
0.07 ± 0.04 vs. -0.9 ± 0.3 µV.s p < 0.05; Figure 2) and became clearly evident in sSNA
(I phase AUC, sSNA, control vs. hypoxia : 0.05 ± 0.04 vs. -0.4 ± 0.1 µV.s P < 0.05).
13
This increase in inspiratory inhibition did not occur in the LPK under hypercapnic
conditions ( I phase AUC, rSNA, control vs. hypercapnia : -0.32 ± 0.2 vs. -0.4 ± 0.2
sSNA: 0.03 ± 0.09 vs. 0.06 ± 0.1 µV.s both P >0.05). There was no significant change
to the inspiratory phase in the Lewis animals under either condition.
3.3 Juvenile baseline parameters
In the working heart brainstem preparation (Figure 6), mean perfusion pressure was not
different between the two strains (Lewis vs. LPK: 73 ± 4 vs. 76 ± 2 mmHg, P = 0.55).
The HR was faster in the LPK (Lewis vs. LPK: 360 ± 13 vs. 302 ± 17 bpm, P < 0.05).
There was no difference observed in PNA frequency (Lewis vs. LPK: 9.5 ± 0.7 vs. 10.2
± 1.6, P = 0.74), although inspiratory duration of PNA was reduced in the LPK (Lewis
vs. LPK: 1658 ± 147 vs. 1242 ± 79 ms, P < 0.05). Tonic levels of tSNA were not
significantly different (Lewis vs. LPK: 4.37 ± 0.47 vs. 4.51 ± 0.58 µV, P = 0.85);
however, respiratory modulation of tSNA was greater in the LPK with a larger AUC
observed (Lewis vs. LPK: 0.42 ± 0.13 vs. 0.86 ± 0.13 µV.s, P < 0.05). No differences
were observed in respSNA burst duration (Lewis vs. LPK: 0.8 ± 0.1 vs. 1.4 ± 0.2 s, P =
0.10) or the magnitude of the peak amplitude of tSNA (Lewis vs. LPK: 6.01 ± 0.56 vs.
6.17 ± 0.85 µV, P = 0.87).
3.4 Juvenile responses to chemoreceptor stimulation
Peripheral chemoreceptor activation with NaCN produced a comparable increase in
perfusion pressure in both strains (Lewis vs. LPK: 5.7 ± 1.4 vs. 6.4 ± 1.2 mmHg, P =
0.71). The accompanying large bradycardia was of greater magnitude in the LPK
(Lewis vs. LPK: -86 ± 18 vs. -176 ± 16 bpm, P < 0.05). RespSNA increased in both
strains in response to peripheral chemoreceptor activation (NaCN, P<0.05; Figure 6), as
14
evidenced by an increase in PA and AUC that was of greater magnitude in both
circumstances in the LPK. (Lewis vs. LPK: delta PA: 3.6 ± 0.6 vs. 6.5 ± 0.9 µV, delta
AUC: Lewis vs. LPK: 7.0 ± 1.2 vs. 15.1 ± 2.7 µV.s, P<0.05).
Central chemoreceptor activation with hypercapnia produced a comparable reduction in
perfusion pressure (Lewis vs. LPK: -7.9 ± 1.9 vs. -8.6 ± 1.1 mmHg, P = 0.75) and HR
(Lewis vs. LPK: -42 ± 9 vs. -53 ± 8 bpm, P = 0.36) in the two strains. RespSNA
increased in both strains (Figure 7) with an increase in PA in the LPK and a comparable
increase in AUC in both strains (delta AUC: 0.98 ± 0.16 vs. 0.99 ± 0.29 µV.s, P =
0.97).
4. Discussion
Our study has determined the characteristic features of respSNA in a rodent model of
CKD. We have directly measured the relationship between inspiratory drive and two
sympathetic nerves in anaesthetized adult animals and to one sympathetic nerve in
juvenile animals using the working heart brainstem preparation. The timing of the peak
of respSNA was observed persistently in the PI period under all conditions tested and,
from the earliest age studied (5 weeks). Adult and juvenile LPK rats demonstrated
increased respSNA compared to Lewis control rats. Under control conditions, the LPK
rat shows an alteration in respSNA patterning with increased inhibition of rSNA during
the inspiratory period. This inspiratory inhibition is augmented in rSNA with hypoxia
and becomes evident in sSNA with this respiratory challenge. In addition, we have
found that peripheral chemoreceptor stimulation has an exaggerated effect on respSNA
in both the juvenile and adult LPK animals, which suggests a role for peripheral
chemoreceptors in the pathophysiology of autonomic dysfunction associated with CKD.
15
Our finding of augmented respSNA in the LPK model of CKD is consistent with
previous work in the SHR model, which also demonstrates amplified respSNA when
examined using the same methodologies in juvenile (Menuet et al., 2017; Simms et al.,
2009) and adult SHRs (Czyzyk-Krzeska and Trzebski, 1990) that we have applied,
though notably, we do not see the marked temporal shift in the peak response from post-
inspiration to inspiration that is described in the SHR during the development of
hypertension. Enhancement of respSNA is also a feature noted in animal models of
chronic intermittent hypoxia (Zoccal et al., 2008), congestive heart failure (Marcus et
al., 2014); and in rats following uteroplacental insufficiency (Menuet et al., 2016).
Common to these diseases is autonomic dysfunction and the development of
hypertension in association with increased SNA.
While we did not see elevated baseline tSNA or perfusion pressures in the juvenile LPK
at 5 weeks of age, our previous work in both young and adult LPK (7 and 14 weeks)
show increased SNA, including recordings from conscious animals (Salman et al.,
2014; Salman et al., 2015), and we observe markedly elevated blood pressure in these
animals from an early age (6 weeks) (Phillips et al., 2007; Sarma et al., 2012). Our
demonstration of increased respSNA at age 5 weeks in the current study indicates that
altered sympathetic activity is a key feature of the autonomic dysfunction in the LPK
model, consistent with clinical data demonstrating that in patients with polycystic
kidney disease, muscle SNA is elevated very early in the disease, often prior to any
significant reduction in renal function (Klein et al., 2001). Our work is supportive of the
hypothesis that in CKD, altered interactions between central respiratory and
sympathetic pathways is a contributor to the development and maintenance of
hypertension.
16
The rostral ventrolateral medulla (RVLM) is an important site in the generation and
regulation of sympathetic vasomotor tone (Guyenet, 2006), and our recent work in the
SHR suggests that the inspiratory pre-Bötzinger complex is a likely site of modulatory
inputs to the RVLM C1 neurons which could drive an increase in SNA (Menuet et al.,
2017). There is also an important contribution from the respiratory pattern generator
that results in phasic modulation of RVLM presympathetic neurons. Recordings from
both anesthetised rats and juvenile rats using the working heart brainstem preparation
show 3 main classes of respiratory-modulated RVLM presympathetic neurons -- I
activated, PI activated and I inhibited -- as well as some none modulated neurons
(Haselton and Guyenet, 1989; Moraes et al., 2013). Whilst not directly tested in this
study, we propose that altered strength of respiratory input to these different neuron
classes could underlie the different patterns of respSNA observed including the
enhanced I inhibition of respSNA in the LPK rat.
While the exact mechanisms underlying amplified respSNA have not yet been
elucidated, it has been suggested that interaction between central (Molkov et al., 2011)
and peripheral chemoreceptor pathways (Guyenet et al., 2009; Moraes et al., 2015) may
contribute to increased sensitivity of retrotrapezoid nucleus (RTN). The RTN has been
shown to play a critical role in amplified respSNA under conditions of chronic
intermittent or sustained hypoxia (Molkov et al., 2014; Moraes et al., 2014). Of
relevance to our work is that peripheral chemoreceptor hypersensitivity has been
suggested to contribute to the development of sympathetic over activity in kidney
disease (Hering et al., 2007). In the current study, stimulation of peripheral
chemoreceptors by NaCN and hypoxia evoked an increase in respSNA parameters in
both the juvenile and adult animals, respectively, that was of a significantly greater
17
magnitude in the LPK. In contrast, exposure to hypercapnia, which drives a centrally
mediated chemoreflex and subsequent sympathoexcitatory and pressor response,
produced changes which were similar between the strains. Moreover, adult LPK rats
showed a higher respSNA even in control condition, and inspiratory motor activity was
similar between strains after both peripheral and central chemoreceptor stimulation.
Given the evidence that chronic intermittent hypoxia induces sensitization of peripheral
chemoreceptors, producing an exaggerated sympathoexcitation and increased
expiratory-related SNA (Braga et al., 2006; Moraes et al., 2012; Zoccal et al., 2008),
and that increased peripheral chemoreceptor sensitivity contributes to sympathetic over
activity and hypertension in SHR rats (Paton et al., 2013b; Tan et al., 2010), we
speculate that there may be at least two mechanisms working centrally in the medulla
oblongata to contribute the augmented respSNA seen in association with hypertension
in CKD; one dependent on peripheral chemoreceptors and another one that is
independent.
Hypoxia is a common contributing factor for amplified respSNA and increased SNA in
disease conditions including hypertension (Calbet, 2003; Somers et al., 1988; Zoccal et
al., 2008), and while we did not investigate directly the mechanism behind increased
respSNA in CKD in the present study, of relevance is our previous published data that
has documented vascular remodelling, hypertrophic changes and calcification in the
aortic arch of LPK animals (Salman et al., 2014). This type of vascular remodelling is
closely associated with ageing, hypertension and CKD in humans and is documented to
occur in the carotid/vertebral arteries, and in association with atherosclerosis, can result
in increased intima-media thickness and ultimately narrowing the lumen of the carotid
and vertebral artery (Tanaka et al., 2012; Yamada et al., 2014). Given that the carotid
18
body and brainstem are highly vascular organs and peripheral/central chemoreceptors
are sensitive to reduced blood flow, remodelling of the carotid body/cerebral arterioles
could be a persistent stimulus for central/peripheral chemoreceptor hypersensitivity
(Moraes et al., 2014; Paton et al., 2013a). Anaemia is also common in CKD, a feature
well documented in the LPK strain (Phillips et al., 2015), which in itself may be cause
of persistant hypoxia. Regardless of the mechanism, we presume that in CKD, a
persistent hypoxic stimulus via peripheral chemoreceptor independent and/or dependent
pathways, can drive premotor sympathetic neurons and respiratory (expiratory) neurons
to produce amplified respSNA and increased sympathetic tone in similar way to that of
other pathological conditions (Wong-Riley et al., 2013; Zoccal and Machado, 2010;
Zoccal et al., 2008).
In conclusion, we provide evidence that respSNA is augmented and associated with
increased SNA and hypertension in CKD. We further demonstrate an exaggerated
respSNA response to peripheral chemoreceptor stimulation, indicating this modulatory
pathway may act as one of the drivering factors of autonomic dysfunction in CKD. The
pathways responsible for modulation of respSNA, from the carotid body
chemoreceptors to central sites such as the RVLM, may well be valuable novel
therapeutic targets in CKD. Our data also show that changes in respSNA are evident
very early in the disease process, suggesting that early intervention may help to reduce
the complex pathogenesis of CKD.
Funding sources
This work was supported by the National Health and Medical Research Council of
Australia (GNT1030301, GNT1030297, GNT1102477) and Macquarie University
19
Australia. M Saha is a recipient of a Macquarie University International Research
Scholarship, and M Menuet was supported by a McKenzie Research Fellowship from
the University of Melbourne, Australia.
References
Abbott, S.B., Pilowsky, P.M., 2009. Galanin microinjection into rostral ventrolateral
medulla of the rat is hypotensive and attenuates sympathetic chemoreflex. Am. J.
Physiol. Regul. Integr. Comp. Physiol. 296, R1019-1026.
Augustyniak, R.A., Tuncel, M., Zhang, W., Toto, R.D., Victor, R.G., 2002. Sympathetic
overactivity as a cause of hypertension in chronic renal failure. J. Hypertens. 20, 3-9.
Boczek-Funcke, A., Habler, H.J., Janig, W., Michaelis, M., 1992. Respiratory
modulation of the activity in sympathetic neurones supplying muscle, skin and pelvic
organs in the cat. J Physiol 449, 333-361.
Braga, V.A., Soriano, R.N., Machado, B.H., 2006. Sympathoexcitatory response to
peripheral chemoreflex activation is enhanced in juvenile rats exposed to chronic
intermittent hypoxia. Exp. Physiol. 91, 1025-1031.
Calbet, J.A., 2003. Chronic hypoxia increases blood pressure and noradrenaline
spillover in healthy humans. J Physiol 551, 379-386.
Campese, V.M., Ku, E., Park, J., 2011. Sympathetic renal innervation and resistant
hypertension. Int. J. Hypertens. 2011, 814354.
Campese, V.M., Massry, S.G., 1983. Effects of acute and chronic treatment with
clonidine. Chest 83, 380-383.
20
Chang, A.J., Ortega, F.E., Riegler, J., Madison, D.V., Krasnow, M.A., 2015. Oxygen
regulation of breathing through an olfactory receptor activated by lactate. Nature 527,
240-244.
Czyzyk-Krzeska, M.F., Trzebski, A., 1990. Respiratory-related discharge pattern of
sympathetic nerve activity in the spontaneously hypertensive rat. J Physiol 426, 355-
368.
Grassi, G., Bertoli, S., Seravalle, G., 2012. Sympathetic nervous system: role in
hypertension and in chronic kidney disease. Curr. Opin. Nephrol. Hypertens. 21, 46-51.
Guyenet, P.G., 2006. The sympathetic control of blood pressure. Nat. Rev. Neurosci. 7,
335-346.
Guyenet, P.G., Bayliss, D.A., Stornetta, R.L., Fortuna, M.G., Abbott, S.B., DePuy,
S.D., 2009. Retrotrapezoid nucleus, respiratory chemosensitivity and breathing
automaticity. Respir. Physiol. Neurobiol. 168, 59-68.
Hanly, P., 2004. DAILY HEMODIALYSIS -- SELECTED TOPICS: Sleep Apnea and
Daytime Sleepiness in End-Stage Renal Disease. Semin Dial 17, 109-114.
Haselton, J.R., Guyenet, P.G., 1989. Central respiratory modulation of medullary
sympathoexcitatory neurons in rat. Am. J. Physiol. 256, R739-750.
Hering, D., Zdrojewski, Z., Krol, E., Kara, T., Kucharska, W., Somers, V.K.,
Rutkowski, B., Narkiewicz, K., 2007. Tonic chemoreflex activation contributes to the
elevated muscle sympathetic nerve activity in patients with chronic renal failure. J.
Hypertens. 25, 157-161.
21
Klein IH, L.G., Neumann J, Oey PL, Koomans HA,Blankestijn PJ., 2003. Sympathetic
nerve activity is inappropriately increased in chronic renal disease. J. Am. Soc. Nephrol
14, 3239 -- 3244.
Klein, I.H.H.T., Ligtenberg, G., Oey, P.L., Koomans, H.A., Blankestijn, P.J., 2001.
Sympathetic activity is increased in polycystic kidney disease and is associated with
hypertension. J. Am. Soc. Nephrol. 12, 2427-2433.
Koshiya, N., Guyenet, P.G., 1996. Tonic sympathetic chemoreflex after blockade of
respiratory rhythmogenesis in the rat. J Physiol 491 ( Pt 3), 859-869.
Levitan, D., Massry, S.G., Romoff, M., Campese, V.M., 1984. Plasma catecholamines
and autonomic nervous system function in patients with early renal insufficiency and
hypertension: effect of clonidine. Nephron 36, 24-29.
Machado, B.H., Zoccal, D.B., Moraes, D.J.A., 2017. Neurogenic hypertension and the
secrets of respiration. Am. J. Physiol. Regul. Integr. Comp. Physiol. 312, R864-R872.
Marcus, N.J., Del Rio, R., Schultz, E.P., Xia, X.H., Schultz, H.D., 2014. Carotid body
denervation improves autonomic and cardiac function and attenuates disordered
breathing in congestive heart failure. J Physiol 592, 391-408.
McCooke, J.K., Appels, R., Barrero, R.A., Ding, A., Ozimek-Kulik, J.E., Bellgard,
M.I., Morahan, G., Phillips, J.K., 2012. A novel mutation causing nephronophthisis in
the Lewis polycystic kidney rat localises to a conserved RCC1 domain in Nek8. BMC
Genomics 13.
22
Menuet, C., Le, S., Dempsey, B., Connelly, A.A., Kamar, J.L., Jancovski, N., Bassi,
J.K., Walters, K., Simms, A.E., Hammond, A., Fong, A.Y., Goodchild, A.K.,
McMullan, S., Allen, A.M., 2017. Excessive Respiratory Modulation of Blood Pressure
Triggers Hypertension. Cell Metab 25, 739-748.
Menuet, C., Sevigny, C.P., Connelly, A.A., Bassi, J.K., Jancovski, N., Williams, D.A.,
Anderson, C.R., Llewellyn-Smith, I.J., Fong, A.Y., Allen, A.M., 2014.
Catecholaminergic C3 neurons are sympathoexcitatory and involved in glucose
homeostasis. J. Neurosci. 34, 15110-15122.
Menuet, C., Wlodek, M.E., Fong, A.Y., Allen, A.M., 2016. Respiratory modulation of
sympathetic nerve activity is enhanced in male rat offspring following uteroplacental
insufficiency. Respir. Physiol. Neurobiol. 226, 147-151.
Molkov, Y.I., Zoccal, D.B., Baekey, D.M., Abdala, A.P., Machado, B.H., Dick, T.E.,
Paton, J.F., Rybak, I.A., 2014. Physiological and pathophysiological interactions
between the respiratory central pattern generator and the sympathetic nervous system.
Prog. Brain Res. 212, 1-23.
Molkov, Y.I., Zoccal, D.B., Moraes, D.J., Paton, J.F., Machado, B.H., Rybak, I.A.,
2011. Intermittent hypoxia-induced sensitization of central chemoreceptors contributes
to sympathetic nerve activity during late expiration in rats. J. Neurophysiol. 105, 3080-
3091.
Moraes, D.J., Bonagamba, L.G., Costa, K.M., Costa-Silva, J.H., Zoccal, D.B.,
Machado, B.H., 2014. Short-term sustained hypoxia induces changes in the coupling of
sympathetic and respiratory activities in rats. J Physiol 592, 2013-2033.
23
Moraes, D.J., da Silva, M.P., Bonagamba, L.G., Mecawi, A.S., Zoccal, D.B., Antunes-
Rodrigues, J., Varanda, W.A., Machado, B.H., 2013. Electrophysiological properties of
rostral ventrolateral medulla presympathetic neurons modulated by the respiratory
network in rats. J. Neurosci. 33, 19223-19237.
Moraes, D.J., Machado, B.H., Paton, J.F., 2015. Carotid body overactivity induces
respiratory neurone channelopathy contributing to neurogenic hypertension. J Physiol
593, 3055-3063.
Moraes, D.J., Zoccal, D.B., Machado, B.H., 2012. Medullary respiratory network drives
sympathetic overactivity and hypertension in rats submitted to chronic intermittent
hypoxia. Hypertension 60, 1374-1380.
Paton, J.F., Ratcliffe, L., Hering, D., Wolf, J., Sobotka, P.A., Narkiewicz, K., 2013a.
Revelations about carotid body function through its pathological role in resistant
hypertension. Curr. Hypertens. Rep. 15, 273-280.
Paton, J.F., Sobotka, P.A., Fudim, M., Engelman, Z.J., Hart, E.C., McBryde, F.D.,
Abdala, A.P., Marina, N., Gourine, A.V., Lobo, M., Patel, N., Burchell, A., Ratcliffe,
L., Nightingale, A., 2013b. The carotid body as a therapeutic target for the treatment of
sympathetically mediated diseases. Hypertension 61, 5-13.
Phillips, J.K., 2005. Pathogenesis of hypertension in renal failure: role of the
sympathetic nervous system and renal afferents. Clin. Exp. Pharmacol. Physiol. 32,
415-418.
24
Phillips, J.K., Boyd, R., Krockenberger, M.B., Burgio, G., 2015. Progression of anemia
and its relationship with renal function, blood pressure, and erythropoietin in rats with
chronic kidney disease. Vet. Clin. Pathol. 44, 342-354.
Phillips, J.K., Hopwood, D., Loxley, R.A., Ghatora, K., Coombes, J.D., Tan, Y.S.,
Harrison, J.L., McKitrick, D.J., Holobotvskyy, V., Arnolda, L.F., Rangan, G.K., 2007.
Temporal relationship between renal cyst development, hypertension and cardiac
hypertrophy in a new rat model of autosomal recessive polycystic kidney disease.
Kidney Blood Press. Res. 30, 129-144.
Salman, I.M., Hildreth, C.M., Ameer, O.Z., Phillips, J.K., 2014. Differential
contribution of afferent and central pathways to the development of baroreflex
dysfunction in chronic kidney disease. Hypertension 63, 804-810.
Salman, I.M., Kandukuri, D.S., Harrison, J.L., Hildreth, C.M., Phillips, J.K., 2015.
Direct conscious telemetry recordings demonstrate increased renal sympathetic nerve
activity in rats with chronic kidney disease. Front. Physiol. 6.
Sarma, K.D., Hildreth, C.M., Phillips, J.K., 2012. Cardiac autonomic dysfunction in
chronic kidney disease. J. Hypertens. 30, 271.
Schlaich, M.P., Socratous, F., Hennebry, S., Eikelis, N., Lambert, E.A., Straznicky, N.,
Esler, M.D., Lambert, G.W., 2009. Sympathetic activation in chronic renal failure. J.
Am. Soc. Nephrol. 20, 933-939.
Seals, D.R., Suwarno, N.O., Joyner, M.J., Iber, C., Copeland, J.G., Dempsey, J.A.,
1993. Respiratory modulation of muscle sympathetic nerve activity in intact and lung
denervated humans. Circ. Res. 72, 440-454.
25
Sekizuka, H., Osada, N., Kida, K., Yoneyama, K., Eguchi, Y., Miyake, F., 2010.
Relationship between chronic kidney disease and sleep blood pressure in patients with
sleep apnea syndrome. Hypertens. Res. 33, 1278-1282.
Simms, A.E., Paton, J.F., Allen, A.M., Pickering, A.E., 2010. Is augmented central
respiratory-sympathetic coupling involved in the generation of hypertension? Respir.
Physiol. Neurobiol. 174, 89-97.
Simms, A.E., Paton, J.F., Pickering, A.E., Allen, A.M., 2009. Amplified respiratory-
sympathetic coupling in the spontaneously hypertensive rat: does it contribute to
hypertension? J Physiol 587, 597-610.
Somers, V.K., Mark, A.L., Abboud, F.M., 1988. Potentiation of sympathetic nerve
responses to hypoxia in borderline hypertensive subjects. Hypertension 11, 608-612.
Spyer, K.M., 1993. Central cardio-respiratory regulation. J. Auton. Nerv. Syst. 43, 47-
48.
Sun, Q.J., Minson, J., Llewellyn-Smith, I.J., Arnolda, L., Chalmers, J., Pilowsky, P.,
1997. Botzinger neurons project towards bulbospinal neurons in the rostral ventrolateral
medulla of the rat. J. Comp. Neurol. 388, 23-31.
Tan, Z.Y., Lu, Y., Whiteis, C.A., Simms, A.E., Paton, J.F., Chapleau, M.W., Abboud,
F.M., 2010. Chemoreceptor hypersensitivity, sympathetic excitation, and
overexpression of ASIC and TASK channels before the onset of hypertension in SHR.
Circ. Res. 106, 536-545.
26
Tanaka, M., Abe, Y., Furukado, S., Miwa, K., Sakaguchi, M., Sakoda, S., Kitagawa, K.,
2012. Chronic Kidney Disease and Carotid Atherosclerosis. J. Stroke Cerebrovasc. Dis.
21, 47-51.
Vanholder, R., Massy, Z., Argiles, A., Spasovski, G., Verbeke, F., Lameire, N., 2005.
Chronic kidney disease as cause of cardiovascular morbidity and mortality. Nephrol.
Dial. Transplant. 20, 1048-1056.
Wilson, R.J., Remmers, J.E., Paton, J.F., 2001. Brain stem PO(2) and pH of the working
heart-brain stem preparation during vascular perfusion with aqueous medium. Am. J.
Physiol. Regul. Integr. Comp. Physiol. 281, R528-538.
Wong-Riley, M.T.T., Liu, Q., Gao, X.-p., 2013. Peripheral-central chemoreceptor
interaction and the significance of a critical period in the development of respiratory
control. Respir. Physiol. Neurobiol. 185, 156-169.
Yamada, S., Oshima, M., Watanabe, Y., Miyake, H., 2014. Arterial location-specific
calcification at the carotid artery and aortic arch for chronic kidney disease, diabetes
mellitus, hypertension, and dyslipidemia. Calcif. Tissue Int. 95, 267-274.
Yao, Y., Hildreth, C.M., Farnham, M.M., Saha, M., Sun, Q.J., Pilowsky, P.M., Phillips,
J.K., 2015. The effect of losartan on differential reflex control of sympathetic nerve
activity in chronic kidney disease. J. Hypertens. 33, 1249-1260.
Zoccal, D.B., Machado, B.H., 2010. Sympathetic overactivity coupled with active
expiration in rats submitted to chronic intermittent hypoxia. Respir. Physiol. Neurobiol.
174, 98-101.
27
Zoccal, D.B., Simms, A.E., Bonagamba, L.G., Braga, V.A., Pickering, A.E., Paton, J.F.,
Machado, B.H., 2008. Increased sympathetic outflow in juvenile rats submitted to
chronic intermittent hypoxia correlates with enhanced expiratory activity. J Physiol 586,
3253-3265.
28
Tables
Table 1: Baseline cardiorespiratory function in adult Lewis and LPK rats
MAP (mmHg)
Lewis
90 ± 4
SBP (mmHg)
117 ± 8
DBP (mmHg)
PP (mmHg)
75 ± 3
41 ± 6
HR (bpm)
453 ± 9
LPK
125 ± 10*
194 ± 22*
99 ± 10*
87 ± 18*
480 ± 3*
PNA amplitude (µV)
18.3 ± 2.1
24.9 ± 5.3
PNA frequency
(cycles/min)
36 ± 1
45 ± 1*
PNA duration (s)
0.82 ± 0.1
MPA
674.9 ± 86.03
sSNA (µV)
rSNA (µV)
3.1 ± 0.5
5.2 ± 0.4
0.72 ± 0.01
1125 ± 232
7.9 ± 1.3*
8.9 ± 0.6*
Measures of cardiorespiratory function in Lewis and LPK rats under control conditions.
MAP: mean arterial pressure, SBP: systolic blood pressure, DBP: diastolic blood
pressure, PP: pulse pressure, HR: heart rate, bpm: beats per min, PNA: phrenic nerve
29
amplitude, MPA: minute phrenic activity, rSNA: renal sympathetic nerve activity,
sSNA: splanchnic sympathetic nerve activity. Results are expressed as mean ± SEM.
*P<0.05 between the Lewis and LPK as determined by Student's t-test. n = 9 Lewis
and n = 8 LPK excepting rSNA where n = 5 LPK and 6 Lewis
30
Table 2: RespSNA parameters in splanchnic and renal sympathetic nerves under
control conditions.
Lewis
LPK
PA (V)
Duration
(sec)
splanchnic
4.1 ± 0.8
renal
8.7 ± 0.6
splanchnic
0.8 ± 0.07
renal
0.8 ± 0.05
AUC (µV.s)
splanchnic
3.5 ± 0.5
renal
7.1 ± 0.2
8.8 ± 1.1*
11.02 ± 0.8*
0.99 ± 0.02
0.96 ± 0.05
8.7 ± 1.1*
10.6 ± 1.1*
Measures of respSNA (phrenic triggered) parameters for integrated sSNA and rSNA in
Lewis and LPK rats under control conditions. PA: peak amplitude, AUC: area under
curve. Results are expressed as mean ± SEM. *P<0.05 between the Lewis and LPK as
determined by Student's t-test. sSNA data: n = 9 Lewis and n = 8 LPK, excepting rSNA
where n = 5 LPK and 6 Lewis.
31
Table 3: Effects of peripheral and central chemoreceptor stimulation on
cardiorespiratory parameters in adult Lewis and Lewis Polycystic Kidney (LPK) rats
∆ MAP (mmHg)
∆ SBP (mmHg)
∆ DBP (mmHg)
∆ PP (mmHg)
Lewis (n = 9)
LPK (n = 8)
4 ± 4
5 ± 5
4 ± 4
1 ± 2
21 ± 5 *
34 ± 11 *
16 ± 4
18 ± 7 *
Hypoxia
∆ HR (bpm)
12 ± 2
10 ± 1
∆ PNA amplitude (µV)
10.9 ± 3.6
9.8 ± 3.9
∆ PNA duration (sec)
0.13 ± 0.02
0.07 ± 0.01
∆ PNA frequency (cycles/min)
5 ± 2
-2 ± 2*
∆ MPA
536.6 ± 152.5
300.7 ± 66.9
∆ MAP (mmHg)
∆ SBP (mmHg)
Hypercapnea
∆ DBP (mmHg)
∆ PP (mmHg)
∆ HR (bpm)
32
19 ± 3
22 ± 4
18 ± 3
4 ± 2
-8 ± 1
16 ± 4
29 ± 8
13 ± 3
16 ± 6
-4 ± 1 *
∆ PNA amplitude (µV)
8.7 ± 2.7
9.1 ± 1.6
∆ PNA duration (sec)
0.15 ± 0.02
0.11 ± 0.02
∆ PNA frequency (cycles/min)
-3 ± 1
-4 ± 1
∆ MPA
178.5 ± 74.08
293.9 ± 71.04
Delta change in phrenic nerve activity (PNA) and blood pressure (mmHg) under
hypoxic or hypercapnic conditions in adult Lewis and Lewis Polycystic Kidney (LPK)
rats. MAP: mean arterial pressure, SBP: systolic blood pressure, DBP: diastolic blood
pressure, PP: pulse pressure; HR: heart rate, MPA: minute phrenic activity. Results are
expressed as mean ± SEM. *P<0.05 between the Lewis and LPK as determined by
Student's t-test. ∆ = Delta change in, n = number of animals per group
33
Figures
Figure 1. Representative traces showing higher sympathetic nerve activity during
different phases of respiration in adult Lewis and LPK rats
Representative data traces showing raw and integrated splanchnic SNA (sSNA and
∫sSNA), raw and integrated renal SNA (rSNA and ∫rSNA), raw phrenic nerve activity
(PNA) and raw arterial pressure (AP) in an adult Lewis (A) and LPK rat (B) under
control conditions. Traces illustrate the higher rSNA, sSNA and AP including pulse
pressure range in the LPK in comparison to Lewis rats. Expanded trace shows raw and
integrated splanchnic and renal SNA over two respiratory cycles, noting different scale
for floating scale bars for Lewis and LPK to allow for clear illustration of respSNA.
34
Figure 2. Tracings of effect of hypoxia on respiratory-related sympathetic nerve activity
in adult Lewis and LPK rats.
Example figure illustrating phrenic triggered integrated rSNA (∫rSNA) and sSNA
(∫sSNA) during different phases of the phrenic cycle (I, inspiration; PI, post inspiration;
E, expiration within trace of phrenic nerve activity PNA) under control (panels A, B, E,
F) and hypoxic conditions (panels C, D, G, H) in a Lewis rat (panels A, C, E, G) and
LPK rat (panels B, D, F, H).
35
Figure 3. Grouped data of effect of hypoxia on respiratory-related sympathetic nerve
activity in adult Lewis and LPK rats.
Grouped data is shown in panels A-D illustrating the change in peak amplitude (V,
A/B), and area under the curve ((V.s, AUC; C/D) of the phrenic triggered ∫sSNA
(panels A/C) and ∫rSNA (panels B/D) under control (SaO2=100%) and hypoxic
(SaO2=82-84%) conditions. Data is expressed as mean ± SEM n ≥5 per group. *
represents P<0.05 versus control within each strain, # represents P<0.05 versus
treatment-matched Lewis.
36
Figure 4. Tracings of effect of hypercapnia on respiratory-related sympathetic nerve
activity in adult Lewis and LPK rats.
Example figure illustrating phrenic triggered integrated rSNA (∫rSNA) and sSNA
(∫sSNA) during different phases of the phrenic cycle (I, inspiration; PI, post inspiration;
E, expiration within trace of phrenic nerve activity PNA) under control (panels A, B, E,
F) and hypercapnic (panels C, D, G, H) conditions in a Lewis rat (panels A, C, E, G)
and LPK rat (panels B, D, F, H).
37
Figure 5. Group data of effect of hypercapnia on respiratory-related sympathetic nerve
activity in adult Lewis and LPK rats.
Grouped data is shown in panels A-F illustrating the change in peak amplitude (A/B),
area under the curve (AUC; C/D) and duration (E/F) of the phrenic triggered ∫sSNA
(panels A/C/E) and ∫rSNA (panels B/D/F) under control and hypercapnic conditions.
Data is expressed as mean ± SEM n ≥5 per group. * represents P<0.05 versus control
within each strain # represents P<0.05 versus treatment-matched Lewis.
38
Figure 6. Effect of stimulation of the peripheral chemoreceptor reflex with NaCN on
respiratory-related sympathetic nerve activity in juvenile Lewis and LPK rats using the
working heart brainstem preparation.
Representative figure illustrating integrated tSNA (∫tSNA) over the respiratory cycle in
a Lewis rat (panel A) and LPK rat (panel B), noting different floating scale bars for
Lewis and LPK (∫VNA) to allow for clear illustration of nerve activity. Grouped data
illustrating effect of peripheral chemoreceptor stimulation on peak amplitude (C) and
area under the curve (AUC; D) of phrenic triggered ∫tSNA. Data is expressed as mean ±
SEM n minimum 5 per group. * represents P<0.05 versus control within each strain #
represents P<0.05 versus treatment-matched Lewis.
39
Figure 7. Effect of stimulation of the central chemoreceptor reflex with hypercapnia on
respiratory-related sympathetic nerve activity in juvenile Lewis and LPK rats using the
working heart brainstem preparation.
Representative figure illustrating phrenic triggered integrated tSNA (∫tSNA) over the
respiratory cycle in a Lewis rat (panels A and B) and LPK rat (panels C and D) under
control (panels A and C) and hypercapnic (panels B and D) conditions. Grouped data
illustrating the effect of central chemoreceptor stimulation on peak amplitude (E) and
area under the curve (AUC; F) of the phrenic triggered ∫tSNA. Data is expressed as
mean ± SEM n ≥5 per group. * represents P<0.05 versus control within each strain, #
represents P<0.05 versus treatment-matched.
40
Table 1: Baseline cardiorespiratory function in adult Lewis and LPK rats
MAP (mmHg)
Lewis
90 ± 4
SBP (mmHg)
117 ± 8
DBP (mmHg)
PP (mmHg)
75 ± 3
41 ± 6
HR (bpm)
453 ± 9
LPK
125 ± 10*
194 ± 22*
99 ± 10*
87 ± 18*
480 ± 3*
PNA amplitude (µV)
18.3 ± 2.1
24.9 ± 5.3
PNA frequency
(cycles/min)
36 ± 1
45 ± 1*
PNA duration (s)
0.82 ± 0.1
MPA
674.9 ± 86.03
sSNA (µV)
rSNA (µV)
3.1 ± 0.5
5.2 ± 0.4
0.72 ± 0.01
1125 ± 232
7.9 ± 1.3*
8.9 ± 0.6*
Measures of cardiorespiratory function in Lewis and LPK rats. MAP: mean arterial
pressure, SBP: systolic blood pressure, DBP: diastolic blood pressure, PP: pulse
pressure, HR: heart rate, bpm: beats per min, PNA: phrenic nerve amplitude, MPA:
minute phrenic activity, rSNA: renal sympathetic nerve activity, sSNA: splanchnic
1
sympathetic nerve activity. Results are expressed as mean ± SEM. *P<0.05 between the
Lewis and LPK as determined by Student's t-test. n = 9 Lewis and n = 8 LPK excepting
rSNA where n = 5 LPK and 6 Lewis.
2
Table 2: RespSNA parameters in splanchnic and renal sympathetic nerves under
control conditions.
Lewis
LPK
PA (V)
Duration
(sec)
splanchnic
4.1 ± 0.8
renal
8.7 ± 0.6
splanchnic
0.8 ± 0.07
renal
0.8 ± 0.05
AUC (µV.s)
splanchnic
3.5 ± 0.5
renal
7.1 ± 0.2
8.8 ± 1.1*
11.02 ± 0.8*
0.99 ± 0.02
0.96 ± 0.05
8.7 ± 1.1*
10.6 ± 1.1*
Measures of respSNA (phrenic triggered) parameters for integrated sSNA and rSNA in
Lewis and LPK rats. PA: peak amplitude, AUC: area under curve. Results are expressed
as mean ± SEM. *P<0.05 between the Lewis and LPK as determined by Student's t-
test. sSNA data: n = 9 Lewis and n = 8 LPK, excepting rSNA where n = 5 LPK and 6
Lewis.
Table 3: Effects of peripheral and central chemoreceptor stimulation on
cardiorespiratory parameters in adult Lewis and Lewis Polycystic Kidney (LPK) rats
∆ MAP (mmHg)
∆ SBP (mmHg)
∆ DBP (mmHg)
∆ PP (mmHg)
Lewis (n = 9)
LPK (n = 8)
4 ± 4
5 ± 5
4 ± 4
1 ± 2
21 ± 5 *
34 ± 11 *
16 ± 4
18 ± 7 *
Hypoxia
∆ HR (bpm)
12 ± 2
10 ± 1
∆ PNA amplitude (µV)
10.9 ± 3.6
9.8 ± 3.9
∆ PNA duration (sec)
0.13 ± 0.02
0.07 ± 0.01
∆ PNA frequency (cycles/min)
5 ± 2
-2 ± 2*
∆ MPA
536.6 ± 152.5
300.7 ± 66.9
∆ MAP (mmHg)
∆ SBP (mmHg)
Hypercapnea
∆ DBP (mmHg)
∆ PP (mmHg)
∆ HR (bpm)
1
19 ± 3
22 ± 4
18 ± 3
4 ± 2
-8 ± 1
16 ± 4
29 ± 8
13 ± 3
16 ± 6
-4 ± 1 *
∆ PNA amplitude (µV)
8.7 ± 2.7
9.1 ± 1.6
∆ PNA duration (sec)
0.15 ± 0.02
0.11 ± 0.02
∆ PNA frequency (cycles/min)
-3 ± 1
-4 ± 1
∆ MPA
178.5 ± 74.08
293.9 ± 71.04
Delta change in phrenic nerve activity (PNA) and blood pressure (mmHg) in hypoxia
(ventilated with only room air) or hypercapnia (ventilated with 5% CO2 with 95% O2)
when switched from control condition (ventilated with oxygen enriched room air) in
adult Lewis and Lewis Polycystic Kidney (LPK) rats under urethane anaesthesia. MAP:
mean arterial pressure, SBP: systolic blood pressure, DBP: diastolic blood pressure, PP:
pulse pressure; HR: heart rate, MPA: minute phrenic activity. Results are expressed as
mean ± SEM. *P<0.05 between the Lewis and LPK as determined by Student's t-test. ∆
= Delta change in, n = number of animals per group
2
|
1902.01535 | 3 | 1902 | 2019-10-09T03:10:06 | Performance of normative and approximate evidence accumulation on the dynamic clicks task | [
"q-bio.NC",
"math.PR"
] | The aim of a number of psychophysics tasks is to uncover how mammals make decisions in a world that is in flux. Here we examine the characteristics of ideal and near-ideal observers in a task of this type. We ask when and how performance depends on task parameters and design, and, in turn, what observer performance tells us about their decision-making process. In the dynamic clicks task subjects hear two streams (left and right) of Poisson clicks with different rates. Subjects are rewarded when they correctly identify the side with the higher rate, as this side switches unpredictably. We show that a reduced set of task parameters defines regions in parameter space in which optimal, but not near-optimal observers, maintain constant response accuracy. We also show that for a range of task parameters an approximate normative model must be finely tuned to reach near-optimal performance, illustrating a potential way to distinguish between normative models and their approximations. In addition, we show that using the negative log-likelihood and the 0/1-loss functions to fit these types of models is not equivalent: the 0/1-loss leads to a bias in parameter recovery that increases with sensory noise. These findings suggest ways to tease apart models that are hard to distinguish when tuned exactly, and point to general pitfalls in experimental design, model fitting, and interpretation of the resulting data. | q-bio.NC | q-bio |
AlanVeliz-Cuba2,∗
ZacharyP.Kilpatrick5,6,†
ORIGINAL ARTICLE
Journal Section
Performanceofnormativeandapproximate
evidenceaccumulationonthedynamicclickstask
AdrianE.Radillo1,∗
KrešimirJosić3,4,†
1DepartmentofNeuroscience,University
ofPennsylvania,Philadelphia,PA19104
2DepartmentofMathematics,Universityof
Dayton,Dayton,OH45469
3DepartmentsofMathematicsandBiology
andBiochemistry,UniversityofHouston,
Houston,TX77204
4DepartmentofBioSciences,Rice
University,Houston,TX77251,USA
5DepartmentofAppliedMathematics,
UniversityofColorado,Boulder,CO80309
6DepartmentofPhysiologyandBiophysics,
UniversityofColoradoSchoolofMedicine,
Aurora,CO80045
Correspondence
ZacharyP.Kilpatrick,Departmentof
AppliedMathematics,Universityof
Colorado,Boulder,CO80309
Email: [email protected]
Fundinginformation
ThisworkwassupportedbyNSF/NIH
CRCNS(R01MH115557)andNSF
(DMS-1517629). ZPKwasalsosupported
byNSF(DMS-1615737). KJwasalso
supportedbyNSF(DBI-1707400). AVCwas
supportedbytheSimonsFoundation
(516088)andOSC(PNS0445-2).
∗co-firstauthors
†co-lastauthors
CodeAvailability: Codesdevelopedtoproducefiguresareavailableat https://github.com/aernesto/NBDT_dynamic_clicks
The aim of a number of psychophysics tasks is to uncover
howmammalsmakedecisionsinaworldthatisinflux. Here
weexaminethecharacteristicsofidealandnear -- idealob-
serversinataskofthistype. Weaskwhenandhowperfor-
mancedependsontaskparametersanddesign,and,inturn,
what observer performance tells us about their decision-
makingprocess. Inthedynamicclickstasksubjectsheartwo
streams(leftandright)ofPoissonclickswithdifferentrates.
Subjectsarerewardedwhentheycorrectlyidentifytheside
withthehigherrate,asthissideswitchesunpredictably. We
showthatareducedsetoftaskparametersdefinesregions
inparameterspaceinwhichoptimal,butnotnear-optimal
observers, maintain constant response accuracy. We also
show that for a range of task parameters an approximate
normativemodelmustbefinelytunedtoreachnear-optimal
performance,illustratingapotentialwaytodistinguishbe-
tweennormativemodelsandtheirapproximations. Inaddi-
tion,weshowthatusingthenegativelog-likelihoodandthe
0/1-lossfunctionstofitthesetypesofmodelsisnotequiva-
1
NBDT} 2
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
lent: the0/1-lossleadstoabiasinparameterrecoverythat
increaseswithsensorynoise. Thesefindingssuggestwaysto
teaseapartmodelsthatarehardtodistinguishwhentuned
exactly,andpointtogeneralpitfallsinexperimentaldesign,
modelfitting,andinterpretationoftheresultingdata.
KEYWORDS
decision-making,Poissonclicks,Bayesianinference,dynamic
environment,modelidentifiability
INTRODUCTION
1
Decision-makingtasksarewidelyusedtoprobetheneuralcomputationsthatunderliebehaviorandcognition(Luce,
1986;GoldandShadlen,2007). Mathematicalmodelsofoptimaldecision-making(normativemodels)1 havebeenkey
inhelpingusunderstandtasksthatrequiretheaccumulationofnoisyevidence(WaldandWolfowitz,1948;Goldand
Shadlen,2002;Bogaczetal.,2006). Suchmodelsassumethatanobserverintegratesasequenceofnoisymeasurements
to determine the probability that one of several options is correct (Wald and Wolfowitz, 1948; Beck et al., 2008;
Veliz-Cubaetal.,2016).
Therandomdotmotiondiscrimination(RDMD)taskisaprominentexample,inwhichtheneuralsubstratesofthe
evidence accumulation process can be identified in cortical recordings (Ball and Sekuler, 1982; Britten et al., 1992;
Roitman and Shadlen, 2002). The associated normative models take the form of tractable stochastic differential
equations(Ratcliff,1978;Bogaczetal.,2006),andhavebeenusedtoexplainbehavioraldata(RatcliffandMcKoon,
2008;KrajbichandRangel,2011). Neuralcorrelatesofsubjects'decisionprocessesdisplaystrikingsimilaritieswith
these models (Shadlen and Newsome, 1996; Huk and Shadlen, 2005), although a clear link between the two is still
underdebate(Latimeretal.,2015;Shadlenetal.,2016).
Poissonclickstasks(Bruntonetal.,2013;Odoemeneetal.,2018)haverecentlybecomepopularinstudyingthe
corticalcomputationsunderpinningmammalianperceptualdecision-making. Neuralactivityduringsuchtasksalso
appearstoreflectanunderlyingevidenceaccumulationprocess(Hanksetal.,2015). Thecorrespondingnormative
modelsandtheirapproximationsarelow-dimensionalandcomputationallytractable. Thismakesthetaskwell-suited
to the analysis of data in high-throughput experiments (Brunton et al., 2013). Piet et al. (2018) have extended the
clickstasktoadynamicenvironmenttounderstandhowanimalsadjusttheirevidenceaccumulationstrategieswhen
olderevidencedecreasesinrelevance. Glazeetal.(2015)carriedoutasimilarstudyinanextensionoftheRDMDtask.
Bothstudiesconcludedthatsubjectsarecapableofimplementingevidenceaccumulationstrategiesthatadapttothe
timescaleoftheenvironment.
However,identifyingthespecificstrategysubjectsusetosolveadecisiontaskcanbedifficultbecausedifferent
strategies can lead to similar observed outcomes (Ratcliff and McKoon, 2008). How to set task parameters to best
identifyasubject'sevidenceaccumulationstrategyhasnotbeenstudiedsystematically,especiallyindynamicenviron-
ments(Ratcliffetal.,2016). Here,wefocusonthedynamicclickstaskandaimtounderstandwhattaskparameters(or
combinationsofparameters)determineperformance,andunderwhatconditionsdifferentstrategiescanbeidentified.
Inthedynamicclickstask,twostreamsofauditoryclicksarepresentedsimultaneouslytoasubject,onestreamper
1Wewillusethephrases'optimalmodel,''optimalobserver,''normativemodel,'and'idealobserver'interchangeably,astheyrefertothebestpossiblemodelfor
agivensetoftaskandobservationconstraints.
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
3
ear(Pietetal.,2018;Bruntonetal.,2013). EachclicktrainisgeneratedbyaMarkov-modulatedPoissonprocess(Fischer
and Meier-Hellstern, 1992) whose click arrival rates switch between two possible values (λlow vs. λhigh). The two
streamshavedistinctrateswhichswitchatdiscretepointsintimeaccordingtoamemorylessprocesswithhazardrate
h. Thusstreamsplayedindifferentearsalwayshavedistinctrates(λL(t) (cid:44) λR(t)). Thesubjectmustchoosethestream
withhighestinstantaneousratewheninterrogatedatatime T,whichendsthetrial. Switchesoccuratrandomtimes
thatarenotsignaledtothesubjects,whomustthereforebasetheirdecisionontheobservedsequencesofPoisson
clicks alone. The rate h at which the environment changes is a latent variable that needs to be learned for optimal
performance. Inthisstudy,however,ourobservermodelsalwaysuseaconstantrateofchangefortheirenvironment2.
Weanalyzethenormativemodelofthedynamicclickstasktoshedlightonhowitsresponseaccuracydependson
taskparameters,asthisisameasurecommonlyusedwhenfittingtobehavioraldata(Pietetal.,2018). Asshownin
Section2,theidealobserveraccumulatesevidencefromeachclicktoupdatetheirloglikelihoodratio(LLR)ofthetwo
choices. EachclickcorrespondstoapulsatileincreaseordecreaseintheLLR.Importantly,evidencemustbediscounted
ataratethataccountsforthetimescaleofenvironmentalchanges.
Themaingoalofthisworkistoidentifyhowtaskparametersshapeanidealobserver'sresponseaccuracy,andthe
identifiabilityofevidenceaccumulationmodels. Wefindeffectiveparametersthatcanbefixedtokeeptheaccuracy
of the ideal observer constant.3 One such parameter is the signal-to-noise ratio (SNR) of the clicks during a single
epochbetweenchangesandtheotheris hT,thetriallength T rescaledbythehazardrate h. Thesetwoparameters
fullydeterminetheaccuracyofanoptimalobserverinterrogatedattheendofthetrial(Section3),aswellasresponse
accuracyconditionedonthetimesincethefinalchangepointofatrial(Section4).
Whilethenormativemodeldeterminestheoptimalstrategy,subjectsmayalsouseheuristicsorapproximations
thatarepotentiallysimplertoimplement. Theaccuracyofapproximatemodelsmayalsobemoresensitivetoparameter
changes, so fitting procedures converge more rapidly. As an example we consider a linear model, which has been
previouslyfittodatafromsubjectsperformingdynamicdecisiontasks(Pietetal.,2018;Glazeetal.,2015),andhasalso
beenstudiedasanapproximationtonormativeevidenceaccumulation(Veliz-Cubaetal.,2016). Toobtainresponse
accuracyclosetothatofthenormativemodel,thediscountingrateofthelinearmodelneedstobetunedfordifferent
clickratesandhazardrates(Section5). Incontrast,thediscountingrateinthenormative(nonlinear)modelequalsthe
hazardrate. Moreover,thelinearmodel'saccuracyismoresensitivetochangesinitsevidence-discountingparameter
thanthenonlinearmodel. 4 ThiseffectismostpronouncedatintermediateSNRvalues,suggestingataskparameter
rangewherethetwomodelscanbedistinguished.
Lastly,weaskhowmodelparameterscanbeinferredfromsubjectresponses. Usingmaximumlikelihoodfitsofthe
modelstochoicedata,weshowthatthefitdiscountingparametersareclosertothetrueparameterinthelinearmodel
compared to the nonlinear model (Section 6). This is expected, since the response accuracy of the nonlinear model
dependsweaklyonitsdiscountingparameter. Wealsoexploretheimpactofthelossfunctiononmodelfitting,and
showthatinthepresenceofsensorynoiseusinga0/1-lossfunctionresultsinasystematicbiasinparameterrecovery
(Section7). The0/1-lossfunctiongivesaoneunitpenaltyontrialsinwhichthedecisionpredictedbythemodelandthe
datadisagree,andnopenaltywhentheyagree. Therefore,minimizingthislossfunctionleadstomodelsthatbestmatch
thetrial-to-trialresponsesinthedataratherthantheresponseaccuracy.
Ultimately, our findings point to ways of identifying task parameters for which subjects' decision accuracy is
2SeeRadilloetal.(2017)foranoptimalobserverthatcanlearnthehazardrate h inadynamicversionoftheRDMDtask. Thisapproachcanbeextendedtothe
caseofthedynamicclickstasksasinRadillo(2018).
3We use the term "accuracy" to refer to the percentage of correct choices for a given model and parameter set. This is our primary measure of a model's
performanceonthetask.
4The 'nonlinear' model here refers to the family of models obtained by tuning the discounting rate away from the value defining the normative model. This
detuningresultsinamodelthatisnotnormative.
4
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
sensitivetothemodeofevidenceaccumulationtheyuseinfluctuatingenvironments. Wealsoshowhowusingdifferent
modelsanddifferentdatafittingmethodscanleadtodivergentresults,especiallyinthepresenceofsensorynoise. We
arguethatsimilarissuescanarisewheneverwetrytointerpretdatafromdecision-makingtasks.
2 NORMATIVE MODEL FOR THE DYNAMIC CLICKS TASK
In the dynamic clicks task an observer is presented with two Poisson click streams, s L(t) and s R(t) (0 < t ≤ T), and
needstodecidewhichofthetwohasahigherrate(Bruntonetal.,2013). Theratesofthetwostreamsarenotconstant,
butchangeaccordingtoahidden,continuous-timeMarkovchain, x(t),withbinarystatespace {x R , x L }. Thefrequency
of the switches is determined by the hazard rate, h, so that P(x(t + d t) (cid:44) x(t)) = h · d t + o(d t). The left and right
rates, λL(t)and λR(t),caneachtakeononeoftwovalues, {λhigh, λlow}with λhigh > λlow > 0. When x(t) = x L,wehave
(λL(t), λR(t)) = (λhigh, λlow),andwhen x(t) = x R theoppositeistrue. Therefore x(t) = x k meansthatstream k hasthe
higherrateattime t: λk (t) = λhigh (Fig.1A).Theobserverispromptedtoidentifythesideofthehigherratestream,
x(T ),atarandomtime T. Theinterrogationtime, T,issampledaheadoftimebytheexperimenterforeachtrialandis
unknowntothesubject. WereferthereadertoPietetal.(2018)andBruntonetal.(2013)formoredetailsaboutthe
experimentalsetup.
ThistaskiscloselyrelatedtothefilteringofaHiddenMarkovModelstudiedinthesignalprocessingliterature(Cappé
etal.,2005;RabinerandJuang,1986). Forasingle,2-stateMarkov-modulatedPoissonprocess(FischerandMeier-
Hellstern,1992),thefilteringproblemwassolvedbyRudemo(1972) -- seealso(Snyder,1975)forreviewandextensions.
Thisfilteringproblemcorrespondstoataskinwhichasingle,variablerateclickstreamispresentedtotheobserverwho
hastoreportwhetheratsometime T therateishighorlow. Inthepresentcase,theobserverispresentedwithtwo
coupledMarkov-modulatedPoissonprocesses. ThenormativemodelreducestothatconsideredbyRudemo(1972)
whenweconsiderasinglestreamversionofthetask,soourapproachcanbeconsideredageneralization.
AssumingthePoissonrates {λhigh, λlow},andthehazardrate, h,areknown,anormativemodelfortheinference
ofthehiddenstate, x(t),hasbeenderivedbyPietetal.(2018). Theresultingmodelcanbeexpressedasanordinary
differentialequation(ODE)describingtheevolutionoftheLLRofthetwoenvironmentalstates:
Forcompleteness,wepresentthederivationinAppendixA,yieldingthesameODEasPietetal.(2018):
d yt
d t
= κ
i =1
yt := log P(x(t) = x R s R(t), s L(t))
P(x(t) = x L s R(t), s L(t)) .
∞
∞
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
δ(t − t L
j )
rightandleftclickstreams
j =1
δ(t − t R
i ) −
−
(cid:124)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:123)(cid:122)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:32)(cid:125)
2h sinh(yt)
,
nonlineardiscounting
(1)
(2)
where κ := log(λhigh/λlow)istheevidencegainedfromaclick, δ(t)istheDiracdeltafunctioncenteredat0,and t R
)isthe i-thrightclick(resp. j-thleftclick).
t L
Eq.(2)hasanintuitiveinterpretation: Aclickprovidesevidencethatthehigherratestreamisonthesideatwhich
j
theclickwasheard. Thus,aclickheardontheright(left)resultsinapositive(negative)incrementintheLLR(Fig.1B).
Since the environment is volatile, as evidence recedes into the past it becomes less relevant. In Eq. (2) each click is
followedbyasuperlineardecaytozero. NotethatthediscountingtermonlydependsonthecurrentLLR, yt,andthe
hazardrate, h,andnotontheclickrates.
(resp.
i
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
5
FIGURE 1 A:SchematicofthedynamicclickstaskfromPietetal.(2018). B:Asingletrajectoryofthelog-likelihood
ratio(LLR), yt ,duringatrial. Theclickstreamsandenvironmentstateareshownabovethetrajectory. C:Response
accuracyoftheidealobserverasafunctionofinterrogationtimefortwodistinctSNRvalues, S/√
h,definedinEq.(6).
Twodistinctpairsofclickratesusedinsimulations(λlow = 30and 60Hz)werechosentomatcheachSNRathazardrate
h = 1Hz,resultinginoverlayingdashed(λlow = 30Hz)andsolid(λlow = 60Hz)lines. For S/√
(λhigh, λlow) = (58.17, 30)and (97.67, 60);for S/√
h = 4, (λhigh, λlow) = (70, 30)and (112.54, 60). Fixing c1 and c2 inEq.(6)
yieldsamatchinaccuracyforbothpairsofclickrates. Astimeevolvesduringthetrial,accuracysaturates. Note hT and
S/√
h jointlydetermineaccuracy. D:Maximalaccuracyoftheidealobserverat T (cid:29) 1isconstantalonglevelcurves
(blacklines)of S/√
h (seenasafunctionof (λlow, λhigh)withconstant h)acrossawiderangeofparameters,consistent
withEq.(6). Weonlyshowthe (λlow, λhigh)regionwhere 0 < λlow < λhigh. Levelcurvesof S areobliqueparabolasinthe
(λlow, λhigh)plane. D(Inset): Levelcurvesslightlyunderestimateaccuracyforsmall λhigh and λlow (black: maximal
accuracy;blue: maximalaccuracyforlargevaluesof λhigh and λlow). SeeAppendixFfordetailsonMonteCarlo
simulations.
h = 2,wetake
Performanceonthetaskmayincreasewiththeinformativenessofeachclick, κ = log(λhigh/λlow). However, κ alone
doesnotpredicttheresponseaccuracy(i.e. thefractionofcorrecttrials)ofthenormativemodel(Bruntonetal.,2013;
Piet et al., 2018). In the next section, we will show that an ideal observer's response accuracy is determined by the
clickfrequencies (λhigh, λlow)andthehazardrate h: Asequenceofafewveryinformativeclicksmayprovideasmuch
evidenceasmanyclicks,eachcarryinglittleinformation. Butiftheenvironmentalhazardrateishigh,eveninformative
clicksquicklylosetheirrelevance.
TheLLR, yt,containsalltheinformationanidealobserverhasaboutthepresentstateoftheenvironment,giventhe
observations(GoldandShadlen,2002). Ifinterrogatedattime t = T,sign(yT ) = ±1,determinesthemostlikelycurrent
state(x R for +1and x L for −1),andthereforetheresponseofanoptimaldecisionmaker. Inthefollowing,wewillshow
thattwoeffectiveparametersgoverntheresponseaccuracyoftheoptimalobserver.
LLRtimeright:left:-22001hhigh rateon left earhigh rateon right earlow-highhigh-lowlow-hightimeλhighλlowinterrogation time0.00.20.4accuracy10203040102030400.90.50.7clicksmaxaccλlow=30λlow=60λlow 4080max acc=4Sh√/=2Sh√/=4Sh√/=3Sh√/=2Sh√/BADC10.80.600.80.850.9h=1 RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
6
3 THE SIGNAL-TO-NOISE RATIO OF DYNAMIC CLICKS
Fourparameterscharacterizethedynamicclickstask: thehazardrate, h,durationofatrial,i.e. interrogationtime, T,
thelowclickrate, λlow,andthehighclickrate, λhigh. However,wenextshowthatonlytwoeffectiveparameterstypically
governanidealobserver'sperformance(Fig.1C,D):theproductoftheinterrogationtimeandthehazardrate, hT,and
thesignal-to-noiseratio(SNR)ofthedynamicstimulus. Theformercorrespondstothemeannumberofswitchesina
trial,andthelattercombinestheclickrates λlow and λhigh intoaSkellam -- typeSNR(Eq.(4)below),scaledbythehazard
rate h (Eq.(6)).
Tomotivateourdefinition,considerfirstthecaseofastaticenvironment, h = 0Hz,forwhichthenormativemodelis
givenbyEq.(2)withoutthenonlinearterm. Since κ doesnotaffectthesignof yt,responseaccuracydependsentirelyon
thedifferenceinclickcounts N R(t) − N L(t),where N j (t)arethecountingprocessesassociatedwitheachclickstream.
Thuswecandefinethedifferenceinclickcountsasthesignal,andtheSNRastheratiobetweenthesignalmeanand
standarddeviationattime T (Skellam,1946),
SNRT
0 :=
where
T λhigh − T λlow
(cid:112)T λhigh + T λlow = S ·
E[N R(T )] − E[N L(T )]
(cid:112)Var[N R(T )] + Var[N L(T )] =
(cid:112)λhigh + λlow .
λhigh − λlow
S :=
√
T ,
(3)
(4)
Inadynamicenvironment,thevolatilityoftheenvironment,governedbythehazardrate, h,alsoaffectsresponse
accuracy. The environment can switch states immediately before the interrogation time, T, decreasing response
accuracy. This suggests that accuracy will not only be determined by the click rates, but also by the length of time
theenvironmentremainsinthesamestatepriortointerrogation. UsingthisfactandthedefinitionofSNRinastatic
environment,wedeterminethestatisticsforthedifferenceinthenumberofclicksbetweenthehigh-andlow-rate
streams during the final epoch preceding interrogation (for derivation details see Appendix B). Averaging over the
Poissondistributionscharacterizingtheclicknumbers,andtheepochlengthdistributionyieldsanonlinearexpression
representingtheSNRthatinvolves S/√
h,andtherescaledtrialtime hT:
(1 − e−hT )S/√
√
h) =
(cid:112)(1 − 2hTe−hT − e−2hT )S2/h + 1 − e−hT
F (hT , S/
(5)
h
.
Theunitlessmeasureoftrialduration, hT ,characterizesthetimescaleoftheevolutionoftheLLR, yt. Asaccuracy
should not depend on the units in which we measure time, this is a natural measure of the evidence accumulation
period 5. As indicated, F (hT , S/√
h. We therefore predict that optimal observer
responseaccuracyisdeterminedbythefollowingtwoparametercombinations,
h) only depends on hT and S/√
√
S/
h =
(cid:112)h(λhigh + λlow) = c1
λhigh − λlow
and
hT = c2.
(6)
Henceforth,wewillreferto S/√
h canalsoberealized
asaSNRofEq.(2)byperformingadiffusionapproximation,andcomputingtheSNRofthecorrespondingdrift-diffusion
5Thisisrelatedtodimensionalanalysisoftenusedwhenstudyingphysicalmodels(Langhaar,1980).
h astheSNRand hT asrescaledtrialtime. Notethattheterm S/√
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
signal(SeeAppendixC).
7
Fig.1Cshowsexamplesinwhichtheidealobserver'sresponseaccuracyisconstantwhenSNRand hT arefixed.
Accuracy is computed as the fraction of trials at which the observer's belief, yt , matches the underlying state, x(t),
at the interrogation time, T, that is the fraction of trials for which sign(yT ) = x(T ). The accuracy as a function of T
and h = 1remainsconstantifwechange λhigh and λlow,butkeep S fixed. Astheinterrogationtime T isincreased,the
accuracysaturatestoavaluebelow1(Fig.1C),consistentwithpreviousmodelingstudiesofdecision-makingindynamic
environments(Glazeetal.,2015;Veliz-Cubaetal.,2016;Radilloetal.,2017;Pietetal.,2018). Evidencediscounting
limitsthemagnitudeoftheLLR, yt. Henceasequenceoflowrateclickscanleadtoerrors,especiallyforlowSNRvalues.
Moreover,onsometrialsthestate, x(t),switchesclosetotheinterrogationtime T. Asitmaytakemultipleclicksfor yt
tochangesignafterachangepoint(SeeFig.1B),thiscanalsoleadtoanincorrectresponse.
In Fig. 1D we show that the maximal accuracy (obtained for T sufficiently large) as a function of λhigh and λlow
(colormap),isapproximatelyconstantalongSNRlevelsets(blackobliquecurves). Thiscorrespondenceisnotexact
when λhigh and λlow aresmall(Fig.1Dinset),andweconjecturethatthisisbecausehigherorderstatisticsofthesignal
determineresponseaccuracyinthisparameterrange. AsdiscussedinAppendixC,forlarge λhigh and λlow wecanusea
diffiusionapproximationforthedynamicsofEq.(2). When λhigh and λlow aresmall,thediffusionapproximationdoes
notapply,andresponseaccuracyischaracterizedbyfeaturesofthesignalbeyonditsmeanandvariance. Sincethe
SNRonlydescribestheratiobetweenthemeanandstandarddeviationofthestimulus,itcannotcapturetheimpactof
higherorderstatisticsonaccuracyatlowclickrates. Nonetheless,theSNRpredictsresponseaccuracywell.
Theconsequencesoftheseobservationsaretwofold:Twoparametercombinationsdetermineoptimalperformance,
potentiallysimplifyingexperimentdesign. Toensurecoverageofdifferentresponseaccuracyregimes,wecaninitially
vary SNR and hT. To increase the accuracy of an ideal observer, it is not sufficient to increase both click rates, for
instance,sincetheSNRstaysconstantif λhigh and λlow followtheparabolasshowninFig.1D.Second,thisapproach
makestestablepredictionsabouttheaccuracyofanoptimalobserver: IfwechangeparameterssothatSNRand hT are
fixed,andasubject'saccuracyisaffected,thisindicatesthatthesubjectmaynothavelearnedthehazardrate, h oris
usingasuboptimaldiscountingmodel.
4 POST CHANGE-POINT DECISIONS DEPEND ON SNR
Tounderstandhowanoptimalobserveradaptstoenvironmentalchanges,wenextaskhowtheirfractionofcorrect
responsesdependsonthefinaltime, Tfin,betweenthelastchangepointprecedingadecisionandthedecisionitself
(Fig. 2A).OverallaccuracyagaindependsonbothSNRandrescaledtrialtime hT. Inaddition,forsufficientlylongtrials,
accuracyasafunctionoftimesincethelastchangepointdependsonlyontherescaledtimesincethechangepoint, hTfin
andtheSNR.
Iftheclick rates, λhigh and λlow, arevaried, but S and h areheldfixed, theaccuracyasafunction of Tfin remains
unchanged(Fig.2B,for h = 1, S = 2). Ontheotherhand,accuracychangesifwefix S/√
h (SNR)butvary h (Fig.2B,left
inset). Withfixed S/√
h, accuracydependsonlyontherescaledtimesincethelastchangepoint, hTfin (Fig.2B,right
inset). Thus,whileabsoluteaccuracydependsonthetotallengthofthetrial, T ,measuredinunitsofaverageepoch
length, 1/h,accuracyrelativetothelastchangepointdependsonlyontheelapsedtime, Tfin,measuredinthesameunits.
Glazeetal.(2015)introducedthenotionofanaccuracycrossovereffectinthedynamicRDMDtask: Thenormative
modelpredictsthatafterachangepointobserversupdatetheirbeliefmoreslowly,buteventuallyreachhigheraccuracy
atlowcomparedtohighhazardrates. Thusplottingthemaximalaccuracyagainsttimesincethelastchangepointfor
differenthazardratesresultsincurvesthatcross. Behavioraldataindicatesthathumanobserversbehaveaccordingto
8
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
FIGURE 2 A:Afterachangepoint,evidenceinfavorofthecorrecthypothesishastobeaccumulatedbefore
optimalobserverschangetheirbelief. Thus,thereisadelaybetweenthechangepointandthesignchangeoftheLLR.
B:Theaccuracyasafunctionoftimefollowingachangepointisthesamewhen S/√
h and h arefixed(h = 1intheplot).
B(inset): Furthermoreaccuracyisconstantforfixed S/√
h andrescaledresponsetime hTfin (left: timewithoutrescaling,
right: timeafterrescaling). C:Forlowerhazardrates, h,accuracyatchangepointtimeislower,andrequiresalonger
timetoreach0.5. D:Asthestrengthofevidence, S,increases,accuracyatthetimeofthechangepointislower,but
increasesmorerapidly. InpanelsCandDthevalueof S/√
h isthesameforcurvesofthesamecolor. Rescalingthe
timescaleofcurvesinpanelCby h wouldyieldthecurvesinpanelD.
thisprediction(Glazeetal.,2015,2018).
Asimilarcrossovereffectalsooccursinthedynamicclickstask: Accuracyjustafterachangepointislowerfor
smallhazardrates, h,(Fig. 2C)andtakeslongertoreach50%,butsaturatesatahigherlevelcomparedtomorevolatile
environments. Inslowenvironments,theoptimalobserverintegratesevidenceoveralongertimescale (1/h),leading
tomorereliableestimatesofthestate, x(t). Butthisincreasedcertaintycomesataprice,asitrequiresmoretimeto
changetheobserver'sbeliefafterachangepoint. Similarly,inenvironmentswithstrongerevidence(larger S,Eq.(4)),
accuracyimmediatelyfollowingachangepointislower,sincestateestimates,andhencethebeliefsaremorereliable
comparedtotrialswithweakevidence(Fig. 2D).However,strongerevidencealsocausesarapidincreaseinaccuracy,
whichthensaturatesatahigherlevelthanontrialswithweakerevidence(lower S). Therefore,bothevidencequality,
andenvironmentalvolatilitydetermineaccuracyafterachangepoint.
fromoptimality.
WeconcludethataccuracyafterachangepointischaracterizedbySNR(S/√
h)andtherescaledtimesincethe
changepoint, hTfin. Thisonlyholdswhentrialsaresufficientlylong,andthebeliefattrialoutsetdoesnotaffectaccuracy.
Inaddition,increasingSNRlowersaccuracyimmediatelyafterachangepoint,andincreasestherecoveryofaccuracy
toahighersaturationpoint(Fig. 2D).Ontheotherhanddecreasingvolatility, h,whilefixing S (Fig. 2C)leadstolower
accuracyimmediatelyafterachangepoint,andhighersaturation. However,therateatwhichaccuracyisrecovered
afterachangepointdecreaseswithdecreasing h.
Theseareagaincharacteristicsofanoptimalobserver,anddeviationsfromthesepredictionsindicatedepartures
time after change point0.5changepoint0150-5-0.5-1time after change pointLLRaccuracy0.50100.40.8time after change pointaccuracytime after change pointaccuracyλlow=10λlow=500.5h=1h=4h=0.25h=1timerescaled time0.20.50.80.5010.501h=1h=2h=0.5=2Sh√/=2Sh√/ACDB0010.510.50010.51S=2=1S=2S=4S RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
9
FIGURE 3 A:Optimallineardiscountingrate, γ∗ inEq.(7),asafunctionof S/√
h. A(inset): Theaccuracyofthe
linearandnonlinearmodelarenearlyidenticaloverawiderangeofSNRvalues, S/√
h,whenthelineardiscountingrate
issetto γ∗ (here h = 1isfixed). B:Responseaccuracyneartheoptimaldiscountingrateforthelinearmodel(dashed),
andassumedhazardratesforthenonlinearmodel(solid)forseveralSNRvalues(h = 1). Thelinearmodelismore
sensitivetorelativechangesinthediscountingrate. Therelativeerrorisdefinedas 100( h − h)/h forthenonlinearmodel
and 100( γ − γ∗)/γ∗ forthelinearmodel. C:Curvature(absolutevalueofthesecondderivative)oftheaccuracyprofiles
inpanelB,evaluatedattheirpeak,asafunctionof S/√
h. Thecurvature,andhencesensitivity,ofthenonlinearmodelis
higherforintermediateandlargevaluesof S/√
h. SincethefunctionsinpanelBdonotdependontheactualvaluesof h
and γ,butrathertherelativedistanceoftheseparametersfromreferencevalues,whatweshowinthisplotarerelative
curvatures. Wecomparerelativecurvaturesas h and γ donothavethesameunits. D:Ratiooftheaccuracyofthelinear
modeltothatofthenormativemodel,asSNRisvaried. Alongeachcurve,thediscountingrate γ ofthelinearmodelis
heldfixedatthevalue γ∗ thatwouldmaximizeaccuracyatthereferenceSNRindicatedbythelegend.
5 A LINEAR APPROXIMATION OF THE NORMATIVE MODEL
FollowingPietetal.(2018)wenextshowthatanapproximationofthenormativemodelgivenbyEq.(2)canbetunedto
givenearoptimalaccuracy,buttheaccuracyoftheapproximationtendstobesensitivetothechangesinthediscounting
parameter. Thisapproximate,linearmodelisgivenby,
∞
i =1
d yt
d t
= κ
δ(t − t R
i ) −
δ(t − t L
j )
∞
j =1
− γ · yt .
(7)
In particular, here the nonlinear sinh term in Eq. (2) is replaced by a linear term proportional to the accumulated
evidence.
Whentunedappropriately,Eq.(7)closelyapproximatesthedynamicsandaccuracyoftheoptimalmodel(Fig.3A)(Piet
etal.,2018;Veliz-Cubaetal.,2016). Moreover,italsoprovidesagoodfittotheresponsesofratsonadynamicclicks
task(Pietetal.,2018). Asthenormativeandlinearmodelsexhibitsimilardynamics,itappearsthattheyaredifficult
todistinguish. However,asweshownext,thelinearmodelismoresensitivetochangesinitsdiscountingparameter,
providingapotentialwaytodistinguishbetweenthemodels.
Weassumethat T islargeenoughsothataccuracyhassaturated(asinFig. 1C),andcomparethemaximalaccuracy
ofthenonlinearandlinearmodel. Forthelineardiscountingratethatmaximizesaccuracy, γ = γ∗, thelinearmodel
max accuracycurvature0240non linear modelmax accuracy024non linlinearnon linlinearγ*relative error in h, γ~~Sh√/Sh√/=4Sh√/=3Sh√/=2Sh√/=1Sh√/Sh√/new0.94max-accuracy ratioACDB0.60.8120406120.90.80.7-100%0%100%0.050.1linear model804120.971=4Sh√/=3Sh√/=2Sh√/=1Sh√/Sh√/ 10
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
obtainsaccuracynearlyequaltothenormativemodel(Fig.3A,inset). Theoptimallineardiscountingrate, γ∗,increases
withSNR(Fig.3A),whereasthediscountingterminthenormative,nonlinearmodelremainsconstantwhenthehazard
rate, h, is fixed. When SNR is large, evidence discounting in the linear model can be stronger (larger γ∗), since each
evidenceincrementismorereliableandcanbegivenmoreweight. WhenSNRislower,linearevidencediscountingis
weaker(smaller γ∗)resultingintheaveragingofnoisyevidenceacrosslongertimescales.
Whatistheimpactofusingthewrong(suboptimal)evidencediscountingrateinthetwomodels? Toanswerthis
questionwecomparetheaccuracyoftwoobservers,oneusingthenonlinearmodelwithawronghazardrate, h (cid:44) h,and
theotherusingthelinearmodelwithasuboptimaldiscountingrate γ (cid:44) γ∗. AsshowninFig.3Baccuracyismoresensitive
torelativechangesin γ inthelinearmodel,thanrelativechangesintheassumedhazardrate, h,inthenonlinearmodel.
Wequantifiedthesensitivityofbothmodelstochangesinevidencediscountingratesbycomputingthecurvatureof
accuracyfunctionsattheoptimaldiscountingvalueoverarangeofSNRs(Fig.3C).
BothmodelsareinsensitivetochangesintheirdiscountingparameteratlowSNR(bottomcurveofFig.3B).This
resultisintuitive,aswhenSNRissmallobserversperformpoorlyregardlessoftheirassumptions. Ontheotherhand,
whenSNRishighobserversreceivestrongevidencefromasingleclick,andthenonlinearmodeladequatelyadapts
acrossabroadrangeofdiscountingparametervalues. Thelinearmodel,however,isstillsensitivetochangesinthe
discountingparameter, γ. AthighSNR,thebelief, yt ,asdescriedbyeithermodelisdriventolargervalues. Whereasthe
nonlinearmodelcanrapidlydiscountextremebeliefsasitincludesasupralinearleakterm,thelinearmodelisnotas
welladapted,andrequiresfinetuning. Note,however,thatatvaluesofSNRhigherthantheonesusedinFig.3B,when,
forinstance,asingleclickissufficientforanaccuratedecision,boththelinearandnonlinearmodelsareinsensitiveto
changesintheirdiscountingparameters. Wealsonotethattheinsensitivityofthenonlinearmodeltochangesinthe
discountingrate, h,suggeststhatthisisamorerobustmodel: Anobserverwhodoesnotlearnthehazardrate, h,exactly
canstillperformwell. Alinearmodelrequiresfinerparametertuningtoachievemaximalaccuracy.
Thenonlinearmodelobtainsmaximalaccuracyaslongastheassumedhazardratematchesthetruehazardrate
h = h. Ontheotherhand,theoptimaldiscountingrateofthelinearmodelisalsosensitivetochangesintheSNRdueto
changesintheclickrates. Toquantifythiseffect,wecomputedtheratiobetweenthemaximalaccuracyofthelinear
modelwithdiscountingrate γ tothemaximalaccuracyofthenonlinearmodelwith h = h astheSNRwasvaried,but
h waskeptfixed(Fig.3D).Tocomputethemaximalaccuracywekept γ fixedat γ∗,theoptimaldiscountingratefora
referenceSNR.Themaximal-accuracyratioforthelinearmodeldecreasesasSNRchangedfromthisreferenceSNR,as
theoptimaldiscountingparameterofthelinearmodeldependsonSNR,andthehazardrate h. Thus,thelinearmodel
canachievemaximalaccuracyveryclosetothatofthenonlinearmodel,butthisrequiresfinetuning.
Thispointstoageneraldifficultyindistinguishingmodelssubjectscouldusetomakeinference: Simplerapproxima-
tionsmaypredictperformancethatisnearidenticaltothatofanormativemodel. However,thismayrequireprecise
tuning of the approximations. If the parameters of the task are changed to differ from those on which the subjects
have been trained, i.e. on tasks where subjects are lead to assume incorrect parameters, the normative model may
behavedifferentlyfromtheapproximations. Inthecaseweconsidered,themodelsmaybedistinguishableifananimal
isextensivelytrainedontrialswithfixedparameters h and S,butsubsequentlyinterrogatedusingoccasionaltrialswith
differenttaskparameters.
Theprecedingpointisillustratedbythefollowingthoughtexperiment. Assumeasubjectisextensivelytrainedon
afixedsetoftaskparameters: Sref = 2, (λlow, λhigh) = (2, 8.5)Hz, h = 1Hz(peakoftheredcurveinFig.3D).Wethen
newhigh) = (12, 53)Hzand h = 1Hz,chosensothat
introducesometrialswithdifferentclickrates,say Snew = 5with (λ
κ = log(λhigh/λlow)isconstantacrossthetwoconditions. WedenotebyAcclin(S)andAccnorm(S)theaccuraciesofan
observerusingthelinearandnormativemodelsontrialswithagiven S. Sincethesubjectwastrainedonclickratesthat
correspondto Sref,theirdiscountingstrategywillbeadaptedtothesevalues. NotethattheratiobetweenAccnorm(S)
newlow , λ
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
11
andAcclin(S)when h = 1istheredcurveinFig.3D.SincetheratiobetweenAccnorm(Sref)andAcclin(Sref)isnear1,the
linearandnormativemodelscannotbedistinguishedat Sref. However,asubjectusingthenormativemodeltunedat
Sref,willstillperformoptimallyat S (cid:44) Sref,if κ and h areheldconstant. Ontheotherhand,alinearmodeloptimizedat
Sref,willnolongerbeoptimalat S (cid:44) Sref. Thisdistinctioniscapturedbythedropintheaccuracyratioalongthered
curveinFig.3D.
Wecanquantifythedistinctionbetweenthetwomodelsbytheirrelativedifference:
Accnorm(Snew) − Acclin(Snew)
Accnorm(Snew)
= 0.05.
Moregenerally,foranydecisionmakingmodel,wemaydefinethequantity
Dmodel(S) :=
Accnorm(S) − Accmodel(S)
Accnorm(S)
,
whichwillequal0ifthemodelusedisthenormativeone. Ifwecompute Dmodel(S)usingresponsesfromarealsubject,
one can generate curves such as those in Fig. 3D. If the curves are not constant (equal to 1), this would suggest the
subjectisnotusinganoptimalmodel. Furthermore,asinglevalueof Snew forwhich Dmodel(Snew) (cid:44) 0providesevidence
thatthemodelisnotoptimal.
Inthenexttwosections,weshowhowthelinearandnonlinearmodelwithaddedsensorynoisedifferwhenfitting
thediscountingparameterstochoicedata.
6 FITTING DISCOUNTING PARAMETERS IN THE PRESENCE OF SENSORY
NOISE
Themodelswehavediscussedsofartranslatesensoryevidenceintodecisionsdeterministically,anddonotaccountfor
thenervoussystem'sinherentstochasticity(Faisaletal.,2008). Wenextaskedwhethertheinclusionofsensorynoise
leadstofurtherdifferencesbetweenthetwomodels,particularlywhenfittochoicedata.
Bruntonetal.(2013)showedthatinthestaticversionoftheclickstaskhumansandratsmakedecisionsthatare
bestdescribedbyamodelinwhichevidenceobtainedfromeachclickisvariable. Inthedynamicversionofthetask,Piet
etal.(2018)showedthatrats'suboptimalaccuracyiswellexplainedbyamodelthatincludessimilarinternalvariability.
Pietetal.(2018)modeledsuch"sensory"noiseeitherbyapplyingGaussianperturbationstotheevidencepulses,orby
attributing,withsomemislocalizationprobability,aclickcomingfromtherightorleftspeakertothewrongside.
Asaminimalmodelofneuralorsensorynoise,wetoointroducedadditiveGaussiannoiseintotheevidencepulseof
eachclick,sothatthenonlinearmodelinEq.(2)takestheform
(cid:16)
∞
i =1
d yt
d t
=
ηi δ
t − t R
i
(cid:17) −
∞
j =1
(cid:16)
t − t L
j
(cid:17) − 2h sinh(yt),
(8)
ζj δ
where ηi , ζj ∼ N(κ, σ)arei.i.d.Gaussianrandomvariableswithmean κ andstandarddeviation σ. Similarly,thelinear
modelfromEq.(7)becomes:
(cid:16)
∞
i =1
(cid:16)
(cid:17) −
∞
j =1
d yt
d t
=
ηi δ
t − t R
i
ζj δ
t − t L
j
(cid:17) − γy .
(9)
12
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Before fitting these models to choice data, we note that an increase in sensory noise, σ, decreases the value of
thediscountingparametersthatmaximizeaccuracyinbothmodels(Pietetal.,2018): Noisierobservationsrequire
integrationofinformationoverlongertimescales(Fig.4A,B).Thus,adaptivitytochangepointsissacrificedinorderto
pooloverlargersetsofobservations. This,inturn,leadstolargerbiases,particularlyafterchangepoints. Asimilar
trade-offbetweenadaptivityandbiashasbeenobservedinmodelsandhumansubjectsperformingarelateddynamic
decisiontask(Glazeetal.,2018).
Wenextfitthediscountingparametersinbothmodelsusingsyntheticchoicedata,treatingtheotherparameters
ofthemodelsasknown. Todosoweproducedresponsesusingafixedreferencemodelfrombothclasses,andfitamodel
fromeachclasstotheresultingdatasets. Specifically,let mref ∈ {L,NL}(L=linear,NL=nonlinear)denotethereference
modelusedtoproducethechoicedata,andlet mfit ∈ {L,NL}denotethemodelthatwasfittotheresultingdata. We
independentlystudiedthefourpossiblemodelpairs (mfit, mref). Inwhatfollows, θ referstothediscountingparameter
thatwasfittodatainanygivenclass,sothat θ := γ when mfit = Land θ := h when mfit = NL. Notealsothatthehazard
rateparameter, θ = h,thatwasfittodatainthecase mfit = NLisdistinguishedbothfromthehazardrate, hstim,used
togenerateclickstimuli,andthehazardrate, href,usedtoproducethereferencechoicedataofthenonlinearmodel.
Therefore,toremoveambiguity,wedenoteby href, γref thetwoconstantdiscountingparametervaluesusedtoproduce
FIGURE 4 A:Accuracyasafunctionofthediscountingparameter γ forthelinearmodelwithsensorynoise
describedbyEq.(9). Asnoiseincreases,maximalaccuracyisachievedatlowerdiscountingvalues(dottedlines). B:
SameasAforthenonlinearmodelwithdiscountingparameter h. C,DThewhiskerplotsrepresentthespreadofthe
posteriormodes(MAPestimates)obtainedacrossthe500fittingprocedures,foreachmodelpairandreference
datasetsize. Oneachbox,theredlineindicatesthemedianestimate. TheMAPestimatesarecloserwithmoretrials,
butarebiasedinthecaseofmodelmismatch. E:Averagerelativeerror,Eq.(11),infittingthediscountingparameterasa
functionofreferencedatasetsize. Eachcolorcorrespondstoaspecificpair (mfit, mref). Inthecaseofmodelmismatch
(L-NLandNL-Lcurves)therelativeerrorwillnotconvergetozeroduetothebiasintheparameterestimate. PanelsC-E
describethesamesetoffits. SeeAppendixFforsimulationdetails.
05100.50.70.9accuracyL=0.1=1=200.511.5h0.50.70.9NL=0.1=1=2maxmaxABrel. error10020030040050010-210-1100NL-NLNL-LL-NLL-L100500100500Number of trials015Fits of hmfit= NLmref=NLmref=L100500100500Number of trials01020Fitsofmfit= Lmref=Lmref=NLCDENumber of trials200400600800100010-210-1100rel. errorNL-NLNL-LL-NLL-LNL-NLNL-LL-NLL-LNumber of trialsE RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
13
thereferencechoicedatawiththenonlinearandlinearmodels,respectively. Topicktheseconstantsinoursimulations,
wetookthevaluesthatwouldmaximizeaccuracyinthecorrespondingnoise-freesystems. Thatis, href := hstim = 1and
γref := γ∗ ≈ 6.75(SeeAppendixFformoredetailsonthesimulations).
Duringasinglefit,wegeneratedstimulusdatafor N i.i.d.trials,
(cid:110)
,
t L
j
(cid:111)NL,k
(cid:18)(cid:8)t R
i
(cid:9)NR ,k
i =1
D := {(Tk , dk ) : 1 ≤ k ≤ N } ,
(cid:19) isthesequenceof NR ,k rightclicksand NL,k leftclicksontrial k,and dk ∈ {0, 1}isthe
where Tk :=
choicedatumforthistrial. WeusedBayesianparameterestimation(SeeAppendixD.3fordetails)toobtainaposterior
probabilitydistributionoverthediscountingparameter,P(θD).
Toaccountforthevariabilityintheposteriorsthatariseduetofinitesizeeffects,weperformed M = 500indepen-
dentfitspermodelpair (mfit, mref),withdifferentdatasetsizes: N ∈ {100, 200, 300, 400, 500}. Toquantifythegoodness
ofthesefits,weusedtherelativemeanposteriorsquarederror,averagedacrossthe M fits,
(10)
j =1
Þ ∞
0
M
i =1
err(p1, . . . , pM , θtrue) :=
1
θ2trueM
pi(θ)(θ − θtrue)2
d θ.
(11)
posteriordensity,Pr(cid:0)θ(cid:12)(cid:12) D(cid:1),fromfit i. Note,thedefinitionof θtrue isnuanced. Ifthereferenceandfitmodelsarethe
This quantity provides a relative measure of how close the posterior distribution is to θtrue. Here pi(θ) denotes the
same,then θtrue issettothegroundtruth,i.e. thediscountingparametervalueusedtoproducethereferencechoice
data (e.g., θtrue = href when mfit = mref = NL). However, when the fit and reference model classes differ (i.e. when
mfit (cid:44) mref), then there is no obvious ground truth, and θtrue must be defined differently. In this case, we used the
correspondence γref ↔ href. Thatis,whenfittingthenonlinearmodelwealwaysset θtrue := href,andwhenfittingthe
linearmodel,wealwaysset θtrue := γref. Thereareotherpossiblewaysofdefining θtrue inthiscase,suchaspickinga
discountingparametervalueforthefitmodelclassthatproducesthesameaccuracyasthereferencemodel. Although
arbitrary, our definition is sufficient to illustrate -- as we show next -- that cross-model fits are feasible and that the
mfit = Lcaseisqualitativelydifferentthanthe mfit = NLcase,regardlessofthereferencemodelclass. However,due
to the model mismatch, we expect a bias in the parameter estimate for these situations (i.e. an error that does not
convergeto 0),unlesswedefinethegroundtruthself-referentiallyasthevalueoftheparameterforwhichtheestimate
isunbiased.
Themaximizer θ∗ ofourBayesianposteriorP(θD)definesthemaximumlikelihoodestimate(MLE)6 ofourdis-
counting parameter. We plot the distribution of these across the 500 independent fits, for each (mfit, mref) pair in
Fig.4C,D.Asthenumberoftrialsusedinthereferencedatasetincreasedfrom100to500,thespreadoftheestimates
diminishes. However,abiasintheestimateappearswhenever mfit (cid:44) mref. Forreferencedatasetsofsize500,98%of
the500MAPestimatesintheL-NLfitsliestrictlyabove γtrue, versus50.4%forthecorrespondingL-Lfits. Similarly,
86.6%oftheestimatesintheNL-Lfitsliestrictlybelow htrue,versus44.2%forthecorrespondingNL-NLfits.
WefoundthattherelativeerrorfromEq.(11)decreasesaslargerblocksoftrialsareusedtofitthediscounting
parameter(Fig.4E).Wenotethefollowingparallelsbetweenthesensitivitytoparameterperturbationofeachmodel
class(exploredinFig.3B,C)andthedecreasingrateoftherelativeerrorsforeachmodelpair. Asexpected,amodelthat
producesresponsesthatarelesssensitivetochangesinitsdiscountingparameterrequiresmoretrialstobefittodata:
Thereductioninrelativeerroristheslowestforthe (NL,NL)and (NL,L)pairs. Thisisconsistentwiththeinsensitivityof
thenonlinearmodeltochangesindiscountingparameter,makingitdifficulttoidentifyitsparameters. Ontheother
6WhichisequaltotheMaximumAPosteriori(MAP)estimateinourcase,aswepickedauniformprioroverawideinterval(SeeAppendixD.3fordetails).
14
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
hand,thelinearmodelfits -- (L,NL)and (L,L) -- convergemorerapidly,likelybecausethelinearmodelissensitiveto
changesinitsdiscountingparameter(SeeFig.3B,C).
Inanticipationofournextsection,wepointoutthatcomputingtheMLEcanbetreatedasastatisticallearning
probleminwhichweminimizeanegativelog-likelihoodlossfunctionoverthedataset D(SeeEq. 7.8inFriedmanetal.
(2001)):
LLL(dk , mfit(Tk )σ, θ) := −logP(dk = mfit(Tk )σ, θ).
(12)
th trial. Asbefore
Here dk and mfit(Tk )arethechoicesgeneratedbythereferenceandfitmodels,respectively,onthe k
thediscountingparameter, θ,andthelevelofsensorynoise, σ,parametrizethefittedmodel. Fittedmodelresponses
mfit(Tk )arenon-deterministiconlybecauseofsensorynoise. ThelikelihoodP(dk = mfit(Tk )σ, θ)istheprobabilitythat
theresponsegeneratedbythefitmodelontrial k matchestheresponseobservedinthedata(SeeAppendixFfordetails
onhowthislikelihoodwascomputedforeachmodelclass),whichmustbeobtainedfrommanyrealizationsof mfit(Tk )
subjecttoclicknoiseofamplitude σ. TheMLE, θ∗,for mfit isthenfoundbyminimizingtheexpectedlossacrossalltrials,
N
k =1
∗ := argmin
θ
E[LLL(dk , mfit(Tk )σ, θ)] = argmin
θ
θ
1
N
LLL(dk , mfit(Tk )σ, θ),
(13)
takingtheexpectationoverall N samplesinthedataset,butconditioningonthefittedmodel'sdiscountingparameter,
θ and noise amplitude, σ. As the MLE is consistent, we expect the fit parameters will converge to the true parame-
ters(Wald,1949)(Fig.4C,D).FramingBayesianparameterestimationinthiswaywillhelpuscomparetoourapproach
offittingbyminimizingthe0/1-lossfunctionweintroducenext.
7 FITTING WITH THE 0/1-LOSS FUNCTION
We next asked how the parameters that define the model whose responses best match the choices of a reference
observercomparetothosethatmaximizethelikelihoodofobservingthesechoices. Aswenotedminimizingthelog-
likelihoodloss LLL giveninEqs.(12)and(13)givestheparametersmostlikelytohaveproducedthedata,andweexpect
thecorrespondingestimatesofthediscountingparametertoconvergetothetruevaluewhenthefitandreference
modelsmatch.
Tofindtheparametersthatmaximizetheprobabilityofmatchingthechoicesofthemodeltothoseobservedina
datasetoneverytrial,wedefinethe0/1-lossfunction,
L0/1(dk , mfit(Tk , Zj )σ, θ) := (cid:177)(dk (cid:44) mfit(Tk , Zj )σ, θ),
where (cid:177)istheindicatorfunction, (Tk , dk )isadatasampleindicatingtheclickstimulusandresponseonatrial k,and
mfit(Tk , Zj )istheresponseofthefittedmodelwithdiscountingparameter θ,clickstimulus Tk,and Zj, j = 1, ..., Q are
realizationsofsensorynoise,i.e. asequenceofi.i.d. Gaussianvariablesthatperturbtheevidenceobtainedfromeach
click. Wewillmarginalizeoverrealizationsofthe(unobserved)sensorynoise,and Q denotesthenumberofrealizations
weuseintheactualcomputation. Fittingthediscountingparameter θtheninvolvesminimizingtheempiricalexpectation
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
ofthelossfunction L0/1 overthedatasamples (Tk , dk )andacrossrealizations, Zj ,ofsensorynoise,
15
E(cid:2)L0/1(dk , mfit(Tk , Zj )σ, θ)(cid:3) = argmin
∗ := argmin
θ
θ
Q
N
1
θ
Q N
j =1
k =1
L0/1(dk , mfit(Tk , Zj )σ, θ).
Forabinarydecisionmodel,thisinvolvesfindingtheparameter θ thatminimizestheexpectednumberofmismatches
(orprobabilityofamismatch)betweenthechoicesofthemodelandthoseobservedbythedata(minimizing0/1-loss),
ormaximizestheexpectednumberofmatches(orprobabilityofamatch)betweenthedataandfitmodel(maximizing
0/1-predictionaccuracy). Inourfits,weused Q = 1,samplingasinglerealizationofclicknoiseperturbationsperclick
stream. Aswesampledfromalargenumberofclickstreams,thiswassufficienttoaveragethelossfunction.
Bothlossfunctions, LLL and L0/1,arecommonlyusedtofitmodelstodata(Friedman,1997;Friedmanetal.,2001).
Minimizing the expectation of L0/1 is reasonable, as it seems likely that the parameters that define the model that
matchesthelargestnumberofchoicesobservedinthedatashouldbeclosetotheonethereferenceobserveractually
uses(assumingthatthereisnomodelmismatch). Theseparameterswillthenalsobestpredictfutureresponses. Onthe
otherhand,minimizing LLL producesthemostlikelyparametersthatproducedtheobserveddata.
However,itiswell-knownthatparametersestimatedusingdifferentcostfunctionscandiffer,evenwhenthemodels
usedtofitandgeneratethedataagree. Toseethedifferencebetweenusing LLL and L0/1 inEq.(13)consideraBernoulli
randomvariable, B withparameter p > 0.5. Givenalargesequenceofobservedoutcomes, N → ∞,theparameterthat
minimizestheexpectedloss ¯LLL convergesto p,astheMLEisconsistentandasymptoticallyefficient(Wald,1949). On
theotherhand,theparameterthatminimizestheexpectedloss ¯L0/1 is p = 1(SeeAppendixE):Theindividualoutcomes
inaseriesofindependenttrialsarebestmatchedbyamodelthatalwaysguessesthemorelikelyoutcome.
Weobservedasimilarbiaswhenweusedthe L0/1 lossfunctiontoinferthediscountingparametersinourevidence
accumulationmodels(Friedman,1997): Wegeneratedasetof 106 click-trainrealizations, Tk ,andtwosetsofresponses,
dk ,fromeachthelinearandnonlinearevidenceaccumulationmodelswithsensorynoise(SeeAppendixF).Nextweused
thesestimulusrealizationsasinputtoanevidence-accumulationmodel(linearornonlinear)withafixeddiscounting
parameter to produce a corresponding set of 106 reference observer responses. We generated a second set of model
responsesusingthesamedatabaseof 106 click-trainrealizations,butallowedthediscountingparameter θ tovary. We
callthefractionoftimethereferenceobserverandmodelresponsesagreethe0/1-predictionaccuracy(PA)ofthemodel,
thecomplementoftheexpected0/1-lossoveratestset,PA = 1 − ¯L0/1. Whenthemodelandreferenceobserveragree
thePAis1intheabsenceofsensorynoise(σ = 0),asthestimulusdeterminesthechoicefully. However,PAdecreases
assensorynoiseincreases.
Somewhatsurprisingly,theparametersthatminimizeexpected0/1-lossarebiased,andthisbiasincreaseswith
sensorynoise(Fig.5). Inparticular,thediscountingparameterthatbestpredictsthereferenceobserverresponsesis
lowerthantheoneusedtogeneratetheseresponses(Fig.5B,D).ThisisconsistentwithourobservationsinSection6,
as integration over longer periods of time decreases response variability (Fig. 4A,B). This tendency is pronounced
whenlargervaluesofthediscountingparametersareusedtoproducethetrainingdata. Largerdiscountingleadsto
shorterintegrationtime,andincreasedvariabilityintheresponses. Furthermore,thenonlinear(NL)modelexhibitsthis
biasmuchmorestronglythanthelinearmodel(L).SeeAppendixGforapossiblemetricofthereportedbias,andits
dependenceonsensorynoiseforeachmodelclass(Fig.6).
Thussensorynoiseisthemainreasontheexpected0/1-lossisminimizedatadiscountingparameterthatdoesnot
matchtheoneusedtogeneratethedata. Suchinternalnoiseintroducesvariabilityintheresponses: eventhesame
modelwillnotmatchitsownresponsesgiventhesamestimulus,andadecreaseinoutputvariabilitycanincreasethe
PAofamodel. Inthepresentcase,suchadecreaseinresponsevariabilityisachievedbydecreasingthediscounting
16
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
FIGURE 5 The0/1-predictionaccuracy(PA)ofafitmodelasafunctionofthediscountingparameterofthefit
model(verticalaxis),andmodelusedtogeneratetrainingdata(horizontalaxis). Eachcolumncorrespondstoanoise
value,andeachrowtoamodel(linearornonlinear). ColorsindicatethevariationofthePAasthediscounting
parametersarevaried(for γ2 ≤ γ1,asPAissymmetricaboutthediagonal). Redcurvesrepresentthefitparameterthat
maximizespercentmatch,asafunctionofthereferenceparameter. Forhighernoisevalues,thislieswellbelowthe
diagonal, γ1 = γ2,whichwouldcorrespondtomatchingtheparametersofthereferencemodelanddata. Thesame 106
clickrealizationswereusedacrossallpanels,buteachdecisionfromthetwomodelswascomputedwithindependent
sensorynoiserealizations. Otherparametersare hstim = 1Hz, (λhigh, λlow) = (20, 5)Hz,and T = 2s.
parameter,andincreasingintegrationtime.
Weexpectthatsimilarbiasesoccurwhenevera0/1-lossfunctionisusedtofitmodelstochoicedata. Sensorynoise,
lapsesinattention,andnumerousothersourcesofnoisenearlyalwaysintroducesomevariabilityintheresponsesof
observers. Insuchcases,modelsthatarelessvariablethantheobservermaybestmatchanobservedsetofresponses,
andbestpredictfutureresponses(Friedmanetal.,2001). However,theseparametersarenotalwaysmostlikelytohave
beenusedbytheobserver. Usinga0/1-lossfunctionmaythusnotalwaysrevealtheprocessthattheobserverusedto
generatetheresponses,evenifthemodeltheobserverusesisclosetotheoneusedtofitthedata.
8 DISCUSSION
Normativemodelsofdecision-makingmakeconcretepredictionsaboutthecomputationsandactionsofexperimental
subjects,andcanbeusedtointerpretbehavioraldata(Geisler,2003). Suchmodelscanalsobeusedtoidentifytask
parameterrangesinwhichobservers'responsesaremostsensitivetotheirassumptionsaboutthetask. Inturn,such
CDPABAPAPAPAPA Psychophysicaltasksusedtoinfersubjects'decision-makingstrategiescanrequireextensivetraininganddata
collection(Hawkinsetal.,2015b,a). Normativeandapproximatelynormativedecision-makingmodelsdivergemostin
theirresponseaccuracywhentasksareofintermediatedifficulty. Aswehaveshown,taskdifficultymaybecontrolled
bycombinationsoftaskparametersrepresentingfewerdimensionsthanthetotalnumberofparameters. Identifying
theseparametercombinationsmaybepossiblebycomputingthesignal-to-noise(SNR)ratioofthestimulusproduced
byaparticularparameterset. However,subjects'responsesarealsosusceptibletonoisefromsensingandprocessing
evidence,soitisimportanttoextenddescriptionsofSNRtoaccountforsuchfactors(Bruntonetal.,2013).
Normative models of evidence accumulation and decision-making can be complex, and simpler, approximately
optimalstrategiesmayperformnearlyaswell(Wilsonetal.,2010;Glazeetal.,2018). Ifsuchapproximatestrategies
areeasiertolearnandtune, subjectsmaypreferthem. Pietetal.(2018)showedrats'performanceonthedynamic
clicks task is well fit by a linear discounting model. However, optimal and well-tuned suboptimal strategies may be
difficulttodistinguish,andthisproblemislikelytoworsenwithincreasingtaskcomplexityandcorrespondingmodel
dimensionality. Wehavedescribedpossiblemodel-guidedtaskdesignapproachesthatmayhelpteaseapartsimilarly
performingmodels.
The addition of noise in our evidence accumulation models provides an extra parameter that can account for
suboptimal performance. What is the best way to distinguishing whether internal noise or suboptimal evidence
accumulationstrategiesbestaccountforunderperformance? Onewaytodothis,assuggestedbyourmodelanalysis,is
tocollectsufficientdataovertrialsinwhichataskparameterwaschangedunbeknownsttotheobserver.
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
17
informationcanthenbeusedtoteaseapartcandidatemodelclassestheexperimentalsubjectmightbeemploying. Here
wehavefocusedonpropertiesofanormative,nonlinearmodel,anditsdifferenceswithaclose,linearapproximation.
Wefoundthatthelinearmodelismoresensitivetochangesinthediscountingparametercomparedtothenonlinear
model,andsuggestthisiswhyfittingalinearmodeltochoicedatarequiresfewertrialsthanfittinganonlinearmodel.
Indynamicenvironments,taskparametersmayhavepredictableeffectsonsubjects'overallaccuracyandaccuracy
relativetochangepoints. WehaveshownthatthereisarangeofintermediatetohighSNRinwhichthelinearmodel
issensitivetochangesinitsdiscountingparameter,butthenonlinearmodelisnot. Thissuggeststhisrangecouldbe
probedtodistinguishtheevidenceaccumulationstrategiessubjectsareusing. Thesestrategiesmayalsobefitbyother
approximate models, like accumulators with no-flux boundaries or sliding-window integrators (Wilson et al., 2013;
Glazeetal.,2015;Barendregtetal.,2019),whichcanalsobesensitivetochangesintheirdiscountingparameters.
Forpurposesofmodelfittingtoexperimentaldata,weexpectthattrial-to-trialvariabilitycanbemorefaithfully
trackedinthedynamicclickstaskthanindynamicdecisiontasksbasedontheRDMDtask. Thisisduetotherelative
simplicityoftheclicksasevidencesources: Theyareeitherontherightorleft,althoughclicksideandevidencestrength
canbemisattributed(Pietetal.,2018). Incontrast,dotmotioncanbeestimatedinmanyways,makingitdifficultto
interpretwhichaspectsofthestimulusananimalobserved,andusedasevidence. Spatiotemporalsamplingmethods
maybetoospatiallycoarseandmayrequirefittingfilterstoeachsubject,whichcouldchangetrial-to-trial(Adelsonand
Bergen,1985;Parketal.,2016). TransformingclicktimestodeltapulsesusingEq.(2)ismorestraightforward. Thus,the
dynamicclicktaskparadigmisapromisingavenueforprobingevidenceaccumulationtocomplementdynamictasks
whichareextensionsofclassicRDMD(Glazeetal.,2015).
Theuse ofdiscrete evidencetasks doescomewith caveats. The neuralcomputations underlyingvisual motion
discriminationinnon-humanprimatesarewellstudied(BornandBradley,2005),andhaveasignificanthistoryofbeing
linkedtodecisiontasks(GoldandShadlen,2007). Asaresult,thereisanextensiveliteratureconnectingneuralsystems
forprocessingvisualmotionandthoseinvolvedindecisiondeliberation(ShadlenandNewsome,1996;Roitmanand
Shadlen,2002). However, onlyrecentlyhavetheneuralunderpinningsofthedecisionsofratsperformingauditory
discriminationtasksbeenexamined(BrodyandHanks,2016). Furthermore,mathematicalissuesmayariseinprecisely
18
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
characterizingdiscountingbetweenclicks,whenevidencearrivesdiscretely. Manydifferentfunctionscouldleadtothe
sameamountofevidencediscountingbetweenclicks,leadingtoambiguityinthemodelselectionprocess.
Parameteridentificationforevidenceaccumulationmodelscanbesensitivetothemethodchosentofitmodel
responses to choice data (Ratcliff and Tuerlinckx, 2002). Glaze et al. (2015) used the approach of minimizing the
cross-entropyerrorfunction,whichmeasuresthedissimilaritybetweenbinarychoicesinthemodelandthedata. Piet
etal.(2018)usedamaximumlikelihoodapproachtoidentifymodelparametersthatmostcloselymatchedchoicedata.
ThisisrelatedtotheBayesianestimationapproachweusedtofitparametersofthenonlinearandlinearmodels. We
obtainedsimilarresultsbyminimizingtheexpected0/1-loss,whichbiasestowardslessvariablemodels,especiallyfor
modelswithstrongsensorynoise(Fig.5). Amorecarefulapproachtofittingmodelparametersshouldalsoconsider
penalizingmorecomplexmodels,whichwouldalsoallowustodistinguishbetweenthenonlinearandlinearmodel.
Glazeetal.(2018)recentlystudiedthestrategieshumansusewhenmakingbinarydecisionsindynamicenviron-
mentswhosehazardrateschangedacrosstrialblocks. Inthiscasetheidealobservermustinferboththestate,and
therateatwhichtheenvironmentischanging(Radilloetal.,2017). Interestingly, Glazeetal.(2018)foundthatthe
model that best accounted for response data was not the full Bayes optimal model, but rather a sampling model in
which a bank of possible hazard rates replaces the full hazard rate distribution. Such sampling strategies can more
easilybeimplementedinspikingnetworks(Buesingetal.,2011),andmayalsoarisewhenconsideringaninformation
bottleneck,whichforcesabalancebetweeninformationrequiredfromthepastandmodelpredictivity(Bialeketal.,
2001). AsinOccam'srazor,thebrainmayfavorsimplermodels,especiallywhentheyperformsimilarlytomorecomplex
models(Balasubramanian,1997).
Analyses of normative models for decision-making are important both for designing experiments that reveal
subjects'decisionstrategiesandfordevelopingheuristicmodelsthatmayperformnear-optimally(Veliz-Cubaetal.,
2016;Pietetal.,2018;Glazeetal.,2018). Ourfindingssuggestsubjectsshouldbetestedmainlyatintermediatelevels
of SNR to provide informative response data. We found that such a level of SNR is between 1 and 2 for an optimal
observer, and between 3 and 4 for an observer that uses linear discounting. Tasks that are too easy or hard allow
subjectstoobtainsimilarperformancewithawidevarietyofstrategies. Interestingly,wealsofoundthatthemodels
thatbestpredictobserverresponses,arenotnecessarilythoseclosesttotheonesthattheobserverisusing. Moreover,
modificationsofnormativemodelscanalsosuggestmorerevealingexperiments,likethosethatincludefeedbackor
signaled change points. Ultimately, data from decision-making tasks that require subjects to accumulate evidence
adaptivelywillprovideabetterpictureofhoworganismsintegratestimulitomakechoicesinthenaturalworld.
A NORMATIVE EVIDENCE ACCUMULATION FOR DYNAMIC CLICKS
Indynamicenvironments,thestate xt ∈(cid:8)x R , x L(cid:9) evolvesaccordingtoacontinuous-timeMarkovchainwithsymmetric
transitionratesgivenbythehazardrate, h. Weconstructasampled-timeapproximation(cid:8) x ∆t
t∈[0,T ] ofthecontinuous-
time Markov process xt, parameterized by ∆t, which is valid for ∆t small enough (Gardiner, 2009). More precisely,
wedefineadiscrete-timeMarkovchain xn ∈ {x R , x L } bythetransitionprobabilities: P(xn (cid:44) xn−1xn−1) = h · ∆t and
P(xn = xn−1xn−1) = 1 − h · ∆t,forall n ∈ (cid:206)andinitialcondition x0. Notethattheseprobabilitiesareatruncationto
firstorderin ∆t ofthetransitionprobabilitiesthatonewouldotherwiseobtainfortheembeddeddiscrete-timeMarkov
chain {xn∆t }n∈(cid:206). Then,weset
(cid:9)
t
:= xn ,
x
∆t
t
(14)
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
19
for all n ∈ (cid:206) and all n∆t ≤ t < (n + 1)∆t. In the following, our discrete-time evidence accumulation equations are
embeddedincontinuous-timeviathecorrespondencegivenbyEq.(14). As ∆t → 0theresultingequationsapplytothe
originalstateprocess xt invirtueofthesampled-timeapproximationjustdescribed.
JustasinEq.(1),thelog-posterioroddsratioindiscrete-timeis:
yn := log P(xn = x R s R(n), s L(n))
P(xn = x L s R(n), s L(n)) .
Hence,equations(A.3)and(B.1)fromtheappendixofVeliz-Cubaetal.(2016)holdinourcontext:
yn − yn−1 = log f R
∆t
f L
∆t
(ξn)
(ξn) + log 1 − h · ∆t + h · ∆t · e−yn−1
1 − h · ∆t + h · ∆t · e yn−1
.
Inaddition,weusetheapproximationlog(1 + z) ≈ z forsmall z ,since 0 < ∆t (cid:28) 1,sothat:
∆yn = log f +
∆t
f −
∆t
d yt
d t
= κ
δ(t − t R
i ) −
j =1
(ξn)
(ξn) − 2h∆t sinh(yn−1).
∞
δ(t − t L
j )
− 2h sinh(yt),
∞
j =1
δ(t − t L
j ) − 2h
κ
sinh(κ · yt),
∞
∞
i =1
i =1
δ(t − t R
i ) −
Takingthelimit ∆t → 0yieldstheODE:
ortheequivalentrescaledversion
d yt
d t
=
whichbothappearinPietetal.(2018).
B DERIVATION OF DYNAMIC CLICKS SNR
Ourderivationconsidersthesignalinthedynamicclickstasktobethedifferenceinthenumberofclicksduringthefinal
epochpriortointerrogationattime t = T. Thedistributionoffinalepochtimes τ ofthetelegraphprocess x(t)is
p(τ) = he−hτ + e−hτ δ(τ − T ),
(15)
Thefirsttermisthedistributionofwaitingtimesbetweenswitches. Wetruncatetheperiodattheinterrogationtime, T,
andthesecondtermaccountsfortheprobabilitythatnoswitchesoccurduringtheentiretrial,andthefinalandonly
epochisoflength T. Forafinalepochofagivenlength τ,wecandescribeboththeconditionalexpectationandvariance
ofthedifferenceinclickcounts ∆N againusingtheresultsofSkellam(1946):
τ ∈ [0, T ].
E[∆N τ] = (λhigh − λlow)τ,
Var[∆N τ] = (λhigh + λlow)τ.
20
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Thereforetoobtaintheunconditionalexpectationandvariancefor ∆N,wemustmarginalizeusingthelawsoftotal
expectationandvariancewithrespecttothedistributionofepochtimes τ giveninEq.(15). Thisyields
E[∆N] = (λhigh − λlow)
(16)
forthetotalexpectation. Noticethatas T → ∞,theexpectednumberofclicksislimitedfromabovebylimT →∞ E[∆N] =
(λhigh − λlow)/h. Usingthelawoftotalvariancewecanthuscompute
· (λhigh − λlow)
τp(τ)dτ =
h
0
1 − e−hT
Þ T
Var[∆N] = Var[E[∆N τ]] + E[Var[∆N τ]] = (λhigh − λlow)2 · Var[τ] +
1 − e−hT
h
· (λhigh + λlow)
1 − 2hTe−hT − e−2hT
=
h2
(λhigh − λlow)2 +
1 − e−hT
(λhigh + λlow).
= E[∆N]/(cid:112)Var[∆N]yields
h
PluggingEq.(16)and(17)intotheexpressionforSNRT
SNRT
h
=
h
(cid:113)(1 − 2hTe−hT − e−2hT )(λhigh − λlow)2 + h · (1 − e−hT )(λhigh + λlow) .
(1 − e−hT )(λhigh − λlow)
Recallingourdefinitionfromequation(4),
S :=
λhigh − λlow
(cid:112)λhigh + λhigh ,
wecanrewriteequation(18)inthemoreconvenientform
(17)
(18)
(19)
(20)
SNRT
√
= F (hT , S/
h) =
(cid:112)(1 − 2hTe−hT − e−2hT )S2/h + 1 − e−hT
scaledbytherootofthetimescale(cid:112)1/h. Indeed,inthelimitas h → 0,wefindthatSNRT
wherewehavehighlightedthefactthattheSNRisafunctionoftherescaledtrialtime hT andtheSkellamSNRrate S
T consistent
withEq.(3). Wealsofindthatinthelimitofinfinitelylongtrials T → ∞,Eq.(20)tendsto
(1 − e−hT )S/√
h → SNRT
0 = S · √
h
h
,
lim
T →∞
SNRT
h
=
sotheSNRissolelydeterminedby S/√
h.
(cid:112)S2/h + 1
S/√
h
,
NotealsothattokeepEq.(20)constantitissufficienttokeepitsconstituentargumentsconstant. Thisisconvenient,
sincewealreadymustkeep hT constanttofixthestatisticsofinformationaccumulatedpriortothefinalepoch,sowe
predictthatperformanceisfixedbythefollowingtwoparameters
S√
h
=
asreportedinEq.(6).
(cid:112)h(λhigh + λlow) = c1
λhigh − λlow
and
hT = c2,
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
C DIFFUSION APPROXIMATION
21
Herewedemonstratethediffusionapproximationofthenormativemodelforthedynamicclickstask,Eq.(2)inthe
limitoflargePoissonrates λhigh and λlow. DiffusionapproximationsforjumpprocesseshavebeenaddressedbyLánsk`y
(1997),andRichardsonandSwarbrick(2010)whostudiedtheimpactofshotnoiseandpulsatilesynapticinputson
integrate-and-fire models. Following this work, we note that the difference of the click streams in Eq. (2) can be
approximatedbyadrift-diffusionprocesswithmatchedmean, ±κ(λhi − λlow),andvariance, κ2(λhi + λlow). Thisresultsin
thefollowingstochasticdifferentialequation(SDE)fortheapproximation yt:
d yt = κg(t)(λhigh − λlow)dt + κ
(21)
where g(t) = sign[λR(t) − λL(t)] and dWt is the increment of a Wiener process. Note the resulting nonlinear drift-
diffusionmodelissimilartothenormativemodelspresentedin(Glazeetal.,2015;Veliz-Cubaetal.,2016). TheSNRof
thesignalinEq.(21)canbeassociatedwiththemeandividedbythestandarddeviationinanaverage-lengthepoch.
FixingthisSNRleadstotherelationsinEq.(6). Importantly,thesignalinEq.(21)ischaracterizedentirelybyitsmean
andvariance,soweexpectthattheperformanceofthemodelcanbedirectlyassociatedwiththeSNR.Note,however,
thatEq.(21)willonlybevalidfor λhigh, λlow (cid:29) 1. Otherwise,onemustconsidertheeffectsofhigherordermomentsof
theclickstreams,andapredictionofperformancepurelybasedontheSNRwillbreakdown(Fig.1D,Inset),sincehigher
orderstatisticslikelyshaperesponseaccuracyinthesecases.
(cid:113)
λhigh + λlowdWt − 2h sinh( yt)dt ,
D MODEL IDENTIFICATION
Wefitparametersofthelinearandnonlinearmodelsintwostages. First,wegeneratedsyntheticresponsedatafroma
model(linearornonlinear)bysolvingthecorrespondingODEorSDE.Wethensolvedasecondsetofmodels(linearor
nonlinear)forarangeofdiscountingparameters(γ forthelinearmodel; h forthenonlinearmodel),andconstructed
a posterior distribution over the discounting parameter. For noisy models, we expect the posterior to be a smooth
function that is peaked around the most likely values of discounting parameter for that trial. We now describe the
detailsoftheseparameterfittingproceduresforeachofthecases: linearvs. nonlinearmodels.
D.1 Linearmodelwithstochasticresponse
WeincorporatenoiseintothelinearEq.(7)byconsideringmultiplicativenoiseontheclickincrements,asdescribedby
Eq.(9). Forafixedrealizationoftheclicktrain,wecansolvethisequationexplicitlyfor yt attheendoftrial k:
R
N k
=
y k
T
L
N k
ηie−γ(T k −t i
R
) −
ζje−γ(T k −t
)
j
L
,
(22)
i =1
j =1
22
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Conditioningontheclicks Tk,then yT k isnormallydistributed(cid:2)yT k Tk , γ(cid:3) withexpectationandvariance
where ηi , ζj ∼ N(κ, σ),revealing yT k issimplythesumofi.i.d. normalrandomvariablesscaledbyexponentialdecay.
N k
R
Ek := E(cid:2)yT k Tk , γ(cid:3) = κ
N k
R
Vk := Var(cid:2)yT k Tk , γ(cid:3) = σ
i =1
i =1
2
c
L
N k
e−γ(T k −t i
R
) −
e−γ(T k −t
)
j
L
,
j =1
e−2γ(T k −t i
R
) +
L
N k
j =1
e−2γ(T k −t
)
j
L
,
so rk (γ) ∈ ±1isaBernoullirandomvariable. Thelikelihoodfunctionwillbeasmoothfunctionof γ,anddeterminedas
anintegraloverthehalf-linecorrespondingtothe ±1decision:
P( rk = ±1γ, Tk ) = ±
p(yT k Tk , γ)dyT k = Φ
Þ ±∞
0
(cid:18)
± Ek√Vk
(cid:19)
,
where Φ(z)isthecumulativedistributionfunctionofastandardnormalrandomvariable. Wecanthuscomputethe
posterioroverthediscountingparameter γ asarescaledproductofthelikelihoodsoneachtrial.
D.2 Nonlinearmodelwithstochasticresponse
Whenclickheightsarenoise-perturbed,wecannotexplicitlysolvetheextendednonlinearmodel. However,wecan
make progress by applying the idea of mapping between clicks. If we draw trains of clicks, Tk, ahead of time, Eq. (8)
defines the nonlinear model with multiplicative noise. We can iteratively define the probability density p(y , t) by
samplingovertheclickamplitudenoisedistributionateachclickaccordingto
2
+
n ) =
p(z , t
p(y , t
∂p(y , t)
1√
2πσ2
= 2h · ∂
∂ y
(23a)
(23b)
whereclicknoiseisdrawnfromthenormaldistribution N(κ, σ), t n isthetimeofthe n-thclick, κn = ±κ accordingto
thesideofthe n-thclick,andwehaveusedtheconvolutiontheoremforindependentrandomvariables. Foranytrains
ofclicks, Tk,Eq.(23)canbesolvediterativelytoobtainthedistribution p(y , T ). Thelikelihoodwillthusbeasmooth
functionof h,determinedbytheintegraloverthehalf-linecorrespondingtothedecision(±1):
t n < t < t n+1,
n )dz ,
−
∂t
Þ ∞
e−(y−z−κn)2/2σ
−∞
[sinh(y)p(y , t)] ,
P( rk = ±1h, Tk ) = ±
p(yT k Tk , h)dyT k .
(24)
Þ ±∞
0
D.3 Bayesianfittingprocedure
Our goal is to compute or estimate the posterior distribution Pr(cid:0)θ(cid:12)(cid:12) D(cid:1), which by Bayes' rule is proportional to the
productofthelikelihoodofthedatawiththepriorovertheparameter7:
P(θD) ∝ P(Dθ)P0(θ).
(25)
7Sincealltheothertaskandmodelparametersareassumedknownandfixed,wemayomitthemfromtheequations.
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
OurmethodfocusesonexploitingthelikelihoodfunctionP(Dθ). Wehave,
23
P(Dθ) = P(T1:N , d1:N θ) = P(d1:N T1:N , θ)P(T1:N θ) = P(d1:N T1:N , θ)P(T1:N ),
wherethelaststepcomesfromthefactthattheclickstrainsareindependentofthediscountingparameter θ usedby
thedecision-makingmodel8. Fromthere,weremarkthatthechoicedataareconditionallyindependentontheclicks
stimulusandthediscountingparameter. Thus,
ThereforewecanrewriteEq.(25)as:
N
k =1
P(dk Tk , θ).
P(d1:N T1:N , θ) =
P(θD) ∝ P0(θ) N
k =1
P(dk Tk , θ).
(26)
N
We use uniform priors for θ, over a finite interval [0, a]. In this context, the problem of computing the posterior
distributionof θ reducestoassessingthelikelihoodsofthedecisiondataoneachtrial,P(dk Tk , θ)(1 ≤ k ≤ N),fora
rangeof θ-valuesspanningtheinterval[0, a]. Inpractice,wepicked a = 40whenfittingthelinearmodeland a = 10when
fittingthenonlinearmodel. Finally,notethatfornumericalstabilityreasons,ouralgorithmsactuallysumlog-likelihood
values,asopposedtomultiplyingprobabilityvalues. Relegatingthe θ-independentpriorintoanormalizationconstant
C,Eq.(26)becomes,inthelog-domain:
logP(θD) = C +
logP(dk Tk , θ),
θ ∈ [0, a].
(27)
k =1
E MINIMIZING 0/1-LOSS IN A BERNOULLI RANDOM VARIABLE
Considerasimplestochasticbinarydecision-makingmodelinwhichweignorethespecificsofevidencesources,as
in Pesaran and Timmermann (1992). We that in this case the 0/1-loss function also leads to biased estimates. This
resulthasbeenpointedoutinpreviousworkinwhichparameterfittingresultshavebeencomparedbetweenBernoulli
randomvariablesfitwiththe0/1-lossfunctionasopposedtomaximumlikelihoodestimators(Friedman,1997;Friedman
etal.,2001).
Consider a Bernoulli random variable B1 with success probability p1 generating the reference choices, and the
fitBernoullimodel B2 withsuccessprobability p2. Minimizingthelog-likelihoodlossfunctionrecovers p∗
2 = p1 inthe
limitofalargenumberoftrials N → ∞: Inthislimit,given p2,wehavethattheexpectedlossmeasuredbythenegative
log-likelihoodis
¯LLL(d p2) = −[p1 log p2 + (1 − p1)log(1 − p2)] ,
(28)
whichisminimized9 at p∗
2 = p1, themeanof B1. Thus, theparameterfromthereferencemodelisrecovered, asthe
8Weremindthereaderthatweoperateadistinctionbetweenthediscountingparameterofthedecisionmakerandthehazardrateusedtoproducethedata.
9NoteEq.(28)isthecross-entropybetween B1 and B2.
24
BernoullirandomvariablesatisfiestherequirementsfortheMLEtobeconsistent(Wald,1949).
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Ontheotherhand,ifwefittheparameter p2 byminimizingtheexpected0/1-lossfunction,inthelimitof N → ∞
trials,theexpectedlossis
¯L0/1(d p2) = 1 − P(B1 = B2) = p1 − p2(2p1 − 1),
whichdecreasesin p2 for p1 > 0.5,sotheminimalexpectedlosswhen p1 > 0.5isachievedwith p2 = 1(for p1 < 0.5itis
minimizedat p2 = 0).
Ofcourse,thesyntheticdataandthefitevidenceaccumulationmodelsweconsideraregeneratedfromthesame
clickstreamsoneachtrial,soarealisticcomparisonshouldaccountforsuchnoisecorrelationsinsimplifiedBernoulli
randomvariablemodels,asanalyzedinDaietal.(2013). Thisanalysisismoreinvolved,andwesavesuchastudyfor
futurework.
F DETAILS ON MONTE CARLO SIMULATIONS FOR FIGURES
Fig. 1C was generated using 105 simulations of Eq. 2 from t = 0 to t = 1s with the parameters shown in the figure.
Thetimeforsaturationwaschosentobe 0.4s. Foreachtimebetween0and0.4stheaccuracywascomputedasthe
percentage of the 105 simulations for which the choices were correct. Fig 1D was generated using 105 simulations
ofEq.2foreachdatapointinthe (λhigh, λlow)plane. Themaximalaccuracyreportedcorrespondstothenumerically
computedaccuracyat t = 1s.
Fig.2B-Dwasgeneratedusing 105 simulationsofEq.2from t = 0to t = 3s withtheparametersshowninthefigure.
Thereferencechangepointwaschosentobethelastchangepointinthesimulation. Foreachtimebetweenthelast
change point and one unit of time later, the accuracy is the fraction of the correct responses, simulations for which
sign(yT ) = x(T ),thesignoftheLLRmatchedthesignofthetelegraphprocess. Sinceintervalsbetweenchangepoints
areexponentiallydistributed,therearemanymoredatapointsforshorttimesthanforlongtimesafterchangepoints.
Sincesomesimulationsdidnotlastafullunitoftimeafterthelastchangepoint,thenumberofsimulationsislessthan
orequalto 105 (decreasingastimeincreases). Simulationsthathadnochangepointwereomittedwhencomputingthe
accuracy.
Fig. 3A was generated as follows. For each value of S/√
h, 106 simulations of Eq. 7 from t = 0 to t = 1s were
generatedoverarangeof γ values. Foreachvalueof γ theaccuracywascomputedat t = 1s and γ∗ wasselectedasthe
valuethatmaximizedaccuracy. Thisresultedinaspecificvalueof γ∗ foreach S/√
h. Fig.3Bwasgeneratedusing 106
simulationsofEq.2(using h insteadof h)andEq.7(using γ insteadof γ)from t = 0to t = 1sforarangeofvaluesof h
and γ. Foreachvalueof h and γ, themaximalaccuracywasestimatedasthevalueoftheaccuracyat t = 1s. Fig.3C
wasgeneratedbyestimatingthesecondderivativeofthecurvesshowninFig.3Bforeachvalueof S/√
h. Fig.3Dwas
generatedasfollowsusingEq.2andEq.7. Foreachofthefourcurves, γ wasfixedtothevalueof γ∗ correspondingto
thereferencevalues S/√
h = i for i = 1, 2, 3, 4(seeFig.3A).Foreachcurve,thisvalueof γ wasnotchangedwhennew
S/√
h valueswereused. Then,foreachcurve,themaximalaccuracyforthelinearandnonlinearmodelswerecomputed
using 106 simulationsforarangeofnew S/√
h values. Thequotientofthemaximalaccuracyofthelinearmodelandthe
maximalaccuracyofthenonlinearmodelisshowninthefigure.
Fig.4C-Dpresentstheresultsoffivehundredindependentfittingprocedures,performedontwodifferentdataset
sizes. The parameters for the reference dataset of trials are: hstim = 1 Hz, (λhigh, λlow) = (20, 5) Hz, and T = 2 s. For
each fitting procedure, the trials (either 100 or 500) were sampled uniformly without replacement from a bank of
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
25
10,000 trials. The fitting algorithm is an implementation of the Bayesian approach leading to equation (27) above.
Whenfittingthelinearmodel,theanalyticalsolutionfromappendixD.1wasusedtocomputethelikelihoodofasingle
trial(P(dk Tk , θ)terminEq.(27)). Whenfittingthenonlinearmodel,MonteCarlosamplingwasusedinstead. More
specifically,thedistributionofthedecisionvariableatdecisiontimeforagivenclicksstimulus, p(yT T),wasestimated
bysimulating800independenttrajectories. Thus,eachtrajectoryhaditsownindependentrealizationofsensorynoise
buttherealizationofthestimulus(timingoftheclicks)wasfrozen. Oncethedensityof yT wasestimated,thelikelihood
term,P(dk Tk , θ)inEq.(27),couldbeestimated. Moredetailsonthismethod,suchashowthenumberof800particles
waschosenandhowthismethodwasvalidatedonthelinearmodelforwhichtheanalyticalsolutionisavailable,maybe
foundinsection3.5.5ofRadillo(2018).
InFig4E,uptotrialnumber500onthex-axis,thesamefitsasinpanelsC-Dwereusedtocomputetherelative
error(y-axis). Becauseofthehighcomputationalcostofourfittingalgorithm(MonteCarlosamplingdescribedabove),
thepointsfor1000trialsonthex-axiswerecomputedwithonly84independentfitspermodelpair(asopposedto500
fortheotherpointsofthefigure).
AllpanelsinFig.5wereproducedwithacommondatasetof 106 trials,generatedbypresentingthesamesetsof
clickstreamstotheevidenceaccumulationmodels. Alltrialshadsametaskparameters: trialduration T = 2s;hazard
rate h = 1Hz; λhigh = 20Hzand λlow = 5Hzso S/√
h = 3;andtheinitialstateoftheenvironmentwasrandomlyassigned
with a uniform prior. For each panel of Fig. 5, we selected a pair (mfit, mref) ∈ {(L,L), (NL,NL)} along with a sensory
noiseamplitude(σ ∈ {0.1, 2}for ηi , ζi ∼ N(κ, σ))tobeappliedtotheevidencepulsesfromtheclicks. Foreachpossible
pairofdiscountingparameters((γ1, γ2)forlinearmodels, (h1, h2)fornonlinearmodels),wecomputedthe 106 decisions
(LeftorRight)anddeterminedwhetherthemodelsagreedornot. Forthelinearmodel,weusedvaluesof γ1, γ2 between
0and10,withincrementsof0.1. Forthenonlinearmodel,weusedvaluesof h1, h2 between0and2.5,withincrements
of0.1. Foreachdecisioncomparisonbetweenreferencemodelandfitmodel,thesameclickstreamswereused,but
independentnoiserealizationsofclickperturbationswereapplied. Thenumberofagreementswasdividedbythetotal
numberofdecisioncomparisonstoproducethecolorofasinglepointintheplot.
G BIAS METRIC AS A FUNCTION OF SENSORY NOISE
In this section, we provide additional information about the bias in parameter recovery with the 0/1-loss function
describedinSection7. Fig.6Aincludesresultsfromsimulationsfornoise= 1inadditiontonoise= 0.1andnoise= 2also
showninFig.5. Biasmagnitudeanditsdependenceonsensorynoiseweredeterminedasfollows. Let θref ∈ {γ1, h1}
denotethediscountingparameterofthereferencemodel -- thisisthemodelusedtoproducetheinitialdecisiondata.
Let θfit denotethefitvalueofthediscountingparameter,using0/1-lossminimization. InFig.6A, θref spansthe x-axis
and θfit asafunctionof θref isdepictedbythegoldencurve. Aftersmoothing θfit withaSavitzky-Golayfilter,weobtain
θsmooth representedbythegreencurvesinthefigure. Pickingafixedreferencevaluefor θref (reddottedline),wethen
plotthebiasasafunctionofsensorynoiselevelsinFig.6B,wherebiasisdefinedas:
(29)
Thefixedvaluesof θref chosenwerethesameasinSection6, γ1 := 6.7457, h1 := 1. AsdescribedinSection7,thebiasin
parameterrecoverywiththe0/1-lossfittingprocedureismorepronouncedforthenonlinearmodelthanforthelinear
model,andincreaseswithsensorynoise.
θref
θsmooth .
bias := log
26
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
FIGURE 6 Biasinparameterrecoveryasafunctionofsensorynoise. A:Recovereddiscountingparameterfromthe
fitsasafunctionofthereferencediscountingparameterusedtoproducetheinitialdecisiondata. Topandbottomrows
arereproductionsofFig.5whilethemiddlerowisforanintermediatelevelofsensorynoise. Theactualfitparameters
(golden)weresmoothed(green)inordertocomputethebiasinpanelBforthereferencediscountingparameters
indicatedbythereddottedlines. Theblackdiagonalindicatestheidentityline,whichwouldcorrespondtoperfect
parameterrecovery. B:BiasinparameterrecoveryEq.(29)asafunctionofsensorynoise,forthetwomodelpairs(L-Lin
blueandNL-NLingolden).
ACKNOWLEDGEMENTS
WethankGaiaTavoni,AlexFilipowicz,andAlexPietforhelpfulfeedbackonanearlierversionofthismanuscript. Some
computationsforthismanuscriptweredoneusingtheOhioSupercomputerCenter(1987).
REFERENCES
Adelson,E.H.andBergen,J.R.(1985)Spatiotemporalenergymodelsfortheperceptionofmotion. JOSAA,2,284 -- 299.
Balasubramanian,V.(1997)Statisticalinference,Occam'srazor,andstatisticalmechanicsonthespaceofprobabilitydistribu-
tions. NeuralComput.,9,349 -- 368.
Ball,K.andSekuler,R.(1982)Aspecificandenduringimprovementinvisualmotiondiscrimination. Science,218,697 -- 698.
Barendregt, N. W., Josić, K. and Kilpatrick, Z. P. (2019) Analyzing dynamic decision-making models using Chapman-
Kolmogorovequations. arXivpreprintarXiv:1903.10131.
Beck, J. M., Ma, W. J., Kiani, R., Hanks, T., Churchland, A. K., Roitman, J., Shadlen, M. N., Latham, P. E. and Pouget, A. (2008)
Probabilisticpopulationcodesforbayesiandecisionmaking. Neuron,60,1142 -- 1152.
Bialek,W.,Nemenman,I.andTishby,N.(2001)Predictability,complexity,andlearning. NeuralComput.,13,2409 -- 2463.
Bogacz,R.,Brown,E.,Moehlis,J.,Holmes,P.andCohen,J.D.(2006)Thephysicsofoptimaldecisionmaking: aformalanalysis
ofmodelsofperformanceintwo-alternativeforced-choicetasks. Psychol.Rev.,113,700 -- 765.
NL-NLL-LABnoise=0.1noise=2noise=1 RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Born,R.T.andBradley,D.C.(2005)Structureandfunctionofvisualareamt. Annu.Rev.Neurosci.,28,157 -- 189.
Britten,K.H.,Shadlen,M.N.,Newsome,W.T.andMovshon,J.A.(1992)Theanalysisofvisualmotion:acomparisonofneuronal
27
andpsychophysicalperformance. JNeurosci.,12,4745 -- 4765.
Brody,C.D.andHanks,T.D.(2016)Neuralunderpinningsoftheevidenceaccumulator. Curr.Op.Neurobiol.,37,149 -- 157.
Brunton, B. W., Botvinick, M. M. and Brody, C. D. (2013) Rats and humans can optimally accumulate evidence for decision-
making. Science,340,95 -- 8.
Buesing,L.,Bill,J.,Nessler,B.andMaass,W.(2011)Neuraldynamicsassampling: amodelforstochasticcomputationinrecur-
rentnetworksofspikingneurons. PLoSComput.Biol.,7,e1002211.
Cappé,O.,Moulines,E.andRydén,T.(2005)InferenceinhiddenMarkovmodels. SpringerSeriesinStatistics,652.
Dai,B.,Ding,S.,Wahba,G.etal.(2013)Multivariatebernoullidistribution. Bernoulli,19,1465 -- 1483.
Faisal,A.A.,Selen,L.P.andWolpert,D.M.(2008)Noiseinthenervoussystem. Nat.Rev.Neurosci.,9,292 -- 303.
Fischer,W.andMeier-Hellstern,K.(1992)TheMarkov-modulatedPoissonprocess(MMPP)cookbook.Perform.Eval.,18,149 --
171.
Friedman,J.,Hastie,T.andTibshirani,R.(2001)Theelementsofstatisticallearning,vol.1,chap.7: ModelAssessmentandSelec-
tion. SpringerseriesinstatisticsNewYork,NY,USA.
Friedman,J.H.(1997)Onbias,variance,0/1 -- loss,andthecurse-of-dimensionality. DataMin.Knowl.Discov.,1,55 -- 77.
Gardiner,C.(2009)Stochasticmethods,vol.4. SpringerBerlin.
Geisler,W.S.(2003)Idealobserveranalysis. InTheVisualNeurosciences(eds.L.ChalupaandJ.Werner),825 -- 837.Boston: MIT
Press.
Glaze,C.M.,Filipowicz,A.L.,Kable,J.W.,Balasubramanian,V.andGold,J.I.(2018)Abias -- variancetrade-offgovernsindivid-
ualdifferencesinon-linelearninginanunpredictableenvironment. Nat.Hum.Behav.,2,213.
Glaze,C.M.,Kable,J.W.andGold,J.I.(2015)Normativeevidenceaccumulationinunpredictableenvironments. Elife,4.
Gold,J.I.andShadlen,M.N.(2002)Banburismusandthebrain: decodingtherelationshipbetweensensorystimuli,decisions,
andreward. Neuron,36,299 -- 308.
-- (2007)Theneuralbasisofdecisionmaking. Ann.Rev.Neurosci.,30.
Hanks,T.D.,Kopec,C.D.,Brunton,B.W.,Duan,C.A.,Erlich,J.C.andBrody,C.D.(2015)Distinctrelationshipsofparietaland
prefrontalcorticestoevidenceaccumulation. Nature,520,220 -- 223.
Hawkins, G., Wagenmakers, E., Ratcliff, R. and Brown, S. (2015a) Discriminating evidence accumulation from urgency signals
inspeededdecisionmaking. JNeurophysiol.,114,40 -- 47.
Hawkins,G.E.,Forstmann,B.U.,Wagenmakers,E.-J.,Ratcliff,R.andBrown,S.D.(2015b)Revisitingtheevidenceforcollapsing
boundariesandurgencysignalsinperceptualdecision-making. JNeurosci.,35,2476 -- 2484.
Huk,A.C.andShadlen,M.N.(2005)Neuralactivityinmacaqueparietalcortexreflectstemporalintegrationofvisualmotion
signalsduringperceptualdecisionmaking. JNeurosci.,25,10420 -- 10436.
Krajbich,I.andRangel,A.(2011)Multialternativedrift-diffusionmodelpredictstherelationshipbetweenvisualfixationsand
choiceinvalue-baseddecisions. Proc.Natl.Acad.Sci.U.S.A.,108,13852 -- 13857.
Langhaar,H.L.(1980)Dimensionalanalysisandtheoryofmodels. R.E.KriegerPublishing: Huntingdon,NY.
28
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Lánsk`y,P.(1997)Sourcesofperiodicalforceinnoisyintegrate-and-firemodelsofneuronaldynamics. Phys.Rev.E,55,2040.
Latimer, K. W., Yates, J. L., Meister, M. L., Huk, A. C. and Pillow, J. W. (2015) Single-trial spike trains in parietal cortex reveal
discretestepsduringdecision-making. Science,349,184 -- 187.
Luce,R.D.(1986)Responsetimes: Theirroleininferringelementarymentalorganization. OxfordUniversityPress: NewYork,NY.
Odoemene, O., Pisupati, S., Nguyen, H.andChurchland, A.K.(2018) Visualevidenceaccumulationguidesdecision-makingin
unrestrainedmice. JNeurosci.,38,10143 -- 10155.
OhioSupercomputerCenter(1987)OhioSupercomputerCenter. http://osc.edu/ark:/19495/f5s1ph73.
Park, H., Lueckmann, J.-M., Von Kriegstein, K., Bitzer, S. and Kiebel, S. J. (2016) Spatiotemporal dynamics of random stimuli
accountfortrial-to-trialvariabilityinperceptualdecisionmaking. Sci.Rep.,6,18832.
Pesaran,M.H.andTimmermann,A.(1992)Asimplenonparametrictestofpredictiveperformance. JBusiness&Econ.Statistics,
10,461 -- 465.
Piet, A.T., ElHady, A.andBrody, C.D.(2018)Ratsadopttheoptimaltimescaleforevidenceintegrationinadynamicenviron-
ment. Nat.Commun.,9,4265.
Rabiner,L.andJuang,B.(1986)AnintroductiontohiddenMarkovmodels. IEEEASSPMag.,3,4 -- 16.
Radillo, A. E. (2018)Optimaldecisionmakinginchangingenvironments. Ph.D. thesis, University of Houston, https://www.math.
uh.edu/~josic/content/04-members/Radillo.pdf.
Radillo, A.E., Veliz-cuba, A., Kilpatrick, Z.P.andJosic, K.(2017)Evidenceaccumulationandchangerateinferenceindynamic
environments. NeuralComput.,29,1561 -- 1610.
Ratcliff,R.(1978)Atheoryofmemoryretrieval. Psychol.Rev.,85,59 -- 108.
Ratcliff,R.andMcKoon,G.(2008)Thediffusiondecisionmodel: theoryanddatafortwo-choicedecisiontasks. NeuralComput.,
20,873 -- 922.
Ratcliff, R., Smith, P.L., Brown, S.D.andMcKoon, G.(2016)Diffusiondecisionmodel: currentissuesandhistory. TrendsCogn.
Sci.,20,260 -- 281.
Ratcliff, R. and Tuerlinckx, F. (2002) Estimating parameters of the diffusion model: Approaches to dealing with contaminant
reactiontimesandparametervariability. Psychon.Bull.Rev.,9,438 -- 481.
Richardson, M. J. and Swarbrick, R. (2010) Firing-rate response of a neuron receiving excitatory and inhibitory synaptic shot
noise. Phys.Rev.Lett.,105,178102.
Roitman,J.D.andShadlen,M.N.(2002)Responseofneuronsinthelateralintraparietalareaduringacombinedvisualdiscrim-
inationreactiontimetask. JNeurosci.,22,9475 -- 9489.
Rudemo,M.(1972)Doublystochasticpoissonprocessesandprocesscontrol. Adv.Appl.Probab.,4,318 -- 338.
Shadlen, M. N., Kiani, R., Newsome, W. T., Gold, J. I., Wolpert, D. M., Zylberberg, A., Ditterich, J., de Lafuente, V., Yang, T. and
Roitman, J. (2016) Comment on "Single-trial spike trains in parietal cortex reveal discrete steps during decision-making".
Science,351,1406 -- 1406.
Shadlen,M.N.andNewsome,W.T.(1996)Motionperception: seeinganddeciding. Proc.Natl.Acad.Sci.U.S.A.,93,628 -- 633.
Skellam,J.G.(1946)Thefrequencydistributionofthedifferencebetweentwopoissonvariatesbelongingtodifferentpopula-
tions. J.RoyalStat.Soc.Ser.A,109,296.
RADILLO,VELIZ-CUBA,JOSIĆ,&KILPATRICK
Snyder,D.L.(1975)Randompointprocesses. JohnWileyandSons.
Veliz-Cuba, A., Kilpatrick, Z. P. and Josić, K. (2016) Stochastic models of evidence accumulation in changing environments.
29
SIAMRev.,58,264 -- 289.
Wald,A.(1949)NoteontheConsistencyoftheMaximumLikelihoodEstimate. Ann.Math.Stat.,20,595 -- 601.
Wald,A.andWolfowitz,J.(1948)Optimumcharacterofthesequentialprobabilityratiotest. Ann.Math.Stat.,326 -- 339.
Wilson,R.C.,Nassar,M.R.andGold,J.I.(2010)Bayesianonlinelearningofthehazardrateinchange-pointproblems. Neural
Comput.,22,2452 -- 2476.
-- (2013) A mixture of delta-rules approximation to Bayesian inference in change-point problems. PLoS Comput. Biol., 9,
e1003150.
REVIEWS AND RESPONSES
PublicationDecision1fromNeurons,Behavior,Dataanalysis,andTheoryonJuly27,2019
Editorialboard'sdetermination: Reviseandresubmit
Commentsfromtheeditor. Sorry again for the long delay. The reviewer is more or less happy with the manuscript,
thereisalistofsuggestions,whichIwouldliketoaskyoutopaycarefulattentiontobeforethepapercanbeaccepted.
Reviewers'commentsareitalicized. Ourresponsesareinplaintext. Changestothemanuscriptareinblue.
Commentstotheauthor.
Summary. Thisisacarefulandthoroughpaperthatpresentsseveralresultsregardinganormativemodelofdecision-making
withindynamicenvironments. ThespecificcasestudiedisthatofanevidenceaccumulationtaskpresentedinPietetal.,2018. In
thistask,thesubjecthearstwostreamsofPoisson-timedclickscomingfromtheirleftandrightsides. Theclickratesonbothsides
mayswitchduringthetrialwithsomehazardrate h. Thesubjectisaskedtoinferwhichsidehadthehigherrateofclickswhen
thetrialends,thusmakingnewerinformationmorerelevantthanolderinformation. Insuchatask,therearefourparameters: h,
T, λhigh, λlow,where h isthehazardrate, T isthedurationofthetrial, λhigh isthehigherPoissonrateofclicks,and λlow isthe
lowerrateofclicks.
h,and hT,where S = (λhigh − λlow)/(cid:112)λhigh + λlow.
Theauthorsfirstshowthat,overabroadrangeaslongashigh,lowaresufficientlylarge,themaximalaccuracyachievableby
thenormativemodelalmostexactlydependsononlytwoparameters, S/√
Theyusethesetwoeffectiveparameterstostudythemodel. Theauthorsfirstcomparethebehaviorofthenormativemodeltoa
linearmodelthatapproximatesthenormativemodel,andfindthatthelinearmodelperformsnear-optimallyifthediscounting
parameterisfinelytuned. Tofurthercomparethemodels,theygeneratedchoicedatafromboththenormativemodelandthe
linearmodel,andfitthechoicesandclicksseparatelytoeitherthenormativemodelorthelinearmodel. Theyfoundthatthe
parametersrecoveredarebiasedwhenthemodelusedtogeneratedataandthemodelusedtofitthedataaredifferent. The
linearmodelrequiredlessnumberoftrialswhenitwasfittodata;thediscountingparameterconvergedtothetruevalueusedto
generatedatafasterforthelinearmodelthanforthenormativemodel. Theauthorsnotethatdifferentcostfunctions(MLEand
0/1-loss)leadtodifferentestimatesoftheparametervaluesandoneshouldbecautiousaboutwhatcostfunctiontouse.
Theauthorssuggestthatthenormativemodelandthelinearmodelfitsonsubjects? performanceatdifferent hT and S
provideaconvenientwayofidentifyingthedecision-makingstrategythatthesubjectisemployinginthedynamicclickstask.
Majorcomment:Theauthorsclaimthat'ifananimalisextensivelytrainedontrialswithfixedparametershandS,butsubse-
quentlyinterrogatedusingoccasionaltrialswithdifferenttaskparameters,'onemaybeabletoidentifywhetherthesubjectis
usingthenormativemodelorasuboptimal(orlinearapproximation)model. Thisisanimportantpoint,anditwouldbegreatly
strengthenedifthisclaimcanbesupporteddirectlywithsimulationsandfurtherelucidatedwithexplanationsofthespecific
stepsthatshouldbetakeninordertoidentifythestrategies.
Specifically, in the case where the clicks are generated with hstim, and the animal performs normatively with href, we may
experimental S/(cid:112)hstim,andcompareittotheaccuracyoftheanimal. However,ifwefitthelinearmodel,howshouldwego
fitthenormativemodelandrecover href reliably. Wecancomputethemaximalaccuracyofthenormativemodelgiventhe
fromhere? Howwouldone,withoutknowing mref,andknowingonlytheclicks/choicedataandtheexperimentalparameters
hstim and S,identifythemodelthattheanimalisusing?
Indeed,youraiseimportantpoints,whichwecanaddressbytakingacloserlookattheaccuracyratioswehaveplotted
inFig.3D.Wenowexplainthesetupofanexperimentwhichcouldberuntotestthesensitivityofasubject'sevidence
accumulationmodeltodeterminewhetheritisnormative-likeormoresensitivelikethelinearmodelsweconsidered.
Ofcourse,inthisidealizedcase,wearenotconsideringinternalnoise,buttheprincipleswouldextendtonoisymodels
1
asdiscussedinPietetal(2018). Ingeneral,thelinearmodelwillnotadaptwelltointerspersedtrialswithdifferent
parameters,whereasthenormativemodelwillbemorerobust. Thisisnowexplainedinthefollowingparagraphswe
haveaddedtotheendofSection4:
Theprecedingpointisillustratedbythefollowingthoughtexperiment. Assumeasubjectisextensivelytrainedona
fixed set of task parameters: Sref = 2, (λlow, λhigh) = (2, 8.5)Hz, h = 1Hz (peak of the red curve in Fig. 3D). We then
introducesometrialswithdifferentclickrates,say Snew = 5with (λ
newhigh) = (12, 53)Hzand h = 1Hz,chosensothat
κ = log(λhigh/λlow)isconstantacrossthetwoconditions. WedenotebyAcclin(S)andAccnorm(S)theaccuraciesofan
observerusingthelinearandnormativemodelsontrialswithagiven S. Sincethesubjectwastrainedonclickratesthat
correspondto Sref,theirdiscountingstrategywillbeadaptedtothesevalues. NotethattheratiobetweenAccnorm(S)
andAcclin(S)when h = 1istheredcurveinFig.3D.SincetheratiobetweenAccnorm(Sref)andAcclin(Sref)isnear1,the
linearandnormativemodelscannotbedistinguishedat Sref. However,asubjectusingthenormativemodeltunedat
Sref,willstillperformoptimallyat S (cid:44) Sref,if κ and h areheldconstant. Ontheotherhand,alinearmodeloptimizedat
Sref,willnolongerbeoptimalat S (cid:44) Sref. Thisdistinctioniscapturedbythedropintheaccuracyratioalongthered
curveinFig.3D.
newlow , λ
Wecanquantifythedistinctionbetweenthetwomodelsbytheirrelativedifference:
Accnorm(Snew) − Acclin(Snew)
Accnorm(Snew)
= 0.05.
Moregenerally,foranydecisionmakingmodel,wemaydefinethequantity
Dmodel(S) :=
Accnorm(S) − Accmodel(S)
Accnorm(S)
whichwillequal0ifthemodelusedisthenormativeone. Ifwecompute Dmodel(S)usingresponsesfromarealsubject,
one can generate curves such as those in Fig. 3D. If the curves are not constant (equal to 1), this would suggest the
subjectisnotusinganoptimalmodel. Furthermore,asinglevalueof Snew forwhich Dmodel(Snew) (cid:44) 0providesevidence
thatthemodelisnotoptimal.
Minorcomments:
Introduction: whentheterm'nonlinearmodel'isfirstused,theauthorshaven'tyetclarifiedthatherethetermissynonymous
with'normativemodel.' Also,somereadersmightnotbefamiliarwithwhata0/1lossfunctionis,makingthatpartoftheIntroa
littleunclearforthem.
Itisimportanttonotethatthenonlinearmodelisonlynormativewhenthediscountingparameteristunedexactly,
andthisiswhyweusedthisspecificphrasing. Toclarifythis,wehaveaddedtheword'nonlinear'inparentheseswhen
mentioningthenormativemodelinthepriorsentence. Wealsoaddedthefollowingfootnote: The'nonlinear'model
herereferstothefamilyofmodelsobtainedbytuningthediscountingrateawayfromthevaluedefiningthenormative
model. Thisdetuningresultsinamodelthatisnotnormative.
Wedescribethe0/1-lossfunctionnowinthesentencefollowingitsintroduction: The0/1-lossfunctiongivesaoneunit
penaltyontrialsinwhichthedecisionpredictedbythemodelandthedatadisagree,andnopenaltywhentheyagree.
Therefore,minimizingthislossfunctionleadstomodelsthatbestmatchthetrial-to-trialresponsesinthedatarather
thantheresponseaccuracy.
Page4,lastparagraph: Youneedtheword'alone': thephrase'kappadoesnotpredicttheresponseaccuracy'shouldprobablybe
'kappaalonedoesnotpredicttheresponseaccuracy.'
2
h jointly determine
h)onlydependson hT and S/√
h.
Thanksforpointingthisout. Wehavemodifiedthetextassuggested.
Figure1,caption:MightbehelpfultotitlepanelCas'h=1'andremindreadersinthecaptionthat hT and S N R jointlydetermine
accuracy? AndforpanelD,perhapswrite'Maximalaccuracyoftheidealobserver,at T (cid:29) 1?
Good suggestion. We have added the title to Fig. 1C and the sentence, "Note that hT and S/√
accuracy." tothecaption,andalsothequalifier,at T (cid:29) 1,tothecaptionofFig.1D.
Equation5: WouldbehelpfultoexplicitlysayherethatFrepresentstheSNR
Wehaveaddedthephrase"representingtheSNR"tothesentenceprecedingequation5.
Paragraphimmediatelyafterequation5: aswritten,itsoundsabitlikeFdependingonlyonthose2parametersfollowsfrom
keeping hT fixed,althoughthatisnotwhatyoumean.
Wehavechangedtheconfusingsentenceto: Asindicated, F (hT , S/√
ParagraphattheendofSection3: 'Toincreasetheaccuracyofanidealobserver,itisnotenoughtoincreasebothclickrates,for
instance.' Ididn'tunderstandthat? Ifeverythingelseiskeptfixed,butbothlambdasgrow,doesn'tSNRgrowandhTstayfixed?
InthissituationSNRdoesnotnecessarilygrow. Indeed,thisdependsonhowtheparametersareincremented. Forin-
stance,ifthelambdasgrowalongtheparabolasshowninFig.1D,thenSNRstaysconstant. Wehaveeditedthesentence
referencedabovetoexplicitlystatethis. Itnowreads: Toincreasetheaccuracyofanidealobserver,itisnotsufficientto
increasebothclickrates,forinstance,sincetheSNRstaysconstantif λhighand λlowfollowtheparabolasshowninFig.1D.
Figure3C,labelonverticalaxis: whyisthis'relative'accuracy?
Theadjective"relative",wasmeanttohighlightthefactthatwecomputedthecurvatureofthegraphsinFig.3B,repre-
sentingfunctionsoftherelativeerrorratherthantheactualvaluesof h and γ. Weagreethatthiswasconfusing,and
haveremovedtheword"relative"fromtheplotandaddedanexplanationtothecaption,whichnowsays: Sincethe
functionsinpanelBdonotdependontheactualvaluesof h and γ,butrathertherelativedistanceoftheseparameters
fromreferencevalues,whatweshowinthisplotarerelativecurvatures. Wecomparerelativecurvaturesas h and γ do
nothavethesameunits.
TheauthorsstatethatforFigure4E,NL-LandL-NLdonotconvergetozero,whereasNL-NLandL-Lconvergetozeroeventually.
This is not very clear in the figure, and it would help the reader if the number of trials was larger to show this more clearly.
Thispanelshowsthatforagivennumberoftrials(inthefigure, < 500trials),thefitsforthelinearmodelwerebetterthanthe
normativemodel. Howsufficientshouldthenumberoftrialsbeforthefitstothenormativemodeltobebetterthanthefitsto
thelinearmodel? Inprinciple,whenthenormativemodelisfittodatasetgeneratedbythenormativemodel,itshouldeventually
havelowerrelativeerrorthanthelinearmodel.
Werananadditional84fittingprocedurespermodelpairfortrainingsetsof1,000trials. Wedidnotrunmoresimu-
lationsasourMonteCarlosamplingmethodiscostly(these84simulationstook24hourstorunonamodern4-core
laptop). WereporttheresultingrelativeerrorsinthenewFig.4E.ThisfigurenowshowsthattheNL-NLcurvedoesfall
belowtheL-NLcurveonaverageafter1000trials. WealsoaddedthefollowingexplanationtoappendixFoftherevised
manuscript: InFig4E,uptotrialnumber500onthex-axis,thesamefitsasinpanelsC-Dwereusedtocomputethe
3
relativeerror(y-axis). Becauseofthehighcomputationalcostofourfittingalgorithm(MonteCarlosamplingdescribed
above),thepointsfor1000trialsonthex-axiswerecomputedwithonly84independentfitspermodelpair(asopposed
to500fortheotherpointsofthefigure).
Figure5needslabelsCandD.WhatwasthegridofvaluesusedforhandingeneratingFig. 5? Therearecomparisonsbetween
0.1and2,butdoweseeamoredeviatingtrendasafunctionofnoiselevel? Thatis,whatisthedeviationlikewhennoiseis
0.5or1,forexample? Itwouldbegreattohaveasummaryplotwithametricofdeviationseparatelyforboththelinearand
normativemodels. Whatdoweseewhenthemetricisshownasafunctionofnoise? Havingthissummaryfigurewillhelpthe
authorsestablishthepointthat'theparametersthatminimizeexpected0/1-lossarebiased,andthisbiasincreaseswithsensory
noise,'andthat'theNLmodelexhibitsthisbiasmuchmorestronglythantheLmodel.'
WehaveaddedthelabelsCandDtotheappropriatepanelsinFigure5. Wehavealsospecifiedthegridvaluesusedin
thesimulationsinappendixF,usingthefollowingaddedpassage: Forthelinearmodel,weusedvaluesof γ1, γ2 between
0and10,withincrementsof0.1. Forthenonlinearmodel,weusedvaluesof h1, h2 between0and2.5,withincrements
of0.1.
Regarding the second suggestion of measuring the bias in parameter recovery for intermediate values of noise, we
haveaddedthefigurethatappearsbelowinthisresponseletter,andisnowcontainedinAppendixGtoourmanuscript.
WereferencethisfigureinthefollowingtextwehaveaddedtoSection7: SeeAppendixGforapossiblemetricofthe
reportedbias,anditsdependenceonsensorynoiseforeachmodelclass(Fig.6).
AppendixBthatderivestheSNRhastypos. InthesentencerightafterEq. (16),thestatementis E[N] = (λhigh − λlow)/(cid:112)(h),as
Tapproachesinfinity. Shouldthisbe E[N] = (λhigh − λlow)/h?
Yes,wehavecheckedAppendixBthoroughly,andcorrectedthistypo.
Eq. (17)seemstocontainanerrorthatmultiplies2infrontofhT*exp(-hT).InthesentencerightafterEq. (17),SNR=E[Delta
N]/Var[DeltaN].ShouldthisbeSNR=E[DeltaN]/sqrt(Var[DeltaN])?Intheequationthatfollowsthissentence,hisnotmultiplied
inthedenominator.
You'reright -- wehavenowcorrectedthesetypos,andalsospottedamissingfactorof2inthedenominatorofthefull
SNRexpression. Theequationsshouldallbecorrectnow.
Forconvenience,onecouldstatewhatSisalsointheAppendix.
WehaveaddedEq.(19)inAppendixBasareminderofEq.(4)forthereader.
AppendixFonthedetailsofthesimulationscouldbemoredetailed. Theauthorsstatethatwhenfittingthenonlinearmodel,
MonteCarlossamplingwasused. Pleaseexplainfurtheronthisfittingmethodtotheextentthatthereadercanreplicate.
We have added the following passage to Appendix F: More specifically, the distribution of the decision variable at
decision time for a given clicks stimulus, p(yT T), was estimated by simulating 800 independent trajectories. Thus,
each trajectory had its own independent realization of sensory noise but the realization of the stimulus (timing of
the clicks) was frozen. Once the density of yT was estimated, the likelihood term, P(dk Tk , θ) in Eq. (27), could be
estimated. Moredetailsonthismethod,suchashowthenumberof800particleswaschosenandhowthismethodwas
validatedonthelinearmodelforwhichtheanalyticalsolutionisavailable,maybefoundinsection3.5.5ofRadillo(2018).
4
Althoughtheconfidenceintervalsdonotseemtocontainthetrueparametervalue,arethemismatchesstatisticallysignificantin
Figure4CD?Thatis,aretheparametervaluessignificantlygreaterthanthetrueparametervalue?
WewouldliketonotethatthewhiskersofeachboxinFig.4C,Ddonotrepresentconfidenceintervals. Instead,they
represent the interval of values that are not considered outliers. More specifically, if q1, q3 are the first and third
quartilesrespectively,thenthewhiskersdefinetheinterval: [q1 − 1.5 × (q3 − q1), q3 + 1.5 × (q3 − q1)].
Instead of testing the hypothesis that the MAP estimates are different than the θtrue value, we provide a summary
statisticforthetrainingdatasetsofsize500trials. Wehaveaddedthefollowingsentencestotheendofsection6: For
referencedatasetsofsize500,98%ofthe500MAPestimatesintheL-NLfitsliestrictlyabove γtrue,versus50.4%for
thecorrespondingL-Lfits. Similarly,86.6%oftheestimatesintheNL-Lfitsliestrictlybelow htrue,versus44.2%forthe
correspondingNL-NLfits. Webelievethisismoreinformativethanap-value.
PublicationDecision2fromNeurons,Behavior,Dataanalysis,andTheoryonAugust26,2019
Editorialboard'sdetermination: Accept
Commentsfromtheeditor. Thanksalotforyourpatiencewiththefirstroundofreviews. Ihavenowcheckedyour
revisionandhavenofurthercomments,whichmeansthemanuscriptwillbetransferredtoa"provisionalacceptance"
stage.
5
|
1508.04241 | 1 | 1508 | 2015-08-18T08:38:24 | Beta Distribution of Human MTL Neuron Sparsity: A Sparse and Skewed Code | [
"q-bio.NC"
] | Single unit recordings in the human medial temporal lobe (MTL) have revealed a population of cells with conceptually based, highly selective activity, indicating the presence of a sparse neural code. Building off previous work by the author and J.C. Collins, this paper develops a statistical model for analyzing this data, based on maximum likelihood analysis. The goal is to infer the underlying distribution of neural response probabilities across the population of MTL cells. The response probability, or neuronal sparsity, is defined as the total probability that the neuron produces an above-threshold firing rate during the presentation of a randomly selected stimulus. Applying the method, it is shown that a beta-distributed neuronal sparsity across the cells of the MTL is consistent with the data. The resulting fits reveal a sparse and highly skewed code, with a huge majority of neurons exhibiting extremely low response probabilities, and a smaller minority possessing considerably higher response probabilities. The distributions are closely approximated by a power law at low sparsity values. Strikingly similar skewed distributions have been found in the statistics of place cell activity in rats, suggesting similar underlying coding dynamics between the human MTL and the rat hippocampus. | q-bio.NC | q-bio |
Beta Distribution of Human MTL Neuron Sparsity: A Sparse and Skewed Code
Physics Department, Pennsylvania State University, University Park, PA 16802, USA
(Dated: 08 Aug 2015)
Andrew Magyar
Single unit recordings in the human medial temporal lobe (MTL) have revealed a population of
cells with conceptually based, highly selective activity, indicating the presence of a sparse neural
code. Building off previous work by the author and J.C. Collins, this paper develops a statistical
model for analyzing this data, based on maximum likelihood analysis. The goal is to infer the
underlying distribution of neural response probabilities across the population of MTL cells. The
response probability, or neuronal sparsity, is defined as the total probability that the neuron produces
an above-threshold firing rate during the presentation of a randomly selected stimulus. Applying
the method, it is shown that a beta-distributed neuronal sparsity across the cells of the MTL
is consistent with the data. The resulting fits reveal a sparse and highly skewed code, with a
huge majority of neurons exhibiting extremely low response probabilities, and a smaller minority
possessing considerably higher response probabilities. The distributions are closely approximated
by a power law at low sparsity values. Strikingly similar skewed distributions have been found in
the statistics of place cell activity in rats, suggesting similar underlying coding dynamics between
the human MTL and the rat hippocampus.
I.
INTRODUCTION
The sparse coding hypothesis states that neural pro-
cessing of sensory information is organized to produce
representations of salient aspects of the environment
(people, objects, landmarks, etc.) using only a small
number of strongly activated neurons [1]. This kind
of representation lies between two theoretical extremes:
dense coding, in which each stimulus is represented in
the activation of a substantial proportion of the avail-
able cells; and local coding, where each object is rep-
resented by the firing of a single neuron [2 -- 4]. Sparse
coding schemes possess many favorable properties includ-
ing a high storage capacity, energy efficiency, and ease of
readability (for references and further discussion, see the
review by Olshausen and Field [5]).
Experimental detection of sparse codes involves identi-
fying numerous cells that are highly selective, responding
to a very small proportion of complex stimuli [6, 7]. Ex-
amples include odor-specific Kenyon cells in locusts [8],
V1 cells in cats [9] and mice [10], and neurons in the
temporal cortex in non-human primates [9, 11 -- 13]. Also,
the RA projecting neurons of the HVC in zebra finches
exhibit extremely sparse activity during song production
[14]. In humans, striking evidence of sparse coding has
been observed in the medial temporal lobe in a series of
experiments [15 -- 18] including the concept cells reported
in Quian Quiroga et al. [17] which were observed to re-
spond to stimuli related to a single concept (e.g.
the
"Jennifer Aniston neuron") out of nearly 100 other con-
cepts presented.
Therefore, characterizing the sparseness of the neu-
ronal representations is an important goal. Towards this
end, one metric used is the total fraction of stimuli that
elicit a response in a particular neuron, that we term the
neuronal sparsity, α [19].
α =
# stimuli triggering a response
total # of possible stimuli
(1)
This definition of sparsity is a property of the neuron it-
self, and is the total probability that the cell responds to
a randomly chosen stimulus. It is not necessarily equal
to the fraction of stimuli that elicit a response during a
particular experiment in which only a small subset of pos-
sible stimuli are presented, as used in Ison et al. [20] for
example. If a particular cell remains unresponsive during
the presentation of 100 randomly selected images, then
its sparsity is not necessarily 0, it may instead have a very
low but non-zero sparsity. In other words, α is a quan-
tity that must be inferred from a particular experiment
rather than directly calculated.
This approach treats neuronal activity as binary
"active-vs-inactive" over a certain post-stimulus time
window. This is appropriate for cells that exhibit highly
elevated firing rates under specific circumstances com-
pared with the baseline rate [19, 21], and few in-between
cases. Place cells, for example, display high firing rates
when the organism is at a particular location in the en-
vironment and much lower firing rates otherwise [22, 23].
The principal goal of this paper is to extend the model
developed by the author and Collins [19] for fitting the
human MTL data presented in Mormann et al.
[24] in
order to estimate the distribution of sparsity across the
population of neurons. In that work, the cells were split
into two discrete populations, each with a characteristic
sparsity value. In this paper, the model is extended to
include continuous sparsity distributions.
Specifically, the neuronal sparsity in the MTL is as-
sumed to follow a beta distribution. Motivation for
choosing the beta distribution comes from its common
usage as a distribution of probabilities, giving it wide
application in Bayesian statistics [25]. The PDF with
parameters a and b is given by
Da,b (α) =
αa−1 (1 − α)b−1
B (a, b)
(2)
where the normalization factor B (a, b) is the beta func-
tion. This produces an overdispersed model for the neu-
ral responses, predicting a heavier tail than if all neurons
had the same sparsity. Such overdispersed, or skewed
models, are suspected to play a prevalent role in many
aspects of neural systems [26, 27], and such distributions
have been found in place cell activity of the rat CA1 [28].
Following the fitting procedure developed by the au-
thor and Collins in [19], we perform inferential statistics,
treating both the neurons and the presented stimuli as
random samples from their respective "universe". This
way, we can infer the response properties of all neurons
in a brain region, rather than merely describing the data,
as is done in previous analyses [20]. The data reported in
Mormann, et al. [24] is fit to the beta model using max-
imum likelihood analysis, yielding the best estimates of
the beta distribution parameters, a and b. Then, the
goodness of fit is assessed using χ2 analysis.
The model is shown to produce acceptable fits in all
four subdivisions of the MTL for which data is available:
the hippocampus (Hipp), the entorhinal cortex (EH),
the amygdala (Amy), and the parahippocampal cortex
(PHC). The estimated beta distributions for the popula-
tion of human MTL neurons indicate:
• The neuronal sparsity is highly non-uniform, with
the least selective 5% of neurons possessing a spar-
sity over 0.01.
• Da,b (α) nearly has a power-law divergence with ex-
ponent ≈ −1 at low sparsity.
• The mean sparsity of MTL neurons is very low, on
the order of 10−3.
This model marks an improvement over the model previ-
ously developed in [19], where the neurons were assumed
to be split into two populations, each characterized by
a sparsity value. This model was able to produce good
fits in the Hipp and EC, but it failed to fit the data from
Amy and PHC. Furthermore, the beta model has only
two fitting parameters, while the two-population model
has three.
Finally, the skewed, highly non-uniform beta-binomial
distributions of responses in the human MTL are com-
pared with strikingly similar results recently observed in
the statistics of place cell activity of the rat hippocam-
pus reported in Rich et al. [28]. Specifically, they observe
that the number of place fields recruited per cell follows
a skewed gamma-poisson distribution. Gamma-poisson
distributions are simply limiting cases of beta-binomial
distributions (shown in Appendix B). This strongly sug-
gests that place cell codes and concept cell codes are two
different manifestations of the same underlying dynam-
ics, in line with previous suspicions [29].
II. BETA-BINOMIAL MODEL
2
neuron is considered responsive to a particular stimulus
if its firing rate exceeds the baseline rate by a statisti-
cal threshold during an appropriate post-stimulus time
frame (see [24] for details).
We define a stimulus as an image of a individual object,
person, building, etc. as used in [24]. The set of stimuli
presented to the patients in [24] constitute a tiny random
sample from the universe of all stimuli. In this section,
we extend our previous procedure to include continuous
distributions of sparsity, D¯θ (α), with fitting parameters
¯θ.
In particular, we postulate the sparseness of a ran-
domly selected neuron from the MTL is sampled from a
beta distribution with pdf given by equation (2).
Examining the data presented in Mormann et al. [24],
we note that there are a large proportion of unresponsive
cells, and that a significant proportion of responsive cells
respond to only one image. Hence, we would expect a
sparsity distribution skewed towards zero sparsity. Note
that Eq. (2) diverges as α → 0 when a < 1, which is what
we would expect for this data. For a ≈ 0, the distribution
behaves like a power law as α → 0, however, at a = 0
the distribution diverges too greatly at zero sparsity to
be normalized.
Let S be the number of stimuli presented to the pa-
tient during an experiment, and let N be the number
of recorded neurons, each with a sparsity sampled from
2. The neurons are assumed statistically independent, as
measured in [17]. For a particular neuron, let K equal
the number of stimuli that were measured to evoke a re-
sponse in that neuron, and let nk for k = 1...S, be the
number of neurons that respond to k out of the S stimuli
presented. Then, we follow earlier work [19] and derive
the likelihood function.
For the data we analyze, we must take into consid-
eration that not all units isolated by the spike sorting
algorithm consist of a single neuron [17, 24]. Limita-
tions in the spike sorting procedure make it so that some
fraction of the recorded units represent the activity of
multiple neurons.
If the activity of a unit is the com-
bined activity of several neurons, then on average that
unit will respond to more stimuli over the course of an
experiment compared with a unit consisting of a single
neuron. Thus, we will carry out the calculation in two
cases: for the first case we assume all units consist of a
single neuron, and for the second case we assume some
fraction of units, p, are comprised of two neurons while
the rest consist of single neurons.
A. Derivation of Beta-Binomial Response
Probability
In this subsection, the relevant results developed by
the author and Collins in [19] are summarized, making
appropriate modifications for the introduction of a con-
tinuous distribution of sparsity.
We define the sparsity, α, of a particular binary neuron
according to Eq. (1). In the experiment we analyze, a
During the presentation of S randomly selected stim-
uli, a single pseudo-binary unit with sparsity, α, responds
to K = k of the stimuli with a conditional probability
given by the binomial distribution ([19]):
P (K = k α) =
αk(1 − α)S−k.
(cid:18)S
(cid:19)
k
If the sparsity, α is sampled from a continuous distri-
bution, D¯θ (α) then, the total probability, k
randomly selected unit responds to k stimuli is given by:
(3)
(cid:0)¯θ(cid:1) that a
3
For a good fit,
χ2 ∼ ( # of data points) − (# of model parameters)
(9)
This procedure only applies for n∗
k larger than a few, so
for our model, we fit only the first five data points (n0,
n1, n2, n3, and n4) in order to include only the bins with
significant responses. Thus, the fit is good when χ2 ∼ 3.
P (K = k) =
dα D¯θ (α) αk(1 − α)S−k
(4)
B. Extension of Model to Multiple-Neuron Units
k
≡ k
0
(cid:19)(cid:90) 1
(cid:18)S
(cid:0)¯θ(cid:1) .
(cid:18)S
Substituting Eq.
integral yields the beta-binomial distribution:
(2) into Eq.
(4) and evaluating the
(cid:19) B (a + k, b + S − k)
k
B (a, b)
k (a, b) =
.
(5)
Eq (5) represents the probability that a single neuron
responds to k out of S presented stimuli. Mixing the bi-
nomial response with a beta distribution over the param-
eter α, produces an overdispersed distribution of neural
responses. This allows the heavy tails present in the data
presented in Mormann et al. [24], displayed in Table (I).
This allows us to fit the outcome of an experiment in
which S stimuli are presented to N neurons, recorded
in parallel. The result of the experiment is the set of
numbers, n0, n1, n2, ..., nS, where nk is the number of
recorded neurons that respond to k of the stimuli. The
expected number of cells per bin, n∗
k assuming the re-
sponses are sampled from Eq. (5), is given by:
k = N k (a, b) ±(cid:112)N k (a, b) (1 − k (a, b)).
n∗
(6)
To find the values of a and b that provide the best n∗
k, we
employ the method of maximum likelihood. This involves
maximizing the likelihood function for the data given the
beta model, L (a, b). The likelihood function is derived
in[19], and the same result applies here, giving:
L (a, b) = W{n0, n1, ..., nS} S(cid:89)
[k (a, b)]nk .
(7)
k=0
where the normalization constant W{n0, n1, ..., nS} is the
multinomial coefficient, i.e. the number of ways of rear-
ranging N objects without changing the nk values. Since
W{n0, n1, ..., nS} is independent of the parameters a and
b, it does not need to be taken into consideration during
maximization. Maximizing L (a, b) gives the parameter
values, a0 and b0 that give the best fit, n∗
k. This was
performed numerically using Mathematica.
The expectation values, n∗
k = N k (a0, b0), are com-
pared with the data, nk, using χ2 analysis to assess the
goodness of fit. The χ2 statistic is defined:
Let (cid:48)
This subsection also follows the procedure developed in
[19] for incorporating the effect of multiple-neuron units.
The experimentalists estimated the fraction of multiple-
neuron units to be p = 0.66 [17]. As in [19], we make
the simplifying assumption that all units consisting of
multiple cells consist of two neurons.
k (a, b) be the total probability that a unit selected
at random responds to k out of S stimuli. We follow [19]
to derive the unit level response, (cid:48)
k (a, b), in terms of the
neuron-level response given by k.
Then,
k (a, b) = (1 − p) k + p(2),k
(cid:48)
(10)
where (2),k is the probability that a randomly chosen
double-neuron unit responds to k stimuli. k is the prob-
ability of a single neuron unit responding to k stimuli
and is given by equation (5).
In order to derive (2),k, we first note that a double-
neuron unit responds to a stimuli if either one of its con-
stituent neurons responds. Let α1, α2 be the sparsities
of the constituent neurons. Then, assuming the two neu-
rons are independent, the unit has an effective sparsity
given by
αunit = 1 − (1 − α1) (1 − α2)
(11)
Similar to the single-neuron unit, the conditional proba-
bility that the double-neuron unit responds to k stimuli
out of S, assuming the unit sparsity is known, is given
by the binomial distribution with probability parameter
αunit.
P (K = k αunit) =
unit(1 − αunit)S−k.
αk
(12)
(cid:18)S
(cid:19)
k
The total probability, (2),k is found by integrating the
conditional probability over the beta distribution, equa-
tion (2) for each constituent neuron:
(cid:90) 1
(cid:90) 1
(2),k =
dα1
dα2
(cid:2)Da,b (α1) Da,b (α2) ×
(cid:18)S
(cid:19)
unit(1 − αunit)S−k(cid:3) (13)
αk
k
χ2(kmax; n1, n2, . . . ) =
k=1
N k
(nk − N k)2
0
0
.
(8)
kmax(cid:88)
(2),k =
1
[B (a, b)]2
(cid:8)(cid:18)k
(cid:19)
j
(cid:18)S
(cid:19) k(cid:88)
× B (a + k − j, b + S − k)(cid:9)
k
× B (a + j, b + S − j)
j=0
Evaluating this integral yields (derivation in appendix):
(14)
The form of the likelihood function is the same as (7)
except that Eq. (10) is used instead of Eq. (5) for the
single unit response probability.
L(cid:48) (a, b) = W{n0, n1, ..., nS} S(cid:89)
[(cid:48)
k (a, b)]nk .
(15)
k=0
Maximizing (15) with respect to a and b provides the best
fit parameters of the model a0 and b0 to the nk data. The
model prediction, n∗
k has the same form as equation (6)
only with (cid:48)
k in place of of k.
III. RESULTS
A. Fits to data
In this section, we use the beta model to analyze data
taken from the human MTL presented in [24]. The data
is recorded from 1194 Hipp units, 844 EC units, 947 Amy
units, and 293 PHC units. The patients were shown on
average 97 images of familiar stimuli, presented in ran-
dom order. The nk for each region are given in Table I.
Maximizing (7) with the respect to the a and b yields the
parameter values given in Table II for the four regions.
The goodness of each fit was assessed using χ2 test. Plots
of the fits against the data are given in figure 1.
All four regions are successfully fit by both the beta
model containing single units and the improved model
containing a mixture of single and double neuron units
(χ2 values given in table II). This suggests that the data
is consistent with the notion that the sparsity of human
MTL neurons follows a beta distribution, given in equa-
tion 2, with a < 1 and b > 1. The sparsity distribution is
far from uniform, consistent with the result in [19] when
the two-population model was used.
Including the double units in the model had little ef-
fect on the goodness of fit, though the value of the a
parameter is brought closer to zero. This results in a
lower mean sparsity compared with assuming only single
neuron units.
Plots of the best-fit sparsity distributions from the
mixture model are given in figure 2. The means of the
distribution in each of the regions are: (cid:104)α(cid:105) = 1.6 × 10−3
in Hipp; 1.5× 10−3 in EC; 1.5× 10−3 in Amy; 4.0× 10−3
in PHC. In each region, a is close to zero, producing a
near-divergence in the neuronal sparsity distribution as
α → 0. This indicates a large population of extremely
sparse neurons, consistent with the results of [19]. For
these fits, roughly 95% of the MTL neurons falls within
4
the power-law regime, as is shown by the linear region of
the log-log plots in figure 2.
The sparsity distribution in the PHC differed quantita-
tively from the fits in the other three regions. Firstly, the
model predicts that neurons of the PHC has the highest
mean sparsity, (cid:104)α(cid:105) = 4 × 10−3, compared to the other
three regions which have mean sparsities clustered near
1.5 × 10−3. Also, the tail of the sparsity distribution ex-
tends to larger k for the PHC than it does for the other
three regions, as shown in figure 2. This is consistent with
the observation that the selectivities of neurons tend to
increase as information proceeds down the ventral visual
pathway from PHC to the EC, Hipp and Amy [24].
B. A Model for Representational Learning by
Preferential Attachment
One advantage of using statistical inference to estimate
the underlying distribution of sparsity, D (α), is that the
form of the distribution can suggest a particular gener-
ating mechanism. Beta distributions with near power
law behavior arise as limiting distributions in numerous
"rich-get-richer" schemes. For example, both preferential
attachment processes on growing networks, where nodes
with a high degree are more likely to receive edges from
newly added nodes [30, 31], and Polya urn schemes [32],
where the proportion of balls of a particular color grows
whenever a ball of that color is sampled from the urn,
both yield beta distributions as t → ∞.
[33].
In this section, we consider the bi-layered network
In their model,
model studied in Peruani et al.
the bottom layer of the network contains N nodes which
remain fixed in number, while the top layer consists of t
nodes that are added one at a time, starting from t = 0.
As nodes are added, they connect with nodes in the bot-
tom layer with a higher probability of attaching to nodes
with a large degree.
We interpret the fixed bottom layer as the binary neu-
rons of the MTL responsible for object encoding, and we
interpret the top layer as stimuli related to familiar con-
cepts which have been previously encoded into memory.
The addition of a node in the top layer indicates a new
concept that is to be coded into memory, e.g. an unfamil-
iar person to whom you have just been introduced. An
edge between stimulus i in the top layer and neuron j in
the bottom layer indicates that neuron j is part of the
code for i (i.e. neuron j activates whenever stimulus i is
presented to the organism). The attachment procedure
for new nodes represents the complex neurological learn-
ing processes by which new concepts are coded onto the
neural substrate of the human MTL. Thus, the growth of
the top layer with the attachment of edges to the bottom
layer is a model for representational learning of a binary
code. The sparsity of neuron j, αj, is given by
αj =
kt
j
t
(16)
5
FIG. 1: Best fits to data recorded in four regions of the human MTL assuming all recorded units consist of a single neuron.
The red circles indicate the experimental values from Table I and blue dots connected by lines indicate the best-fit predictions
for the expectation values of nk with double-neuron units included predicted by the beta-binomial model. The dotted lines
indicate best fits for a pure binomial model i.e. assuming all cells have the same sparsity. Note the log scale on the y-axis.
FIG. 2: Sparsity distributions across neurons in each region predicted by the multi-unit model consisting of single units and
double units. The plots are shown using log-log axes to indicate power law behavior at low sparsity. The shaded region indicates
the upper 5% tail of the sparsity distribution, i.e. the model predicts 95% of the neurons in each region have a sparsity left of
the shaded area. The verticle dotted lines indicate the mean sparsities in each region.
(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)02468101101001000numberofresponsesnumberofneuronsHippdatabeta(cid:45)binomialbinomial(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)02468101101001000numberofresponsesnumberofneuronsECdatabeta(cid:45)binomialbinomial(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)02468101101001000numberofresponsesnumberofneuronsAmydatabeta(cid:45)binomialbinomial(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)(cid:45)0246810110100numberofresponsesnumberofneuronsPHCdatabeta(cid:45)binomialbinomial10(cid:45)40.0010.010.10.010.11101001000104ΑDΑHipp10(cid:45)40.0010.010.10.010.11101001000104ΑDΑEC10(cid:45)40.0010.010.10.010.11101001000104ΑDΑAmy10(cid:45)40.0010.010.10.010.11101001000104ΑDΑPHC n0 n1 n2 n3 n4 n5 n6 n7 n8 n9 n10 n11 n12 n13 n14
0
0
2
0
Hipp 1019 113 30 17 7 4 1 2 0 0
EC 761 45 15 9 4 8 0 0 1 0
Amy 842 61 17 15 3 3 1 0 1 1
PHC 244 13 11 7 3 0 4 1 2 4
0
1
1
0
0
0
0
3
1
0
0
1
0
0
0
0
TABLE I: Number of units nk responding to k images as reported by [24] in four MTL regions.
Single-Unit Model
Multi-Unit Model
6
Hipp EC Amy PHC
0.17 0.08 0.09 0.08
12
66
χ2(5) 6.1
3.3
36
1.9
34
8.6
a
b
Hipp EC Amy PHC
0.11 0.05 0.05 0.05
13
67
2.7
34
χ2(5) 2.1 0.56 5.2
a
b
36
TABLE II: Best fit parameters of the beta distribution a and b for the four MTL regions. Left-hand table gives values assuming
all recorded units consist of a single neuron. Right-hand table gives values assuming a mixture of single and double neuron
units. χ2 values evaluated for k = 1 to k = 10.
where kt
have been added in the top layer.
j is the node degree of neuron j after t stimuli
j
j
(17)
1
C (t)
(cid:1) =
j is defined by
A(cid:0)kt
to neuron j with node degree kt
Following Peruani, et al. [33], the network is grown by
adding a node to the top layer at each time step, t, and
then attaching it to µ nodes in the bottom layer. In our
model, µ is the code word length of each stimulus and is
assumed to be fixed. We assume µ << N . As each of the
where γ is a parameter that determines the influence of
node degree on the attachment process and C (t) is a nor-
malization constant. For γ = 0, the edges are attached
In the case where µ << N , the probability that neuron j
connects to the stimulus added at time t after all µ edges
have been connected is approximately
µ edges is added, the probability A(cid:0)kt
(cid:1) that it attaches
(cid:0)γkt
j + 1(cid:1)
at random to the neurons with A(cid:0)kt
(cid:1) = 1N for each j.
j + 1(cid:1)
µ(cid:0)γkt
(cid:1) = µN .
the sparsity distribution, D(cid:0) k
(cid:1) of neurons in the bottom
(cid:19)r−1(cid:18)
(cid:18) k
The network is grown starting from t = 0, with no
stimuli in the top layer and all node degrees in the bottom
layer equal zero. Equation (18) defines the attachment
process as each stimulus is added. We are interested in
layer after a large number of stimuli have been learned.
Peruani et al. [33] showed that for γ > 0, in the limit of
large t,
When γ = 0, a newly added stimulus attaches to neuron
j with probability Aµ
(cid:1) =
(cid:0)kt
(cid:19)s−1
(cid:0)kt
j
(cid:18) k
(cid:19)
C (t)
(18)
Aµ
1
j
j
t
D
=
1
B (r, s)
t
γ and s = N
γµ − 1
γ .
t
where r = 1
1 − k
t
(19)
distribution D(cid:0) k
(cid:1) → δ(cid:0) k
t
t − µN
(cid:1) as t → ∞, i.e. the neu-
In the case of purely random attachment ( γ = 0), the
ronal sparsity is the same for all neurons. Purely random
attachment is inconsistent with the data fit above, as can
be seen from the dotted lines in 1.
If we treat the data analyzed above as a random sample
of N neurons from the bottom layer and S stimuli from
the top layer, then we can match the beta parameters r
and s with the fits reported above. This gives an estimate
of roughly 10 ≤ γ ≤ 20 and 1
250 . These
values for γ suggest from equation 18 that preferential
attachment plays a large role in assigning new stimuli
to MTL neurons. In other words, it is indirect evidence
that neurons are not randomly assigned to new concepts,
but rather new concepts are more likely to be coded onto
neurons that have previously been assigned to previously
learned concepts.
600 ≤ µN ≤ 1
IV. DISCUSSION
A. Relation to Previous Work
This paper builds upon previous work [19, 34] within
which neurons were assumed to be split into two discrete
populations: a sparse population comprising roughly
5% − 10% of the MTL neurons, with a sparsity on the
order of 10−2; and an ultra-sparse population compris-
ing the remaining 90% − 95%, with a sparsity on the
order of 10−3. There, the neurons in each population
were assumed to have the same sparsity value. This two-
population model produced good fits in the Hipp and the
EC, but produced poorer fits in the Amy and the PHC.
The beta model developed in this paper is a continuous
distribution, and it fits all four regions adequately, in-
cluding the Hipp and the EC.
Thus, there are two radically different sparsity distri-
butions that are consistent with the single-unit responses
from the Hipp and EC. In order to produce good fits to
the data, which shows very large n0 and n1 bins com-
pared with the bins at higher k, the sparsity distribu-
tions at small α are largely determined by these two bins.
There is insufficient statistical power to distinguish be-
tween different sparsity distributions in this low sparsity
regime.
To illustrate the connection between the beta model
and the two-pop model, the shaded region seen in figure
2, roughly corresponds to the sparse population (pop-
ulation labeled "D" in [19]), while the neurons in the
unshaded region are analogous to the ultra-sparse pop-
ulation (labeled "US" in [19]).
In the two-population
model, the distribution would have a Dirac-delta func-
tion located in the shaded region and another located in
the unshaded region, representing the two populations.
Consequently, the exact form of the beta distribution
should not be taken too literally, as there are likely to be
other continuous distributions that are consistent with
the data. However, one would still expect similar overall
behaviors regardless of which distribution is chosen. The
advantage of the beta distribution is that it yields com-
pact analytical results in the likelihood analysis for both
the single unit model and the multi-unit model.
B. Similarity to Place Cell Statistics
Recent experiments have revealed that place cells in
the rat hippocampus display skewed activity, in which
the number of place fields recruited by place cells ex-
hibits a heavier tail than what would be predicted if the
cells all recruited place fields at the same rate [28]. In
the experiment performed by Rich et al.
[28], the re-
cruitment of place fields among the cells of CA1 obeys
a skewed gamma-poisson process, in which each cell ac-
quires place fields according to a poisson process, but the
rate parameter for each cell is sampled from a gamma dis-
tribution. This results in a poisson distribution of place
fields per cell mixed by a gamma distribution.
In this
paper, the number of concpet fields of MTL neurons are
shown above to be be distributed by a binomial distribu-
tion mixed with a beta distribution. The gamma-poisson
distribution and beta-binomial distributions are closely
related, with the gamma-poisson being a limiting case of
the beta-binomial.
Place cells of the rat and concept cells in humans share
many characteristics [29], and observing nearly identicle
distributions of activity in both populations suggests that
the place cell code and the concept cell code likely arise
from similar underlying processes.
In other words, the
results suggest that the recruitment of place fields by the
place cells of the rat hippocampus and the recruitment of
"concept fields" by the concept cells of the human MTL
are two manifestations of the same neural mechanism,
despite coding for spatial information vs conceptual in-
formation.
7
Furthermore, finding nearly identicle distributions in
different species across cells with strikingly different re-
ceptive fields lends more credence to the idea that skewed
distributions in general play a fundamental role in neural
functioning [26, 27].
One possible mechanism for generating skewed distri-
butions, especially distributions that have that have ap-
proximate power-law behavior, are the various cumula-
tive advantage or "rich get richer" schemes such as pref-
erential attachment processes on growing networks ex-
plored above. Another possibility, briefly explored in [28]
is that the non-uniform recruitment of receptive fields
arises from intrinsic cell differences, such as non-uniform
excitability and pre-synaptic inputs. I suggest that these
two hypotheses are not mutually exclusive, and that both
may play a role in generating the observed sparse, skewed
distributions.
C.
Impact of Silent Cells
The spike sorting techniques for single unit-recordings
only detect a small fraction of the neurons within range
of the electrode. The remaining neurons, constituting
perhaps as much as 90% of the overall cells, remain com-
pletely silent during the experiment and thus are missed
completely by the spike sorting algorithm [35 -- 38]. This
means that the sample of neurons is biased in favor of
more active cells.
To clarify, these silent cells, or "neural dark matter,"
are distinct from the n0 cells reported in [24] which did
not produce above threshold firing rates for any of the
presented stimuli. The population of n0 cells still emit-
ted spikes and thus were detected by the spike sorting
techniques used. The silent cells on the other hand, emit-
ted no spikes, and thus could not be detected. In effect,
these cells should be included in the n0 bin to give a more
biased estimate of the sparsity distribution. The result
on the fits would be to lower the value of a, bringing it
even closer to zero.
Some studies estimate the silent cell population to be
as high as a factor of ten [35]. That is, there are perhaps
ten silent cells for each recorded cell. Using this factor of
cells added to the n0 bin, maximum likelihood analysis
yields a = 0.007 and b = 55 in the hippocampus, with
χ2 = 1.2. This brings the sparsity distribution much
closer to a power law, indicating a considerably sparser
code.
Acknowledgments
I would like to thank John Collins and Mike Skocik
for valuable commentary and discussions regarding the
content of this manuscript.
Appendix A: Derivation of multi-unit response
probability
In this section we derive equation (14) starting from
the integral given in equation (13):
(cid:90) 1
(cid:90) 1
(2),k =
dα1
dα2
0
0
(cid:2)Da,b (α1) Da,b (α2) ×
(cid:19)
(cid:18)S
unit(1 − αunit)S−k(cid:3) (A1)
αk
k
where Da,b (α) is the beta distribution PDF given in
equation (2), and αunit is the effective sparsity for the
double unit given by equation (11).
To evaluate the integral, we first note that:
(1 − αunit)S−k = [(1 − α1) (1 − α2)]S−k
(A2)
Then, we expand the factor αk
nomial theorem:
unit in (A1) using the bi-
unit = (α1 + α2 (1 − α1))k
αk
(cid:18)k
(cid:19)
k(cid:88)
j
j=0
=
1[α2 (1 − α1)]k−j
αj
(A3)
Using these results and equation (2), we can write equa-
tion (A1) as a sum of separable integrals
(cid:18)S
(cid:19) k(cid:88)
(cid:8)(cid:18)k
(cid:19)
k
j=0
j
(2),k =
[B (a, b)]2
1
(cid:90) 1
(cid:90) 1
0
0
×
×
(2),k =
1
[B (a, b)]2
dα1 αa−1+j
1
dα2 αa−1+k−j
2
(1 − α1)b−1+S−j
(1 − α2)b−1+S−k(cid:9)
(cid:19)
(cid:8)(cid:18)k
(cid:18)S
(cid:19) k(cid:88)
× B (a + k − j, b + S − k)(cid:9) (A5)
k
× B (a + j, b + S − j)
j=0
j
The integrals can now be evaluated as beta functions
yielding the result given by equation (14).
Appendix B: Gamma-poisson distribution as a
limiting case of the beta-binomial distribution
8
(B1)
(B2)
0 < α < 1, where the probability parameter α is sam-
pled from a beta distribution with parameters a > 0 and
b > 0. The PMF for the binomial distribution and the
PDF for the beta distributions are given respectively:
(cid:18)S
(cid:19)
k
P (K = k) =
αk (1 − α)S−k
and
respectively.
D (α) =
αa−1 (1 − α)b−1
B (a, b)
(cid:90) 1
(cid:0)S
(cid:1)
(cid:19) B (a + k, b + S − k)
(cid:18)S
B (a, b)
k
0
k
B (a, b)
P (K = k) =
=
Thus, the PMF of the beta-binomial distribution with
parameters S, a, and b is given by:
dα αk (1 − α)S−k αa−1 (1 − α)b−1
.
(B3)
The gamma-poisson distribution, more commonly called
the negative binomial distribution, is a Poisson distribu-
tion with parameter λ > 0 where λ is sampled from a
gamma distribution with parameters r > 0 and β > 0.
The PMF of the Poisson distribution and the PDF of the
gamma distribution is given:
P (K = k) =
λke−λ
k!
g (λ) =
rβ
Γ(β)
λβ−1e−rλ.
(B4)
(B5)
and
(A4)
and the gamma-poisson distribution with parameters r
and β is then given by
P (K = k) =
=
=
=
1
k!
1
k!
1
k!
1
k!
rβ
Γ(β)
rβ
Γ(β)
rβ
Γ(β)
rβ
Γ(β)
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
0
0
dλ (cid:2)λke−λ(cid:3)(cid:2)λβ−1e−rλ(cid:3)
(cid:19)k+β−1
dλ λk+β−1e−λ(1+r)
(cid:18) x
dx
r + 1
r + 1
0
Γ(k + β)
(r + 1)k+β
e−x
(B6)
In this appendix, I show that the gamma-poisson dis-
tribution is a limiting case of the beta-binomial distri-
bution.
I show it here because a convenient reference
showing this result could not be located.
The beta-binomial distribution is a mixture distribu-
tion of a binomial distribution, with parameters S and
To show that B3 yields B6 as a limiting case, we begin by
observing that the Poisson distribution is a limiting case
of the binomial distribution as S → ∞ while holding the
mean, (cid:104)k(cid:105) = Sα constant. The Poisson parameter λ is
the expected response, i.e. λ = (cid:104)k(cid:105).
Similarly, the gamma-Poisson distribution is the lim-
iting case of the beta-binomial distribution as S → ∞
while ensuring the mean (cid:104)k(cid:105) = S(cid:104)α(cid:105) for the beta-binomial
distribution is finite. From equation (B2), we see that
(cid:90) 1
αa (1 − α)b−1
(cid:104)α(cid:105) =
=
=
1
B(a, b)
0
B(a + 1, b)
B(a, b)
a
a + b
(B7)
(B8)
So,
(cid:104)k(cid:105) = S(cid:104)α(cid:105) =
Sa
a + b
.
One way to ensure (cid:104)k(cid:105) is finite as S → ∞, is for the
parameter b to approach infinity as a linear function of
S by setting b = rS. The constant r > 0 is arbitrary. It
will be shown that that r matches the parameter of the
gamma distribution above when the limit is taken.
Before taking the limit by setting b = rS in equation
(B3), it is useful to write the beta functions and bino-
mial coefficient in terms of gamma functions using the
identities
(cid:18)S
(cid:19)
k
and
=
Γ(S + 1)
Γ(k + 1)Γ(S − k + 1)
(B9)
Using these identities, we get write equation (B3) as
9
P (K = k) =
Γ(S + 1)
Γ(k + 1)Γ(S − k + 1)
× Γ(a + k)Γ(S − k + b)
Γ(a + b + S)
× Γ(a + b)
Γ(a)Γ(b)
Γ(S + 1)
=
Γ(k + 1)Γ(S − k + 1)
× Γ(a + k)Γ(S(1 + r) − k)
Γ(a + S(1 + r))
× Γ(a + rS)
Γ(a)Γ(rS)
(B11)
Now, to take the limit as S → ∞, we use Stirling's ap-
proximation applied to ratios of gamma functions:
Γ(x + γ)
Γ(x + ω)
≈ xγ−ω
(B12)
for large x. We get:
P (K = k) =
Γ(k + 1)
Sk
[S(1 + r)]a+k × (rS)a
× Γ(a + k)
(1 + r)a+k × ra
× Γ(a + k)
Γ(a)
Γ(a)
=
1
k!
(B13)
B (x, y) =
Γ(x)Γ(y)
Γ(x + y)
(B10)
Comparing equations (B13) and (B6), we see that they
match, and that a = β. This concludes the derivation.
[1] H. B. Barlow, Perception pp. 795 -- 8 (1972).
[2] P. Foldi´ak and D. Endres, Scholarpedia 3(1), 2984
[13] W. E. Vinje and J. L. Gallant, Science 287, 1273 (2000).
[14] R. Hahnloser, A. Kozhevnikov, and M. Fee, Nature 419,
(2008).
65 (2002).
[3] P. Foldi´ak and M. Young, in Handbook of brain theory
and neural networks, edited by M. Arbib (MIT Press,
Cambridge, MA, 2002), pp. 1064 -- 1068.
[4] J. Bowers, Psychol. Rev. 116, 220 (2009).
[5] B. A. Olshausen and D. J. Field, Current opinion in neu-
robiology 14, 481 (2004).
[15] I. Fried, K. A. MacDonald, and C. L. Wilson, Neuron
18, 753 (1997).
[16] G. Kreiman, C. Koch, and I. Fried, Nature neuroscience
3, 946 (2000).
[17] R. Quian Quiroga, L. Reddy, G. Krieman, C. Koch, and
I. Fried, Nature 435, 1102 (2005).
[6] B. Willmore and D. J. Tolhurst, Network: Computation
[18] R. Q. Quiroga, A. Kraskov, C. Koch, and I. Fried, Cur-
in Neural Systems 12, 255 (2001).
rent Biology 19, 1308 (2009).
[7] B. D. Willmore, J. A. Mazer, and J. L. Gallant, Journal
[19] A. Magyar and J. Collins, Phys. Rev. E. 92, 012712
of neurophysiology 105, 2907 (2011).
(2015).
[8] J. Perez-Orive, O. Mazor, G. C. Turner, S. Cassenaer,
R. I. Wilson, and G. Laurent, Science 297, 359 (2002).
[9] R. Baddeley, L. F. Abbott, M. C. Booth, F. Sengpiel,
T. Freeman, E. A. Wakeman, and E. T. Rolls, Proceed-
ings of the Royal Society of London B: Biological Sciences
264, 1775 (1997).
[10] E. Froudarakis, P. Berens, A. S. Ecker, R. J. Cotton,
F. H. Sinz, D. Yatsenko, P. Saggau, M. Bethge, and A. S.
Tolias, Nature neuroscience 17, 851 (2014).
[11] E. T. Rolls and M. J. Tovee, Journal of Neurophysiology
73, 713 (1995).
[12] N. C. Rust and J. J. DiCarlo, The Journal of Neuro-
science 32, 10170 (2012).
[20] M. J. Ison, F. Mormann, M. Cerf, C. Koch, I. Fried, and
R. Quian Quiroga, Journal of Neurophysiology 106, 1713
(2011).
[21] R. Quian Quiroga, L. Reddy, C. Koch, and I. Fried, Jour-
nal of neurophysiology 98, 1997 (2007).
[22] J. O'Keefe and J. Dostrovsky, Brain research 34, 171
(1971).
[23] M. A. Wilson and B. L. McNaughton, Science 265, 676
(1994).
[24] F. Mormann, S. Kornblith, R. Q. Quiroga, A. Kraskov,
M. Cerf, I. Fried, and C. Koch, The Journal of Neuro-
science 28, 8865 (2008).
[25] D. J. MacKay, Information theory, inference and learning
10
algorithms (Cambridge university press, 2003).
[34] J. Collins and D. Z. Jin, arXiv preprint q-bio/0603014
[26] G. Buzs´aki and K. Mizuseki, Nature Reviews Neuro-
(2006).
science 15, 264 (2014).
[27] G. Buzs´aki et al., science 347, 612 (2015).
[28] P. D. Rich, H.-P. Liaw, and A. K. Lee, Science 345, 814
(2014).
[29] R. Quian Quiroga, Nat. Rev. Neurosci. 13, 587 (2012).
[30] A.-L. Barab´asi and R. Albert, science 286, 509 (1999).
[31] M. E. Newman, Contemporary physics 46, 323 (2005).
[32] H. Mahmoud, P´olya urn models (CRC press, 2008).
[33] F. Peruani, M. Choudhury, A. Mukherjee, and N. Gan-
guly, EPL (Europhysics Letters) 79, 28001 (2007).
[35] S. Waydo, A. Kraskov, R. Q. Quiroga, I. Fried, and
C. Koch, The Journal of Neuroscience 26, 10232 (2006).
[36] D. A. Henze, Z. Borhegyi, J. Csicsvari, A. Mamiya, K. D.
Harris, and G. Buzs´aki, Journal of neurophysiology 84,
390 (2000).
[37] L. Thompson and P. Best, The Journal of neuroscience
9, 2382 (1989).
[38] B. Olshausen and D. Field, pp. 1665 -- 1699 (2005).
|
1507.00125 | 1 | 1507 | 2015-07-01T06:59:14 | Optimal prediction and natural scene statistics in the retina | [
"q-bio.NC"
] | Almost all neural computations involve making predictions. Whether an organism is trying to catch prey, avoid predators, or simply move through a complex environment, the data it collects through its senses can guide its actions only to the extent that it can extract from these data information about the future state of the world. An essential aspect of the problem in all these forms is that not all features of the past carry predictive power. Since there are costs associated with representing and transmitting information, a natural hypothesis is that sensory systems have developed coding strategies that are optimized to minimize these costs, keeping only a limited number of bits of information about the past and ensuring that these bits are maximally informative about the future. Another important feature of the prediction problem is that the physics of the world is diverse enough to contain a wide range of possible statistical ensembles, yet not all motion is probable. Thus, the brain might not be a generalized predictive machine; it might have evolved to specifically solve the prediction problems most common in the natural environment. This paper reviews recent results on predictive coding and optimal predictive information in the retina and suggests approaches for quantifying prediction in response to natural motion. | q-bio.NC | q-bio | J Stat Phys manuscript No.
(will be inserted by the editor)
Jared Salisbury1,2 · Stephanie E. Palmer2
Optimal prediction and natural scene
statistics in the retina
5
1
0
2
l
u
J
1
]
.
C
N
o
i
b
-
q
[
1
v
5
2
1
0
0
.
7
0
5
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract Almost all neural computations involve making predictions. Whether
an organism is trying to catch prey, avoid predators, or simply move through a
complex environment, the data it collects through its senses can guide its actions
only to the extent that it can extract from these data information about the future
state of the world. An essential aspect of the problem in all these forms is that
not all features of the past carry predictive power. Since there are costs associ-
ated with representing and transmitting information, a natural hypothesis is that
sensory systems have developed coding strategies that are optimized to minimize
these costs, keeping only a limited number of bits of information about the past
and ensuring that these bits are maximally informative about the future. Another
important feature of the prediction problem is that the physics of the world is di-
verse enough to contain a wide range of possible statistical ensembles, yet not all
motion is probable. Thus, the brain might not be a generalized predictive machine;
it might have evolved to specifically solve the prediction problems most common
in the natural environment. This paper reviews recent results on predictive cod-
ing and optimal predictive information in the retina and suggests approaches for
quantifying prediction in response to natural motion.
Keywords Prediction · Neural computation · Natural statistics · Retina · Motion
processing
Support for this work was provided by: The Alfred P. Sloan Foundation, a FACCTS grant from
the France Chicago Center, and a Chateaubriand Fellowship to JS.
1 Graduate Program in Computational Neuroscience
2 Department of Organismal Biology and Anatomy
University of Chicago
1027 E 57th St., Chicago, IL 60637, USA
E-mail: [email protected]
2
1 Introduction
What are the predictable components of the input to an animal’s visual system in
its natural environment? While the characteristics of static images have been ex-
plored in large image repositories [1,21,31,36,6,23,43], some, but not as much
is known or measured in the temporal domain [21,12]. One interesting feature
of scaling in static images is the power-law distribution of spatial variation in
local contrast [21,36]. This scaling implies that natural images are scale-free, dis-
playing the same basic structure on all length scales. Power-law behavior in the
frequency distribution of temporal fluctuations in total scene luminance have also
been observed in a variety of natural contexts, and scenes display slightly different
exponents depending on their specific content [12]. This paper will review recent
attempts to connect natural motion statistics to efficient prediction in the visual
system, focusing on the retina. Tying temporal statistics of natural scenes to neu-
ral prediction will reveal what types of motion the brain can efficiently represent
and therefore constrain the types of predictions the brain can perform.
The concept of efficient coding for prediction in the brain has been developed
in two main ways: via theories of predictive coding [41,42,35,26,29,5] that elim-
inates redundancy in the temporal response of the brain, and through analytical
work to characterize the optimal trade-offs between representing the past and fu-
ture sensory input [10,11] (via information bottleneck calculations[44,16,18]). In
this paper, we will review and relate these two approaches to neural coding in
the retina, and propose methods for extending this work to the context of natural
motion statistics.
It has been shown that retinal ganglion cells (RGCs), the output cells of the
retina whose axons form the optic nerve, display a whole host of nonlinear pro-
cessing characteristics that may be connected to prediction. RGCs respond differ-
entially to object versus background motion [32]. Ganglion cells have also been
shown to code for a variety of motion features in ways that cannot be accounted
for by a simple receptive field picture of encoding. This includes: motion antici-
pation, the coding in the retina for the anticipated position of an object moving at
constant velocity [8]; the omitted-stimulus response, in which ganglion cells fire
after the cessation of a sequence of visual flashes at the appropriate delay where
the next flash in the sequence would be expected [38]; and reversal responses,
where neurons in the retina fire a synchronous burst of activity after the reversal
of a moving bar, irrespective of their relative receptive field positions [39]. All of
these adaptive motion-processing features speak to the retina’s complexity as an
encoding device, and have a component of predictability of the future state of vi-
sual stimuli. Recent work has also shown that the retina solves a general prediction
problem in a near-optimal way[33].
We will review this background material and discuss how extensions of this
work could reveal how an organism’s ecological niche shapes its predictive pro-
cessing. In particular, it may be that the retinas of different species possess the
capacity to solve different suites of motion prediction problems tailored to their
natural environment. Evolution may shape which problems are hardwired in the
early visual system, and exploring that can uncover just how far the retina is able
to tune its predictive power to the statistics of its inputs.
3
2 Theories of optimal prediction
Sitting at the front end of the visual system and with a limited number of fibers to
transmit all the visual information the brain receives, the retina has long been hy-
pothesized to be an efficient and perhaps even optimal encoder of the visual world
[3,4,1,26,20]. This notion of efficiency dictates that the retina’s representation
of the visual inputs to the photoreceptor layer should be as loss-less as possible,
given the number of cables the retina has along which to transmit information to
the brain and the fact that metabolic constraints limit the firing rate of neurons.
To make the best use of each fiber, these signals should be independent in space
and in time. Recent work from a variety of researchers expand on this simple no-
tion of efficiency. Natural inputs to the retina are non-Gaussian[21,36], the noise
spectrum in neural data is not white, retinal firing is certainly highly redundant
[34], and not all information about the input is equally relevant to the organism.
The concept of optimizing the predictive capacity of the retina assigns value to
particular bits of information: it says that compression is only successful when
the transmitted bits convey information about the future input [11]. The informa-
tion bottleneck method [44] is a way of defining relevant information, in this case
information about the future, as the distortion measure.
2.1 Efficient coding in the time domain: Predictive coding
The efficient coding hypothesis states that all of the information about the input
should be retained, while minimizing the entropy of the response [40]. If not all
bits of information about the input are retained, the problem can be formulated
using rate distortion theory [7],
min
p(rtst )
I(Rt;St) + β D(Rt,St),
(1)
where D is the average distortion, R is the neural response to the stimulus S, and
the minimization finds the lowest transmitted bit rate, given D.
The core concept in predictive coding, in the time domain, is that temporal
correlations in the output stream should be eliminated, so that only deviations
from expected response, or those that are ‘surprising’ are encoded [41]. If the
input statistics are stationary, predictive coding aims to minimize the response of
the system. The role of neurons in a predictive coding paradigm is to code for
changes in response statistics, not the ongoing predictable events in a stationary
world.
Predictive coding has been postulated to be achieved through feedback con-
nections from higher areas onto sensory input streams [35,29,45,5], and early
[41] as well as recent work in the retina hypothesizes feedforward adaptive mech-
anisms at the sensory periphery may result in predictive coding [28]. As such,
predictive coding is highly efficient, because redundancy in time is eliminated.
Mechanisms have been proposed by which the retina could implement predic-
tive coding, via inhibitory interactions at bipolar terminals [26]. Also, the work of
Den`eve shows how predictive coding may be self-organized in neural networks
[13].
4
2.2 Predictive information
Information theoretic treatments of prediction in the brain focus on defining not
just the code that retains the most stimulus information for a given output bit rate,
but the one that retains the most information about the future stimulus. This ad-
dition of the notion of relevant information has sharpened discussions of early
sensory processing in the context of prediction [11]. The theory for retaining the
optimal amount of predictive information has been well-developed by Tishby and
colleagues [44,16,18], and leads not only to elegant but also testable results. Re-
cent experimental and theoretical work draws on these results and has shown that
the salamander retina may be optimized for prediction [33].
The efficient representation of predictive information that we see in [33] adds
the notion of relevant information to the classical ideas of efficient coding. The
simplest version of the efficient coding hypothesis is that the retina processes vi-
sual inputs to remove redundancy, allowing the array of retinal ganglion cells to
make fuller use of their limited capacity to transmit information [4,2,1]. The re-
sults in [33] suggest that the retina is not designed to represent all of the input light
patterns impinging on its photoreceptors, but instead to represent those parts of the
input that are most predictive of the future. The retina clearly throws away some
aspects of the input light patterns, but perhaps only those parts that are irrelevant
for the task of prediction.
Information bottleneck approaches The maximal amount of predictive informa-
tion a system can possibly encode can be found by solving the following informa-
tion bottleneck problem:
I(Rt;St,t−∆t,t−2∆t,...)− β I(Rt;St+∆t).
min
p(rtst )
(2)
This can also be understood as a rate distortion problem where the distortion met-
ric is the predictive information. Here we can see how predictive information and
predictive coding are really solving complementary optimization problems for sta-
tionary stimuli.
Providing an efficient representation of predictive information is nearly oppo-
site of what one would expect from neurons doing predictive coding. In that type
of code, signals are decorrelated in time so that predictable components are elim-
inated and neurons encode only the deviations from expectation, or surprise. The
responses of neurons implementing a truly optimal predictive code in a stationary
input environment would thus carry no predictive information about their own re-
sponses. In contrast, recent results [33] suggest that the retina has a large amount
of response predictive information, and that these responses efficiently separate
predictive from non–predictive bits, and transmit the predictive bits preferentially.
Predictive coding does not explicitly preserve information about the future
stimulus. As such, it is hard to compare to optimal predictive information schemes.
The prescription for predictive coding is if S is the input, and R∗ is the expected re-
sponse and R is the observed response: when p(rs) is much different from p(r∗s),
respond. Somewhere in the brain must live the model(s) that generates r∗. Higher
cortical areas that inhibit early sensory areas, such as the lateral geniculate nu-
cleus, might provide precisely this type of feedback, and have been implicated in
5
experimental evidence for predictive coding, se e.g.[35]. In the retina, with little
to no feedback from any downstream area, these models must be wired into the
retinal circuit.
Coding for surprising stimuli Predictive coding and optimal predictive informa-
tion are mostly opposed processing theories. We illustrate this by way of a few toy
examples: If we imagine a world that is wholly static, there is nothing to predict,
no predictive information, and a predictive coder would have no response. If the
world is instead completely stochastic, however, predictive coding would dictate
that all noise signals that deviate strongly from the prior on the input generate a
response, since each one is unexpected and therefore surprising. These surprising
inputs are, however, uninformative about the future, since they are a pure noise sig-
nal. The predictive information present is zero and no response modulation should
be encoded. Predictive coding is, however, designed explicitly for non-stationary
stimuli. Predictive information optimization would code for a surprising change
in the inputs, since that will have maximum information about the future state.
In that sense the two methods are aligned and preserving predictive information
preferentially over other bits of past information is ‘efficient’.
Where models of the input statistics are stored Predictive coding, in its clear for-
mulation by Rao and Ballard, postulates that a set of models of the input statistics
are present in higher order areas that feedback onto sensory areas to suppress re-
sponse to predictable events [35]. Recent work has elegantly shown how this can
be implemented in a hierarchical (Bayesian) framework [19]. With limited feed-
back impinging on the retina, however, the retina itself must store the models of
the input statistics it will receive. It has been demonstrated how adaptive gain
mechanisms in the retina, that have presumably been encoded over evolutionary
time, can instantiate predictive coding in the retina [26]. To further test these ideas
in the context of natural scene statistics, one needs to define the set of motion
models an organism encounters in its natural environment.
3 LNP models fail to capture motion processing in the retina
In many contexts, our simplest models of retinal ganglion cell firing fail to recapit-
ulate the actual response properties of the retina. This is most clearly demonstrated
for motion stimuli. So called LN, or linear-nonlinear models, of neural processing
fail to recapitulate the motion processing properties of the retina for a variety of
moving stimuli, including motion anticipation, the reversal response, object mo-
tion detection, and the omitted-stimulus response. Sometimes complex adaptive
gain controls mechanisms need to be added to these models to explain the motion
processing of the retina. In LN models, the probability of spiking is an instanta-
neous, nonlinear function of a linearly filtered version of the sensory input. In the
case of retinal ganglion cells that we study here, the inputs are the image or con-
trast as a function of space and time, s(x,t). Thus, if we write the probability per
unit time of a spike (the firing rate), we have
rLN(t) = r0g(z),
(3)
6
where r0 sets the scale of firing rates, g(z) is a dimensionless nonlinear function,
and
(cid:90)
(cid:90) t
0
z(t) =
dτ
d2x f (x,τ)s(x,t − τ);
(4)
where the function f (x,τ) is the receptive field.
If we deliver stimuli that are drawn from a Gaussian white noise ensemble,
then
f (x,τ) ∝ (cid:104)s(x,t − τ)δ (t −tspike)(cid:105),
(5)
where tspike is the time of a spike and (cid:104)···(cid:105) denotes an average over the stimulus
ensemble [17]. If we take the LN model derived from the random checkerboard
stimuli and use it to produce neural responses to the moving bar stimulus, the
predictive information carried by the neurons is drastically wrong. In recent re-
views from Gollisch and Meister [22] and Berry and Schwartz [9], the myriad
ways in which a simple linear-nonlinear-Poisson (LNP) model fails to reproduce
known retinal response properties are described in detail. When simple bars of
light move across the retina, reverse their path, blink on and off, or move in more
complex ways, many kinds of nonlinear processes in the retina are activated. None
of these effects are captured by this basic version of the LN model for retinal fir-
ing. Additionally, results in [33] reveal that the LN model fails to recapitulate the
near–optimal behavior of the real data, as illustrated for a slightly different LN
model here in Figure 1.
All groups fall away from the bound determined by ∆t = 1/60 s, the delay
between the current response and the onset of the common future. When we com-
pute information about the future, we assume that the future starts now, and do
not make any allowances for processing delays. We could, instead, compare the
performance of the LN model with bounds calculated assuming that there is a
delay between past and future, so that ∆t∗ = ∆t + tdelay. The bound for ∆t∗ is
chosen to be tdelay = 117ms, comparable to the delay one might estimate from the
peak of the information about position, or from the structure of the receptive fields
themselves. Interestingly, the model neurons do come close to this less restrictive
bound.
Of course, these model cells might not be optimal at any delay, instead they
could fail to represent all of the predictable components of the stimulus, such as
the velocity. Real salamander RGCs have a delay of at least 50ms, as measured by
the time to the peak firing rate induced by a flash. The data reveal that the retina has
a mechanism that allows it to saturate the bound on the predictive information with
almost zero effective processing delay when responding to a predictable moving
stimulus.
This work only scratches the surface of optimal coding for the future stimu-
lus in the retina. The motion statistics chosen here were chosen to include both
two time scales of predictable motion as well as a purely stochastic motion com-
ponent, while still being soluble via Gaussian information bottleneck techniques
[16]. Important extensions of this work will test other parametric motion models
with different statistics, but one of the most important directions of future research
here is to explore motion models that mimic the properties of natural scenes.
7
Fig. 1 The predictive information present in groups of 5 retinal ganglion cells (RGCs, black
dots), as well as for model neurons fit with linear-nonlinear (LN) models (red dots). The bound
on the maximal amount of information about the future a group response with a given about of
information about the past stimulus is denoted with the black line. The input is a moving bar
stimulus as in [33]. Some model groups have more information about the past because the LN
model does not capture the low-level noise in the response of the RGC’s. Other model groups
have less info than observed in real data because they fail to capture the stimulus driven response
of these cells. No model groups capture as much information about the future of the stimulus as
the core predictive groups of the cells in the retina.
4 Towards naturalistic motion stimuli
Natural scenes have heavy-tailed distributions of many quantities of interest, in-
cluding intensity, contrast, and temporal modulation frequency [31,21,36,12].
This means that there exists no single length scale in space or time one can use
to coarse-grain natural scenes without sacrificing large amounts of structure in
the data, and that potentially salient fluctuations exist on all scales. We illustrate
some basic statistics of natural scenes in Figure 2, taken from our own database of
natural movies.
To fully test whether the retina conveys information to the brain in a way that
is optimized for prediction, we must consider what prediction problems the retina
evolved to solve. In particular, if the retina solves only a subset of the possible
spatiotemporal prediction problems that could possibly confront its input, it should
solve those that are present in the natural environment. It could even be the case
that the retinas of different animals evolved to encode most efficiently prediction
on the scales and correlation structures present in its own ecological niche. To
test this, we need a framework for quantifying the prediction problems present in
natural visual stimuli.
4.1 Quantifying optimal coding for natural motion
It is not possible to measure the information content of a neuron’s response to
the pixel-by-pixel representation of an image sequence of any useful size. Proxies
for this calculation include computing information about time within a long and
ergodic stimulus sequence [14], but a better approach is to find some reduced and
00.10.20.30.400.050.10.150.20.250.3 data, N = 5, 1000 groupsLN model, N = 5, 1000 groups(bits)Ipast(bits)Ifuture 8
Fig. 2 (a) Representative frames from 3 natural movies: bees on a honeycomb (a plate of glass
exposes the hive from the side); pack ice flowing in Lake Michigan; tree blowing in the wind. (b)
Contrast distribution for an ensemble of 8 natural movie clips of 20 s each, filmed at 60 Hz. The
contrast of each pixel is defined as C = I/I0 where I is the 8-bit intensity value for that pixel and
I0 is chosen for each frame such that the average contrast for that frame is 0. Distributions were
estimated at different scales by averaging contrast over N × N blocks. Contrast distributions are
normalized to have unit variance. (c) Temporal power spectra for 8 natural movie clips. Spectra
were computed for each pixel using a sliding Hamming window of 256 frames with 50% overlap,
then averaged across pixels.
parametrizable (and therefore sample-able) representation of the motion present
in the natural environment that retains the features relevant to the organism. The
heavy-tailed nature of natural scene statistics do not offer up a coarse-graining
length scale over which we might smooth our inputs. Instead, we are left with
the difficult task of quantifying and summarizing the key statistics of the natural
world. One useful extraction in the context of prediction is to track salient objects
in natural scenes and analyze the statistics of those trajectories. Another useful
experiment would be to show the retina of one animal object trajectories from a
different ecological niche and determine whether it fails at this predictive compu-
tation while excelling at those drawn from its native environment.
Higher-order statistics of motion Deciding how to quantify a natural scene can be
challenging. The high dimensionality of natural inputs to the visual system means
that direct approaches to quantifying the information transmission is wrought with
sampling error pitfalls or completely impossible. Making educated guesses about
what features of natural motion to quantify, or searching directly for a lower di-
mensional representation of the structure of natural scenes are two promising ap-
a)b)c)normalized contrast-4-2log p(contrast)-12-11-10-9-8-7-6-5-4-31 2 4 8 16 32N204temporal frequency (s-1)100101power (arbitrary units)100101102103104105beesbushesfishicelarvaeleavestreewater 9
proaches to this problem. Recent work has defined a set of motion primitives,
correlations structures in space and time that find their basis in early theories of
motion processing, and extend and complete a local motion structure. The Fourier,
non-Fourier, expander and glider components of local motion can be readily com-
puted from natural movies [27,30]. This approach reveals that certain ratios of
these components may be prevalent in natural scenes [30]. The brain might make
use of this fact to tailor its motion processing to just these types of input. If so,
deviations from this natural ratio should lead to noisier, less efficient coding for
the future stimulus. Downstream of the retina, this could lead to motion perception
deficits.
Such local motion signatures only characterize part of the total motion signal
in natural scenes. Machine learning efforts have been launched to find longer-
range, collective components of natural image and motion statistics [15,24,25,
37]. Work from these groups has shown that some physical models of long range
fluctuations may be applicable to natural motion. This is exciting because it could
lead to a generative model for such motion, opening up the possibility of more
stringent, parametrized tests of optimal coding for motion in the retina.
5 Conclusions
Testing theories of optimal prediction in the visual stream requires an integration
of existing theories of optimal coding in the retina and beyond, with a careful
quantification of the motion statistics present in the natural environment. By ex-
amining how well information processing in the brain is tuned to natural inputs,
we may discover new static and adaptive features of the predictive part of the
neural code.
Acknowledgements We thank D. J. Schwab for comments on the manuscript.
References
1. Atick, J.J., Redlich, A.N.: What does the retina know about natural scenes? Neural Com-
2. Attneave, F.: Some informational aspects of visual perception. Psychological review 61(3),
3. Barlow, H.B.: Summation and inhibition in the frog’s retina. The Journal of physiology
putation 4(2), 196–210 (1992)
183–93 (1954)
119(1), 69–88 (1953)
4. Barlow, H.B.: Possible principles underlying the transformation of sensory messages. Sen-
sory Communication pp. 217–234 (1961)
5. Bastos, A.M., Usrey, W.M., Adams, R.A., Mangun, G.R., Fries, P., Friston, K.J.: Canonical
microcircuits for predictive coding. Neuron 76(4), 695–711 (2012)
6. Bell, A.J., Sejnowski, T.J.: The independent components of natural scenes are edge filters.
Vision research 37(23), 3327–3338 (1997)
7. Berger, T.: Rate-distortion theory. Encyclopedia of Telecommunications (1971)
8. Berry, M.J., Brivanlou, I.H., Jordan, T.A., Meister, M.: Anticipation of moving stimuli by
the retina. Nature 398(6725), 334–8 (1999)
9. Berry, M.J., Schwartz, G.: The retina as embodying predictions about the visual world.
Predictions in the Brain: Using Our Past to Generate a Future p. 295 (2011)
10. Bialek, W., Nemenman, I., Tishby, N.: Predictability, complexity, and learning. Neural
computation 13(11), 2409–63 (2001)
10
11. Bialek, W., de Ruyter van Steveninck, R.R., Tishby, N.: Efficient representation as a design
principles for neural coding and computation. Proceedings of the International Symposium
on Information Theory 2006(arXiv:0712.4381 [q–bio.NC] (2007)) (2006)
12. Billock, V.A., de Guzman, G.C., Kelso, J.S.: Fractal time and 1/f spectra in dynamic images
and human vision. Physica D: Nonlinear Phenomena 148(1), 136–146 (2001)
13. Boerlin, M., Machens, C.K., Den`eve, S.: Predictive coding of dynamical variables in bal-
anced spiking networks. PLoS Computational Biology 9(111), e1003,258 (2013)
14. Brenner, N., Strong, S.P., Koberle, R., Bialek, W., de Ruyter van Steveninck, R.R.: Synergy
in a neural code. Neural computation 12(7), 1531–52 (2000)
15. Cadieu, C.F., Olshausen, B.A.: Learning intermediate-level representations of form and mo-
tion from natural movies. Neural computation 24(4), 827–866 (2012)
16. Chechik, G., Globerson, A., Tishby, N., Weiss, Y.: Information bottleneck for gaussian vari-
able. JMLR 6, 165–188 (2005)
17. Chichilnisky, E.: A simple white noise analysis of neuronal light responses. Network: Com-
putation in Neural Systems 12(2), 199–213 (2001)
18. Creutzig, F., Globerson, A., Tishby, N.: Past-future information bottleneck in dynamical
systems. Physical review. E, Statistical, nonlinear, and soft matter physics 79(4 Pt 1),
041,925 (2009)
19. Den`eve, S.: Bayesian spiking neurons i: inference. Neural computation 20(1), 91–117
20. Doi, E., Gauthier, J.L., Field, G.D., Shlens, J., Sher, A., Greschner, M., Machado, T.A.,
Jepson, L.H., Mathieson, K., Gunning, D.E., Litke, A.M., Paninski, L., Chichilnisky, E.J.,
Simoncelli, E.P.: Efficient coding of spatial information in the primate retina. The Journal
of neuroscience : the official journal of the Society for Neuroscience 32(46), 16,256–64
(2012)
21. Dong, D.W., Atick, J.J.: Statistics of natural time-varying images. Network: Computation
in Neural Systems 6(3), 345–358 (1995)
22. Gollisch, T., Meister, M.: Eye smarter than scientists believed: neural computations in cir-
cuits of the retina. Neuron 65(2), 150–164 (2010)
23. van Hateren, J.H., van der Schaaf, A.: Independent component filters of natural images
compared with simple cells in primary visual cortex. Proceedings of the Royal Society of
London B: Biological Sciences 265(1394), 359–366 (1998)
24. Hausler, C., Susemihl, A.: Temporal autoencoding restricted boltzmann machine. arXiv
preprint arXiv:1210.8353 (2012)
25. Hausler, C., Susemihl, A., Nawrot, M.P.: Natural image sequences constrain dynamic re-
ceptive fields and imply a sparse code. Brain research 1536, 53–67 (2013)
26. Hosoya, T., Baccus, S.A., Meister, M.: Dynamic predictive coding by the retina. Nature
436(7047), 71–77 (2005)
(2008)
27. Hu, Q., Victor, J.D.: A set of high-order spatiotemporal stimuli that elicit motion and
reverse-phi percepts. Journal of vision 10(3), 9 (2010)
28. Kastner, D.B., Baccus, S.A.: Spatial segregation of adaptation and predictive sensitization
in retinal ganglion cells. Neuron 79(3), 541–554 (2013)
29. Kilner, J.M., Friston, K.J., Frith, C.D.: Predictive coding: an account of the mirror neuron
system. Cognitive processing 8(3), 159–166 (2007)
30. Nitzany, E.I., Victor, J.D.: The statistics of local motion signals in naturalistic movies. Jour-
nal of vision 14(4) (2014)
32.
31. Olshausen, B.A., Field, D.J.: Natural image statistics and efficient coding*. Network: com-
putation in neural systems 7(2), 333–339 (1996)
Olveczky, B.P., Baccus, S.A., Meister, M.: Segregation of object and background motion in
the retina. Nature 423(6938), 401–408 (2003)
33. Palmer, S.E., Marre, O., Berry, M.J., Bialek, W.: Predictive information in a sensory popu-
lation. Proceedings of the National Academy of Sciences 112(22), 6908–6913 (2015)
34. Puchalla, J.L., Schneidman, E., Harris, R.A., Berry, M.J.: Redundancy in the population
code of the retina. Neuron 46(3), 493–504 (2005)
35. Rao, R.P., Ballard, D.H.: Predictive coding in the visual cortex: a functional interpretation
of some extra-classical receptive-field effects. Nature neuroscience 2(1), 79–87 (1999)
36. Ruderman, D.L.: Origins of scaling in natural images. Vision research 37(23), 3385–3398
37. Saremi, S., Sejnowski, T.J.: Hierarchical model of natural images and the origin of scale
invariance. Proceedings of the National Academy of Sciences 110(8), 3071–3076 (2013)
(1997)
38. Schwartz, G., Harris, R., Shrom, D., Berry, M.J.: Detection and prediction of periodic pat-
terns by the retina. Nature neuroscience 10(5), 552–4 (2007)
39. Schwartz, G., Taylor, S., Fisher, C., Harris, R., Berry, M.J.: Synchronized firing among
retinal ganglion cells signals motion reversal. Neuron 55(6), 958–969 (2007)
40. Shannon, C.E.: A mathematical theory of communication. Bell Sys. Tech. J 27, 379–423,
11
623–656 (1948)
41. Srinivasan, M.V., Laughlin, S.B., Dubs, A.: Predictive coding: a fresh view of inhibition in
the retina. Proceedings of the Royal Society of London B: Biological Sciences 216(1205),
427–459 (1982)
42. Srinivasan, R., Rao, K.: Predictive coding based on efficient motion estimation. Communi-
cations, IEEE Transactions on 33(8), 888–896 (1985)
43. Stephens, G.J., Mora, T., Tkacik, G., Bialek, W.: Statistical thermodynamics of natural im-
ages. Physical review letters 110(1), 018,701 (2013)
44. Tishby, N., Pereira, F.C., Bialek, W.: The information bottleneck method. Proceedings
of the 37th Annual Allerton Conference on Communication, Control and Computing,
37(arXiv:physics/0004057 (2000)), 368–377 (1999)
45. Wacongne, C., Changeux, J.P., Dehaene, S.: A neuronal model of predictive coding account-
ing for the mismatch negativity. The Journal of neuroscience 32(11), 3665–3678 (2012)
|
1707.05182 | 1 | 1707 | 2017-07-17T14:33:14 | A probabilistic model for learning in cortical microcircuit motifs with data-based divisive inhibition | [
"q-bio.NC"
] | Previous theoretical studies on the interaction of excitatory and inhibitory neurons proposed to model this cortical microcircuit motif as a so-called Winner-Take-All (WTA) circuit. A recent modeling study however found that the WTA model is not adequate for data-based softer forms of divisive inhibition as found in a microcircuit motif in cortical layer 2/3. We investigate here through theoretical analysis the role of such softer divisive inhibition for the emergence of computational operations and neural codes under spike-timing dependent plasticity (STDP). We show that in contrast to WTA models - where the network activity has been interpreted as probabilistic inference in a generative mixture distribution - this network dynamics approximates inference in a noisy-OR-like generative model that explains the network input based on multiple hidden causes. Furthermore, we show that STDP optimizes the parameters of this model by approximating online the expectation maximization (EM) algorithm. This theoretical analysis corroborates a preceding modelling study which suggested that the learning dynamics of this layer 2/3 microcircuit motif extracts a specific modular representation of the input and thus performs blind source separation on the input statistics. | q-bio.NC | q-bio |
A probabilistic model for learning in cortical microcircuit
motifs with data-based divisive inhibition
Robert Legenstein∗, Zeno Jonke∗, Stefan Habenschuss, Wolfgang Maass
Institute for Theoretical Computer Science
Graz University of Technology
October 3, 2018
∗ These authors contributed equally to the work.
Abstract
Previous theoretical studies on the interaction of excitatory and inhibitory neu-
rons proposed to model this cortical microcircuit motif as a so-called Winner-Take-All
(WTA) circuit. A recent modeling study however found that the WTA model is not
adequate for data-based softer forms of divisive inhibition as found in a microcircuit
motif in cortical layer 2/3. We investigate here through theoretical analysis the role of
such softer divisive inhibition for the emergence of computational operations and neu-
ral codes under spike-timing dependent plasticity (STDP). We show that in contrast
to WTA models -- where the network activity has been interpreted as probabilistic
inference in a generative mixture distribution -- this network dynamics approximates
inference in a noisy-OR-like generative model that explains the network input based on
multiple hidden causes. Furthermore, we show that STDP optimizes the parameters
of this model by approximating online the expectation maximization (EM) algorithm.
This theoretical analysis corroborates a preceding modelling study which suggested
that the learning dynamics of this layer 2/3 microcircuit motif extracts a specific
modular representation of the input and thus performs blind source separation on the
input statistics.
1
1
Introduction
Winner-take-all-like (WTA-like) circuits constitute a ubiquitous motif of cortical microcir-
cuits [Douglas and Martin, 2004]. Previous models and theories for competitive Hebbian
learning in WTA-like circuit from [Rumelhart and Zipser, 1985] to [Nessler et al., 2013]
were based on the assumption of strong WTA-like lateral inhibition. Several theoretical
studies showed that spike-timing dependent plasticity (STDP) supports the emergence of
Bayesian computation in such winner-take-all (WTA) circuits [Nessler et al., 2013, Haben-
schuss et al., 2013b, Klampfl and Maass, 2013]. These analyses were based on a probabilis-
tic generative model approach. In particular, it was shown that the network implicitly
represents the distribution of input patterns through a generative mixture distribution
and that STDP optimizes the parameters of this mixture distribution. But this analysis
assumed that the input to a WTA is explained at any point in time by a single neuron,
and that strong lateral inhibition among pyramidal cells ensures a basically fixed total
output rate of the WTA. These assumptions, however, may not be suitable in the context
of more realistic activity dynamics in cortical networks.
In fact, recent modeling results [Avermann et al., 2012, Jonke et al., 2017] show that
the WTA model is not adequate for a softer form of inhibition that has been reported
for cortical layer 2/3. This softer form of inhibition is often referred to as feedback in-
hibition, or lateral inhibition, and has been termed more abstractly based on its influ-
ence on pyramidal cells as divisive inhibition [Wilson et al., 2012, Carandini and Heeger,
2012]. It stems from dense bidirectional interconnections between layer 2/3 pyramidal cells
and nearby Parvalbumin-positive (PV+) interneurons (often characterized as fast-spiking
interneurons, in particular basket cells), see e.g. [Packer and Yuste, 2011, Fino et al.,
2012, Avermann et al., 2012]. The simulations results in [Jonke et al., 2017] also indicate
that blind source separation emerges as the computational function of this microcircuit
motif when STDP is applied to the input synapses of the circuit.
The results of [Jonke et al., 2017] raise the question whether they can be understood
from the perspective of a corresponding probabilistic generative model, that could replace
the mixture model that underlies the analysis of emergent computational properties of
microcircuit motivs with hard WTA-like inhibition. We propose here such a model that
is based on a Gaussian prior over the number of active excitatory neurons in the network
and a noisy-OR-like likelihood term. We develop a novel analysis technique based on the
neural sampling theory [Buesing et al., 2011] to show that the microcircuit motif model
approximates probabilistic inference in this probabilistic generative model. Further, we
derive a plasticity rule that optimizes the parameters of this generative model through
online expectation maximization (EM), the arguably most powerful tool from statistical
learning theory for the optimization of generative models. We show that this plasticity
rule can be approximated by an STDP-like learning rule.
This theoretical analysis strengthens the claim that blind source separation [Foldiak,
1990] -- also referred to as independent component analysis [Hyvarinen et al., 2004] --
emerges as a fundamental computation on assembly codes through STDP in this microcir-
cuit motif. This computational operation enables a network to disentangle and separately
represent superimposed inputs that result from independent assembly activations in dif-
ferent upstream networks. Furthermore, our theoretical analysis reveals that the ability of
this cortical microcircuit motif to perform blind source separation is facilitated either by
the normalization of activity patterns in input populations, or by homeostatic mechanisms
that normalize excitatory synaptic efficacies within each neuron.
2
Figure 1: Data-based network model M for a microcircuit motif. A) Network
anatomy. Circles denote excitatory (black) and inhibitory (red) pools of neurons. Black
arrows indicate excitatory connections. Red lines with dots indicate inhibitory connec-
tions. Numbers above connections denote corresponding connection probabilities. B)
Network physiology. Same as in (A), but connection delays δ are indicated. All synapses
are modeled with the same PSP shape using a decay time constant of τf = 10 ms as
indicated on top right. Input synapses are subject to STDP.
2 Results
A data-based microcircuit motif model for the interaction of pyramidal cells with PV+
inhibitory neurons in layer 2/3 has been introduced in [Avermann et al., 2012]. Based
on this study, [Jonke et al., 2017] analyzed the computational properties that emerge in
this microcircuit motif from synaptic plasticity. We first briefly introduce the microcircuit
motif model analyzed in [Jonke et al., 2017] and discuss its properties. Subsequently, we
present a theoretical analysis of this network motif based on a probabilistic generative
model P.
2.1 A data-based model for a network motif consisting of excitatory and
inhibitory neurons
[Jonke et al., 2017] proposed a specific model for interacting populations of pyramidal cells
with PV+ inhibitory neurons in cortical layer 2/3 based on data from the Petersen Lab
[Avermann et al., 2012], see Fig. 1A, B. We refer to this specific model as the microcircuit
motif model M.
The model M consists of two reciprocally connected pools of neurons, an excitatory
pool and an inhibitory pool. M stochastic spiking neurons constitute the excitatory pool.
Their dynamics is given by a stochastic version of the spike response model that has been
fitted to experimental data in [Jolivet et al., 2006]. The instantaneous firing rate ρm of a
neuron m depends exponentially on its current membrane potential um,
ρm(t) =
1
τ
exp(γ · um(t)) ,
(1)
where τ = 10 ms and γ = 2 are scaling parameters that control the shape of the response
function. After emitting a spike, the neuron enters a refractory period.
The excitatory neurons are reciprocally connected to a pool of recurrently connected
inhibitory neurons. All connection probabilities in the model were taken from [Avermann
et al., 2012]. Excitatory neurons receive excitatory synaptic inputs y1(t), .., yN (t) with
corresponding synaptic efficiencies wim between the input neuron i and neuron m. These
afferent connections are subject to a standard form of STDP. Thus, the membrane po-
3
tential of excitatory neuron m is given by the sum of external inputs, inhibition from
inhibitory neurons, and its excitability α
um(t) =Xi
wim yi(t) − Xj∈Im
wIEIj(t) + α,
(2)
where Im denotes the set of indices of inhibitory neurons that project to neuron m, and
wIE denotes the weight of these inhibitory synapses. Ij(t) and yi(t) denote synaptic input
from inhibitory neurons and input neurons respectively, see above.
Inhibitory contributions to the membrane potential of pyramidal cells have in this
neuron model a divisive effect on the firing rate. This can be seen by by substituting
eq. (2) in eq. (1), see also eq. (23) in Methods, thus implementing divisive inhibition (see
[Carandini and Heeger, 2012] for a recent review). Divisive inhibition has been shown to be
a ubiquitous computational primitive in many brain circuits (see [Carandini and Heeger,
2012] for a recent review).
In mouse visual cortex, divisive inhibition is implemented
through PV+ inhibitory neurons [Wilson et al., 2012]. Although the inhibitory signal
is common to all neurons in the pool of excitatory neurons, contrary to the inhibition
modeled in [Nessler et al., 2013] it does not normalize the firing rates of neurons exactly
and therefore the total firing rate in the excitatory pool is variable and depends on the
input strength. Importantly, in contrast to [Nessler et al., 2013] where inhibition strictly
enforced that only a single neuron in the excitatory pool is active at any given time, the
data-based model M allows several neurons to be active concurrently.
2.2 Emergent properties of the data-based network model: From WTA
to k-WTA
The computational properties of this data-based network model M were extensively stud-
ied through simulations in [Jonke et al., 2017].
In order to compare the properties of
this data-based network model M to previously considered WTA models, they examined
the emergence of orientation selectivity, which we briefly discuss here. For details, please
see [Jonke et al., 2017]. Pixel-representations of noisy bars in random orientations were
provided as external spike inputs (Fig. 2A). Input spike trains were generated from these
pixel arrays by converting pixel values to Poisson firing rates of input neurons (black pixel:
75 Hz; white pixel: 1 Hz). Randomly oriented bars were presented to the network for 400 s
where each bar was presented for 50 ms, see Fig. 2B (STDP was applied to synapses from
input neurons to excitatory neurons). The resulting network response (Fig. 2D) shows the
emergence of assembly codes for oriented bars. The resulting Gaussian-like tuning curves
of excitatory neurons (Fig. 2E) densely cover all orientations, resembling experimental data
from orientation pinwheels (see Fig. 2 d,e in [Ohki et al., 2006]). Also consistent with
experimental data [Kerlin et al., 2010, Isaacson and Scanziani, 2011], inhibitory neurons
did not exhibit orientation selectivity (not shown).
In contrast, previously considered models with idealized strong inhibition in WTA-
circuits [Nessler et al., 2013] show a clearly distinct behavior, see Fig. 2F. For this model,
at most a single neuron could fire at any moment of time, and as a result at most two
neurons responded after a corresponding learning protocol with an increased firing rate
to a given orientation (see Fig. 2F and Fig. 5 in [Nessler et al., 2013]). In the simulations
of the data-based model M, on average k = 17 neurons responded to each orientation
with an increased firing rate. This suggests that the emergent computational operation
of the layer 2/3 microcircuit motif with divisive inhibition is better described as k-WTA
computation, where k winners may emerge simultaneously from the competition. This
4
Figure 2: Emergent computational properties of the data-based network model
M. A) Network inputs are given by images of randomly oriented bars (inputs arranged
in 2D for clarity; pixel gray-level indicates effective network input yi(t), see eq. (22)). B)
Input neuron spike patterns (every 4th neuron shown). Presence of a bar in the input with
orientation indicated in panel A indicated by gray shading. C, D) Spike responses of a
subset of excitatory neurons in M to the input in (B) before (C) and after (D) learning
(neurons sorted by preferred orientation). E) Tuning curves of excitatory neurons in with
preferred orientations between 90 and 120 degrees. F) Orientation tuning curves in a
WTA model [Nessler et al., 2013, Habenschuss et al., 2013b]. Figure modified from Fig. 2
in [Jonke et al., 2017].
5
number k is however not a strict constraint in the data-based model M. The actual
number of winners depends on synaptic weights and the external input. Its computation
is thus better describe as an adaptive k-WTA operation. The k-WTA characterisitcs of
the layer 2/3 microcircuit motif is quite attractive, since it is known from computational
complexity theory that the k-WTA computation is more powerful than the simple WTA
computation (for k > 1) [Maass, 2000].
2.3 Theoretical framework for understanding emergent computational
properties the layer 2/3 microcircuit motif
Fig. 2 demonstrates that significantly different computational properties emerge in the
data-based model M through STDP as compared to previously considered WTA mod-
els [Nessler et al., 2013, Habenschuss et al., 2013b]. The main aim of this article is to
understand this different emergent computational capability theoretically, in particular
since the analysis from [Nessler et al., 2013] and [Habenschuss et al., 2013b] in terms of
mixture distributions is only applicable to WTA circuits. The novel analysis technique
that we will use is summarized as follows. First, using some simplifications on the network
dynamics, we formulate the network dynamics in the neural sampling framework [Buesing
et al., 2011]. This allows us to deduce the distribution p(zy, W ) of activities z of excita-
tory neurons in the network for a given input y and for the given network weights W . We
then show that this distribution approximates the posterior distribution of a generative
probabilistic model P. This generative model is not a mixture distribution as in the WTA
case [Nessler et al., 2013], but a more complex distribution that is based on a noisy-OR-
like likelihood. We make the nature of the approximation explicit and evaluate its severity
through simulations. Finally, we derive a plasticity rule that implement online EM in this
generative model, thus implementing blind source separation. We find that this plasticity
rule can be approximated by an STDP-like learning rule.
2.3.1 Formulation of the network dynamics of M in the neural sampling
framework
The neural sampling framework [Buesing et al., 2011] provides us with the ability to
determine the stationary distribution (defined in the following) of network states for the
given network parameters and a given network input. In order to be able to describe the
probabilistic relationships between input and network activity, we describe network inputs
by binary vectors y(t) and responses of excitatory neurons in M by binary vectors z(t).
The vectors y(t) and z(t) capture the spiking activity of ensembles of spiking neurons in
continuous time according to the common convention introduced in [Berkes et al., 2011]
and [Buesing et al., 2011]: A spike of the ith neuron in the ensemble at time t sets the
corresponding component yi(t) (zi(t)) of the bit vector y(t) (z(t)) from its default value
0 to 1 for some duration τ (that can be chosen for example to reflect the typical time
constant of an EPSP), see Methods. Note the difference between the vectors y(t), z(t)
and the output traces y(t), z(t) used in eq. (2) (and defined in eq. (22) in Methods). The
former constitute an abstract convention to describe the momentary state of the network
based on its current firing activity, while the latter describe the impact that the neurons
have on their postsynaptic targets in terms of real-valued double-exponential EPSPs.
We want to describe the distribution of network states z(t) for given inputs y(t)
In this
and network parameters W in terms of a probability distribution p(zy, W ).
distribution, the activities of network inputs and excitatory neurons in M are represented
by two vectors of binary random variables: y = (y1, . . . , yN ) (termed input variables in the
6
following) and z = (z1, . . . , zM ) (termed hidden variables or hidden causes). The network
state y(t), z(t) at time t is interpreted as one specific realization of these random variables.
In order to make this mapping between network activity in M and the distribution of
network states feasible, one has to make three simplifying assumptions about the dynamics
of the neural network models similar as in [Buesing et al., 2011]. First, PSPs of inputs
and network neurons are rectangular with length τ (chosen here to be τ = 10 ms) and
network neurons are refractory for the same time span τ after each spike. Second, synaptic
connections are idealized in the sense that the synaptic delay is 0 (i.e., a presynaptic spike
leads instantaneously to a PSP in the postsynaptic neuron). And finally, the weights of
recurrent synaptic connections are symmetric (i.e., the weight from neuron i to neuron j is
identical to the weight from neuron j to neuron i). This necessitates that lateral inhibition
is not implemented through a pool of inhibitory neurons. Instead, the network dynamics
is defined by only one pool of M network neurons (the same number as the number of
excitatory network neurons in M). Since the inhibitory neurons in M show linear response
properties, the inhibition in the network depends linearly on the activity of excitatory
neurons in the network. One can therefore model the inhibition in the network by direct
inhibitory connections between excitatory neurons (where synaptic delays are neglected)
with weight β. For clarity, we provide the full description of the approximate dynamics
in the following.
The approximate dynamics is described by M network neurons. Network neurons
have instantaneous firing rates that depend exponentially on their membrane potential,
as given in eq. (1). Whenever neuron m spikes, the output trace zm(t) of neuron m is set
to 1 for a period of duration τ (this corresponds to a rectangular PSP; the same definition
applies to output traces ym(t) of input neurons). After emitting a spike, the neuron enters a
refractory period of duration τ = 10 ms, during which its instantaneous spiking probability
is zero. Note that for this definition of the output trace, the state vector z(t) is identical
to the vector of output traces z(t). Lateral inhibition in the network is established by
direct inhibitory connections between excitatory neurons, leading to membrane potentials
um(t) =
N
Xi
wim yi(t) − Xj6=m
β zj(t) + cm,
(3)
where cm denotes some neuron-specific excitability of the neuron that is independent of
the input and network activity. Each network neuron receives feedforward synaptic inputs
y1(t), . . . , yN (t) whose contribution to the membrane potential of a neuron m at time t
depends on the synaptic efficiency wim between the input neuron i and the network neuron
m. Network neurons are all-to-all recurrently connected. The second term in (3) specifies
this recurrent input where β is the inhibitory recurrent weight of the connection between
network neuron j and network neuron m. It has been shown in [Buesing et al., 2011] that
for such membrane potentials, the distribution of network states is given by the Boltzmann
distribution:
pNetwork(zy, W ) =
1
Z
where Z is a normalizing constant.
exp
γ ·
Xi,m
wimyizm +
1
2 Xm6=l
βzmzl +Xm
,
(4)
cmzm
7
Figure 3: Relationship between the data-based model M and the generative
probabilistic model P. A-C) Schema of the response of model M to superimposed
bars. Network inputs (left) are schematically arranged in a 2D array for clarity of the
argument. Black indicates highly active inputs neurons. A) A vertical bar with added
noise is presented to M. This input activates an excitatory neuron (filled circle, spiking
activity is indicated), similar as in hard WTA models. B) Another neuron is activated
by a horizontal bar, also similar as in hard WTA models. C) A combination of these
two basic input patterns activates both neurons in M. This response is inconsistent with
a WTA model, and with any generative model based on mixture distributions. But it
can still be viewed as approximate inference of the posterior distribution p(zy, W ) over
hidden causes z for the given inputs y in the probabilistic model P shown in D. D)
Schema of probabilistic model P. The joint p(z, yW ) is defined by the prior p(z) and the
likelihood p(yz, W ). Synaptic efficacies W implicitly define the likelihood over inputs y
for given hidden causes z (probability values for inputs yi indicated by shading of squares;
σLS denotes the logistic sigmoid function). In the likelihood model, a given input yi is 1
(corresponding to a black pixel in this example) with high probability if it has a large wim
to at least one active hidden cause zm. In the depicted example, yi belongs to two bars
(see A, B) with corresponding active hidden causes. Due to the nonlinear behavior of the
likelihood, its probability is comparable to one where only one of the hidden causes zm is
active. The inset on the right depicts the Gaussian prior p(z) on hidden causes z with
µ = 4 and σ = 2.5. The prior implicitly incorporates in P the impact of the inhibitory
feedback in the data-based model M (therefore indicated with dashed lines).
8
2.3.2 A probabilistic model P for the layer 2/3 microcircuit motif:
Fig. 3A-C illustrates the putative stochastic computation performed by the model M,
i.e., how network input leads to network activity in the model. Assume that we have a
probabilistic model P for the inputs y defined by a prior p(z) and a likelihood p(yz, W ).
These distributions describe how one can generate input samples y by first drawing a
hidden state vector z from p(z) and then drawing an input vector y from p(yz, W ).
Therefore, such a probabilistic model P is also called a generative model (for the inputs).
If the distribution of network states (4) is the posterior distribution given by
p(zy, W ) =
,
(5)
p(z)p(yz, W )
Pz′ p(y, z′W )
then the network performs probabilistic inference in this probabilistic model P. The
inference task described by eq. (5) assumes that y is given and the hidden causes z (such
as the basic components of a visual scene) have to be inferred. This inference can intuitively
also be described as providing an "explanation" z for the current observation y according
to the generative model P. In the following, we describe a probabilistic model P and show
that eq. (4) approximates the posterior distribution of this model. This implies that the
simplified dynamics of the data-based model M approximate probabilistic inference in the
probabilistic model P.
The probabilistic model P is defined by two distributions, the prior over hidden vari-
ables p(z) (that captures constraints on network activity imposed for example by lateral
inhibition) and the conditional likelihood distribution over input variables p(yz, W ) that
describes the probability of input y for a given network state z in a network with pa-
rameters W . These two distributions define the joint distribution over hidden and visible
variables since p(z, yW ) = p(yz, W )p(z), see Fig. 3D. The specific forms of these two
distributions in the probabilistic model P considered for M are discussed in the following
and defined by eqs. (6)-(8) below.
In previously considered hard WTA models [Nessler et al., 2013, Habenschuss et al.,
2013b], strong lateral inhibition was assumed. This corresponded to a prior where only a
single component zj of the hidden vector z can be active at any time. The biologically
more realistic divisive inhibition in M allows several of them to fire simultaneously. This
corresponds to a prior that induces sparse activity in a soft manner ("adaptive" k-WTA):
It does not enforce a strict ceiling k on the number of z-neurons that can fire within a
time interval of length τ , but only tries to keep this number within a desired range. Hence
we use as prior in P a Gaussian distribution
p(z) =
1
Zprior
exp
−
1
2σ2 M
Xm=1
zm − µ!2
,
(6)
where Zprior is a normalizing constant, µ ∈ R is a parameter that shifts the mean of the
distribution, and σ2 > 0 defines the variance (see Fig. 3D). Note that the Gaussian is
restricted to integers as the sum runs over binary random variables z1, . . . , zM .
As in other generative models we assume for the sake of theoretical tractability that
the conditional likelihood p(yz, W ) factorizes, so that each input yi is independently
9
(A) Likelihood that an input yi is 1 (red) or 0
Figure 4: Likelihood model of P.
(black) for the proposed likelihood model (8) (full lines) and for the noisy-OR likelihood
(9) (broken lines). The likelihood of an input yi depends on the hidden causes z through
the weighted contribution ai = γ ¯wT
i z of the hidden causes to this input. For large ai, the
likelihood of yi = 1 approaches 1. While for ai = 0, the likelihood of yi = 1 is 0.5 in the
proposed likelihood model, whereas it is 0 in the noisy-OR model. B Approximation of
log(1 + exp(ai)) (black full line) by ai (red broken line) as used in Eq. (11).
explained by the current network state z:
p(yz, W ) =
N
Yi=1
p(yiz, W ).
(7)
A probabilistic model for hard WTA circuits can only explain each input variable yi by
a single hidden cause zm.
In contrast, in probabilistic models with soft inhibition and
the prior (6), several hidden causes can be active simultaneously and explain together an
input variable yi. We define ¯wi as the vector of weights ¯wi = (wi1, . . . , wiM )T that define
the likelihood for variable yi. We consider the following likelihood model
p(yiz, W ) =
i z)yi
exp(γ ¯wT
1 + exp(γ ¯wT
i z)
=
exp(ai)yi
1 + exp(ai)
= σLS ((−1)yiai) ,
(8)
1
i z, and σLS is the logistic sigmoid σLS(u) =
where we have defined ai = γ ¯wT
1+exp(−u) .
This likelihood function is shown in Figure 4A together with the often used noisy-OR
likelihood. Note that if none of the hidden causes zm is active, i.e., zm = 0 for all m, then
yi = 1 with probability 0.5. Each active hidden cause zm = 1 with wim > 0 increases the
probability that input variable yi assumes the value 1, see also Fig. 3D). This likelihood,
allows the generative model to deal with situations where an input neuron can fire in the
context of different hidden causes, for example with pixels in the network inputs that lie
in the intersection of different patterns, see Fig. 3). The soft Gaussian prior (6) allows the
internal model to develop modular representations for different components of complex
input patterns. This likelihood is quite similar to the frequently used noisy-OR model
(see e.g. [Neal, 1992, Saund, 1995]):
pnOR(yi = 0z, W ) = exp(−ai)1−yi(1 − exp(−ai))yi
(9)
10
One difference is that (for purely excitatory weights), the probability of an input yi being
zero is at most 0.5 in the proposed likelihood, while it can become 0 in the noisy-OR
model. Such a model may reflect the situation that network inputs are noisy, so their
firing rates are never zero.
We now analyze the relationship between the probabilistic model P and the descrip-
tion of the data-based model M in the neural sampling framework. We will see that M
approximates probabilistic inference in P. Finally, we show that adaptation of network pa-
rameters in M through STDP can be understood as an approximate stochastic expectation
maximization (EM) process in the corresponding probabilistic model P.
2.3.3
Interpretation of the dynamics of M in the light of P:
Using the likelihood and prior of P, the posterior of hidden states z for given inputs y
and given parameters W is
p(zy, W ) =
1
Z
exp
γ ·
Xi,m
wimyizm −
1
2 Xm6=l
βzmzl +Xm
αzm −
1
γ Xi
log(1 + exp(ai))
(10)
,
where Z is a normalizing constant, β = 1
2γσ2 . The terms including β and α
stem from the prior, while the last term stems from the normalization of the likelihood.
When we compare this posterior to the posterior of the network model eq. (4), we see that
they are quite similar with β denoting the strength of inhibitory connections and α being
the neural excitabilities.
γσ2 , and α = 2µ−1
The last term in eq. (10) is problematic since the ai's depend on z and thus the whole
posterior is not a Boltzmann distribution and can therefore not be computed by the model
M. It turns out however that this last term can be approximated quite well by a term that
is linear in z. Note that for zero ai (i.e., for zero weights or for the zero-z-vector), this
last term evaluates to log(2). But as ai increases, one can neglect the 1 in the logarithm
and the expression quickly approaches ai. We can thus write
Xi
log(1 + exp(ai)) ≈Xi
ai = γXi Xm
wimzm = γXm
zm Xi
wim! ,
(11)
where the term in the brackets on the right is just the L1-norm of the weight vector of
neuron j. Note that for a given weight matrix, an increased γ leads to a better approxi-
mation. The approximation of log(1 + exp(ai)) by ai is illustrated in Fig. 4B. Hence, the
first approximate posterior we consider is given by
pA1(zy, W ) =
1
Z
exp
γ ·
Xi,m
wimyizm −
1
2 Xm6=l
βzmzl +Xm
αzm −Xm
zmXi
wim
.
(12)
This is a Boltzmann distribution of the form (4) and the last term accounts to a neuron-
specific homeostatic bias that depends on the sum of incoming excitatory weights. Match-
ing the terms of this equation to the terms in eq. (4) and performing the same match
in the membrane potential (3), we see that the membrane potential of neurons in this
11
approximation is given by
N
um(t) =
Xi
wim yi(t) − Xj6=m
β zj(t) + α −Xi
wim.
(13)
If excitatory weight vectors are normalized to an L1-norm of wnorm = Pi wim for all m,
this simplifies to
N
um(t) =
Xi
wim yi(t) − Xj6=m
β zj (t) + α − wnorm.
(14)
Note that wnorm can be incorporated into α. Such constant weight sum could be enforced in
a biological network by a synaptic scaling mechanism [Turrigiano and Nelson, 2004, Savin
et al., 2010] that normalizes the sum of incoming weights to a neuron. For the data-based
model M, [Jonke et al., 2017] used uniform excitabilities α for all excitatory neurons and
no homeostasis for simplicity, see eq. (2). We will argue below under which conditions this
approximation is justified. We consider the posterior distribution of such circuits as our
second approximation of the exact posterior:
wimyizm −
1
2 Xm6=l
βzmzl +Xm
zm(α − wnorm)
. (15)
pA2(zy, W ) =
1
Z
exp
γ ·
Xi,m
Note that in this case, wnorm effectively leads to a smaller mean of the Gaussian prior.
A detailed discussion about how parameters of the generative probabilistic model P can
be mapped to parameters of the data-based microcircuit motif model M is provided in
Network parameter interpretation in Methods.
We evaluated the impact of these two approximations in Foldiak's superposition-of-
bars problem [Foldiak, 1990]. This is a standard blind-source separation problem that has
also been used in [Jonke et al., 2017] to evaluate the data-based network model M. In
this problem, input patterns y are two-dimensional pixel arrays on which horizontal and
vertical bars (lines) are superimposed, see Fig. 5A. Input patterns were generated with a
superposition of 1 to 3 bars from the distribution that was used in [Jonke et al., 2017].
We performed inference of hidden causes by sampling from the approximate posterior
distribution pA1(zy, W ) given by eq. (12) for a network of 20 hidden-cause neurons. We
performed approximate stochastic online EM in order to optimize the parameters of the
model. The synaptic update rule (20) used for parameter updates is discussed in detail
below. We compared the approximated posterior to the exact one (10) by computing
the Kullback-Leibler (KL) divergence DKL(p(zy, W )pA1(zy, W )). The KL divergence
was small throughout learning, with a slight decrease during the process, see Fig. 5B
(mean KL divergence during the second half of training was 0.55). To evaluate what that
KL-divergence means for the inference, we considered the hidden state vector zmax with
the maximum posterior probability after learning in the exact and approximate posterior
and used this to reconstruct the input pattern by computing y = σLS(W T zmax). The
reconstructed inputs for the 8 example inputs of Fig. 5A are shown in Fig. 5C for the exact
posterior (top) and the approximate posterior (bottom). The approximate reconstructions
resemble the exact ones in many cases, with occasional misses of a basic pattern. The final
weights of the 20 neurons are shown in Fig. 5D in the two-dimensional layout of the input
to facilitate interpretability. Note that all basic patterns were represented by individual
neurons with additional neurons that specialized on combined patterns.
12
A
B
e
c
n
e
g
r
e
v
d
-
L
K
i
10
5
0
0
C
exact
appr. A1
D
E
e
c
n
e
g
r
e
v
d
-
L
K
i
s
t
h
g
e
w
i
80
60
40
20
0
0
uniform
approx. A1
5
10
weight update (x1000)
15
uniform
approx. A2
adjusted prior A2
5
10
weight update (x1000)
15
Figure 5: Empirical evaluation of approximations in a superposition-of-bars
task. A) Sample input patterns, depicted on the 8×8 grid. Patterns consist of a varying
number of superimposed horizontal or vertical bars. B) Evolution of the Kullback-Leibler
divergence between the exact posterior and the posterior of approximation A1 during
learning (red). As a comparison, the KL-divergence to a uniform distribution is indicated
in blue. C) Example reconstruction of inputs from panel A from the posterior according
to the hidden states with maximum probability in the exact posterior (top row) and the
posterior of approximation A1 after learning (bottom). Scale between 0 (white) and 1
(black). D) Weights vectors of network neurons depicted on the 8×8 grid as the input
in panel A. Scale between 0 (white) and 6 (black). E) Evolution of the Kullback-Leibler
divergence between the exact posterior, the posterior of approximation A2 during learning
(red), and the posterior of approximation A2 with an adjusted sparsity prior d (yellow)
during learning. As a comparison, the KL-divergence to a uniform distribution is indicated
in blue.
13
The approximate posterior pA2 is equivalent to the approximate posterior pA1 if the
synaptic weights of each neuron are normalized to a common L1 norm.
It turned out
in [Jonke et al., 2017] that such a normalization is not strictly necessary. In this work, the
data-based model M managed to perform blind source separation with a posterior that
can be best described by pA2 without normalization of synaptic efficacies. We found that
for the superposition-of-bars problem, the posterior distribution pA2 differs significantly
from the exact posterior if we set wnorm = 0 in eq. (15), see red line in Fig. 5E. This
difference is mostly induced by the tendency of pA2 to prefer many hidden causes due to
the missing last term in eq. (15). If the prior was corrected to reduce the number of hidden
causes, we found that the approximation was significantly improved, in particular as the
network weight vectors approached their final norm values, see yellow line in Fig. 5E. A
closer inspection of eq. (15) in comparison with eq. (12) shows that this approximation
is effective if basic patterns consist of a similar number of active units, because otherwise
patterns with strong activity are preferred over weakly active ones (this is exactly what
the last term in eq. (12) compensates for).
We conclude from this analysis that the microcircuit motif model M approximates
probabilistic inference in a noisy-OR-like probabilistic model of its inputs. The prior on
network activity favors sparse network activity, but does not strictly enforce a predefined
activity level. Such a more flexible regulation of network activity is obviously important
when the network input is composed of a varying number of basic component patterns. We
show below that this network behavior in combination with STDP allows the microcircuit
motif model M to perform blind source separation of mixed input sources. Our analysis
above has shown that the computation of blind source separation in M can be facilitated
either by the normalization of activity in input populations, or by homeostatic mechanisms
that normalize excitatory synaptic efficacies within each neuron.
2.3.4 STDP in the microcircuit motif model M creates an internal model of
network inputs:
After we have established a link between a well-defined probabilistic model P and the
spiking dynamics of microcircuit motif model M, we can now analyze plasticity in the
network. The probabilistic model P defines a likelihood distribution over inputs y that
depends on the parameters W through
p(yW ) =Xz
p(z)p(yz, W ),
(16)
where the sum runs over all possible hidden states z.
We propose that STDP in M can be viewed as an adaptation of the parameters W so
that p(yW ) as defined by the probabilistic model P with parameters W approximates
the actually encountered distribution of spike inputs p∗(y) within the constraints of the
prior p(z). Since the prior typically is defined to favor sparse representations, this tends to
extract the hidden sources of these patterns, an operation called blind source separation
[Foldiak, 1990].
More precisely, we show that STDP in M approximates stochastic online EM [Sato,
1999, Bishop, 2006] in P. Given some external distribution p∗(y) of synaptic inputs,
EM adapts the model parameters W such that the model likelihood distribution p(yW )
approximates the given distribution p∗(y). More formally, the Kullback-Leibler divergence
between the likelihood of inputs in the internal model p(yW ) and the empirical data
distribution p∗(y) is brought to a local minimum. The theoretically optimal learning
14
rule for EM contains non-local terms which are hard to interpret from a biological point
of view. In the following, we derive a local approximation to yield a simple STDP-like
learning rule.
The goal of the EM algorithm is to find parameters that (locally) minimize the
Kullback-Leibler divergence between the likelihood p(yW ) of inputs in the probabilis-
tic model P and the empirical data distribution p∗(y), that is, the distribution of inputs
experienced by the network M. This is equivalent to the maximization of the average
data log-likelihood Ep∗[log p(yW )]. For a given set of training data Y and corresponding
unobserved hidden variables Z, this corresponds to maximizing log p(Y W ). The opti-
mization is done by iteratively performing two steps. For given parameters W old, the
posterior distribution over hidden variables p(ZY , W old) is determined (the E-step). Us-
ing this distribution, one then performs the M-step where Ep(ZY ,W old)[log p(Y , ZW )]
is maximized with respect to W to obtain better parameters for the model. These
steps are guaranteed to increase (if not already at a local optimum) a lower bound
L = Ep(ZY ,W old)hlog p(Y ,ZW )
p(Y ,ZW old)i on the data log likelihood, that is, L ≤ log p(Y W ).
These steps are iterated until convergence of parameters to a local optimum [Bishop,
2006]. Computation of the M-step in the probabilistic model P is hard. In the generalized
EM algorithm, the M-step is replaced by a procedure that just improves the parameters,
without necessarily obtaining the optimal ones for a single M-step. This can be done for
example by changing the parameters in the direction of the gradient
∆wim ∝
∂
∂wim
Ep(ZY ,W old)[log p(Y , ZW )].
(17)
Since in our model, we assume that synaptic efficacy changes are instantaneous for each
pre-post spike pair, we need to consider an online-version of the generalized EM algorithm.
In stochastic online EM, for each data example y(k), a sample z(k) from the posterior
is drawn (the stochastic E-step) and parameters are changed according to this sample-
pair. As shown above, the M network implements an approximation of the stochastic
E-step. In the M-step, each parameter wim is then updated in the direction of the gradient
∆wim ∝ ∂
log p(y(k), z(k)W ). As the prior p(z) in P does not depend on W , this is
log p(y(k)z(k), W ). For the likelihood given by eq. (8), this
equivalent to ∆wim ∝ ∂
derivative is given by
∂wim
∂wim
∂
∂wim
log p(y(k)z(k), W ) = γzm(cid:18)yi −
exp(ai)
1 + exp(ai)(cid:19) = γzm (yi − σLS(ai)) ,
(18)
where σLS denotes the logistic sigmoid function. Hence, the synaptic update rule for
weight wim is given by
∆wim = ηzm (yi − σLS(ai)) ,
(19)
where η > 0 is a learning rate. This learning rule is not local as it requires information
about the activation of all output neurons as well as values of all synaptic weights origi-
nating from input neuron i. In order to make this biologically plausible we approximate
rule (19) by
∆wim = ηzm (yi − σLS(γwim)) ,
(20)
This rule uses only locally available information at the synapse. What are the consequences
of this approximation during learning? If only a single neuron in the network is active, then
the approximation is exact. Otherwise, the approximation ignores what other neurons
contribute to the explanation of input component yi. This means that for yi = 1, the
15
weight will further be increased even if yi = 1 is already fully explained by the network
activity. For yi = 0, the decrease will in general be smaller than in the exact rule (since
weights are non-negative). Note however that only the magnitudes of weigh changes are
affected, but not which weights change and the sign of the change. Hence, we can conclude
that the angle between the approximate parameter change vector and the exact parameter
change vector is between 0 and 90 degrees. In other words, the inner product of these
two vectors is always non-negative and the updates are performed in the correct direction.
This was confirmed in simulations.
In the learning experiment described in Fig. 5, we
compared the approximate update (that was used to optimize the model) with the update
that was proposed by the exact rule (18) at every 50th update step. The angle between
the exact and approximate update vector was between 0◦ and 84◦ with a mean of 57◦.
In the simplified dynamics, a value of z(k)
m = 1 is indicated by a spike in network neuron
m when pattern y(k) is presented as input. We therefore map this update to the following
synaptic plasticity rule: For each postsynaptic spike, update weight wim according to
∆wim = η (yi(t) − σLS(γwim)) .
(21)
According to this learning rule, when the presynaptic neuron i spikes shortly before the
postsynaptic neuron this results in long-term potentiation (LTP) which is weight depen-
dent according to the term σLS(wim). Due to the weight dependence, large weights lead to
small weight changes, with vanishing changes for very large weights. When a post-synaptic
spike by neuron m is not preceded by a presynaptic spike by neuron i (e.g. when the pre-
synaptic spike comes after the post-synaptic spike), this results in long term depression
(LTD). LTD is also weight dependent, but to a much lesser extent as the weight-dependent
factor varies only between 0.5 and 1. This behavior is mimicked by the standard STDP
rule implemented in the data-based model M that a standard weight dependence where
updates exponentially decreased with wim for LTP and did not depend on wim for LTD.
Hence, the dynamics and synaptic plasticity of the data-based model M can be under-
stood as an approximation of EM in the probabilistic model P, that creates an internal
model for the distribution p∗(y) of network inputs. This internal probabilistic model is
defined by a noisy-OR-like likelihood term and a sparse prior on the hidden causes of the
current input pattern. Hence, STDP can be understood as optimizing model parameters
such that the observed distribution of input patterns can be explained through a set of
basic patterns (hidden causes). It is assumed that the input at each time point can be
described by a combination of a sparse subset of these patterns. In other words, STDP in
the microcircuit motif model M performs blind source separation of input patterns.
3 Discussion
We have provided a novel theoretical framework for analyzing and understanding com-
putational properties that emerge from STDP in a prominent cortical microcircuit motif:
interconnected populations of pyramidal cells and PV+ interneurons in layer 2/3. The
computer simulations in [Jonke et al., 2017], that were based on the data from [Avermann
et al., 2012], indicate that the computational operation of this network motif cannot be
captured adequately by a WTA model. Instead, this work suggests a k-WTA model, where
a varying number of the most excited neurons become active. Since the WTA circuit model
turns out to be inadequate for capturing the dynamics of interacting pyramidal cells and
PV+ interneurons, one needs to replace the probabilistic model that one had previously
used to analyze the impact of STDP on the computational function of the network motif.
16
Mixture models such as those proposed by [Nessler et al., 2013] and [Habenschuss et al.,
2013b] are inseparably tied to WTA dynamics: For drawing a sample from a mixture model
one first decides stochastically from which component of the mixture model this sample
should be drawn (and only a single component can be selected for that). We have shown
here that a quite different generative model, similar to the noisy-OR model, captures the
impact of soft lateral inhibition on emergent network codes and computations much bet-
ter (Fig. 3). The noisy-OR model is well-known in machine learning [Neal, 1993, Saund,
1995], but has apparently not previously been considered in computational neuroscience.
Our probabilistic model P further suggests that the varying number of active neurons in
the circuit may depend both on a prior that is encoded by the network parameters and
the familiarity of the network input.
We have shown that the evolution of the dynamics and computational function of
the network motif under STDP can be understood from the theoretical perspective as an
approximation of expectation maximization (EM) for fitting a noisy-OR based generative
model to the statistics of the high dimensional spike input stream. This link to EM is
very helpful from a theoretical perspective, since EM is one of the most useful theoretical
principles that are known for understanding self-organization processes.
In particular,
this theoretical framework allows us to elucidate emergent computational properties of
the network motif for spike input streams that contain superimposed firing patterns from
upstream networks. It disentangles these patterns and represents the occurrence of each
pattern component by a separate sparse assembly of neurons, as already postulated in
[Foldiak, 1990].
The established relationship between the network M and the probabilistic model P
allows us to relate the network parameters α and wIE (eq. (2)) of M to the parameters
of the generative model P. Briefly (for a detailed discussion, see Network parameter
interpretation in Methods), the excitability α of pyramidal cells is proportional to 2µ−1
2σ2 ,
see eq. (26). The strength of inhibitory connections wIE to the pool of pyramidal cells is
proportional to 1
σ2 , see eq. (30). Hence, a large µ in combination with a small σ2 (i.e.,
a sharp activity prior), leads to a large spontaneous activity that is tightly regulated by
strong inhibitory feedback. On the other hand, a broad prior (larger σ2) leads to weaker
inhibitory feedback, thus allowing the network to attain a broader range of activities.
Related work
A related theoretical study for WTA circuits was performed in [Nessler et al., 2013, Haben-
schuss et al., 2013a, Kappel et al., 2014] and extended to sheets of WTA circuits in [Bill
et al., 2015]. It was assumed in these models that inhibition normalizes network activity
exactly, leading to a strict WTA behavior. The analysis in the present work is much more
complex and necessarily has to include a number of approximations. Out analysis reveals
that the softer type of inhibition that we studied provides the network with additional com-
putational functionality. There exists also a structural similarity of the proposed learning
rule (21) to those reported in [Nessler et al., 2013, Habenschuss et al., 2013a, Kappel et al.,
2014]. This is insofar significant as it raises the question why the application of almost
the same learning rule in one motif leads to learning and extraction of a single hidden
cause and in another to the extraction of multiple causes. The answer most likely lies in
the interplay between "prior knowledge" in the model (e.g. in the form of the intrinsic
excitability of neurons), the learning rule and inhibition strength: As there are multiple
neurons in the proposed microcircuit motif model which can spike in response to the same
input, each one of them can adapt its synaptic weights to increase the likelihood of spiking
17
again whenever the same or a similar input pattern is presented in the future, possibly in
conjunction with other different input components. This is manifested through increased
total input strength to those neurons when the pattern is seen again. But this results also
in increased total inhibition to all other neurons, thereby effectively limiting the number
of winners. As there is no fixed normalization of firing rates (probabilities), as soon as the
input strength caused by a single feature component is strong enough to trigger the spike
in some neuron, the neuron will respond to each pattern which consists of that particular
feature. On average this will force neurons to specialize on a single feature component.
Therefore, after learning, each spike can be interpreted as indication of a particular feature
component.
The noisy-OR model eq. (9) is tightly related to the likelihood model used in this
article. It is one of the most basic likelihood models that allows to combine basic patterns.
Noisy-OR and related models have previously been used in the machine learning literature
as models for nonlinear component extraction [Saund, 1995, Lucke and Sahani, 2008], or
as basic elements in belief networks [Neal, 1992], but they have so far not been linked to
cortical processing.
The extraction of reoccurring components of input patterns is closely related to blind
source separation and independent component analysis (ICA) [Hyvarinen et al., 2004].
Previous work in this direction includes implementations of ICA in artificial neural net-
works [Hyvarinen, 1999], see also [Lucke and Eggert, 2010]. These abstract models are
only loosely connected to computation in cortical network motifs. [Savin et al., 2010] in-
vestigated ICA in the context of spiking neurons. Theoretical rules for intrinsic plasticity
were derived which enable neurons in combination with input normalization, weight scal-
ing, and STDP, to extract independent components of inputs. An interesting difference
is that the inhibition in [Savin et al., 2010] acts to decorrelate neuronal activity. Intrinsic
plasticity on the other hand enforces sparse activity (this sparsening has to happen on the
time scale of input presentations). In our probabilistic model P, sparse network activity
is enforced by a prior over network activities, implemented in M through the inhibitory
feedback that models experimentally found network connectivity [Avermann et al., 2012].
This inhibition naturally acts on a fast time scale [Okun and Lampl, 2008], while the time
scale for intrinsic plasticity is unclear [Turrigiano and Nelson, 2004].
4 Methods
4.1 Definition of M: Data-based network model for a layer 2/3 micro-
circuit motif
The layer 2/3 microcircuit motif was modeled in [Jonke et al., 2017] by the data-based
model M. The model is described here briefly for completeness. See [Jonke et al., 2017]
for a thorough definition. The model M consists of two reciprocally connected pools
of neurons, an excitatory pool and an inhibitory pool. Inhibitory network neurons are
recurrently connected. Excitatory network neurons receive additional excitatory synaptic
input from a pool of N input neurons. Fig. 1A summarizes the connectivity structure of
the data-based model M together with connection probabilities. Connection probabilities
have been chosen according to the experimental data described in [Avermann et al., 2012].
, . . . denote the spike times of input neuron i. The output trace yi(t) of
input neuron i is given by the temporal sum of unweighted postsynaptic potentials (PSPs)
Let t(1)
i
, t(2)
i
18
arising from input neuron i:
ǫ(t − t(f )
i
),
yi(t) =Xf
(22)
where ǫ is the synaptic response kernel, i.e., the shape of the PSP. It is given by a double-
exponential function with a rise time constant τr = 1 ms and a fall time constant τf = 10
ms. For given spike times, output traces of excitatory network neurons and inhibitory
network neurons are defined analogously and denoted by zm(t) and Ij(t) respectively.
The network consists of M = 400 excitatory neurons, modeled as stochastic spike
response model neurons [Jolivet et al., 2006], see eqs. (1) and (2). See Sec. 4.2 for a
motivation of network parameter values from a theoretical perspective.
The instantaneous firing rate ρm of neuron m can be re-written (by substituting eq. (2)
in eq. (1)) as:
ρm(t) =
1
τ
exp (γPi wim yi(t) + γα)
exp(cid:16)γPj∈Im wIEIj(t)(cid:17)
.
(23)
Here, the numerator includes all excitatory contributions to the firing rate ρm(t). The
denominator in this term for the firing rate describes inhibitory contributions, thereby
reflecting divisive inhibition [Carandini and Heeger, 2012].
Apart from excitatory neurons there are Minh = 100 inhibitory neurons in the net-
work. Inhibitory neurons are also modeled as stochastic spike response neurons with an
instantaneous firing rate given by
ρinh
m (t) = σrect(uinh
m (t)),
(24)
where σrect denotes the linear rectifying function σrect(u) = u for u ≥ 0 and 0 otherwise.
The membrane potentials of inhibitory neurons are given by
uinh
m (t) = Xi∈Em
wEI zi(t) − Xj∈IIm
wIIIj(t),
(25)
where zi(t) denotes synaptic input (output trace) from excitatory network neuron i, Em
(IIm) denotes the set of indices of excitatory (inhibitory) neurons that project to inhibitory
neuron m, and wEI (wII) denotes the excitatory (inhibitory) weight to inhibitory neurons.
Synaptic connections from input neurons to excitatory network neurons are subject to
STDP. A standard version of STDP is employed with an exponential weight dependency
for potentiation [Habenschuss et al., 2013b], see [Jonke et al., 2017].
The simulations for Fig. 2 are described in detail in [Jonke et al., 2017].
4.2 Network parameter interpretation:
In the section Interpretation of the dynamics of M in the light of P, we have established
a relationship between the parameters of the probabilistic model P and network param-
eters. This relationship was however derived based on a simplified network model that
included for example rectangular EPSPs and direct inhibitory connections without explicit
inhibitory neurons. Nevertheless, one can also determine reasonable parameter settings
for the data-based model M based on a prior on network activity that is defined in the
probabilistic model P. These parameters are the excitability α and the synaptic weights
19
between and within excitatory and inhibitory network neurons.
In this section we start by assuming such a prior eq. (6) with parameters µ = −3.4
(this includes already a correction of the prior for the missing wnorm) and σ2 = 0.35 as well
as a fitting parameter γ = 2 (eq. 1) and deduce the parameters used in the simulations.
2µ−1
As shown above, the neural excitability α is then given by α = 1
2σ2 . We obtain for the
γ
chosen γ = 2:
α =
1
γ (cid:18) 2µ − 1
2σ2 (cid:19) = −5.57.
(26)
The inhibition strength β of the approximate dynamics eq. (3) is replaced by the weight
wIE from inhibitory neurons to excitatory neurons in M. From the probabilistic model,
we determined β as β = 1
γσ2 under the assumption of rectangular inhibitory PSPs. In M
we use double-exponential PSPs instead of rectangular ones. One therefore has to correct
for differences in the PSP integrals. Using this correction, one obtains
β′ = cPSPβ =
cPSP
γσ2 = 1.114,
(27)
where cPSP is the ratio between the integrals over the rectangular PSPs used in the ap-
proximate dynamics and the double-exponential PSPs used in M (cPSP = 0.78 for the
shapes used for M).
The weights wIE can be determined by comparing eq. (2) with eq. (3) with corrected
inhibition strength
wIEIj(t) =
Xj∈Im
β′zj(t).
M
Xj=1
(28)
Under the assumption that the number of spikes in the pool of inhibitory neurons is at
any time (with a slight delay) approximately equal to the number of spikes in the pool of
excitatory neurons, we obtain
pIEwIE = β′,
(29)
where pIE denotes the connection probability from inhibitory to excitatory network neu-
rons. This yields
(30)
wIE = β′ 1
pIE =
cPSP
γσ2pIE = 1.86.
We now first consider the weights wEI from excitatory to inhibitory neurons under the
assumption of no I-to-I connections. In this case, in order to obtain the same number of
spikes in the inhibitory neurons as in the excitatory neurons, each spike from an excitatory
neuron should induce on average one spike within all inhibitory neurons, that is,
wEI,no II¯ǫMinhpEI = 1,
(31)
where ¯ǫ = 0.01/cPSP is the integral over the alpha-PSPs, pEI is the connection probability
from excitatory to inhibitory neurons, and Minh = 100 is the number of inhibitory neurons.
We obtain
wEI,no II = cPSP/pEI = 1.357.
(32)
Without I-to-I connections, this guarantees that excitation is balanced by inhibition. How-
ever, the single spike (on average) will occur on average with a delay of 5 ms. Interestingly,
the I-to-I connections can help to decrease this delay. In particular, if one demands that
the inhibitory spike is elicited with a delay of less than one ms on average, then one can
20
simply increase the weights wEI by some factor cEI = 10, leading to
wEI = wEI,no IIcEI = cPSPcEI/pEI = 13.57.
(33)
Now, each spike in the excitatory population induces in the inhibitory population for
approximately 10 ms a total rate of 1000 Hz, leading to an average delay of 1 ms. Without
I-to-I connections this would however lead to too many successive spikes within these 10
ms. The I-to-I connections can compensate this too large excitation. For an approximately
correct compensation, the first inhibitory spike has to balance out this excitation, which is
approximately achieved by providing exactly the same amount of inhibition to inhibitory
neurons, leading to
wII = cPSPcEI/pII.
(34)
Since pII ≈ pEI, we used wII = wEI for simplicity. These are the parameter values used for
the M model in [Jonke et al., 2017].
4.3 Details to simulations for Figure 5
Creation of basic patterns: Basic patterns were 64-dimensional vectors, each repre-
senting a horizontal or vertical bar on an 8 × 8 two-dimensional pixel array. We defined 16
basic patterns x(1), . . . , x(16) in total, corresponding to all possible horizontal and vertical
bars of width 1 in this pixel array. For a horizontal (vertical) bar, all pixels of a row
(column) in the array attained the value 1 while all other pixels were set to 0. The entries
of the basic pattern vectors were then defined by the values of the corresponding pixels in
the array.
Superposition of basic patterns: To generate an input pattern, basic rate patterns
were superimposed as follows. The number of superimposed basic patters nsup was chosen
between 1 and 3 drawn from the distribution p(nsup = k) = 0.9k0.13−k
l=1 0.9l0.13−l . Then each basic
pattern to be superimposed was drawn uniformly from the set of basic patterns without
replacement. This corresponds to the distribution used in [Jonke et al., 2017]. The input
i=1 x(bp(i))}, where bp(i) denotes the ith basic
P3
vector y was then given by y = max{1,Pnsup
pattern to be superimposed and the max operation is performed element-wise.
Optimization of the generative model: The generative model (6) -- (8) with 20 hidden
causes z was fitted to this data in an iterative manner. One iteration of the fitting
algorithm was performed as follows:
1. draw an input vector y as described above;
2. draw a sample from the approximate posterior pA1(zy, W ), eq. (12);
3. update the parameters W of the model according to eq. (20).
Since the posterior in step (2) is intractable, it was approximated by assuming that a
maximum of 4 hidden causes are active in the posterior distribution (state vectors with
more active hidden causes usually had negligible probabilities). This allowed us to compute
the partition function and therefore to sample hidden state vectors in a straight-forward
manner. Further, we did not consider hidden state vectors with no active hidden state
since those would not lead to any parameter changes.
Parameters of the model and learning rule: A prior distribution p(z) with param-
eters µ = 6 and σ2 = 0.35 was used. The scaling parameter γ was set to 1. Weights wij
21
were initialized with values drawn from a uniform distribution in [0, 0.1]. Weights were
clipped between a minimal value of 0 and a maximal value of 6. A constant learning rate
of η = 0.1 was used. Training was performed for 15000 updates. Network characteristics
(such as KL divergences) were computed every 50th update.
Figure 5B: Every 50th update, we computed the KL-divergence between the true pos-
terior and the approximate posterior DKL(p(zy, W )pA1(zy, W )). In addition we also
computed the KL-divergence between p(zy, W ) and a uniform distribution over state
vectors. In all these divergences, we only considered the distribution over vectors with
at most 4 active hidden causes for tractability (see above). For Fig. 5B, the data was
smoothed using a box-car filter of size 10.
Figure 5C: We considered the input patters given in panel A and computed the hidden
state zmax with the maximum posterior probability (computed as described above) after
learning in the exact and approximate posterior. This hidden state vector was then used
to reconstruct the input pattern by computing y = σLS(W T zmax). Note that this is not a
sample of y but it defines the probability of each individual pixel to be 1.
Figure 5E: Every 50th update of the simulation described above, we computed the KL-
divergence DKL(p(zy, W )pA2(zy, W )) between the true posterior and the posterior
according to approximation A2. For the red curve we used pA2 with wnorm = 0 and the
same µ = 6 as given for the original model. For the yellow curve, we corrected the prior of
the model to have µ = −12. The KL-divergence to the uniform distribution was computed
as described above.
Acknowledgements: Written under partial support by the Human Brain Project of the
European Union #604102 and #720270, and the Austrian Science Fund (FWF): I 3251-
N33.
References
[Avermann et al., 2012] Avermann, M., Tomm, C., Mateo, C., Gerstner, W., and Pe-
tersen, C. (2012). Microcircuits of excitatory and inhibitory neurons in layer 2/3 of
mouse barrel cortex. Journal of neurophysiology, 107(11):3116 -- 3134.
[Berkes et al., 2011] Berkes, P., Orban, G., Lengyel, M., and Fiser, J. (2011). Spontaneous
cortical activity reveals hallmarks of an optimal internal model of the environment.
Science, 331:83 -- 87.
[Bill et al., 2015] Bill, J., Buesing, L., Habenschuss, S., Nessler, B., Maass, W., and Leg-
enstein, R. (2015). Distributed bayesian computation and self-organized learning in
sheets of spiking neurons with local lateral inhibition. PloS one, 10(8):e0134356.
[Bishop, 2006] Bishop, C. M. (2006).
Pattern Recognition and Machine Learning.
Springer, New York.
[Buesing et al., 2011] Buesing, L., Bill, J., B., and Maass, W. (2011). Neural dynamics as
sampling: A model for stochastic computation in recurrent networks of spiking neurons.
PLoS Computational Biology, 7(11):e1002211.
22
[Carandini and Heeger, 2012] Carandini, M. and Heeger, D. J. (2012). Normalization as
a canonical neural computation. Nature Reviews Neuroscience, 13(1):51 -- 62.
[Douglas and Martin, 2004] Douglas, R. J. and Martin, K. A. (2004). Neuronal circuits
of the neocortex. Annual Reviews of Neuroscience, 27:419 -- 451.
[Fino et al., 2012] Fino, E., Packer, A. M., and Yuste, R. (2012). The logic of inhibitory
connectivity in the neocortex. The Neuroscientist, 19(3):228 -- 237.
[Foldiak, 1990] Foldiak, P. (1990). Forming sparse representations by local anti-hebbian
learning. Biological cybernetics, 64(2):165 -- 170.
[Habenschuss et al., 2013a] Habenschuss, S., Jonke, Z., and Maass, W. (2013a). Stochas-
PLoS Computational Biology,
tic computations in cortical microcircuit models.
9(11):e1003311.
[Habenschuss et al., 2013b] Habenschuss, S., Puhr, H., and Maass, W. (2013b). Emer-
gence of optimal decoding of population codes through STDP. Neural computation,
25(6):1371 -- 1407.
[Hyvarinen, 1999] Hyvarinen, A. (1999). Fast and robust fixed-point algorithms for inde-
pendent component analysis. Neural Networks, IEEE Transactions on, 10(3):626 -- 634.
[Hyvarinen et al., 2004] Hyvarinen, A., Karhunen, J., and Oja, E. (2004). Independent
Component Analysis. John Wiley & Sons.
[Isaacson and Scanziani, 2011] Isaacson, J. S. and Scanziani, M. (2011). How inhibition
shapes cortical activity. Neuron, 72(2):231 -- 243.
[Jolivet et al., 2006] Jolivet, R., Rauch, A., Luscher, H., and Gerstner, W. (2006). Predict-
ing spike timing of neocortical pyramidal neurons by simple threshold models. Journal
of Computational Neuroscience, 21:35 -- 49.
[Jonke et al., 2017] Jonke, Z., Legenstein, R., Habenschuss, S., and Maass, W. (2017).
Feedback inhibition shapes emergent computational properties of cortical microcircuit
motifs. arXiv preprint arXiv:1705.07614.
[Kappel et al., 2014] Kappel, D., Nessler, B., and Maass, W. (2014). STDP installs in
winner-take-all circuits an online approximation to hidden markov model learning. PLoS
Comput. Biol, 10:e1003511.
[Kerlin et al., 2010] Kerlin, A. M., Andermann, M. L., Berezovskii, V. K., and Reid, R. C.
(2010). Broadly tuned response properties of diverse inhibitory neuron subtypes in
mouse visual cortex. Neuron, 67(5):858 -- 871.
[Klampfl and Maass, 2013] Klampfl, S. and Maass, W. (2013). Emergence of dynamic
memory traces in cortical microcircuit models through stdp. The Journal of Neuro-
science, 33(28):11515 -- 29.
[Lucke and Eggert, 2010] Lucke, J. and Eggert, J. (2010). Expectation truncation and the
benefits of preselection in training generative models. The Journal of Machine Learning
Research, 9999:2855 -- 2900.
[Lucke and Sahani, 2008] Lucke, J. and Sahani, M. (2008). Maximal causes for non-linear
component extraction. The Journal of Machine Learning Research, 9:1227 -- 1267.
23
[Maass, 2000] Maass, W. (2000). On the computational power of winner-take-all. Neural
Computation, 12(11):2519 -- 2535.
[Neal, 1992] Neal, R. M. (1992). Connectionist learning of belief networks. Artificial
intelligence, 56(1):71 -- 113.
[Neal, 1993] Neal, R. M. (1993). Probabilistic inference using markov chain monte carlo
methods. Technical report, University of Toronto Department of Computer Science.
[Nessler et al., 2013] Nessler, B., Pfeiffer, M., Buesing, L., and Maass, W. (2013). Bayesian
computation emerges in generic cortical microcircuits through spike-timing-dependent
plasticity. PLoS Computational Biology, 9(4):e1003037.
[Ohki et al., 2006] Ohki, K., Chung, S., Kara, P., Hubener, M., Bonhoeffer, T., and Reid,
R. C. (2006). Highly ordered arrangement of single neurons in orientation pinwheels.
Nature, 442(7105):925 -- 928.
[Okun and Lampl, 2008] Okun, M. and Lampl, I. (2008).
Instantaneous correlation of
excitation and inhibition during ongoing and sensory-evoked activities. Nature Neuro-
science, 11(5):535 -- 537.
[Packer and Yuste, 2011] Packer, A. M. and Yuste, R. (2011). Dense, unspecific connec-
tivity of neocortical parvalbumin-positive interneurons: a canonical microcircuit for
inhibition? The Journal of Neuroscience, 31(37):13260 -- 13271.
[Rumelhart and Zipser, 1985] Rumelhart, D. E. and Zipser, D. (1985). Feature Discovery
by Competitive Learning. Cognitive Science, 9(1):75 -- 112.
[Sato, 1999] Sato, M. (1999). Fast learning of on-line EM algorithm. Technical report,
ATR Human Information Processing Research Laboratories, Kyoto, Japan.
[Saund, 1995] Saund, E. (1995). A multiple cause mixture model for unsupervised learn-
ing. Neural Computation, 7(1):51 -- 71.
[Savin et al., 2010] Savin, C., Joshi, P., and Triesch, J. (2010). Independent component
analysis in spiking neurons. PLoS computational biology, 6(4):e1000757.
[Turrigiano and Nelson, 2004] Turrigiano, G. G. and Nelson, S. B. (2004). Homeostatic
plasticity in the developing nervous system. Nature Reviews Neuroscience, 5(2):97 -- 107.
[Wilson et al., 2012] Wilson, N. R., Runyan, C. A., Wang, F. L., and Sur, M. (2012).
inhibitory networks in vivo. Nature,
Division and subtraction by distinct cortical
488(7411):343 -- 348.
24
|
1104.1090 | 1 | 1104 | 2011-04-06T13:20:03 | The Narrow Escape problem in a flat cylindrical microdomain with application to diffusion in the synaptic cleft | [
"q-bio.NC",
"cond-mat.stat-mech",
"physics.bio-ph",
"q-bio.QM"
] | The mean first passage time (MFPT) for a Brownian particle to reach a small target in cellular microdomains is a key parameter for chemical activation. Although asymptotic estimations of the MFPT are available for various geometries, these formula cannot be applied to degenerated structures where one dimension of is much smaller compared to the others. Here we study the narrow escape time (NET) problem for a Brownian particle to reach a small target located on the surface of a flat cylinder, where the cylinder height is comparable to the target size, and much smaller than the cylinder radius. When the cylinder is sealed, we estimate the MFPT for a Brownian particle to hit a small disk located centrally on the lower surface. For a laterally open cylinder, we estimate the conditional probability and the conditional MFPT to reach the small disk before exiting through the lateral opening. We apply our results to diffusion in the narrow synaptic cleft, and compute the fraction and the mean time for neurotransmitters to find their specific receptors located on the postsynaptic terminal. Finally, we confirm our formulas with Brownian simulations. | q-bio.NC | q-bio |
The Narrow Escape problem in a flat cylindrical microdomain
with application to diffusion in the synaptic cleft
Jurgen Reingruber and David Holcman ∗
Abstract
The mean first passage time (MFPT) for a Brownian particle to reach a small target in cellular
microdomains is a key parameter for chemical activation. Although asymptotic estimations of the
MFPT are available for various geometries, these formula cannot be applied to degenerated structures
where one dimension of is much smaller compared to the others. Here we study the narrow escape
time (NET) problem for a Brownian particle to reach a small target located on the surface of a
flat cylinder, where the cylinder height is comparable to the target size, and much smaller than the
cylinder radius. When the cylinder is sealed, we estimate the MFPT for a Brownian particle to hit
a small disk located centrally on the lower surface. For a laterally open cylinder, we estimate the
conditional probability and the conditional MFPT to reach the small disk before exiting through
the lateral opening. We apply our results to diffusion in the narrow synaptic cleft, and compute
the fraction and the mean time for neurotransmitters to find their specific receptors located on the
postsynaptic terminal. Finally, we confirm our formulas with Brownian simulations.
1
Introduction
The problem of computing the mean first passage time (MFPT) for a Brownian particle to reach a small
target located on a surface of a microdomain, also referred to as the Narrow Escape Time (NET) [36, 14],
is ubiquitous in biophysics and cellular biology because it corresponds to determining the forward binding
rate of chemical reactions [39, 2, 38, 10]. Applications of the NET ranges from quantitative analysis for
the resident time of receptors in the postsynaptic density [14, 27, 35, 12], a fundamental microdomain
associated to synaptic transmission and plasticity [8], to scaling laws in physics [7], early steps of viral
infection [11, 19, 18], or the hydrolysis rate of activated phosphodiesterase in rod photoreceptors [23, 24].
Recent analytical approaches lead to asymptotic formula for the NET in a confined geometry [36, 37,
30, 29]. For example, in a three dimensional domain of volume V with isoperimetric ratio of order 1, and
with no bottlenecks, the overall NET to an absorbing circular hole of (dimensionless) radius a centered
at xs on the surface is [30]
τ = V
4aD (cid:18)1 +
L(xs) + N (xs)
2π
a ln a + O(a ln a))(cid:19)−1
,
(1)
where D is the diffusion constant, L(xs) and N (xs) are the principal curvatures at xs. In the case of
a sphere, a precise asymptotic expression with the first three terms was recently obtained in [6], where
the O(1) term depends on the regular part of the Green's function. The NET computations were further
generalized to the case of several holes [15, 21, 6], and to stochastic dynamics with a potential well [28, 35].
However, the NET formula (1) cannot be directly applied to degenerated microdomains where one
dimension is much smaller than the others. This is for example the case for the synaptic cleft separating
pre- and post-synaptic neuronal terminals (Fig. 1a), which can be approximated as a flat cylinder with
height much smaller compared to its width [3]. Furthermore, in retinal rod photoreceptors sustaining
∗Department of Computational Biology (IBENS) and Mathematics, Ecole Normale Sup´erieure, 46 rue d'Ulm 75005
Paris, France.
1
night vision, the outer segment contains thousands of piled flat cylinders that define the photoresponse
and the fidelity of the vision under dim light conditions [25, 13, 23].
The goal of this paper is to extend the NET analysis to degenerate domains. More specifically, we
study the NET of a Brownian particle in a flat cylinder, where the cylinder height h is much smaller
compared to the cylinder radius R (h ≪ R), with a small circular hole of radius a centered on the bottom
cylinder surface (Fig. 1b). In the first part, we will analyze the NET to exit the cylindrical domain when
the boundary is reflecting everywhere except at the small hole, where it is absorbing. Due to the radial
symmetry, the solution of the the mixed boundary value problem can be expanded in terms of Bessel
functions. For a flat cylinder with h ≪ R and R ≫ a, we find that the NET is given by
τ ≈ V
aD
a0(cid:0) h
a(cid:1)√2
R2
+
8D (cid:18)4 ln(cid:18) R
a(cid:19) − 3(cid:19) ,
(2)
where the function a0(cid:0) h
expect that it remains a valid approximation until h ∼ R, in which case a0( h
4 and the leading order
terms in (2) and (1) coincide. We note that the log-contribution in (1) comes from the local property of
the boundary at the hole, whereas in (2) it originates from the degenerated geometry.
a(cid:1)/√2 ∈ [0.07, 0.25] is depicted in Fig. 2a. Although we derive (2) for h ≪ R, we
a ) ∼ 1
In the second part of the paper, we study a cylinder that is open at lateral boundary, and we present
asymptotic estimates for the conditional probability p and the conditional mean time τc that a Brownian
particle reaches the small hole before leaving the domain through the lateral boundary. For example,
for a flat cylinder with h ∼ a and R ≫ a, when the particle starts at the upper boundary at position
(r = 0, z = h) opposite to the small hole, the conditional probability p(0, h) and the conditional mean
time τc(0, h) are ((40), (52) and (67))
τ (0, h) ≈ V
aD
a0(h/a)
I0( πa
2h )
,
p(0, h) ≈ 1 −
2D
R2 ln(R/a)
τ (0, h),
τc(0, h) ≈
√2D
1 −
R2 ln(R/a)
2D
1 −
R2 ln(R/a)
τ (0, h)
I0(
πa
2h
)τ (0, h)
(3)
τ (0, h)
2(ln(R/a))2 ,
where α = R/a ≫ 1, β = h/a ∼ 1 and τ (0, h) is the mean time to reach the small hole when the
cylinder is closed. These asymptotic expressions can be applied to study diffusion in the synaptic cleft,
where synaptic transmission depends on neurotransmitters that are released at the presynaptic terminal
from vesicles and activate receptors located on the opposite post-synaptic neuron (Fig. 1). The trans-
mission efficiency depends crucially on the conditional probability for a diffusing neurotransmitter to hit
a functional receptor before leaving the synaptic cleft.
2 Mean time to find a small target in a bounded cylindrical
compartment
We shall now present our analysis to estimate the MFPT τ (x) for a Brownian molecule, initially located
at position x = (x1, x2, z), to escape a cylinder of radius R and height h (Fig.1) through a small circular
hole of radius a located centrally on the lower surface at z = 0. The cylindrical surface is reflecting,
except for the small hole where it is absorbing. Due to the radial symmetry, the MFPT is a function of
2 and the height z. Using the small hole radius a, we define the dimensionless
1 + x2
parameters and variables
the radius r = px2
x =
r
a
,
y =
z
a
, α =
R
a
,
β =
h
a
,
Ω = V
a3 = πβα2 ,
2
and the scaled MFPT
which is a solution of [26]
τ (x, y) =
τ (r, z) =
D
πR2β
τ (r, z) ,
aD
V
∂
∂x
x
∂
∂x
+
(cid:18) 1
x
∂2
∂y2(cid:19) τ (x, y) = −
τ (x, y) = 0 ,
1
Ω
,
x ∈ Ω
y = 0 , x < 1
∂
∂y
∂
∂y
τ (x, y)(cid:12)(cid:12)(cid:12)y=β
τ (x, y) = 0 ,
y = 0 , 1 < x < α
= 0 ,
∂
∂x
= 0 .
τ (x, y)(cid:12)(cid:12)(cid:12)x=α
(4)
(5)
Our goal is to obtain a solution for (5) and to clarify its dependency on the parameters α and β. To
study the shape of the boundary layer, we note that (5) corresponds to a heat equation where the total
amount of heat produced in Ω is one, independent of α and β. Furthermore, because the scaled radius
of the hole through which the heat dissipates is one, it follows that τ (x, y) has a finite asymptotic limit
in the neighborhood of the hole for α → ∞ and β → ∞.
h
(a)
z
a
(b)
x
r
R
Figure 1: (a) EM picture of a synapse showing a synaptic cleft and the two pre and post-synaptic
terminals. (b) Schematic representation of a Brownian trajectory in a cylinder of height h and radius R
with reflecting boundaries, except at the small absorbing disk of radius a ≪ R, located centrally on the
lower surface. We are interested in the case of a flat cylinder with R ≫ a and h ∼ a.
2.1 Equation for the scaled MFPT τ (x, y)
To derive the solution τ (x, y) of (5), we consider the two domains
and obtain for τ (x, y) the representation
Ωi = {xx < 1} and Ωo = {x1 < x < α},
τ (x, y) =
τi(x, y) ,
τo(x, y) ,
(x, y) ∈ Ωi
(x, y) ∈ Ωo .
(6)
To ensure that τ (x, y) is a solution of (5) in Ω, τ (x, y) and the flux ∂
x = 1, leading to the conditions
∂x τ (x, y) have to be continuous at
τi(1, y)
= τo(1, y) ,
0 < y < β
∂ τi(x, y)
∂x
(cid:12)(cid:12)(cid:12)x=1
=
∂ τo(x, y)
∂x
3
(cid:12)(cid:12)(cid:12)x=1
,
0 < y < β .
(7)
Using a separation of variable method, we expand τi(x, y) and τo(x, y) in series
τi(x, y) =
τo(x, y) =
where
kn =
nπ
β
un(y) = cos (kny)
u0 =
1
√2
,
bnqn(x)vn(y) + wi(x, y)
anpn(x)un(y) + wo(x, y)
∞
Xn=0
Xn=0
∞
2 )π
(n + 1
β
vn(y) = sin(lny)
,
,
ln =
(n ≥ 1) ,
(8)
(9)
(10)
(n ≥ 0) ,
wi(x, y) and wo(x, y) are the inhomogeneous solutions of (5) that vanish at x = 1, and pn(x) and qn(x)
will be derived below in terms of the modified Bessel functions I0(x) and K0(x), and are normalized such
that pn(1) = qn(1) = 1.
The functions vn(y) and un(y) satisfy the orthogonality relations
Z β
0
un(y)um(y)dy =Z β
0
vn(y)vm(y)dy =
β
2
δnm
Z β
0
vn(y)um(y)dy =
β
2
ξnm ,
where
ξnm =
is an orthogonal matrix satisfying
m
ln
2
l2
n − k2
β
√2
√2
π
βln
=
=
2
π
(n + 1
2 )
2 )2 − m2
(n + 1
, m ≥ 1
1
n + 1
2
, m = 0
(11)
(12)
ξpnξpm =
∞
Xp=0
∞
Xp=0
ξnpξmp = δnm .
Using the orthogonality relations, we obtain the expansions
un(y) =
∞
Xm=0
ξmnvm(y) ,
vn(y) =
∞
Xm=0
ξnmum(y) .
(13)
2.1.1 Derivation of pn(x) and wo(x, y)
The equation for τo(x, y) in Ωo is
(cid:18) 1
x
∂
∂x
∂
∂y
x
∂
∂x
+
∂2
∂y2(cid:19) τo(x, y) = −
1
Ω
= 0 ,
τo(x, y)(cid:12)(cid:12)(cid:12)y=0,β
∂
∂x
τo(x, y)(cid:12)(cid:12)(cid:12)x=α
= 0
,
(x, y) ∈ Ωo
and we choose wo(x, y) to satisfy
(14)
∂
∂x
x
(cid:18) 1
x
wo(1, y) = 0 ,
∂
∂x
+
∂
∂y
∂2
∂y2(cid:19) wo(x, y) = −
wo(x, y)(cid:12)(cid:12)(cid:12)y=0,β
= 0 ,
4
,
1
Ω
∂
∂x
(x, y) ∈ Ωo
wo(x, y)(cid:12)(cid:12)(cid:12)x=α
= 0.
The solution for wo(x, y) is
wo(x, y) =
ln x
2πβ −
x2 − 1
4Ω
Inserting τo(x, y) from (9) into (14) yields for pn(x) the equations
.
(15)
(cid:18) 1
x
∂
∂x
x
∂
∂x − k2
pn(1) = 1 ,
1 < x < α
= 0 .
n(cid:19) pn(x) = 0 ,
pn(x)(cid:12)(cid:12)(cid:12)x=α
∂
∂x
F0(knx, knα)
F0(kn, knα)
Using the modified Bessel functions I0(x) and K0(x) and the relations [5] (I′0(x) = I1(x), K′0(x) =
−K1(x), we obtain
,
(16)
with
pn(x) =
F0(x, y) = I0(x)K1(y) + K0(x)I1(y) .
2.1.2 Derivation of qn(x) and wi(x, y)
Proceeding similarly to the previous paragraph, the equation for τi(x, y) in Ωi is
∂
∂x
x
∂
∂x
+
(cid:18) 1
x
τi(x, 0) = 0 ,
∂2
,
∂
∂y
1
Ω
∂y2(cid:19) τi(x, y) = −
τi(x, y)(cid:12)(cid:12)(cid:12)y=β
∂y2(cid:19) wi(x, y) = −
1
Ω
∂2
,
We choose wi(x, y) to satisfy
∂
∂x
x
∂
∂x
+
(cid:18) 1
x
wi(1, y) = 0 , wi(x, 0) = 0 ,
with solution
(x, y) ∈ Ωi
= 0 .
(x, y) ∈ Ωi
= 0 ,
∂
∂y
wi(x, y)(cid:12)(cid:12)(cid:12)y=β
wi(x, y) =
=
cosh(znβ)
cnJ0(znx)
cosh(zn(β − y))
−
cnJ0(znx)(cid:18) cosh(zn(β − y))
cosh(znβ)
x2 − 1
4Ω
− 1(cid:19) ,
∞
1
Ω
1
Ω
∞
Xn=1
Xn=1
J′0(zn)2 Z 1
2
0
where zn are the positive zeros of the Bessel function J0(x), and the coefficients cn are given by
cn =
J0(znx)
x2 − 1
4
xdx .
To derive expression (19), we used the orthogonality relation [5]
Z 1
0
J0(znx)J0(zmx)xdx = δnm
1
2(cid:0)J0(zn)2 + J′0(zn)2(cid:1) .
Inserting τi(x, y) from (8) into (17) gives for qn(x) the equation
∂
∂x
x
(cid:18) 1
x
∂
∂x − l2
n(cid:19) qn(x) = 0 ,
qn(1) = 1 ,
x < 1
and the solution that is regular at x = 0 is
qn(x) =
I0(lnx)
I0(ln)
.
5
(17)
(18)
(19)
(20)
2.1.3 General expression for τ (x, y)
Using the expressions for pn(x), qn(x), wi(x, y) and wo(x, y), the NET solution is
τ (x, y) =
τi(x, y) =
τo(x, y) =
∞
Xn=0
Xn=0
∞
bn
I0(lnx)
I0(ln)
vn(y) + wi(x, y) ,
x ≤ 1
an
F0(knx, knα)
F0(kn, knα)
un(y) +
ln x
2πβ −
x2 − 1
4Ω
,
1 ≤ x ≤ α,
(21)
where the unknown coefficients an and bn will be determined by patching the two expressions at x = 1.
The continuity condition for τ (x, y) at x = 1 gives
and using the expansions in (13), we obtain that an and bn are related by
anun(y) =
∞
Xn=0
bnvn(y) ,
∞
Xn=0
bn =
ξnmam ,
am =
∞
Xm=0
The continuity condition for the flux at x = 1 gives
ξnmbn .
∞
Xn=0
bnln
I1(ln)
I0(ln)
vn(y) −
∞
Xn=0
∞
Xn=1
ankn
F1(kn, knα)
F0(kn, knα)
un(y) =
1
2πβ −
1
2Ω −
∂
∂x
,
wi(x, y)(cid:12)(cid:12)(cid:12)x=1
with
F1(x, y) =
∂
∂x
F0(x, y) = I1(x)K1(y) − K1(x)I1(y) .
This can be rewritten as
where
bnβnvn(y) +
∞
Xn=0
∞
Xn=0
anαnun(y) =
γnun(y) .
∞
Xn=0
α0 = 0 , αn = −kn
F1(kn, knα)
F0(kn, knα)
(n ≥ 1) ,
βn = ln
I1(ln)
I0(ln)
and the γn are implicitly defined by the equation
∞
Xn=0
γnun(y) =
1
2πβ −
1
2Ω −
By using the expansions in (13) we obtain from (23)
.
∂
∂x
wi(x, y)(cid:12)(cid:12)(cid:12)x=1
Xm=0
ξnmγm
∞
Finally, using the relations between an and bn given in (22), we obtain the matrix equations
αmξnmam + βnbn =
αnan +
βmbmξmn = γn .
∞
Xm=0
(βn + αm)ξnmam =
(βm + αn)ξmnbm = γn .
ξnmγm
∞
Xm=0
∞
Xm=0
∞
∞
Xm=0
Xm=0
6
(22)
(23)
(24)
(25)
(26)
(27)
For given α and β, by truncating and numerically solving these equations we find approximated values
for an and bn, and from this we obtain an approximation for τ (x, y). We will analyze the equations for
an and bn in more detail later on.
2.2 MFPT with a uniform initial distribution
We shall first consider the average MFPT τ (x) when the Brownian particle is initially uniformly dis-
tributed at radial position x. Using (21) we obtain
τ (x) =
1
β Z β
0
τ (x, y)dy =
τi(x) =
τo(x) =
∞
1
β
Xn=0
a0√2
+
bn
ln
I0(lnx)
I0(ln)
+
1
β Z β
0
wi(x, y)dy ,
x ∈ Ωi
(28)
ln x
2πβ −
x2 − 1
4Ω
,
x ∈ Ωo.
Expression (28) shows that a0√2
is the averaged MFPT for Brownian particles that are initially uniformly
distributed at x = 1. The expression for τo(x) has an intuitive interpretation: the escape time starting
at x ≥ 1 is the sum of the average time to reach x = 1, plus the escape time starting at x = 1.
The average MFPT τ for particles that are initially uniformly distributed in Ω is
τ =
1
ΩZΩ
= Ωi
Ω
1
τ (x, y)dV = Ωi
Ω
1
α2 τi +
ΩiZΩi
α2 − 1
α2
τo =
τi + Ωo
Ω
τo ,
τ (x, y)dV + Ωo
Ω
1
ΩoZΩo
τ (x, y)dV
where
τi =
τo =
∞
2
β
Xn=0
a0√2
+
wi(x, y)xdxdy
bn
l2
n
I1(ln)
I0(ln)
α2
α2 − 1
2
+
β Z 1
0 Z β
4
4 ln α − 3 +
α2 −
8πβ
0
1
α4
.
(29)
(30)
The time τi is the average MFPT for particles starting uniformly distributed in the inner cylinder Ωi,
and τo is the average MFPT for particles starting uniformly distributed in the annulus Ωo. We shall now
derive asymptotic limits for τi and τo under various conditions.
2.3 Asymptotic expressions for a cylinder with R ≫ a, and a flat cylinder
with R ≫ a and h ≪ R
We will first derive asymptotic expressions for τ for a cylinder with R ≫ a (α ≫ 1) and arbitrary height
h, and we will then focus on a flat cylinder with h ≪ R (β ≪ α). We show that a0(α, β) is the leading
order contribution to τ for α ≫ 1. For a flat cylinder with α ≫ 1 and β/α ≪ 1, we further have that
a0(α, β) ≈ a0(β). To derive a0(β) as a function of β, we consider the limit α → ∞ while β stays bounded
(R → ∞ with finite h). We show that τ (x, y) and τ have finite asymptotic limits for α → ∞ that depend
only on β.
We start by considering the limit α ≫ 1. The function wi(x, y) in (18) is of the order Ω−1 ∼ α−2
and can be neglected, and we have
τi(x, y) ≈
∞
Xn=0
bn
I0(lnx)
I0(ln)
vn(y).
(31)
Because the average time τi starting uniformly distributed in Ωi is similar to the the average time
τ (1) = a0√2
starting uniformly distributed at x = 1, the contribution of τi in (29) is by a factor α−2
7
smaller compared to the contribution of τo, and we arrive at the asymptotic expression
τ ≈ τo ≈
a0(α, β)
√2
+
4 ln α − 3
8πβ
.
The dimensional time τ is
τ ≈ V
aD
τ ≈ V
aD
a0(α, β)
√2
+
R2
8D (cid:18)4 ln(cid:18) R
a(cid:19) − 3(cid:19) .
In particular, for β ≫ ln α and α ≫ 1 we obtain the result
=⇒ τ ≈ V
a0(α, β)
τ ≈
aD
√2
a0(α, β)
√2
.
(32)
(33)
(34)
Equations (33)-(34) show that a0(α, β) is the leading term that determines the average MFPT for α ≫ 1.
To further evaluate τ , we shall now estimate a0(α, β) for a flat cylinder with a small hole, when β ≪ α
and α ≫ 1, by considering the limit α → ∞ while β remains finite (R → ∞ with fixed h). For α → ∞,
the scaled times τ (x, y) and τ have have finite limits that depend on β, and only the dimensional times
τ (r, z) and τ diverge ∼ R2. In this limit, the coefficients αn, βn and γn in (24) and (25) are given by
α0 = 0, αn = kn
K1(kn)
K0(kn)
(n ≥ 1) ,
βn = ln
I1(ln)
I0(ln)
,
γn =
δn0√2πβ
= γ0δn0 ,
(35)
with γ0 = 1√2πβ
, and (26) simplifies to
(βn + αm)ξnmam = ξn0γ0 .
∞
Xm=0
(36)
αn, βn and γn are functions of β only, and hence, also an and bn depend only on β. For α ≫ 1 and
β/α ≪ 1, τ (x, y) in (21) is in first order given by
τ (x, y) =
where we used
∞
Xn=0
a0(β)
√2
bn(β)
I0(lnx)
I0(ln)
vn(y) ,
x ≤ 1
+
∞
Xn=1
an(β)
K0(knx)
K0(kn)
un(y) +
ln x
2πβ
,
1 ≤ x ≪ α ,
F0(knx, knα)
F0(kn, knα) ≈
K0(knx)
K0(kn)
, α ≫ 1 and x ≪ α .
We conclude that the NET for α ≫ 1 and β/α ≪ 1 is in leading order
a(cid:19) − 3(cid:19) .
8D (cid:18)4 ln(cid:18) R
τ ≈ V
a0(β)
√2
aD
R2
+
(37)
(38)
In the next section we shall analyze the behavior of a0 as a function of β.
2.3.1 Behavior of a0(β) as a function of β
To evaluate a0(β) as a function of β, we solve numerically (36) by truncating the series at sufficiently high
values n: in Fig. 2(a) we plot the analytic approach result for a0(β)/√2 and confirm that it agrees well
with results from Brownian simulations that were performed with α = 50. Interestingly, Fig. 2a shows
that the simulation result for β = 40 (when β is comparable to α = 50) still agrees very well with the
8
analytic result derived with the assumption α ≫ β, suggesting that a0(β) remains a good approximation
until values β ∼ α (h ∼ R). As a consequence, this suggests that (38) is an acceptable approximation
for τ until values h ∼ R. Fig. 2a shows that a0/√2 approaches the value 1
4 for large β, thus, from (34)
we recover the narrow escape formula τ ≈ V 4aD [36, 10, 31] derived for a volume with isoperimetric ratio
of order 1. Conversely, (32) shows that the validity of the narrow escape formula V 4aD is not limited to
the range where h ∼ R, but it is already a valid approximation when ln α ≪ β and β & 40 (Fig. 2a).
Hence, we conclude that τ = V 4aD is a good approximation even for an oblate volume with R ≫ h (and
h ≫ a). In the opposite limit β → 0, we find that a0(β)/√2 does not converge towards zero (Fig. 2a), but
limβ→0 a0(β)/√2 ≈ 0.071 (in appendix A we derive an analytical approximation for a0(β) for β → 0). In
Fig. 2b we show the effect of the truncation level n on the value of a0(β)/√2: after n ∼ 100 the steady
state regime is achieved. Finally, in Fig. 2c-d, we compare the values of the coefficients a0(β) and b0(β),
where b0(β) is obtained using (22). The graphs show that b0(β) ≈ a0(β) is a good approximation, and
we will use this in section 2.3.3.
0.24
0.22
0.2
a0(β)
√2
0.14
0.12
0.1
0.08
0.06
0
0.25
0.2
0.15
0.1
0.05
0
0.4
0.35
a0(β)
√2
0.3
0.25
0.2
0
0.08
0.06
0.04
0.02
0
−0.02
−0.04
0
β = 10
β = 20
β = 40
300
400
500
100
200
n
(b)
a0−b0
30
a0
40
10
20
(d)
β
10
20
β
(a)
30
40
a0
√2
b0
√2
10
20
β
30
40
(c)
Figure 2: (a) Graph of a0(β)
as a function of β for α ≫ 1. The values for a0 are obtained by truncating
√2
and numerically solving (36) (for β > 10 we truncate at n = 500). a0(β)/√2 is the average MFPT for
Brownian particles starting uniformly distributed at x = 1. The analytic estimations for a0(β)/√2 are
compared to results from Brownian simulations (data points) obtained for 10 000 Brownian trajectories,
starting uniformly distributed at x = 1 ( for α = 50). (b) Dependency of a0(β)
on the truncation level
√2
n for various β. (c) Comparison between a0(β) and b0(β), where b0 is obtained from (22), showing that
a0(β) ≈ b0(β). (d) Relative difference (a0(β) − b0(β))/a0(β).
9
2.3.2 Boundary layer analysis: particles starting near the absorbing hole
In the neighborhood of the small absorbing window (for x ∼ 1 and y ∼ 1), there is a boundary layer (BL)
where the behavior of τ (x, y) is very different compared to large x and y. In Fig. 3 we study numerically
the shape of the BL using (37). The different panels depict τ (x, y) in the neighborhood of the absorbing
window for various β. The plots show that a boundary layer starts to evolve around β ∼ 0.3, and the
evolution is almost finished for β ∼ 10 (there is no significant difference between the plots for β = 10 and
β = 50). Furthermore, the approximate extent of the boundary layer for large β is ∆x ∼ ∆y ∼ 10.
In Fig. 4a, we show τ (x, β) for particles released on the upper surface at y = β as a function of
x and for various β. Such a situation is relevant at synapses where neurotransmitters are released at
the presynaptic terminal, located opposite to the surface with the postsynaptic density (PSD) where
receptors are clustered [20, 4]. The hole radius a would correspond to the radius of the PSD. Fig. 4(a)
shows that when the height of the synaptic cleft is comparable to the radius of the PSD (β ∼ 1), τ (x, β)
changes considerably as a function of the radial release position x. In contrast, for β ≫ 1 the release site
is outside the boundary layer and the NET is almost independent of x and is well approximated by 1/4.
0.25
0.2
τ (x, y)
0.15
0.1
0.05
0
0
0.25
τ (x, y)
0.2
0.18
0.16
0.14
0.12
0
y=0.01,0.05,0.1,0.2,0.3
β=0.3
y=0.1,0.2,0.3,0.4,0.6,0.8,1
β=1
0.25
0.2
τ (x, y)
0.15
0.1
0.05
0.5
x
1
1.3
0
0
0.5
1
x
1.5
2
0.25
τ (x, y)
0.2
0.18
0.16
0.14
0.12
0
β=10
y=1,2,3,4,6,10
2
4
x
6
8
β=50
y=1,2,3,4,6,10,50
2
4
x
6
8
Figure 3: Shape of the boundary layer at the absorbing hole for α ≫ 1 and different values of β. The
NET τ (x, y) (from (37)) as a function of x for different y and β (as displayed in each panel).
2.3.3 Impact of truncating the series for τ (x, y) in(37)
We now study the error induced by truncating the sum in (37) at levels n ∼ 1 by considering the truncated
series
n
Xi=0
a0(β)
√2
τ (n)(x, y) =
bi(β)
I0(lix)
I0(li)
vi(y) ,
x ≤ 1
+
n
Xi=1
ai(β)
K0(kix)
K0(ki)
ui(y) +
ln x
2πβ
,
1 ≤ x ≪ α ,
10
(39)
β=0.4
β=0.7
β=1
β=2
β=5
0.9
0.8
0.7
0.6
τ (x, β)
0.4
0.3
0.2
0.1
0
0
0.4
0.3
0.2
0.1
2
4
x
(a)
6
8
0
0
5
τ (0, β)
τ (0, β)ap
15
20
10
β
(b)
Figure 4: (a) The NET τ (x, β) computed from (37) for particles that are released at the upper cylinder
surface at radial position x for various β. We further compare the analytic computations with results
from Brownian simulations obtained with 10000 particles and α = 50. (b) Comparison of τ (0, β) with
the approximation τ (0)(0, β)ap ≈ a0(β)/I0( π
2β ) given in (40).
To evaluate the error induced by the truncation, we first compute the coefficients an and bn with high
precision (using a truncation level n ∼ 200), and then use these values in (39). In Fig. 5, we show the
effect of the truncations for various n and β: interestingly, the numerical analysis reveals that for β . 1,
truncating at n = 0 or n = 1 already provides a very good approximation. The accuracy of the truncation
depends on β, and n has to be increased for larger β in order to maintain a similar accuracy (Fig. 5c).
In Fig. 4a-c, we plot the effect of the truncation as a function of x for y = β (particles are released at the
upper surface), and in Fig. 4d-f, the starting position y is reduced to y = 0.7β, y = 0.4β and y = 0.1β.
a0(β)
Due to the truncation, at the patching boundary x = 1, (39) has a small discontinuity ∆(n)(y) =
τ (n)(1+, y) − τ (n)(1−, y). For example, for n = 0 and y = β (see Fig. 5a-c) we obtain ∆(0)(β) =
√2 − b0(β) ≈ a0(β)(1 − 2√2
β), using the truncation n = 0, we obtain from (39)
Finally, for β . 1, when a Brownian particle is released at the center of the upper surface (x = 0, y =
), where we used a0(β) ≈ b0(β) (see Fig. 2c-d).
τ (0, β) ≈ τ (0)(0, β) =
b0(β)
I0( π
2β ) ≈
a0(β)
I0( π
2β )
=
√2
I0( π
2β )
τ (1) ,
(40)
where we additionally used that a0(β ≈ b0(β) and τ (1) = a0/√2. In Fig. 4b, we test this approximation
as a function of β by comparing it with τ (0, β), computed with high accuracy (n ∼ 200). We find that
this approximation is valid until β ∼ 1.
2.3.4 Analogy with the electrified disk problem
For β ≫ 1 and α ≫ 1, an asymptotic solution for τ (x, y) can be obtained by considering the analogy with
the electrified disk problem: the total outflux −1 through the hole leads to the electrified disk problem
with a disk charge Q = −1, and using the capacitance C = 4 of the unit disk [33, 32], we find that
the disk potential is Q/C = −1/4. Using the solution U (x, y) of the electrified disk problem with disk
potential −1/4 [33], we obtain the asymptotic correspondence
τ (x, y) = U (x, y) +
1
4
, α ≫ 1 , β ≫ 1 .
(41)
Hence, for large β, U (x, y) determines the shape of the boundary layer. Furthermore, by comparing (41)
with (37), we recover that a0√2
= 1
4 for β ∼ α ≫ 1.
11
)
β
,
x
(
)
n
(
τ
0.5
0.4
0.3
0.2
0.1
0
0
)
y
,
x
(
)
n
(
τ
0.3
0.25
0.2
0.15
0.1
0.05
0
0
3
n=0
n=1
n=2
1
x
2
(a) β = 0.5
n=0
n=1
n=2
3
1
2
x
)
β
,
x
(
)
n
(
τ
)
y
,
x
(
)
n
(
τ
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.3
0.25
0.2
0.15
0.1
0.05
0
0
3
n=0
n=1
n=2
1
2
x
(b) β = 1
n=0
n=1
n=2
3
1
2
x
)
β
,
x
(
)
n
(
τ
)
y
,
x
(
)
n
(
τ
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.3
0.25
0.2
0.15
0.1
0.05
0
0
3
n=0
n=1
n=2
n=5
1
2
x
(c) β = 10
n=0
n=1
n=2
n=5
3
1
2
x
(d) β = 1 and y = 0.7 × β
(e) β = 1 and y = 0.4 × β
(f) β = 1 and y = 0.1 × β
Figure 5: The truncated NET τ (n)(x, y) from (39) for various truncation levels n and several β. The
figure shows that for β . 1 very good results can be obtained by considering a low truncation level n ∼ 2.
(a)-(c): τ (n)(x, β) for various β. (d)-(f): τ (n)(x, y) for β = 1 and various y.
3 Conditional probability to reach the small target before leav-
ing a laterally open cylinder
When the cylinder is open at the lateral boundary, we shall now compute the conditional probability
p(x, y) that a Brownian particle, initially at position (x, y), reaches the small target disk before leaving
the cylinder through the lateral opening. Because the geometry of the synaptic cleft can be approximated
by a laterally open cylinder [1, 22, 34], we will use our computations to estimate the efficiency of receptor
activation at a synapse.
Indeed, at the presynaptic site, vesicles release neurotransmitters into the
synaptic cleft, and the diffusing neurotransmitter either bind to and thereby activate receptors located
on postsynaptic terminal, or they leave the synaptic cleft without activating a receptor. We will first
derive a general expression for the conditional probability p(x, y) to hit the small target before exiting,
and then compute average values for uniform initial distributions. Finally, we will determine the leading
order behavior in a flat cylinder with R ≫ a and h ≪ R.
The conditional probability p(x, y) satisfies the Laplace equation [26, 16]
∂
∂x
x
(cid:18) 1
∂
∂x
+
x ∈ Ω
∂2
∂y2(cid:19) p(x, y) = 0 ,
x
p(x, 0) = 1 for x < 1 ,
p(x, α) = 0
(42)
∂
∂y
p(x, 0) = 0 for 1 < x < α ,
∂
∂y
p(x, β) = 0 .
Similarly to the analysis of τ (x, y) in the previous section, we solve (5) in the subdomains Ωi and Ωo,
and then patch the two solutions pi(x, y) and po(x, y) at the boundary x = 1. The general expressions
12
are
where
p(x, y) =
vn(y) ,
x ≤ 1
∞
bp
n
1 −
I0(lnx)
I0(ln)
Xn=0
0√2(cid:19) ln(cid:0) α
x(cid:1)
(cid:18)1 −
ln α −
ap
ap
n
G0(knx, knα)
G0(kn, knα)
un(y) ,
1 ≤ x ≤ α
∞
Xn=1
(43)
and the unknown coefficients ap
(22). The coefficient ap
n resp. bp
G0(x, y) = I0(x)K0(y) − K0(x)I0(y),
n and bp
n are given by
n are functions of α and β and are related by a relation similar to
(βn + αp
m)ξnmap
m = ξn0γp
0
(βm + αp
n)ξmnbp
m = γp
0 δn0 ,
∞
∞
Xm=0
Xm=0
with
and
αp
0 =
1
ln α
, αp
n = −kn
G1(kn, knα)
G0(kn, knα)
(n ≥ 1) ,
βn = ln
I1(ln)
I0(ln)
,
γp
0 =
√2
ln α
,
(44)
(45)
G1(x, y) =
∂
∂x
G0(x, y) = I1(x)K0(y) + K1(x)I0(y) .
For the βn we omitted the superscript p because they coincide with the βn already defined in (24).
3.1 Conditional probabilities with uniform initial distributions
The fraction p(x) of Brownian particles that eventually reach the target starting initially uniformly
distributed at x is
p(x) =
1
β Z β
0
p(x, y)dy =
,
∞
bp
n
βln
I0(lnx)
I0(ln)
1 −
Xn=0
0√2(cid:19) ln(cid:0) α
x(cid:1)
(cid:18)1 −
ln α
ap
x ≤ 1
ln(cid:0) α
x(cid:1)
ln α
= p(1)
(46)
,
1 ≤ x ≤ α
where p(1) = 1−ap
is
0/√2. For particles starting initially uniformly distributed in Ω, the average probability
p =
2
α2 Z α
0
p(x)xdx = p(1)
α2 − 2 ln α − 1
2α2 ln α
+
1
α2 −
2
βα2
∞
Xn=0
bn
l2
n
I1(ln)
I0(ln)
.
(47)
13
When the particles are initially uniformly released at the upper surface within an area x ≤ x0, the fraction
pβ(x0) that will reach the target before escaping through the lateral opening is
pβ(x0) =
2
x2
0 Z x0
0
p(x, β)xdx =
2
x0
1 −
(−1)n bp
nβn
l2
n
∞
Xn=0
I1(lnx0)
I1(ln)
,
x0 ≤ 1
2
x2
0 (cid:18)1 −
nαp
(−1)n ap
n
k2
ap
0(cid:16)1 + 2 ln( α
0√2(cid:19) x2
n (cid:18) x0G1(knx0, knα)
G1(kn, knα) − 1(cid:19) ,
x0
4 ln α
)(cid:17) − (1 + 2 ln α)
1 ≤ x0 ≤ α .
(48)
pβ(1)
x2
0
2
x2
0
+
+
∞
Xn=1
3.2 Asymptotic expressions for the conditional probability in a flat cylinder
with α ≫ 1 and β ≪ α
To obtain an asymptotic expression for the conditional probability p in the limit α ≫ 1 and β ≪ α, we
estimate ap
n and bp
n. For α ≫ 1 and β/α ≪ 1 we have
αp
0 =
1
ln α
, αp
n ≈ kn
K1(kn)
K0(kn)
= αn, (n ≥ 1) ,
n = ln α
2πβ ap
where αn are given in (35). For the scaled coefficients ap
equations
n we obtain from (44) the asymptotic
(βn +
1
ln α
)ξn0ap
0 +
(βn + αm)ξnmap
m =
1
√2πβ
ξn0 .
(49)
∞
Xm=1
1
ln α ≪ βn, (49) reduces to (36) for the coefficients an of τ (x, y), hence, ap
When α is large such that
Because βn are monotonically increasing with n, if
by setting n = 0, we obtain the condition 1
ln α ≪ π
large β when ln α ≫ β2. In this limit, the asymptotic expressions for ap
n ≈ an.
ln α ≪ βn for n = 0, this is also valid for n > 0. Hence,
2β )
2β ) , satisfied for small β when ln α ≫ β, and for
n and bp
I1( π
I0( π
n are
2β
1
ap
n =
an
and bp
n =
2πβ
ln α
2πβ
ln α
bn .
Finally, by inserting ap
n and bp
n into (43) we obtain for for ln α ≫ 2β
π
I0( π
I1( π
2β )
2β ) the asymptotic formula
p(x, y) =
1 −
p(1)
∞
2πβ
ln α
Xn=0
ln(cid:0) α
x(cid:1)
ln α −
a0√2
bn
I0(lnx)
I0(ln)
vn(y) ,
x ≤ 1
2πβ
ln α
∞
Xn=1
an
K0(knx)
K0(kn)
un(y) ,
1 ≤ x ≪ α
where p(1) = 1 − ap
0√2
= 1 − 2πβ
ln α
(46) and we used
G0(knx, knα)
G0(kn, knα) ≈
K0(knx)
K0(kn)
, α ≫ 1 and x ≪ α .
By considering (37) for τ (x, y), this can be rewritten as
p(x, y) =
2πβ
ln α
1 −
τ (x, y) ,
1 +
2πβ
ln α 1 −
x ≤ 1
ln α ! τ (1) −
ln(cid:0) α
x(cid:1)
14
2πβ
ln α
τ (x, y) ,
1 ≤ x ≪ α
(50)
(51)
(52)
showing an interesting relation between the conditional probability and the MFPT. As an example, for
particles that start uniformly distributed at x = 1 we have
p(1) = 1 −
2πβ
ln α
τ (1) ,
(53)
and using that τ (1) . 1
particles that start initially uniformly distributed in Ω we obtain from (47)
4 , we find that the probability p(1) approaches one for large α. In contrast, for
p(1)
2 ln α
p ≈
, α ≫ 1 ,
(54)
which shows that p tends to zero for large α, in contrast to p(1).
3.3 Numerical evaluations
In Fig. 6a we plot p(1) = 1 − ap
as a function of α for various β (we numerically solve (44) with high
0√2
accuracy up to n ∼ 400). As α → 1, the probability p(1) tends to zero because the particles start close
to the lateral boundary. The asymptotic limit of p(1) for large α is one, as shown in (53). However, the
convergence is only logarithmical and therefore very slow. In Fig. 6b we display p(1) as a function of β
for various α: for a fixed value α > 1, the asymptotic limit of p(1) for β → 0 is one, and zero for β → ∞.
The limit for β → 0 is intuitive, because the particles start next to the target. To obtain the asymptotic
limit for β → ∞ we consider (44) for ap
n (n ≥ 0)
0 = γp
tend to zero for large β, the non-vanishing part of (44) is αp
0 ξn,0, from which it follows that
0/√2 converges to zero in this limit.
the asymptotic limit of ap
Indeed, for fixed α and increasing β, it becomes less probable that a particle reaches the disk before the
lateral boundary.
0ξn0ap
0 for large β is √2, and thus p(1) = 1 − ap
n and (45): because the coefficients αp
n (n ≥ 1) and βp
1
0.8
p(1)
0.6
0.4
0.2
0
0
β=1
β=5
β=10
200
400
α
600
800
1000
(a)
1
0.8
p(1)
0.6
0.4
0.2
0
0
α=2
α=10
α=50
α=100
α=1000
15
20
5
10
β
(b)
Figure 6: Average conditional probability p(1) = 1 − ap
for Brownian particles starting uniformly
0√2
distributed at x = 1 to reach the target before leaving the cylinder through the lateral boundary. The
coefficient ap
0 are obtained by truncating and numerically solving (44) with high accuracy. (a) p(1) as a
function of α for different β. (b) p(1) as a function of β for different α.
We now study the conditional probability for particles that are released at the upper surface, which
is relevant for synaptic transmission. In Fig. 7a-c we plot p(x, β) as a function of x for α = 10, α = 5
and α = 100, and various β between 0.1 and 10, and in Fig. 7a we further show that our analytical
computations agree with results from Brownian simulations. In Fig. 7d we further depict the average
probability pβ(x0) when particles are released at the upper surface within an area of radius x0 for similar
values α and β as in panel (c). In general, Fig. 7 shows that p(x, β) is very sensitive to the cylinder
height β and the release position x. For very flat cylinders with β . 1, when the particles are released in
15
1
0.8
0.6
p(x, β)
0.4
0.2
0
0
1
0.8
0.6
p(x, β)
0.4
0.2
0
0
β=0.1
β=0.5
β=1
β=2
β=3
β=5
β=10
1
0.8
0.6
0.4
0.2
β=0.1
β=0.5
β=1
β=2
β=3
β=5
5
x
(a) α = 10
10
0
0
1
2
x
3
(b) α = 5
4
5
1
0.8
0.6
pβ(x0)
0.4
0.2
0
0
5
x
(c) α = 100
10
5
x0
(d) α = 100
10
Figure 7: Panels a-c: Conditional probability p(x, β) (from (43)) for a Brownian particle to reach the
target at (x < 1, y = 0) when released at the upper surface at y = β and radial position x. The data
points in (a) are obtained from Brownian simulations with 103 particles. (c) The values for β are as
in panel (a).
(d) Average conditional probability pβ(x0) (from (48)) when the particles are released
uniformly on the upper surface in an area of radius x0 with values β as in (c).
the area opposite to the hole (for x < 1), the probability is ∼ 1, whereas when they are released outside
this region (for x > 1), the probability that they reach the target before exiting decreases considerably
(Fig. 7a-c). For example, Fig. 7b, obtained for β = 0.5 and α = 5, shows that the conditional probability
is larger than 90% when the particles are released at the upper surface within the area x < 1, whereas it
decreases to around 60% when they are released at x ∼ 2. For cylinder with large α the impact of the
release position is much less pronounced (panel (d) with α = 100). In addition, Fig. 7 shows that the
conditional probability to reach the target is more sensitive to changes in the cylinder height β compared
to the radius α.
3.4
Impact of truncating the series for p(x, y) in (43)
We now proceed similarly to section 2.3.3 and estimate the error induced by truncating the series in (43)
at small n. We first compute the coefficients ap
i using (44) with a high accuracy, and then use
i and bp
16
these coefficients to define the truncated probability
vi(y) ,
x ≤ 1
p(n)(x, y) =
ap
n
bp
i
1 −
I0(lix)
I0(li)
Xi=0
0√2(cid:19) ln(cid:0) α
x(cid:1)
(cid:18)1 −
ln α −
ap
i
G0(kix, kiα)
G0(ki, kiα)
ui(y) ,
1 ≤ x ≤ α
n
Xi=1
(55)
In Fig. 8, we plot p(n)(x, β) for various α and β as a function of x, and show that truncating at n = 0, 1, 2
already gives good approximations (we plot p(n)(x, β) for n = 100 to show the error induced by the low
truncations). Similarly as in section 2.3.3, we conclude that truncation at n = 1 or n = 2 already provides
a good approximation for β . 1.
1
β = 0.5
)
β
,
x
(
)
n
(
p
0.8
0.6
0.4
0.2
β = 1
β = 2
β = 5
0
0
n=100
n=0
n=1
n=2
)
β
,
x
(
)
n
(
p
2
3
1
x
(a) α = 5
β = 0.5
1
0.8
β = 1
0.6
β = 2
n=100
n=0
n=1
n=2
0.4
0.2
0
0
β = 5
1
2
x
(b) α = 100
3
4
Figure 8: The truncated conditional probability p(n)(x, β) from (55) for various n, β and α. Similar to
Fig. 5, for β . 1, we already obtain good approximations for n ∼ 2.
4 Conditional mean time to reach a small target before escaping
a laterally open cylinder
After having estimated the conditional probability p(x, y), we are now in a position to study the con-
ditional MFPT τc(x, y) to reach a target before escaping through the lateral cylinder boundary. This
analysis will estimate the time scale of the synaptic response. For example, the conditional time τc(0, β)
in (67) provides an estimate for the mean time until postsynaptic receptors become activated after trans-
mitter release into the synaptic cleft at the upper cylinder surface. In a wide cylinder with α ≫ 1, (67)
shows that the conditional time τc(0, β) is roughly by a factor 1/(ln α)2 faster than the mean time τ (0, β)
for a laterally closed cylinder, because in a closed cylinder all trajectories that are reflected at the lateral
boundary contribute to the mean time, thereby increasing the mean compared to an open cylinder. Thus,
to obtain a realistic estimation of the synaptic activation time, we have to consider the conditional time
τc for an open cylinder.
To determine the mean conditional time τc(x, y), we use the known conditional probability p(x, y) to
solve an equation for the function A(x, y) = τc(x, y)p(x, y), and then we obtain τc(x, y) from τc(x, y) =
17
A(x, y) =
∞
Xn=0
aA
0√2
bA
n
I0(lnx)
I0(ln)
vn(y) + wA
i (x, y) ,
x ≤ 1
ln(cid:0) α
x(cid:1)
ln α
+
∞
Xn=1
aA
n
G0(knx, knα)
G0(kn, knα)
un(y) + wA
o (x, y) ,
1 ≤ x ≤ α,
where wA
n and bA
aA
i (x, y) and wA
n are related as an and bn in (22), and aA
n satisfies
o (x, y) are the inhomogeneous solutions of (56) that vanish at x = 1. The coefficients
(βA
n + αA
m)ξnmaA
m =
∞
Xm=0
ξnmγA
m ,
∞
Xm=0
where
αA
0 =
1
ln α
, αA
n = −kn
G1(kn, knα)
G0(kn, knα)
(n ≥ 1) ,
βA
n = ln
I1(ln)
I0(ln)
,
and the γA
n are implicitly defined through
(56)
(57)
(58)
(59)
(60)
A(x,y)
p(x,y) . The function A(x, y) satisfies [9, 35, 17]
∂
∂x
x
∂
∂x
+
(cid:18) 1
x
∂2
∂y2(cid:19) A(x, y) = −
A(x, y) = 0 ,
A(x, y) = 0 ,
,
p(x, y)
Ω
y = 0 , x < 1
x = α
(x, y) ∈ Ω
∂
∂y
∂
∂y
A(x, y) = 0 ,
y = 0 , x > 1
A(x, y) = 0 ,
y = β .
Proceeding as in the previous sections, we expand A(x, y) as
∂
∂x(cid:0)wA
o (x, y) − wA
i (x, y)(cid:1)(cid:12)(cid:12)(cid:12)x=1
=
∞
Xn=0
γA
n un(y) .
To determine the coefficients γA
i (x, y) is of the
order α−2 and we neglect its contribution in first approximation. Using p(x, y) from (43), the equation
for wA
o (x, y). When α ≫ 1, wA
n , we first evaluate wA
i (x, y) and wA
o (x, y) is
∂
∂x
x
∂
∂x
+
(cid:18) 1
x
∂2
∂y2(cid:19) wA
o (x, y) = −
wA
o (x, y) = 0 ,
wA
o (x, y) = 0 ,
∂
∂y
1
Ω p(1)
ln(cid:0) α
x(cid:1)
ln α −
x = 1 and x = α
∞
Xn=1
ap
n
G0(knx, knα)
G0(kn, knα)
un(y)!
(61)
y = β and y = 0 .
To solve (61) we expand wA
o (x, y) in terms of un(y),
wA
o (x, y) = wA(0)
o
(x) +
∞
Xn=1
wA(n)
o
(x)un(y) ,
and inserting this expansion into (61) gives for wA(0)
o
(x) the solution
wA(0)
o
(x) = −
p(1)
4Ω ln(α)(cid:18)x2(ln(cid:16) α
x(cid:17) + 1) − (α2 − ln α − 1)
ln x
ln α − (ln α + 1)(cid:19) .
(62)
18
The higher order functions wA(n)
o
∂
∂x
x
(cid:18) 1
x
∂
∂x − k2
(x) (n ≥ 1) satisfy the equation
n(cid:19) wA(n)
G0(knx, knα)
G0(kn, knα)
(α) = 0 .
(1) = 0 , wA(n)
o
wA(n)
ap
n
Ω
(x) =
o
o
,
1 < x < α
To proceed, we now truncate the series for wA
o (x, y) in (62) at n = 0 and use only the first order
o (x, y) ≈ wA(0)
approximation wA
(x). We expect that this already provides a good approximation because
the coefficients ap
n are small for for n ≥ 1 and large α, and, as shown in section 3.4, truncation at n = 0
already gives a very good approximation for p(x, y when β . 1. Hence, we expect that our analysis is a
valid approximation for large α and small β ∼ 1. With truncation at n = 0, the parameters γA
n defined
in (60) are
o
γA
n = γA
0 δn0 , with γA
0 = p(1)
α2 − 2(ln α)2 − 2 ln α − 1
2√2πβα2(ln α)2
1
√2πβ
p(1)
2(ln α)2 .
≈
(63)
Similar to (50), for ln α ≫
2β
π
I0( π
I1( π
2β )
2β )
, the asymptotic solutions for aA
n and bA
n in terms of an and bn are
aA
n =
p(1)
2(ln α)2 an
and bA
n =
p(1)
2(ln α)2 bn .
Finally, we obtain for τc(x, y) the approximation
(64)
(65)
τc(x, y) =
A(x, y)
p(x, y)
=
1
p(x, y)
∞
Xn=0
aA
0√2
bA
n
I0(lnx)
I0(ln)
vn(y) =
p(1)
2(ln α)2 τ (x, y) ,
x ≤ 1
ln(cid:0) α
x(cid:1)
ln α
+
∞
Xn=1
aA
n
K0(knx)
K0(kn)
un(y) + wA(0)
o
(x) ,
1 ≤ x ≪ α
where p(x, y) is given in (51). To estimate how much τc(x, y) is faster compared to τ (x, y), we consider
particles that are released centrally on the upper surface at position (0, β). By taking into account (52)
for p(x, y), we obtain
τc(0, β) =
1 − 2πβ
1 − 2πβ
ln α τ (1)
ln α τ (0, β)
τ (0, β)
2(ln α)2 .
Furthermore, using from (40) the approximation τ (0, β) ≈ a0(β)
approximation
I0( π
2β ) =
(66)
√2
2β ) τ (1), we obtain for β . 1 the
I0( π
1 −
τc(0, β) ≈
2β )τ (0, β)
√2πβ
ln α I0( π
1 − 2πβ
ln α τ (0, β)
τ (0, β)
2(ln α)2 .
(67)
In Fig. 9 we plot a τc(x, β) as a function of x for various β, and we confirm that our analysis agrees well
with results from Brownian simulations. Comparing Fig. 9 with Fig. 4 for τ (x, y) shows that τc(x, β) is
roughly by a factor (2 ln α)2 faster compared to τ (x, y).
5 Discussion and conclusion
We generalized here the narrow escape problem to a degenerated geometry defined by a flat cylindrical
compartment of height h and radius R, where the absorbing hole is a small circular disk of radius a
located centrally on the lower surface. We analyzed the problem for a laterally closed and open cylinder.
19
0.05
0.04
τc(x, β)
0.03
β=0.4
β=0.7
β=1
β=2
β=5
β=8
0.02
0.01
0
0
2
4
x
6
8
Figure 9: Conditional MFPT τc(x, β) to reach the hole in a laterally open cylinder for particles that are
released on the surface opposed to the hole. The numerical values for τc(x, β) are calculated using (65)
and α = 50. The data points are the results of Brownian simulations with 10000 particles and α = 50.
Because a uniform analytic expansion of the solution in the whole domain is not possible, we derived two
different expansions in the two subregions Ωi and Ωo, and then matched them at the boundary between
the two subcompartments.
We first analyzed the narrow escape time (NET) τ (r, z) to reach the small hole in a laterally closed
cylinder. For a flat cylinder with h ≪ R and R ≫ a, we obtained that the NET (38) is given by
τ ≈ V
aD
a0(cid:0) h
a(cid:1)√2
+
R2
8D (cid:18)4 ln(cid:18) R
a(cid:19) − 3(cid:19) ,
(68)
where the function a0(β) ∈ [0.07, 0.25] (Fig. 2a). Although (68) was derived in the condition that h ≪ R,
In particular, for a
our numerical results suggest that it remains valid until h ∼ R (section 2.3.1).
a(cid:1) we recover the well known NET approximation τ ≈ V 4aD [36, 10, 14], which
cylinder with h/a ≫ ln(cid:0) R
was derived for non degenerated geometries (isoperimetric ratio of order 1 and no bottle neck). However,
our analysis for the cylinder revealed that τ = V 4aD is already a very good approximation for an oblate
volume with an isoperimetric ratio that can be very different from 1 (although h/a ≫ ln(cid:0) R
a(cid:1), the ratio
h/R can still be small). Formula 68 can also be used to estimate the rate constant of a key chemical
reaction during the early stage of phototransduction, which is the rate for diffusing cGMP molecules to
reach the phosphodiesterase enzyme located on the surface of a narrow cylinder, located in the outer
segment of a rod photoreceptor [23, 24].
In a next step, we used our method to analyze the narrow escape problem for a flat and laterally open
cylinder, which is relevant for synaptic transmission. In many cases, the geometry of the synaptic cleft is
well approximated by a laterally open cylinder [1, 22, 34], where neurotransmitters are released into the
synaptic cleft from the presynaptic terminal, located on the upper surface. The neurotransmitters move
in the synaptic cleft (Fig. 1) by Brownian diffusion, and they either activate receptors clustered in the
postsynaptic density located on the postsynaptic terminal (corresponding to the lower cylinder surface),
or they leave the synaptic cleft through the lateral boundary without binding to a receptor.
We estimated the conditional probability p(r, z) that a particle starting at position (r, z) reaches the
small hole before leaving the cylinder. By identifying the small hole with the postsynaptic density where
receptors are clustered, we estimated the fraction of released neurotransmitters that reach the receptor
area before leaving the synaptic cleft. Using our analytic solution, we studied the impact of the synaptic
cleft geometry as well as the location of neurotransmitter release. In Fig. 7, we plotted the probability
to reach the postsynaptic density before leaving as a function of the release position for various cylinder
height, and we found that it is very sensitive to the release position and the width of the cylinder. We
conclude that, in order to achieve an efficient activation of postsynaptic receptors, the presynaptic and
postsynaptic densities should be properly aligned such that the neurotransmitters are released opposite
20
to the receptors. Finally, we computed the conditional mean time to reach the small hole before leaving
through the lateral boundary (see (65)). For a wide cylinder, we found that the conditional mean time
is roughly by a factor (2 ln α)2 faster compared to the NET in a closed cylinder (see (67)).
We shall now present some numerical estimates for neurotransmitters that need to activate receptors
clustered in the postsynaptic density with a radius a = 50nm, when the synaptic cleft has a height
h = 20nm and a total radius of R = 500nm, so that α = R/a = 10 and β = h/a = 0.4. When the
transmitter are released at distance r away from the center, the conditional probability is approximately
given by truncating (43) at n = 0
p(r, h) =
0
1 − bp
(cid:18)1 −
I0( πr
2h )
I0( aπ
2h )
,
x ≤ 1
0√2(cid:19) ln(cid:0) R
r(cid:1)
ln R/a
ap
,
1 ≤ r ≤ R
(69)
which is a very good approximation, as shown in Fig. 8. Using a diffusion constant D = 200µm2/s, we
obtain for the mean times τ (0, h) ≈ 17µs and τc(0, h) = 1µs.
The exact solution for the mean time and the conditional probability were obtained here by using
the patching eigenfunction expansion approach, and from these approximations, we derived in the limit
of large aspect ratio R/h ≫ 1 the asymptotic behavior. Our approach works well because of the radial
symmetry due to the absorbing trap that is is located at the center of the cylinder. This situation accounts
well for the postsynaptic density located at the center of the post-synaptic terminal. However, using our
method it would be difficult to treat the case of multiple non-concentric traps, and in this case a different
approach based on matching asymptotic analysis should be more appropriate [21, 6]. The analysis should
start with an explicit representation of the Green's function for a cylinder. Once an inner solution is
determined near each trap, it should be matched to the outer solution [21]. This method should allow to
study the effect of the trap positions and trap clustering on the synaptic current, which was only partially
discussed in [34, 15].
Acknowledgements
D.H. research is supported by an ERC Starting Grant.
21
APPENDIX
A Equation for the parameters an in the limit β ≪ 1
In order to find the asymptotic equations for the coefficients an for α ≫ 1 and β → 0, we introduce the
scaled quantities
βn = ββn = ln
ln = βln =
I1(ln)
I0(ln)
,
(2n + 1)π
2
,
αn = βαn = kn
kn = βkn = nπ ,
K1(kn)
K0(kn)
,
γ0 = βγ0 =
1
√2π
.
In the limit β → 0 we have ln → ∞ and kn → ∞ for every n > 0, and the asymptotic behaviour of βn
and αn is
Using (36), the asymptotic equation for the coefficients an in the limit β → 0 is
βn ≈ ln , nαn ≈ kn
(ln + km)ξnmam =
∞
Xm=0
1
√2π
ξn0 .
Truncating (70) at various levels n gives the approximations
a0√2
a0√2
a0√2
=
=
=
1
π2 ≈ 0.10 , n = 0
5
1
π2 ≈ 0.085 , n = 1
6
47
1
π2 ≈ 0.078 , n = 2 .
60
a0√2 ≈ 0.071
Numerically we find from (70)
References
(70)
(71)
[1] B. Barbour. An evaluation of synapse independence. J. Neuroscience, 21(20):7969 -- 7984, 2001.
[2] H. C. Berg and M. Purcell. Physics of chemoreception. Biophys. J., 20:193 -- 219, 1977.
[3] J.N. Bourne and K.M. Harris. Balancing structure and function at hippocampal dendritic spines.
Annu. Rev. Neurosci., 31:47 -- 67, 2008.
[4] Harris K.M. Bourne J.N. Balancing structure and function at hippocampal dendritic spines. Annu
Rev Neurosci., 31:47 -- 67, 2008.
[5] H.S. Carslaw and J.C. Jaeger. Conduction of Heat in Solids. Oxford University Press, USA, 2 edition
edition, 1986.
[6] A.F. Cheviakov, M.J. Ward, and R. Straube. An asymptotic analysis of the mean first passage
time for narrow escape problems: Part ii: The sphere. SIAM Multiscale Modeling and Simulation,
8:836 -- 870, 2010.
[7] S. Condamin, O. B´enichou, V. Tejedo, R. Voituriez, and J. Klafter. First-passage times in complex
scale-invariant media. Nature, 450(7166):77 -- 80, 2007.
22
[8] G.M. Elias and Nicoll R.A. Synaptic trafficking of glutamate receptors by maguk scaffolding proteins.
Trends Cell Biol., 7:343 -- 52, 2007.
[9] C.W. Gardiner. Handbook of Stochastic Methods. Springer, third edition, 2003.
[10] I. V. Grigoriev, Y. A. Makhnovskii, A. M. Berezhkovskii, and V. Yu. Zitserman. Kinetics of escape
through a small hole. J. Chem. Phys., 116:9574 -- 77, 2002.
[11] D Holcman. Modeling viral and dna trafficking in the cytoplasm of a cell. J. of Stat. Phys.,
127(3):471 -- 494, 2007.
[12] D Holcman. Computational challenges in synaptic transmission. Book Series: Contemporary Math-
ematics; Source: Imaging Microstructures: Mathematical and Computational challenges, 494:1 -- 26,
2009.
[13] D. Holcman and J.I. Korenbrot. Longitudinal diffusion in retinal rod and cone outer segment cyto-
plasm: The consequence of the cell structure. Biophys. J., 86:2566 -- 2582, 2004.
[14] D. Holcman and Z. Schuss. Escape through a small opening: receptor trafficking in a synaptic
membrane. J. Stat. Phys., 117(5/6):975 -- 1014, 2004.
[15] D. Holcman and Z. Schuss. Diffusion through a cluster of small windows and flux regulation in
microdomains. Phys. Lett. A., 372(21):3768 -- 72, 2008.
[16] S. Karlin and H.M. Taylor. A First Course in Stochastic Processes. Academic Press, 2. edition
edition, 1975.
[17] S. Karlin and H.M. Taylor. A Second Course in Stochastic Processes. Academic Press, 1 edition
edition, 1981.
[18] T. Lagache, E. Dauty, and D. Holcman. Physical principles and models describing intracellular virus
particle dynamics. Curr. Opin. Microbiol., 12(4):439 -- 445, 2009.
[19] T. Lagache and D. Holcman. Quantifying intermittent transport in cell cytoplasm. Phys. Rev. E,
77(3):030901, 2008.
[20] T.M. Newpher and M.D. Ehlers. Glutamate receptor dynamics in dendritic microdomains. Neuron,
58(4):472 -- 97, 2008.
[21] S. Pillay, M.J. Ward, A. Peirce, and T. Kolokolnikov. An asymptotic analysis of the mean first
passage time for narrow escape problems: Part i: Two-dimensional domains. SIAM Multiscale
Modeling and Simulation, 8:803 -- 835, 2010.
[22] S. Raghavachari and J.E. Lisman. Properties of quantal transmission at ca1 synapses. J. Neuro-
physiology, 92(4):2456 -- 2467, 2004.
[23] J. Reingruber and D Holcman. Estimating the rate of cgmp hydrolysis by phosphodiesterase in
photoreceptors. J. Chem. Phys., 129:145192, 2008.
[24] J. Reingruber and D Holcman. Diffusion in narrow domains and application to phototransduction.
Phys. Rev. E, 79(3):030904, 2009.
[25] F. Rieke and D. Baylor. Single photon detection by rod cells of the retina. Rev. of Mod. Phys.,
70(3):1027 -- 1036, 1998.
[26] Z. Schuss. Theory and Applications of Stochastic Differential Equations. Wiley Series in Probability
and Statistics, John Wiley Sons, Inc., New York, 1980.
[27] Z. Schuss, A. Singer, and D Holcman. The narrow escape problem for diffusion in cellular mi-
crodomains. Proc. Natl. Acad. Sci. USA, 104(41):16098 -- 103, 2007.
23
[28] A. Singer and Z. Schuss. Activation through a narrow opening. Phys. Rev. E, 74(2):020103, 2006.
[29] A. Singer, Z. Schuss, and D Holcman. Narrow escape iii. J. Stat. Phys., 122(3):491 -- 509, 2006.
[30] A. Singer, Z. Schuss, and D. Holcman. Narrow escape and leakage of brownian particles. Phys. Rev.
E, 78(5):051111, 2008.
[31] A. Singer, Z. Schuss, D. Holcman, and B. Eisenberg. Narrow escape i. J. Stat. Phys., 122(3):437 -- 536,
2006.
[32] W. R. Smythe. Static and dynamic electricity. McGraw-Hill Book Company, Inc., New York Toronto
London, 1950.
[33] I. Sneddon. Mixed boundary value problems in potential theory. John Wiley Sons, Inc., New York,
1966.
[34] A. Taflia and D Holcman. Estimating the synaptic current in a multi-conductance ampa receptor
model. arXiv:1009.0867v1 [q-bio.NC].
[35] A. Taflia and D Holcman. Dwell time of a brownian molecule in a microdomain with traps and a
small hole on the boundary. J. Chem. Phys., 126(23):234107, 2007.
[36] M.J. Ward and J.B. Keller. Strong localized perturbations of eigenvalue problems. SIAM J. Appl.
Math., 53:770 -- 798, 1993.
[37] M.J. Ward and E. Van De Velde. The onset of thermal runaway in partially insulated or cooled
reactors. IMA J. Appl. Math., 48:53 -- 83, 1992.
[38] G. Wilemski and M. Fixman. General theory of diffusion-controlled reactions. J. Chem. Phys.,
58:4009, 1973.
[39] R. Zwanzig. Diffusion-controlled ligand binding to spheres partially covered by receptors: An effective
medium traetment. Proc. Natl. Acad. Sci. USA, 87:5856 -- 5857, 1990.
24
1
0.8
0.6
pβ(x0)
0.4
0.2
0
0
β=0.1
β=0.5
β=1
β=2
β=3
β=5
β=10
5
x0
10
|
1904.12481 | 1 | 1904 | 2019-04-29T08:00:59 | Towards a Passive BCI to Induce Lucid Dream | [
"q-bio.NC"
] | Lucid dreaming (LD) is a phenomenon during which the person is aware that he/she dreaming and is able to control the dream content. Studies have shown that only 20% of people can experience lucid dreams on a regular basis. However, LD frequency can be increased through induction techniques. External stimulation technique relies on the ability to integrate external information into the dream content. The aim is to remind the sleeper that she/he is dreaming. If this type of protocol is not fully efficient, it demonstrates how sensorial stimuli can be easily incorporated into people's dreams. The objective of our project was to replicate this induction technique using material less expensive and more portable. This material could simplify experimental procedures. Participants could bring the material home, then have a more ecological setting. First, we used the OpenBCI Cyton, a low-cost EEG signal acquisition board in order to record and manually live-score sleep. Then, we designed a mask containing two LEDs, connected to a microcontroller to flash visual stimulation during sleep. We asked two volunteers to sleep for 2 hours in a dedicated room. One of the participants declared having a dream during which the blue lights diffused by the mask were embedded into the dream environment. The other participant woke up during the visual stimulation. These results are congruent with previous studies. This work marked the first step of a larger project. Our ongoing research includes the use of an online sleep stage scoring tool and the possibility to automatically send stimuli according to the sleep stage. We will also investigate other types of stimulus induction in the future such as vibro-tactile stimulation that showed great potentials. | q-bio.NC | q-bio |
TOWARDS A PASSIVE BCI TO INDUCE LUCID DREAM
JJC-ICON '19
Morgane Hamon
Ullo
La Rochelle, France
[email protected]
Emma Chabani
Institut du Cerveau et de la Moelle épinière
Paris, France
[email protected]
Philippe Giraudeau
Inria
Bordeaux, France
[email protected]
May 21, 2019
ABSTRACT
Lucid dreaming (LD) is a phenomenon during which the person is aware that he/she dreaming and is
able to control the dream content. Studies have shown that only 20% of people can experience lucid
dreams on a regular basis. However, LD frequency can be increased through induction techniques.
External stimulation technique relies on the ability to integrate external information into the dream
content. The aim is to remind the sleeper that she/he is dreaming. If this type of protocol is not fully
efficient, it demonstrates how sensorial stimuli can be easily incorporated into people's dreams. The
objective of our project was to replicate this induction technique using material less expensive and
more portable. This material could simplify experimental procedures. Participants could bring the
material home, then have a more ecological setting. First, we used the OpenBCI Cyton, a low-cost
EEG signal acquisition board in order to record and manually live-score sleep. Then, we designed a
mask containing two LEDs, connected to a microcontroller to flash visual stimulation during sleep.
We asked two volunteers to sleep for 2 hours in a dedicated room. One of the participants declared
having a dream during which the blue lights diffused by the mask were embedded into the dream
environment. The other participant woke up during the visual stimulation. These results are congru-
ent with previous studies. This work marked the first step of a larger project. Our ongoing research
includes the use of an online sleep stage scoring tool and the possibility to automatically send stimuli
according to the sleep stage. We will also investigate other types of stimulus induction in the future
such as vibro-tactile stimulation that showed great potentials.
Keywords Lucid Dream · Induction techniques · Sleep · Visual stimulation · EEG
1
Introduction
During a normal night's sleep, people enter on a cyclical basis different vigilance states: Wakefulness, N1, N2, N3
and REM sleep [Silber et al., 2007]. A sleep cycle lasts around 90 minutes, and occurs 4 to 6 times a night. The
N1, N2 and N3 sleep stages are usually grouped as N-REM (Non-REM) as an opposition to the REM sleep. It was
long believed that dreams exclusively occur during REM sleep [Dement and Kleitman, 1957] but this hypothesis has
been refuted [Foulkes, 1962]. Awakenings during N-REM sleep showed dreams recall as well [Nielsen et al., 2001].
Lucid dreaming (LD) is a phenomenon during which the person is aware that he/she dreaming and is able to
control the dream content [LaBerge, 1985]. LD can be used to enhance the dream content, train physical skills
[Erlacher and Ehrlenspiel, 2017] or avoid nightmares. That's why authors have studied the ability to induce LD or in-
crease the LD frequency [Gackenbach and LaBerge, 1988]. Thus, this project aims to reproduce induction techniques
to help people experience LD. As EEG amplifiers became less expensive and more portable, we aim at proposing a
solution that could be deployed at home.
JJC-ICON '19
Figure 1: Left: EEG cap with electrodes connected to an OpenBCI Cyton, Right: The mask with LEDs and the
micro-controller (Arduino UNO)
2 Visual induction to Augment LD
This pilot study has been conducted during a hackathon, a 48-hours event that occurred in December 2017 1. We
used a electroencephalogram (EEG) headset to acquire sleep data like sleep stages. First, we built a EEG cap out
of swim cap on which electrodes are sewed (see Fig.2, Left). They are connected to the OpenBCI Cyton Board,
that is sending data to a computer via Bluetooth. The OpenBCI software allows to visualize brainwaves obtained
from the electrodes. Sleep stages, arousal, eye movement were deducted from the neural oscillations and were scored
according to international criteria [Rechstchaffen and Kales, 1968]. Finally, a mask has been created with two LED
placed above each eye to send flash light (see Fig.2, Right). Lights were controlled by an Arduino Uno connected to
the same computer which runs the openBCI software. We sent visual stimulation when the participant enters in REM
sleep. REM stage for Rapid Eye Movements has an EEG activity mainly branded by theta waves (5-7 Hz). There is a
complete muscle atonia.
3 Pilot Study
Two volunteers (1 male 24y, 1 female, 26y), were not naive about lucid dreaming and had already experienced it. They
were also considered as recallers, who are subjects with high Dream Recall Frequency (DRF): remembering dreams
more than two nights a week [Cory et al., 1975]. The volunteers were equipped with our EEG headset, our mask and
earplugs to help them fall asleep. Subsequently visual stimulations were presented to determine the intensity of the
LEDs. Then participants were told they could sleep for about 2 hours. As the system wasn't live-scoring sleep, sleep
stages were monitored by our team and stimuli were sent when needed (REM sleep). The method is pictured in Fig.1.
4 Results
We observed that the blue flashing light was integrated in dream content for both participants. One participant reported
to be behind a fish tank window and saw blue flashes wondering if it was natural or not but without understand it was
a signal. However it did not trigger lucid dreaming. The second participant woke up right after the first stimulation.
These results are congruent with previous studies [LaBerge, 1985], [Paul et al., 2014].
1https://mindlabdx.github.io/hack1cerveau/
2
JJC-ICON '19
Figure 2: Cerebral activity, eye movements and movements are monitored and visualized in real-time to determine the
sleep stage the volunteer is going through. Visual stimulations are sent when the participant reaches REM sleep.
5 Discussion and Future plan
This pilot study allowed us to better know how and when sending stimuli, an important step towards a passive BCI
[Zander and Kothe, 2011]. Our ongoing research will explore different modalities of visual stimulation such as the
color, the intensity and the frequency in order to determine the best combination to induce LD. Vibro-tactile stimu-
lation showed great potentials as well [Stumbrys et al., 2012]. We will also compare different devices to detect sleep
stages (an EEG headband with less electrodes and an actimeter). This will be the first step to build a device able to
automatically discriminate sleep stages.
References
[Cory et al., 1975] Cory, T. L., Ormiston, D. W., Simmel, E., and Dainoff, M. (1975). Predicting the frequency of
dream recall. Journal of Abnormal Psychology, 84(3):261.
[Dement and Kleitman, 1957] Dement, W. and Kleitman, N. (1957). The relation of eye movements during sleep to
dream activity: an objective method for the study of dreaming. Journal of experimental psychology, 53(5):339.
[Erlacher and Ehrlenspiel, 2017] Erlacher, D. and Ehrlenspiel, F. (2017). Sleep, dreams, and athletic performance. In
Sport, Recovery, and Performance, pages 182 -- 196. Routledge.
[Foulkes, 1962] Foulkes, W. D. (1962). Dream reports from different stages of sleep. The Journal of Abnormal and
Social Psychology, 65(1):14.
[Gackenbach and LaBerge, 1988] Gackenbach, J. and LaBerge, S. (1988). Lucid dreaming: New research on con-
sciousness during sleep.
[LaBerge, 1985] LaBerge, S. (1985). Lucid dreaming.
[Nielsen et al., 2001] Nielsen, T., Kuiken, D., Hoffmann, R., Moffitt, A., et al. (2001). Rem and nrem sleep mentation
differences: a question of story structure? Sleep and Hypnosis, 3:9 -- 17.
[Paul et al., 2014] Paul, F., Schädlich, M., and Erlacher, D. (2014). Lucid dream induction by visual and tactile
stimulation: An exploratory sleep laboratory study. International Journal of dream research, 7(1):61 -- 66.
[Rechstchaffen and Kales, 1968] Rechstchaffen, A. and Kales, A. (1968). A manual of standardized terminology,
techniques and scoring system for sleep stages of human subjects. los angeles: Ucla brain information service.
Brain Research Institute, 57.
[Silber et al., 2007] Silber, M. H., Ancoli-Israel, S., Bonnet, M. H., Chokroverty, S., Grigg-Damberger, M. M., Hir-
shkowitz, M., Kapen, S., Keenan, S. A., Kryger, M. H., Penzel, T., et al. (2007). The visual scoring of sleep in
adults. Journal of Clinical Sleep Medicine, 3(02):22 -- 22.
3
[Stumbrys et al., 2012] Stumbrys, T., Erlacher, D., Schädlich, M., and Schredl, M. (2012). Induction of lucid dreams:
A systematic review of evidence. Consciousness and Cognition, 21(3):1456 -- 1475.
[Zander and Kothe, 2011] Zander, T. O. and Kothe, C. (2011). Towards passive brain -- computer interfaces: apply-
ing brain -- computer interface technology to human -- machine systems in general. Journal of neural engineering,
8(2):025005.
JJC-ICON '19
4
|
1301.0050 | 1 | 1301 | 2013-01-01T03:06:46 | Higher Order Correlations within Cortical Layers Dominate Functional Connectivity in Microcolumns | [
"q-bio.NC"
] | We report on simultaneous recordings from cells in all layers of visual cortex and models developed to capture the higher order structure of population spiking activity. Specifically, we use Ising, Restricted Boltzmann Machine (RBM) and semi-Restricted Boltzmann Machine (sRBM) models to reveal laminar patterns of activity. While the Ising model describes only pairwise couplings, the RBM and sRBM capture higher-order dependencies using hidden units. Applied to 32- channel polytrode data recorded from cat visual cortex, the higher-order mod- els discover functional connectivity preferentially linking groups of cells within a cortical layer. Both RBM and sRBM models outperform Ising models in log- likelihood. Additionally, we train all three models on spatiotemporal sequences of states, exposing temporal structure and allowing us to predict spiking from network history. This demonstrates the importance of modeling higher order in- teractions across space and time when characterizing activity in cortical networks. | q-bio.NC | q-bio |
Higher Order Correlations within Cortical Layers
Dominate Functional Connectivity in Microcolumns
Urs K oster
Redwood Center for Theoretical Neuroscience
UC Berkeley
Jascha Sohl-Dickstein
Department of Applied Physics
Stanford University
Charles M Gray
Cellular Biology and Neuroscience
Montana State University Bozeman
Bruno A Olshausen
Redwood Center for Theoretical Neuroscience
UC Berkeley
Abstract
We report on simultaneous recordings from cells in all layers of visual cortex
and models developed to capture the higher order structure of population spiking
activity. Specifically, we use Ising, Restricted Boltzmann Machine (RBM) and
semi-Restricted Boltzmann Machine (sRBM) models to reveal laminar patterns
of activity. While the Ising model describes only pairwise couplings, the RBM
and sRBM capture higher-order dependencies using hidden units. Applied to 32-
channel polytrode data recorded from cat visual cortex, the higher-order mod-
els discover functional connectivity preferentially linking groups of cells within
a cortical layer. Both RBM and sRBM models outperform Ising models in log-
likelihood. Additionally, we train all three models on spatiotemporal sequences
of states, exposing temporal structure and allowing us to predict spiking from
network history. This demonstrates the importance of modeling higher order in-
teractions across space and time when characterizing activity in cortical networks.
1
Introduction
Electrophysiology is rapidly moving towards high density recording techniques capable of capturing
the simultaneous activity of large populations of neurons. This raises the challenge of understanding
how networks encode and process information in ways that go beyond feedforward receptive field
models for individual neurons. Modeling the distribution of states in a network provides a way to
discover communication patterns and functional connectivity between cells.
The Ising model, originally developed in the 1920s to describe magnetic interactions [1], has been
used to describe electrophysiological data, particularly in the retina [2], and more recently for corti-
cal recordings [3]. This model treats spikes from a population of neurons binned in time as binary
vectors and captures dependencies between cells with the maximum entropy distribution for pair-
wise correlation. However, this only provides a good approximation for small groups of cells in the
retina [4].
In this work, we apply maximum entropy models to data from the visual cortex. Cortical networks
have proven to be more challenging to model than the retina: even the existence of significant
pairwise correlations between cortical cells is controversial [5, 6] and higher order correlations play
an important role [7, 8, 9]. One of the challenges with current recording technologies is that we
cannot simultaneously record from more than a tiny fraction of the cells that make up a cortical
circuit. Sparse sampling together with the complexity of the circuit mean that the majority of a
cell’s input will be from cells that are not part of the recording. In adult cat visual cortex, direct
synaptic connections have been reported to occur between 11% - 30% of nearby pairs of excitatory
1
neurons in layer IV [10], while a larger fraction of cell pairs show “polysynaptic” couplings [11],
defined by a broad peak in the cross-correlation between two cells. This type of coupling can be
due to common inputs (either from a different cortical area or in the form of lateral connections) or
a series of monosynaptic connections. A combination of these is believed to give rise to most of the
statistical interactions between the recorded cells. The Ising model, which assumes only pairwise
couplings, is well suited to model direct synaptic coupling, but cannot deal with interactions that
include many cells simultaneously. Therefore, we propose a new approach, that addresses both
incomplete sampling and tentative inputs from larger scale cell assemblies. We extend the Ising
model with a layer of hidden units or latent variables. The resulting model is a semi-Restricted
Boltzmann Machine (sRBM), which combines pairwise connections between visible units with an
additional set of connections to hidden units.
Estimating the parameters of Ising models and Boltzmann machines is hard because probabilities
are only defined up to a normalization constant. For both Ising models and Boltzmann machines
with hidden units, this normalization constant is computationally intractable, requiring a sum over
the exponential number of states of the system. This makes exact maximum likelihood estima-
tion impossible for all but the smallest systems and has previously necessitated approximate and/or
computationally expensive estimation methods. In this work, we use Minimum Probability Flow
(MPF [12, 13], in the context of neural decoding see [14]) to estimate parameters efficiently without
computing the intractable partition function. This allows us to estimate Ising models on higher-
dimensional data than is otherwise feasible, and to estimate the sRBM in a straightforward way.
Similar to parameter estimation, model evaluation is also hampered by the fact that the learned mod-
els are not normalized. To compute probabilities and compare the likelihood of different models,
annealed importance sampling [15] was used to estimate the partition function.
Combining these two methods for model estimation and evaluation, we show that the sRBM can
capture the distribution of states in a cortical network of tens of cells recorded from cat visual cortex
significantly better than a pairwise model. The higher order structure discovered by the model is
spatially organized and specific to cortical layers, indicating that networks within individual layers
of a microcolumn are the dominant source of correlations. Applied to spatiotemporal patterns of
activity, the model captures temporal structure in addition to dependencies across different cells,
allowing us to predict spiking activity based on the history of the network.
2 Methods
2.1 Recording and experimental procedures
Data were recorded from anesthetized cat visual cortex in response to a custom set of full field
natural movie stimuli. Movies of 8 to 30 minutes duration were captured at 300 frames per second
and 512 × 384 pixel resolution with a Casio F1 camera. Care was taken to avoid scene changes and
camera motion, so the spatiotemporal statistics of the movie correspond to those of natural scene
motion. Movies were converted to grayscale and down-sampled to 150 Hz for presentation. The
high frame rate was chosen to prevent cells phase locking to the frame rate, and scene changes were
minimized to avoid evoked potentials due to sudden luminance changes. Fig 1c) shows an example
frame crop from one of the natural scene movies.
Recordings were made with single shank 32 channel polytrodes (Fig. 1a) with a channel spacing
of 50µm, spanning all the layers of visual cortex. Individual datasets had on the order of of 20 to
40 simultaneously recorded neurons. The data was spike sorted offline using k-means clustering
(KlustaKwik, [16]) with a manual cleanup step (MClust, [17]). Unless noted otherwise, spikes were
binned at 20ms where bins with a single spike (2.4% of bins) and multiple spikes (0.9% of bins) were
both treated as spiking and the rest as non-spiking. The bin size was chosen to span the width of the
central peak in cross-correlograms between pairs of cells such as shown in 1e), where the dashed
vertical lines at ± 20ms envelope the central peak. An example of binned data for 23 isolated single
units is shown in 1b) with black marks indicating spiking.
To register individual recording channels with cortical layers, recording locations were reconstructed
from Nissl-stained histological sections, and current source density analysis in response to flashed
full-field stimuli was used to infer the location of cortical layer IV on the polytrode [18]. In 1b) we
use horizontal lines to indicate the upper and lower boundaries of layer IV.
2
(a) Recordings were made from columns of cat visual cortex. The 32 channel probe is
Figure 1:
shown to scale. (b) Example data from 23 cells, binned in 20ms windows, 2s of data. Columns
of this matrix are the input to our algorithm. For the spatiotemporal version of the model, we
concatenate several adjacent columns. (c) Example of one of the natural movie stimuli presented,
showing a group of ducks. (d) Histogram of instantaneous correlations between pairs of cells. (e)
Cross-correlation as a function of time lag between a pair of cells binned at 6.7ms shows a peak in
the correlation about 20ms wide (dashed lines). This determines the window size for binning the
data.
The surgical methods are described in detail elsewhere [19]. The protocol used in the experiments
was approved by the Institutional Animal Care and Use Committee at Montana State University
and conformed to the guidelines recommended in Preparation and Maintenance of Higher Mam-
mals During Neuroscience Experiments, National Institutes of Health Publication 913207 (National
Institutes of Heath, Bethesda, MD 1991).
The models were estimated on a data set of 180, 000 data vectors corresponding to 60 minutes
of recording time. The data were split into two subsets of 90, 000 data points: a training set for
parameter estimation and a test set to compute cross-validated likelihoods.
We also analyzed spatiotemporal patterns of data, which were created by concatenating consecutive
state vectors. For the spatiotemporal experiments a bin width of 6.7 ms, corresponding to the frame
rate of the stimulus, was used. This bin size is a compromise capturing more detailed structure in
the data without leading to an undue increase in dimensionality and complexity. Up to 10 time bins
were concatenated in order to discover spatiotemporal patterns and predict spiking given the history
over the prior 67ms. These models were trained on 100, 000 samples and likelihoods were computed
on a set of equal size.
2.2 Model and Estimation
The sRBM consists of a set of binary visible units x ∈ {0, 1}N corresponding to observed neurons
in the data and a set of hidden units h ∈ {0, 1}M that capture higher order dependencies. Weights
between visible units, corresponding to an Ising model or fully visible Boltzmann machine, capture
pairwise correlations in the data. Weights between visible and hidden units, corresponding to an
RBM, learn to describe higher order structure.
The Ising model with visible-visible coupling weights J ∈ RN ×N and biases b ∈ RN has an
energy function
EI (x) = −xT Jx − bT x,
(1)
3
II/IIIIVV/VIWM500µm0.10.30.5020Correlation coefficientCountd) Pairwise corr., 20ms binsb) Example data, 2s of data in 20ms binsa) Recording methodc) Stimulus frames−20002000.060.10.14e) Temporal cross−corr.Lag / msCorr. coef40 stant, or partition function, ZI = (cid:80){x(cid:48) } exp [−EI (x(cid:48) )] consists of a sum over all 2N system states.
exp [−EI (x)], where the normalization con-
with associated probability distribution pI (x) = 1
ZI
The RBM with visible-hidden coupling weights W ∈ RN ×M and hidden and visible biases bv ∈
RN and bh ∈ RM has an energy function
ER (x, h) = −xT Wh − bT
v x − bT
(2)
h h,
exp [−ER (x, h)]. Since the there are no
with associated probability distribution pR (x, h) = 1
ZR
(cid:90)
connections between hidden units (hence “restricted” Boltzmann machine), the latent variables h
can be analytically marginalized out of the distribution (see Appendix A) to obtain
exp [−ER (x)] ,
1
p (x) =
dh p (x, h) =
ZR
where the energy for the marginalized distribution over x (sometimes referred to as the free energy
ER (x) = − (cid:88)
in machine learning literature) is
i
where wi are rows of the matrix W. The energy function for an sRBM combines the Ising model
and RBM energy terms,
i x+bh,i ) − bT
log(1 + ewT
v x,
(3)
(4)
(5)
(6)
(7)
v x − bT
ES (x, h) = −xT Jx − xT Wh − bT
h h.
As with the RBM, it is straightforward to marginalize over the hidden units for an sRBM,
ES (x) = −xT Jx − (cid:88)
exp [−ES (x)] ,
1
pS (x) =
ZS
i x+bh,i ) − bT
log(1 + ewT
v x.
i
A hierarchical Markov Random Field based on the sRBM has previously been applied as a model for
natural image patches [20], with the parameters estimated using contrastive divergence (CD, [21]).
Instead of CD, which is based on sampling, we train the models using Minimum Probability Flow
(MPF, [12]), a recently developed estimation method for energy based models. MPF works by
minimizing the KL divergence between the data distribution and the distribution which results from
moving slightly away from the data distribution towards the model distribution. This KL divergence
will be uniquely zero in the case where the model distribution is identical to the data distribution.
While CD is a stochastic heuristic for parameter update, MPF provides a deterministic and easy
(cid:19)
(cid:18) 1
to evaluate objective function. Second order gradient methods can therefore be used to speed up
(cid:88)
(cid:88)
optimization considerably. The MPF objective function
[E (x) − E (x(cid:48) )]
g (x, x(cid:48) ) exp
2
x∈D
x(cid:48) /∈D
measures the flow of probability out of data states x into neighboring non-data states x(cid:48) , where the
connectivity function g (x, x(cid:48) ) ∈ {0, 1} identifies neighboring states, and D is the list of data states.
We consider the case where the connectivity function g (x, x(cid:48) ) is set to connect all states which differ
(cid:26) 1
by a single bit flip
H (x, x(cid:48) ) = 1
0
otherwise
where H (x, x(cid:48) ) is the Hamming distance between x and x(cid:48) . See Appendix B1 for a derivation of the
MPF objective function and gradients for the sRBM, RBM, and Ising models. In all experiments,
parameters. This was done by adding a term of the form λ (cid:80)
i,j Jij + λ (cid:80)
minimization of K was performed with the MinFunc implementation of L-BFGS [22].
To prevent overfitting all models were estimated with an L1 sparseness penalty on the coupling
j,k Wjk to the objective
1Code is available for download at http://github.com/Sohl-Dickstein
g (x, x(cid:48) ) =
K =
(8)
(9)
,
4
function, summing over the absolute values of the elements of both the visible and hidden weight
matrices. The optimal sparseness λ was chosen by cross-validating the log-likelihood on a holdout
set.
Since MPF learning does not give an estimate of the partition function, we use annealed importance
sampling (AIS, [15]) to compute normalized probabilities. AIS is a sequential Monte Carlo method
that works by gradually morphing a distribution with a known normalization constant (in our case a
uniform distribution over x) into the distribution of interest. See Appendix C for more detail. AIS
applied to RBM models is described in [23], which also highlights the shortcomings of previously
used deterministic approximations.
pendent and characterized by their firing rate pr (x) = (cid:81)
Normalizing the distribution via AIS allows us to compute the log likelihood of the model pm and
compare it to the likelihood gained over a baseline model. This baseline assumes cells to be inde-
i (rixi + (1 − ri )(1 − xi )) with rates ri
(cid:80)
for individual cells i. The independent model is easily estimated and normalized, and is commonly
used as a reference for model comparison. The excess log likelihood over this baseline is defined in
terms of an expectation per unit time as L = 1
x∈D [log2 pm (x) − log2 pr (x)], where τ is the
width of a time bin. The excess log likelihood rate L, measured in bits/s, is used as the basis for
N τ
model comparisons.
3 Results
We estimated Ising, RBM and sRBM models for populations of cortical cells simultaneously
recorded across all cortical layers in a microcolumn. Here we present data from two animals, one
with 23 single units (B4), another with 36 units (T6), as well as a multiunit recording with 28 units
(B4MU). In each case the population was verified to be visually responsive and the majority of cells
were orientation selective simple or complex cells.
The estimated model parameters for the three different types of models (Ising, RBM and sRBM)
are shown in Fig. 2. The sparseness penalty, chosen to optimize likelihood on a validation dataset,
results in many of the parameters being zero. For the Ising model (a) we show the coupling as a
matrix plot, with horizontal lines indicating anatomical layer boundaries. The diagonal contains the
bias terms (there is no self-coupling), which are negative since all cells are off the majority of the
time. The matrix has many small positive weights that encourage positive pairwise correlations by
lowering the energy of connected states being active simultaneously.
In (b) we show the hidden units of the RBM as individual bar plots, with the bars representing con-
nection strengths to visible units. The topmost bar corresponds to the hidden bias of the unit. Hidden
units are ordered by variance. The units are highly selective in connectivity: The first unit almost
exclusively connects to cells in the deep (granular and subgranular) cortical layers. The second and
third unit capture correlations between cells in the superficial (supergranular) layers. The correla-
tions are of high order, with 10 and more cells receiving input from a hidden unit. The remaining
units connect fewer cells, but still tend to be location-specific. Unit five captures a predominantly
pairwise correlation that is also visible in the Ising model coupling matrix. Only six out of the total
23 hidden units have non-zero couplings and are shown. Additional hidden units are disabled by
the L1 sparseness penalty, which was chosen to maximize likelihood on the cross-validation dataset.
The interpretation of hidden units is quite similar to the pairwise terms of the Ising model: positive
coupling to a group of visible units encourages these units to become active simultaneously, as the
energy of the system is lowered if both the hidden unit and the cells it connects to are active. Thus
the hidden units become switched on when cells they connect to are firing (activation of hidden units
not shown).
The sRBM combines both pairwise and hidden connections and hence is show with a pairwise
coupling matrix and bar plots for hidden units. Due to the larger number of parameters, the optimal
model is even more sparse. The remaining pairwise terms now predominantly encode negative
interactions, whereas much of the positive coupling has been explained away by the hidden units.
These still provide strong positive couplings within either superficial (II/III) or intermediate (IV)
and deep (V/VI) layers, which explain the majority of structure in the data.
It is noteworthy that the preferred explanation for dependencies between recorded neurons is via
connections to shared hidden units, rather than direct couplings between visible units. The RBM
5
Figure 2: Connectivity patters of the three models estimated for a recording session with 23 spike
sorted single units. The horizontal lines indicate approximate boundaries between cortical layers
II/III, layer IV and layers V/VI. (a) Ising model with the bias terms on the diagonal. The model has
many small coupling terms that encode positive correlations. (b) The RBM coupling weights are
plotted as a bar chart for each hidden unit, ordered by activity from left to right. The first bar chart
is the bias for all the visible units, and the extra bar at the top of each plot corresponds to the bias
of that hidden unit. Blue bars indicate a sign flip (the bias terms are predominantly negative, but
plotted with reversed sign for easier comparison with the remaining terms). (c) The sRBM weights
are shown in the same way, with the pairwise couplings on the left and hidden units on the right.
The pairwise connections are qualitatively very different from those of the Ising model, as most of
the structure is better captured by hidden units.
(a) Model comparison using likelihood gain over the independent (firing rate) model.
Figure 3:
Likelihoods are normalized to bits/s, but not for model size, explaining the large difference between
data sets of different size: 23 cells for session B4, 28 cells for MU, and 36 cells for T6. B4 and T6
are spike sorted, MU is a multiunit dataset. All three models outperform the independent model by
10-20 bits/s. The higher order models with hidden units give a further improvement of about 2 bit/s
over the Ising model.
6
b) RBM hidden unitsc) sRBM pairwise couplings and hidden unitsa) Ising model1 2 3 4 5 −2201050050bias 1 2 3 4 5 6 B4MUT605101520RBML(bits/s) Figure 4: Scatter plot of test data set showing empirical probabilities against model probabilities.
Different models are encoded by color, the number of simultaneously spiking cells in each pattern
by different symbols. (a) shows the independent model compared to the Ising model, (b) shows the
Ising model compared to the RBM. The sRBM is omitted as it is very similar to the RBM. The RBM
significantly outperforms simpler models.
and sRBM in this comparison were both estimated with 25 hidden units, but we show only units
that did not go to zero due to the sparseness constraint. In this case, λ = 2 × 10−3 was found to be
optimal. Since at this level of sparseness, many of the hidden units turn off entirely, it was deemed
unnecessary to further increase the number of hidden units.
For a quantitative comparison between models, we computed normalized likelihoods using AIS to
estimate the partition function. For each model, we generated 500 samples through a chain of 105
annealing steps. To ensure convergence of the chain, we use a series of chains varying the number
of annealing steps and verify that the estimate of Z stabilizes to within at least 0.02 bits/bin (see
supplemental Fig. 7 in the Appendix). For models of size 20 and smaller we furthermore computed
the partition function exactly to compare against the AIS estimate.
Fig. 3 a) shows a comparison of excess log likelihood L over the independent firing rate model
for the three different models and on all three datasets. L is computed in units of bits/s for the full
population. Both higher-order models outperform the Ising model in fitting the datasets, significantly
so for two datasets. Error bars are standard error on the mean, computed from 10 models with
different random subsets of the data used for learning and validation, and different random seeds
for the parameters and the AIS sampling runs. Each of the models was estimated for a range of
sparseness parameters λ = [0, 1, 2, 4, 6, 8, 10] × 10−3 bracketing the optimal λ, and the results are
shown for the optimal choice of λ for each model.
Additional insight into the relative performance of the models can be gained by comparing model
probabilities to empirical probabilities for the various types of patterns. Fig. 4 shows scatter plots of
model probabilities under the different models against pattern frequencies in the data. Patterns with
a single active cell, two simultaneously active cells, etc. are distinguished by different symbols. As
expected from the positive correlations, the independent model (yellow) shown in a) has to greatly
overestimate the probabilities of cells being active individually, so these patterns fall above the iden-
tity line, while all other patterns are underestimated. For comparison the Ising model is shown in the
same plot (blue), and does significantly better, indicated by the points moving closer to the identity
line. It still tends to fail in a similar way though, with many of the triplet patterns being underesti-
mated as the model cannot capture triplet correlations. In b), this model is directly compared against
the RBM (green). Except for very rare patterns, most points are now very close to the identity line,
as the model can fully capture higher order dependencies. Hidden units describe global dependen-
cies that greatly increase the frequency of high order patterns compared to individually active cells.
The 5% and 95% confidence intervals for the counting noise expected in the empirical frequency of
system states are shown as dashed lines. The solid line is the identity.
7
10−810−610−410−210010−410−310−210−1100Empirical Pattern FrequencyModel Probability10−410−310−210−1100Empirical Pattern Frequencya) Independent vs. Ising modelb) Ising vs. RBM modelnonesingledoubletriplequadhigherIsingIndep.RBM Figure 5: Likelihood comparison as a function of model size, for spatiotemporal models with 10
cells and a varying number of concatenated time steps. The log-likelihood per neuron increases
as each neuron is modeled as part of a longer time sequence. This effect holds both for Ising and
higher order models, but since the Ising model cannot capture many of the relevant dependencies,
the increase in likelihood saturates at about 3 timesteps.
Note that any error in estimating the partition function of the models would lead to a vertical offset
of all points. Thus visually checking the alignment of the data cloud around the identity line pro-
vides an intuitive verification that there are no catastrophic errors in the estimation of the partition
function. Unfortunately we cannot use this alignment (e.g. of the most frequent all zeros state) as
a shortcut to compute the partition function without sampling: L1 regularization tends to reduce
model probabilities of the most frequent states, so this estimate of Z would systematically overesti-
mate the likelihood of regularized models. We note, however, that for models with no regularization
this estimate does indeed agree well with the AIS estimate.
3.1 Spatiotemporal models
The same models can be used to capture spatiotemporal patterns by treating previous time steps as
additional cells. Consecutive network states binned at 6.7ms are concatenated in blocks of up to
10 time steps , for a total network dimensionality of 100 with 10 cells. These models were cross-
validated and the sparseness parameters optimized in the same way as for the instantaneous model.
This allows us to learn kernels that describe the temporal structure of interactions between cells.
In Fig. 5 we compare the relative performance of spatiotemporal Ising and higher order models as
a function of the number of time steps included in the model. To create the datasets, we picked
a subset of 10 cells with the highest firing rates from the B4 dataset (4 cells from subgranular, 2
from granuar and 4 from supergranular layers) and concatenated blocks of up to 10 subsequent data
vectors. This way models of any dimensionality divisible by 10 can be estimated. The number of
parameters of the RBM and Ising model were kept the same by fixing the number of hidden units
in the RBM to be equal to the number of visible units, the sRBM was also estimated with a square
weight matrix for the hidden layer. As before, the higher order models consistently outperform the
Ising model. The likelihood per unit increases with the network size for all models, as additional
information from network interactions leads to an improvement in the predictive power of the model.
However, the curve levels off for the Ising model after a dimensionality of about 30 is reached, as
higher order structure that is not well captured by the Ising model becomes increasingly important.
A similar observation has been made in [4], where Ising and higher order models for 100 retinal
ganglion cells were compared to models for 10 time steps of 10 cells. It is noteworthy that temporal
dependencies are similar to dependencies between different cells, in that there are strong higher order
correlations not well described by pairwise couplings. These dependencies extend surprisingly far
across time (at least 67ms, corresponding to the largest models estimated here) and are of such a
form that including pairwise couplings to these states does not increase the likelihood of the model.
This has implications e.g. for GLMs that are typically estimated with linear spike coupling kernels
which will miss these interactions.
8
12345681001234Number of time slicesRBML(bits/unit/s) (a) Hidden units of spatiotemporal sRBM model. For each hidden unit, the horizontal
Figure 6:
axis is time and the vertical axis cells with horizontal bars separating the subgranular, granuar and
supergranular cortical layers. (b) Log-likelihood gain for one cell conditioned on the network state
for all three models. The remaining population carries more information about cells with higher
firing rates. (c) Spike prediction from network history. For one of the cells, we show 1s of predicted
activity given the history of the network state. In each case when a spike occurs in the data, there is
an elevated probability under the models.
To predict spiking based on the network history, we can compute the conditional distribution of
single units given the state of the rest of the network. This is illustrated for a network with 15 time
steps for a dimensionality of 150. Fig. 6 a) shows the learned weights of 18 randomly chosen hidden
units for a spatiotemporal RBM model with 150 hidden units. Each subplot corresponds to one
hidden unit, which connects to 10 neurons (vertical axis) across 15 time steps or 100ms (horizontal
axis). Some units specialize in spatial coupling across different cells at a constant time lag. The
remaining units describe smooth, long-range temporal dependencies, typically for small groups of
cells. Both of these subpopulations capture higher order structure connecting many neurons that
cannot be approximated with pairwise couplings.
By conditioning the probability of one cell at one time bin on the state of the remaining network, we
can compute how much information about a cell is captured by the model over a naive prediction
based on the firing rate of the cell. This conditional likelihood for each cell is plotted in 6 b) in a
similar way to excess log likelihood for the entire population in Fig. 5. While the result here reflects
our previous observation that Boltzmann machines with hidden units outperform Ising models, we
note that the conditional probabilities are easily normalized in closed form since they describe a
one-dimensional state space. Thus we can ensure that the likelihood gain holds independent of the
estimation of Z and cannot be due to systematic errors in sampling from the high-dimensional mod-
els. Fig. 6 c) provides a more intuitive look at the prediction. For 1s of data from one cell, where 5
spikes occur, we show the conditional firing probabilities for the 3 models given 100ms of history of
itself and the other cells. Qualitatively, the models perform well in predicting spiking probabilities,
suggesting it might compare favorably to prediction based on GLM-type or Ising models [24].
9
a) Subset of hidden unit filters123456789100.05.1.15CellL / bits s−1b) Log-likelihood gain, cells sorted by spike ratesRBMRBMIsingTime / secondstimecells10P00.20.40.60.81c) Spike prediction for cell 1 IsingRBMsRBMdata 4 Discussion
We explored the utility of Boltzmann machines with hidden units as models for neural population
data, arguing that it provides a better model for cortical data than Ising models and previous exten-
sions. While there has been a resurgence of interest in these maximum entropy models for describing
neural data, progress has been hampered mainly by two problems: Estimation of energy based mod-
els is difficult since these models cannot be normalized in closed form. Evaluating the likelihood
of the model thus requires approximations or a numerical integral over the exponential number of
states of the model, making maximum likelihood estimation computationally intractable. Therefore
even the pairwise Ising model is hard to estimate in general, and various approximations have been
used to overcome this problem. This is a purely computational difficulty, but there is a more funda-
mental issue with generalizing models to include higher order dependencies. While this does not by
itself make the model estimation any more difficult, in general the number of model parameters to
be estimated is now also exponential in the size of the data. This can be dealt with either by cutting
off dependencies at some low order, estimating only a small number of higher order coupling terms,
or by imposing some specific form on the dependencies.
We attempted to address both of these problems here. Parameter estimation was made tractable using
MPF, and latent variables were shown to be an effective way of capturing high order dependencies.
This addresses several shortcomings that have been identified with the Ising model.
As Macke argues in [25], models with direct (pairwise) couplings are not well suited to model data
recorded from cortical networks. Since only a tiny fraction of the neurons making up the circuit are
recorded, most input is likely to be common input to many of the recorded cells rather than direct
synapses between them. Whiles his work compares Generalized Linear Models (GLMs) such as
models of the retina [26] and for LGN [27] to linear dynamical systems (LDS) models, the argument
applies equally for the models presented here.
Another shortcoming of the Ising model and some previous extensions is that the number of parame-
ters to be estimated does not scale favorably with the dimensionality of the network. The number of
pairwise coupling terms in GLM and Ising models scales with the square of the number of neurons,
so with the amounts of data typically collected in electrophysiological experiments it is only possible
to identify the parameters for small networks with a few tens of cells. This problem is aggravated by
including higher order couplings: for example the number of third order coupling parameters scales
with the cube of the data dimensionality. Therefore attempting to estimate these coupling parameters
directly is a daunting task that usually requires approximations and strong regularization.
An alternative is to focus on higher order structure in very small networks. Ohiorhenuan noted that
Ising models fail to explain structure in cat visual cortex [7] and was able to model triplet correla-
tions [8] by considering very small populations of no more than 6 neurons. However, Schneidman
and Ganmor caution [4] that trying to model small subsets (10 cells) of a larger network to infer
properties of the full network may lead to wrong conclusions and show that for retinal networks
higher order correlations start to dominate only once a certain network size is reached. Therefore
they address the same question as the present paper, i.e. how to capture nth order correlations with-
out the accompanying dn growth in the number of free parameters in a larger network. In their
proposed reliable interaction model, they exploit the sparseness of the neural firing patterns to ar-
gue that most higher order coupling terms will be zero. Therefore the true distribution can be well
approximated from a small number of these terms, which can be calculated using a simple recursive
scheme. In practice, the main caveat is that only patterns that appear in the data many times are used
to calculate the coupling terms. While the model by construction assigns correct relative probabil-
ities to observed patterns, the probability assigned to unobserved patterns is unconstrained, and the
most probable states may therefore be ones which never occur in the data.
In contrast to these models focussed on retinal data however, in visual cortex higher order corre-
lations play an important role even in small networks. Yu et al. [3, 9] show that over the scale of
adjacent cortical columns of anesthetized cat visual cortex, small subnetworks of 10 cells are better
characterized with a dichotomized Gaussian model than the pairwise maximum entropy distribution.
While the dichotomized Gaussian [28] is estimated only from pairwise statistics, it carries higher
order correlations that can be interpreted as common Gaussian inputs [29]. However these correla-
tions are implicit in the structure of the model and not directly estimated from the data as with the
RBM, so it is not clear that the model would perform as well on different datasets.
10
Finally, GLMs [26] can be used to model each cell conditioned on the rest of the population. While
mostly used for stimulus response models including stimulus terms, they are easily extended with
terms for cross-spike coupling, which capture interactions between cells. A major limitation of
GLMs is that current implementations can only be estimated efficiently if they are linear in the
stimulus features and network coupling terms, so they are not easily generalized to higher order
interactions. Two approaches have been used to overcome this limitation for stimulus terms. The
GLM can be extended with additional nonlinearities, preserving convexity on subproblems [27].
Alternatively, the stimulus terms can be packaged into nonlinear features which are computed in
preprocessing and usually come with the penalty of a large increase in the dimensionality of the
problem [30]. However, we are not aware of any work applying either of these ideas to spike history
rather than stimulus terms. Another noteworthy drawback of GLMs is that instantaneous coupling
terms cannot be included as this causes probabilities to diverge [25], so instantaneous correlations
cannot be modeled and have to be approximated using very fine temporal discretization.
In conclusion, the RBM provides a parsimonious model for higher order dependencies in neural
population data. Without explicitly enumerating a potentially exponential number of coupling terms
or being constrained by only measurements of pairwise correlations, it provides a low-dimensional,
physiologically interpretable model that can be easily estimated for populations of 100 and more
cells.
The connectivity patterns the RBM learns from cells simultaneously recorded from all cortical lay-
ers are spatially localized, showing that small neural assemblies within cortical layers are strongly
coupled. This suggests a shared computation these local networks perform on common input, while
cells across different cortical layers participate in distinct computations and have much less coupled
activity. This novel observation is made possible by the RBM: because each of the hidden units re-
sponds to (and therefore learns on) a large number of recorded patterns, it can capture dependencies
that are too weak to extract with previous models. In particular, the connectivity patterns discovered
by the RBM and sRBM are by no means obvious from the covariance of the data or by inspecting
the coupling matrix of the Ising model. This approach, combining a straightforward estimation pro-
cedure and a powerful model, can be extended from polytrode recordings to capture physiologically
meaningful connectivity patterns in other types of multi-electrode data.
Acknowledgments
BAO, CMG and UK were supported by National Eye Institute grant #EY019965.
References
[1] Ernst Ising. Beitrag zur Theorie des Ferromagnetismus. Zeitschrift f ur Physik A, 31(1):253–
258, 1925.
[2] Elad Schneidman, Michael J Berry, Ronen Segev, and William Bialek. Weak pairwise correla-
tions imply strongly correlated network states in a neural population. Nature, 440(7087):1007–
12, April 2006.
[3] Shan Yu, Debin Huang, Wolf Singer, and Danko Nikolic. A Small World of Neuronal Syn-
chrony. Cerebral Cortex, 18(12):2891–2901, 2008.
[4] Elad Ganmor, Ronen Segev, and Elad Schneidman. Sparse low-order interaction network
underlies a highly correlated and learnable neural population code. Proceedings of the National
Academy of Sciences of the United States of America, 108(23):9679–84, June 2011.
[5] Alexander S Ecker, Philipp Berens, Georgios Keliris, Matthias Bethge, Nikos K Logothetis,
and Andreas S Tolias. Decorrelated neuronal firing in cortical microcircuits. Science (New
York, N.Y.), 327(5965):584–7, January 2010.
[6] Alfonso Renart, J. De la Rocha, Peter Bartho, Liad Hollender, N. Parga, Alex Reyes, and K.D.
Harris. The asynchronous state in cortical circuits. Science, 327(5965):587–590, 2010.
[7] Ifije E Ohiorhenuan, Ferenc Mechler, Keith P Purpura, Anita M Schmid, Qin Hu, and
Jonathan D Victor. Sparse coding and high-order correlations in fine-scale cortical networks.
Nature, 466(7306):617–21, July 2010.
11
[8] Ifije E Ohiorhenuan and Jonathan D Victor. Information-geometric measure of 3-neuron fir-
ing patterns characterizes scale-dependence in cortical networks. Journal of computational
neuroscience, 30(1):125–41, March 2011.
[9] Shan Yu, Hongdian Yang, Hiroyuki Nakahara, Gustavo Santos, Danko Nicolic, and Dietmar
Plenz. Higher-order interactions characterized in cortical activity. The Journal of Neuro-
science, 31(48):17514–17526, 2011.
[10] K J Stratford, K Tarczy-Hornoch, K A Martin, N J Bannister, and J J Jack. Excitatory synaptic
inputs to spiny stellate cells in cat visual cortex. Nature, 382(6588):258–261, July 1996.
[11] GM Ghose, RD Freeman, and I Ohzawa. Local intracortical connections in the cat’s visual
cortex: postnatal development and plasticity. Journal of Neurophysiology, 72(3):1290–303,
1994.
[12] Jascha Sohl-Dickstein, Peter Battaglino, and Michael DeWeese. Minimum Probability Flow
Learning. In International Conference on Machine Learning, 2011.
[13] Jascha Sohl-Dickstein, Peter Battaglino, and Michael DeWeese. New Method for Parameter
Estimation in Probabilistic Models: Minimum Probability Flow. Physical Review Letters,
107(22):11–14, November 2011.
[14] MT Schaub. The Ising decoder: reading out the activity of large neural ensembles. Journal of
computational neuroscience, 2012.
[15] RM Neal. Annealed importance sampling. Technical report, University of Toronto, Toronto,
2001.
[16] KD Harris, DA Henze, Jozsef Csicsvari, Hajime Hirase, and Gyorgy Buzsaki. Accuracy of
Tetrode Spike Separation as Determined by Simultaneous Intracellular and Extracellular Mea-
surements. Journal of Neurophysiology, 84(1):401–414, 2000.
[17] David Redish. MClust. Technical report, University of Minnesota, Minneapolis, 2012.
[18] U Mitzdorf. Current source-density method and application in cat cerebral cortex: investigation
of evoked potentials and EEG phenomena. Physiological reviews, 65(1):37–100, January 1985.
[19] Charles M Gray, Pedro E Maldonado, M Wilson, and Bruce McNaughton. Tetrodes markedly
improve the reliability and yield of multiple single-unit isolation from multi-unit recordings in
cat striate cortex. Journal of neuroscience methods, 63(1-2):43–54, December 1995.
[20] Simon Osindero and Geoffrey Hinton. Modeling image patches with a directed hierarchy of
Markov random fields. Advances in neural information processing, pages 1–8, 2008.
[21] Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural
computation, 14(8):1771–800, August 2002.
[22] Mark Schmidt. MinFunc. Technical report, Laboratoire d’Informatique de l’ ´Ecole Normale
Sup ´erieure, Paris, 2005.
[23] Ruslan Salakhutdinov and Iain Murray. On the quantitative analysis of deep belief networks.
In ICML, 2008.
[24] Wilson Truccolo, Leigh R Hochberg, and John P Donoghue. Collective dynamics in human
and monkey sensorimotor cortex : predicting single neuron spikes. Nature Neuroscience,
13(1):105–111, 2009.
[25] Jakob H Macke, John P Cunningham, Krishna V Shenoy, and Maneesh Sahani. Empirical
models of spiking in neural populations. In Advances in Neural Information Processing, pages
1–9, 2011.
[26] Jonathan W Pillow, Jonathon Shlens, Liam Paninski, Alexander Sher, Alan M Litke, E J
Chichilnisky, and Eero P Simoncelli. Spatio-temporal correlations and visual signalling in
a complete neuronal population. Nature, 454(7207):995–9, August 2008.
[27] D. Butts, C. Weng, J. Jin, J.-M. Alonso, and L. Paninski. Temporal Precision in the Visual
Pathway through the Interplay of Excitation and Stimulus-Driven Suppression. Journal of
Neuroscience, 31(31):11313–11327, August 2011.
[28] Jakob H Macke, Philipp Berens, Alexander S Ecker, Andreas S Tolias, and Matthias Bethge.
Generating spike trains with specified correlation coefficients. Neural computation, 21(2):397–
423, February 2009.
12
[29] J Macke, Iain Murray, and PE Latham. How biased are maximum entropy models? Advances
in neural information processing, pages 1–11, 2011.
[30] Sebastian Gerwinn, Jakob H Macke, and Matthias Bethge. Bayesian inference for general-
ized linear models for spiking neurons. Frontiers in computational neuroscience, 4(May):12,
January 2010.
[31] Aapo Hyv arinen. Estimation of non-normalized statistical models by score matching. Journal
of Machine Learning Research, 6:695–709, 2006.
[32] Aapo Hyv arinen. Some extensions of score matching. Computational statistics & data analy-
sis, 51(5), 2007.
[33] Michael Gutmann. Noise-contrastive estimation : A new estimation principle for unnormalized
statistical models. Journal of Machine Learning Research, pages 297–304, 2009.
.
(12)
p(x, h) =
p(x) =
e
p(x) =
e
A Marginalizing over hidden units
In this section we review how the joint probability for visible and hidden units in the RBM can be
(cid:88)
integrated in closed form to obtain the marginal distribution for the visible units. First we rewrite
(cid:88)
(cid:88)
the sum over hidden variables as a product
1
(cid:32)(cid:88)
(cid:33)
xihj wij +
hj bj
xiai +
exp
M(cid:89)
Z
j
i
i,j
(cid:80)
1
and integrate p(x) = (cid:80){h} p(x, h), where the sum is over all configurations of the hidden units.
(11)
i xi ai
xihj wij + hj bj
exp
e
=
Z
i
j=1
(cid:32)(cid:88)
(cid:33)
(cid:33) M(cid:89)
(cid:32)(cid:88)
i xi ai (cid:88)
For the purpose of this derivation, we move the first term outside the product
(cid:80)
1
exp
xihj wij + hj bj
xih1wi1 + h1 b1
exp
Z
(cid:32)(cid:88)
(cid:33)
{h}
i
j=2
i
(cid:88)
i xi ai (cid:88)
We can write the sum over all hidden states as nested sums over every state for each hidden unit
(cid:80)
1
(cid:32)(cid:88)
(cid:33)
xih1wi1 + h1 b1
exp
M(cid:89)
Z
hm∈{0,1}
h1∈{0,1}
i
xihj wij + hj bj
exp
i
j=2
(cid:32)(cid:88)
(cid:33)
Noting that the term we have singled out appears only in one of the sums, we rearrange to isolate
i xi ai (cid:88)
the sum,
(cid:80)
(cid:32)(cid:88)
e
exp
xih1wi1 + h1 b1
(cid:88)
(cid:88)
M(cid:89)
h1∈{0,1}
i
xihj wij + hj bj
exp
...
(cid:32)(cid:88)
(cid:33)
hm∈{0,1}
h2∈{0,1}
i
j=2
(cid:88)
M(cid:89)
Doing this for all terms we obtain a product over individual one-dimensional sums
(cid:80)
(cid:33)(cid:35)
(cid:32)(cid:88)
(cid:34)
i xi ai
xihj wij + hj bj
exp
M(cid:89)
hj ∈{0,1}
j=1
i
(cid:80)
(cid:88)
i xi ai
(cid:34)
i
j=1
1 + exp
i
(cid:32)(cid:88)
xiwij + bj
i
(cid:33)(cid:35) .
1 + exp
M(cid:88)
j=1
log
xiwij + bj
1
Z
1
Z
1
Z
p(x) =
e
=
exp
(cid:33)
.
(15)
(16)
(13)
(14)
(17)
(18)
p(x) =
1
Z
=
e
(10)
(19)
...
.
xiai
13
(cid:34)
(cid:32)(cid:88)
(cid:33)(cid:35)
The marginal distribution over x for the RBM is now in the form of a standard energy based model,
E (x) = − (cid:88)
xiai − M(cid:88)
with energy function
i
j=1
i
The sRBM follows the same logic, since additional terms in the energy that do not depend on the
hidden units stay outside the sum over hidden configurations in the same fashion as for the visible
bias.
xiwij + bj
1 + exp
(20)
log
.
B MPF objective
B.1 Ising model
,
K (Θ) =
Here we review the derivation of the MPF objective for an Ising model, where the objective function
consists of terms connecting the data states to all states which differ by a single bit flip.
(cid:19)
(cid:18) 1
(cid:88)
(cid:88)
The general MPF objective function is given by
[E (x; Θ) − E (x(cid:48) ; Θ)]
g (x, x(cid:48) ) exp
(21)
2
x∈D
x(cid:48) /∈D
where g (x, x(cid:48) ) = g (x(cid:48) , x) ∈ {0, 1} is the connectivity function, E (x; Θ) is an energy function
parameterized by Θ, and D is the list of data states. We consider the case where the connectivity
(cid:26) 1
x and x(cid:48) differ by a single bit flip, (cid:80)
function g (x, x(cid:48) ) is set to connect all states which differ by a single bit flip,
n = 1
n xn − x(cid:48)
g (x, x(cid:48) ) =
0
otherwise
(cid:18) 1
(cid:19)
N(cid:88)
(cid:88)
The MPF objective function in this case is
[E (x; Θ) − E (x + d(x, n); Θ)]
(23)
K (Θ) =
exp
2
x∈D
n=1
where the sum over n is a sum over all data dimensions, and the bit flipping function d(x, n) ∈
(cid:26)
{−1, 0, 1}N is
(22)
d(x, n)i =
For the Ising model, the energy function is
0
−(2xi − 1)
i (cid:54)= n
i = n
(24)
K =
(25)
E = xT Jx
where x ∈ {0, 1}N , J ∈ RN ×N , and J is symmetric (J = JT ). The bias terms have been absorbed
into the diagonal of the matrix J which is possible since x2 = x holds for binary x.
(cid:2)xT Jx − (x + d(x, n))T J(x + d(x, n))(cid:3)(cid:19)
(cid:18) 1
(cid:88)
(cid:88)
Substituting this energy into the MPF objective function, it becomes
(cid:18) 1
(cid:2)xT Jx − (cid:0)xT Jx + 2xT Jd(x, n) + d(x, n)T Jd(x, n)(cid:1)(cid:3)(cid:19)
(cid:88)
(cid:88)
2
x∈D
n
(cid:18)
(cid:2)2xT Jd(x, n) + d(x, n)T Jd(x, n)(cid:3)(cid:19)
(cid:88)
(cid:88)
2
x∈D
n
− 1
(cid:34)
(cid:32)
(cid:35)(cid:33)
(cid:88)
(cid:88)
(cid:88)
2
x∈D
n
− 1
xiJin (1 − 2xn ) + Jnn
(cid:32)(cid:34)
(cid:35)(cid:33)
2
(cid:88)
(cid:88)
(cid:88)
2
x∈D
n
i
(2xn − 1)
x∈D
n
i
xiJin − 1
2
(26)
(27)
(28)
(29)
(30)
=
=
Jnn
.
=
=
exp
exp
exp
exp
exp
14
Assume the symmetry constraint on J is enforced by writing it in terms of a another possibly asym-
metric matrix J(cid:48) ∈ RN ×N ,
∂K
∂ J (cid:48)
J(cid:48) T .
J(cid:48) +
1
1
J =
2
2
(cid:32)(cid:34)
(cid:35)(cid:33) (cid:20)
The derivative of the MPF objective function with respect to J(cid:48) is
(cid:88)
(cid:88)
(2xm − 1)
xiJim − 1
(2xm − 1) xl − δlm
(cid:35)(cid:33) (cid:20)
(cid:32)(cid:34)
exp
(cid:88)
(cid:88)
2
x∈D
i
(2xl − 1)
(2xl − 1) xm − δml
x∈D
i
where the second term is simply the first term with indices l and m reversed.
xiJil − 1
2
Jmm
exp
1
2
1
2
Jll
=
lm
+
(31)
,
(32)
(cid:21)
(cid:21)
1
2
1
2
B.2 RBM
log(1 + e−Wi x )
After marginalizing out the hidden units, the energy function over the visible units for an RBM is
E (x) = − (cid:88)
given by:
i
where Wi is a vector of coupling parameters and x is the binary input vector. Bias terms have been
omitted for readability.
1 + e−Wi x+Wid(x,n)(cid:17)(cid:35)(cid:33)
(cid:34)
(cid:32)
(cid:16)
(cid:88)
− (cid:88)
(cid:88)
(cid:88)
log (cid:0)1 + e−Wi x (cid:1) +
As previously, we substitute into the objective function Eq. 23 to obtain
x∈D
n
i
i
Unlike for the Ising model there is no cancellation of data and non-data energy terms, so evaluating
the function and derivative requires looping over all bit flips for the data set.
(34)
(33)
K =
exp
log
1
2
.
B.3 sRBM
(35)
log(1 + e−WT
i x )
The energy function over the visible units for an sRBM obtained by marginalizing out the condi-
E (x; J, W) = xT Jx − (cid:88)
tionally independent hidden units is
i
where x ∈ {0, 1}N is the visible state, J = JT ∈ RN ×N is a symmetric coupling matrix, and
W ∈ RM ×N is a weight matrix to M hidden units. Equation 35 consists of a term capturing
connections between visible units (an Ising model), and a term capturing connections to hidden
units (an RBM).
The MPF objective we use again consists of energy differences between data and non-data states
(cid:104)
(cid:105)
n = − (cid:88)
differing by a single bit. For the RBM this energy difference with the nth bit flipped is
i x ) − log(1 + e−wT
log(1 + e−wT
= − (cid:88)
(cid:2)log(1 + ezi ) − log(ezi + ewin bn )(cid:3)
i (x+d(x,n)) )
dER
i
i
i x and b = 2x − 1. The energy difference
where for notational simplicity we have defined zi = wT
contributed by connections between visible units (the Ising model) is
n = 2bn yn − 1
dE I
2
(36)
(37)
(38)
Jnn
15
K =
(cid:20) 1
(cid:21)
where we define the shorthand y = Jx for simplicity. The total objective function is then given by
(cid:88)
(cid:88)
a sum over samples and bit flips as
exp
2
x∈D
(cid:88)
(cid:88)
n
To compute the gradient of this objective w.r.t. the parameters W and J we not that
∂
∂K
(cid:88)
(cid:88)
∂ J
∂ J
x∈D
n
x∈D
n
n + dER
(dE I
n )
∂K
∂W
dER
n
dE I
n
(39)
(40)
(41)
Kn
=
these terms are computed as
∂
∂ J
for the pairwise terms, and
∂
∂Wab
Kn
∂
∂ J
dE I
n =
∂
=
(cid:18)
(cid:19)
∂W
2bn yn − 1
= 2bnxn − 1
Jnn
(cid:88)
2
2
(cid:2)log(1 + ezi ) − log(ezi + ewin bn )(cid:3)
n = − ∂
(cid:88)
dER
∂Wab
i
eza
(cid:88)
xb
1 + eza
n
eza
(cid:88)
eza + ewan bn
n
j
1
1 + eza−wab bb
1
2
1
2
1
2
xb
bb
=
+
(42)
(43)
(44)
(45)
(46)
+
for the higher order terms.
C Annealed importance sampling
Zp /Zq =
Estimating the normalization constant, also referred to as partition function, of an energy based
probabilistic model, remains a challenging task [23]. Many approaches to make learning the pa-
rameters of energy based models tractable, such as Contrastive Divergence, MPF, Score and Ratio
Matching [31, 32] do not attempt to estimate the partition function. A notable exception is Noise
Contrastive Estimation [33] which treats the partition function as a parameter to be estimated, but has
only been applied for continuous-valued data. Most commonly the partition function is estimated
by sampling.
(cid:28) p(x)
(cid:29)
Using importance sampling, the partition function can be estimated by
q(x)
q(x)
where Zq is the known partition function of the proposal distribution q(x), Zp is the partition func-
tion of interest for p(x) and the symbol indicates a non-normalized distribution. The angle brack-
ets indicate a sample expectation over samples from the distribution p(x). However, if q(x) is not a
good match to the target distribution, it takes a very large number of samples to get a good estimate.
AIS uses an annealing process to gradually transform a simple proposal distribution, such as the
uniform distribution, into the target distribution, leading to an accurate estimate of Z from only a
small number of samples.
To assure convergence of the estimator, we run several annealing chains, increasing the number of
steps in factors of 2 up to a size of 105 steps. We check that the final estimate of log2 (Z ) does
not deviate more than 0.02 from the previous estimates. This criterium was chosen since log2 (Z )
appears as an additive term to L, and at a bin size τ = 50ms an error of 1 bits / second in the
final estimate of the likelihood was seen as an acceptable trade-off between estimation speed and
accuracy. In Fig. 7, we show this convergence plot for a small 20-dimensional model, where the
normalization constant was computed exactly. For larger models, where the partition function could
not be calculated analytically, we monitored that the estimate stabilized to within this tolerance.
(47)
16
Figure 7: Monitoring the convergence of the AIS estimate for the partition function, here for a
20-dimensional Ising model. Each entry on the horizontal axis corresponds to an annealing chain
with a different number of steps. Points correspond to the 500 individual samples, the blue line
is the log2 of the average from the samples. The solid green line is the true value of the partition
function computed numerically by summing over the 220 states. The dashed lines correspond to our
convergence criterion of 0.02 deviation from the true partition function.
17
102103104105106Number of intermediate distributions0.70.80.6log Z estimate |
1112.2072 | 1 | 1112 | 2011-12-09T11:09:30 | Long Brief Pulse Method for Pulse-wave modified Electroconvulsive Therapy | [
"q-bio.NC",
"physics.med-ph"
] | Modified-Electroconvulsive Therapy (m-ECT) is administered for the treatment of various psychiatric disorders. The Seizure Generalization Hypothesis holds that propagation of the induced seizure throughout the whole brain is essential for the effective ECT intervention. However, we encounter many clinical cases where, due to high thresholds, seizure is not induced by the maximum dose of electrical charge. Some studies have indicated that the ultrabrief pulse method, in which pulse width is less than 0.5millisecond (ms), is more effective at inducing seizure than conventional brief pulse (0.5ms-2.0ms). Contrary to the studies, we experienced a case of schizophrenia in which m-ECT with 1.0 and 1.5 ms width pulse (referred to as 'long' brief pulse as 0.5ms width pulse is the default in Japan) succeeded in inducing seizure, whereas ultrabrief pulse failed to induce seizure. This case is described in detail. Moreover, we discuss the underlying mechanism of this phenomenon. | q-bio.NC | q-bio |
Long Brief Pulse Method for Pulse-wave
modified Electroconvulsive Therapy
Hiroaki Inomata,1, 2, 3, ∗ Hirohiko Harima,1 and Masanari Itokawa1, 3
1Tokyo Metropolitan Matsuzawa Hospital
2Yokohama City University
3Tokyo Metropolitan Institue of Medical Science
Modified-Electroconvulsive Therapy (m-ECT) is administered for the treatment of various psy-
chiatric disorders. The Seizure Generalization Hypothesis holds that propagation of the induced
seizure throughout the whole brain is essential for the effective ECT intervention. However, we
encounter many clinical cases where, due to high thresholds, seizure is not induced by the maximum
dose of electrical charge. Some studies have indicated that the ultrabrief pulse method, in which
pulse width is less than 0.5millisecond (ms), is more effective at inducing seizure than conventional
brief pulse (0.5ms-2.0ms). Contrary to the studies, we experienced a case of schizophrenia in which
m-ECT with 1.0 and 1.5 ms width pulse (referred to as long brief pulse as 0.5ms width pulse is the
default in Japan) succeeded in inducing seizure, whereas ultrabrief pulse failed to induce seizure.
This case is described in detail. Moreover, we discuss the underlying mechanism of this phenomenon.
Introduction
Modified-Electroconvulsive Therapy (m-ECT) is administered for the treatment of various psychiatric disorders.
The Seizure Generalization Hypothesis, which underlies the mechanism of ECT, holds that propagation of induced
seizure throughout the entire brain is essential for effective ECT intervention[1] . However, there are many clinical
cases where, due to high thresholds, seizure is not induced by the maximum dose of electrical charge. In these cases,
the following procedures are considered options for inducing seizure; (i) using the older method of sine-wave ECT
(ii) promoting hyperventilation in patients [2] (iii) using anesthetic agents such as ketamine with ECT [2]. However,
these are not standard methods as sine-wave ECT induces more severe side effects than pulse-wave ECT, and not all
anesthesiologists are fully trained in the latter two procedures.
Recently randomized control trials focusing on pulse width have been conducted[3, 4]. Sackeim et al [3] reported
that the ultrabrief pulse method, in which pulse width is less than 0.3millisecond (ms), induces more therapeutic
effects and fewer side effects and requires less electrical charge to induce seizure compared to conventional brief pulse
(1.5ms).
It could be predicted then that the ultrabrief pulse would be more effective in inducing seizure in patients with high
thresholds. Contrary to this, we experienced a case of schizophrenia in which m-ECT with 1.0 and 1.5ms width pulse
(referred to as long brief pulses as 0.5ms width pulse is the default in Japan) succeeded in inducing seizure, whereas
ultrabrief pulse failed. We present this case in detail and discuss the possible underlying mechanisms.
Written informed consent was obtained from the patient and his wife (legal guardian). All personal information
has been anonymized.
Case presentation
The patient is a 35-year-old schizophrenic Japanese male. His history of illness started at age 23 with symptoms of
auditory hallucinations, persecutory delusions and psychomotor excitement. Subsequently these symptoms relapsed
every one to three years. He was discharged from hospital three years ago and was being followed as an outpatient.
One month ago, he complained to his doctor that he was being watched by strangers. The atypical antipsychotic
blonanserine and the mood stabilizer valproate were added to his medication but these produced the persecutory
delusion that the new therapy was part of an experiment instigated by his doctor. He became aggressive with verbal
threats and was admitted to our hospital due to a lack of available beds in his regular treating hospital.
On admission, he was psychotic and extremely agitated, and needed to be restrained. He initially refused treatment
but eventually agreed to his old treatment regime excluding blonanserine and valproate. His medical history revealed
∗Electronic address: [email protected]
2
that medication was of little effect for his relapsed symptoms while m-ECT was effective, although sine-wave ECT
was necessary due to his high threshold.
The presence of brain disease such as tumours was excluded after review of Computed-Tomography images. ECT
treatment was instigated three days post admission. For the ECT interventions, we used a Somatic Thymatron ECT
device, placing electrodes in a bitemporal configuration and administering propofol at 1.0mg/kg as an anesthetic
induction and succinylcholin at 1.0 mg/kg as a muscle relaxant.
Figure 1 shows the clinical course of this patient. In trials 1 to 3, we used a LOW 0.5 setting, in which the pulse
width is fixed at 0.5ms, but this failed to induce seizure at the maximum dose of electrical charge (504 milicoulomb).
We then changed the Thymatron setting from LOW 0.5 to LOW 0.25, with pulse width fixed at 0.25 ms, for trial
4, performed immediately after trial 3. Contrary to expectations, this did not induce seizure. In trial 5, a setting of
LOW 0.5 succeeded in inducing 9-second seizure, recognized by EEG and EMG charts. This seizure, however, was
deemed insufficient due to its short duration time. We then changed pulse width settings to 1.0ms and administered
trial 6. This setting successfully induced seizure with the desirable waveform and of sufficient duration (Fig 2).
For the remainder of the treatment we administered ECT, decreasing the electrical charge to avoid side effects. At
the 1.0ms pulse width, we achieved desirable seizures with 60 percent of the maximal charge. At the 1.5ms pulse
width, we succeeded in inducing seizure at 40 percent of the maximal charge. We stopped ECT treatment after 11
trials as his psychiatric condition had improved considerably from the day of admission. He was then transferred to
a chronic ward and prepared for discharge.
Discussion
For this patient, all trials at 0.25ms width pulse failed to induce seizures while 0.5ms width pulse was successful
only once (in trial 5).
It is likely that falling serum valproate concentrations enabled seizure induction in trial 5
whereas the same pulse width failed to induce convulsions in the earlier trials (1 through 3). The 1.0ms width pulse
succeeded in inducing therapeutic seizure with desirable waveform, while the 1.5ms width pulse trials induced seizure
of unknown clinical effect.
In summary, long brief pulse was more effective for inducing seizure than ultrabrief pulse for this patient. Taken
together with recent RCT studies, this case suggests that seizure threshold depends on pulse width. However, results
are contradictory. In the RCT studies, ultrabrief pulse was more effective than long brief pulse for inducing seizures
but, in this case, the reverse was true.
Recent RCT studies of ultrabrief pulses are based on the electrophysiological fact that the chronaxie (the most
effective pulse width in firing neuron) of neurons in the mammalian central nervous system lies within 0.1-0.3ms
[5]. However, West et al reported that the strength-duration curve of one-thirds of neurons is right-shifted, even in
normal subjects, so that their chronaxie is prolonged [6]. We speculate that in our patient the strength-duration curve
involved in ECT-induced seizures might be right-shifted resulting in prolonged chronaxie through to about 1.0ms.
Figure 3 shows a possible mechanistic explanation for this apparent contradiction. Moreover, this view accords well
with the fact that there is little clinical difference between a 0.25ms and 0.5ms width pulse [7]. If the strength-duration
curve is not right-shifted (in an individual, or where sample size is such that this can be assumed), then the difference
in electrophysiological response produced by a pulse width of a 0.25ms compared to 0.50ms is negligible relative to
the difference in response between 0.30ms compared to 1.5ms.
Although our hypothesis is likely to bridge clinical usage of ECT and relevant fundamental research, it is first
necessary to confirm whether our observations are limited to this case or are applicable to a wider group of patients
with careful ECT treatment.
We thank the medical staff at Matsuzawa Hospital especially Dr Okazaki.
Acknowledgments
[1] Abrams R. Electroconvulsive Therapy. 4th ed. Oxford University Press, NY, 2002
[2] Loo C et al. "Augmentation strategies in electroconvulsive therapy." J ECT, 26; 202-207, 2010
[3] Sackeim HA et al. "Effects of pulse width and electrode placement on the efficacy and cognitive effects of electroconvulsive
therapy." BRAIN STIMULATION, 1; 71-83, 2008
[4] Loo C et al. "A comparision of RUL ultrabrief pulse (0.3ms) ECT and standard RUL ECT." Int J Neuropsychopharmacol,
11; 883-890, 2008
[5] Peterchev AV et al. "ECT Stimulus Parameters: Rethinking dosage." J ECT, 26; 159-174, 2010
[6] West DC et al. "Strength-duration characteristics of myelinated and non-myelinated bulbospinal axon in the cat spinal
cord". J Physiol, 337; 37-50, 1983
[7] Niemantsverdriet L et al. "The efficacy of ultrabrief-pulse (0.25ms) versus brief-pulse (0.50ms) bilateral electroconvulsive
therapy in major depression." J ECT, 27; 55-58, 2011
3
Figure
4
FIG. 1: Clinical course
FIG. 2: EEG and EMG charts of the 5th, 6th, and 11th trials
5
FIG. 3: Strength-duration curve This graph is modified from West et al[6]
|
1704.03941 | 2 | 1704 | 2017-04-15T13:38:15 | Analysis of pacemaker activity in a two-component model of some brainstem neurons | [
"q-bio.NC"
] | Serotonergic, noradrenergic and dopaminergic brainstem (including midbrain) neurons, often exhibit spontaneous and fairly regular spiking with frequencies of order a few Hz, though dopaminergic and noradrenergic neurons only exhibit such pacemaker-type activity in vitro or in vivo under special conditions.
A large number of ion channel types contribute to such spiking so that detailed modeling of spike generation leads to the requirement of solving very large systems of differential equations. It is useful to have simplified mathematical models of spiking in such neurons so that, for example, features of inputs and output spike trains can be incorporated including stochastic effects for possible use in network models.
In this article we investigate a simple two-component conductance-based model of the Hodgkin-Huxley type. Solutions are computed numerically and with suitably chosen parameters mimic features of pacemaker-type spiking in the above types of neurons. The effects of varying parameters is investigated in detail, it being found that there is extreme sensitivity to eight of them. Transitions from non-spiking to spiking are examined for two of these, the half-activation potential for an activation variable and the added (depolarizing) current and contrasted with the behavior of the classical Hodgkin-Huxley system. The plateaux levels between spikes can be adjusted, by changing a set of voltage parameters, to agree with experimental observations. Experiment has shown that in, in vivo, dopaminergic and noradrenergic neurons' pacemaker activity can be induced by the removal of excitatory inputs or the introduction of inhibitory ones. These properties are confirmed by mimicking opposite such changes in the model, which resulted in a change from pacemaker activity to bursting-type phenomena. | q-bio.NC | q-bio | Analysis of pacemaker activity in a two-component model of
some brainstem neurons
Henry C. Tuckwell1,2∗, Ying Zhou3,, Nicholas J. Penington 4,5
1 School of Electrical and Electronic Engineering, University of Adelaide,
Adelaide, South Australia 5005, Australia
2 School of Mathematical Sciences, Monash University, Clayton, Victoria 3800, Australia
3 Department of Mathematics, Lafayette College, 1 Pardee Drive, Easton, PA 18042,
USA
5 Program in Neural and Behavioral Science and Robert F. Furchgott Center for Neural
4 Department of Physiology and Pharmacology,
and Behavioral Science
State University of New York, Downstate Medical Center,
Box 29, 450 Clarkson Avenue, Brooklyn, NY 11203-2098, USA
∗ Corresponding author: Henry C. Tuckwell, School of Electrical and Electronic
Engineering, University of Adelaide, North Terrace, Adelaide, SA 5005, Australia;
Tel: 61-481192816; Fax: 61-8-83134360: email: [email protected]
September 11, 2018
7
1
0
2
r
p
A
5
1
]
.
C
N
o
i
b
-
q
[
2
v
1
4
9
3
0
.
4
0
7
1
:
v
i
X
r
a
1
Abstract
Serotonergic, noradrenergic and dopaminergic brainstem (including midbrain) neu-
rons, often exhibit spontaneous and fairly regular spiking with frequencies of order a
few Hz, though dopaminergic and noradrenergic neurons only exhibit such pacemaker-
type activity in vitro or in vivo under special conditions. A large number of ion channel
types contribute to such spiking so that detailed modeling of spike generation leads
to the requirement of solving very large systems of differential equations. It is use-
ful to have simplified mathematical models of spiking in such neurons so that, for
example, features of inputs and output spike trains can be incorporated including
stochastic effects for possible use in network models. In this article we investigate a
simple two-component conductance-based model of the Hodgkin-Huxley type. Solu-
tions are computed numerically and with suitably chosen parameters mimic features
of pacemaker-type spiking in the above types of neurons. The effects of varying pa-
rameters is investigated in detail, it being found that there is extreme sensitivity to
eight of them. Transitions from non-spiking to spiking are examined for two of these,
the half-activation potential for an activation variable and the added (depolarizing)
current and contrasted with the behavior of the classical Hodgkin-Huxley system.
The plateaux levels between spikes can be adjusted, by changing a set of voltage pa-
rameters, to agree with experimental observations. Experiment has shown that in,
in vivo, dopaminergic and noradrenergic neurons' pacemaker activity can be induced
by the removal of excitatory inputs or the introduction of inhibitory ones. These
properties are confirmed by mimicking opposite such changes in the model, which re-
sulted in a change from pacemaker activity to bursting-type phenomena. The article
concludes with a brief review of previous modeling of these types of neurons and a
brief discussion of their involvement in some pathological pysychiatric conditions.
Keywords: Dorsal raphe nucleus, serotonergic neurons, locus coeruleus, noradrenergic
neurons, dopaminergic neurons, reduced computational models, pacemaker activity.
Contents
1 Introduction
2 A reduced physiological model with two component currents
2.1 Examples of pacemaker-like firing . . . . . . . . . . . . . . . . . . . . . . .
3 Effects of varying parameters
3.1 More detailed analysis for Ve1 and µ . . . . . . . . . . . . . . . . . . . . .
3.1.1 Half-activation potential for Ie
. . . . . . . . . . . . . . . . . . . .
3.1.2 Magnitude µ of the added current: comparison with usual Hodgkin-
Huxley parameters . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.2 Changing plateau level and the minimum of the AHP . . . . . . . . . . .
4 Firing frequency and bursting
4.1 Firing rates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Bursting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5 Discussion
5.1 LC NA neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.2 DRN SE neurons . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Midbrain DA neurons
. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Acknowledgements
3
6
8
12
14
14
15
17
20
20
21
22
23
23
24
24
2
7 References
24
Abbreviations
5-HT, 5-hydroxytryptamine (serotonin); AC, adenylate cyclase; AHP, afterhyperpolar-
ization; cAMP, cyclic adenosine monophosphate; CREB, cAMP response element bind-
ing protein; CRH, corticotropin-releasing hormone; CRN, caudal raphe nucleus; DA,
dopamine or dopaminergic; DRN, dorsal raphe nucleus; GABA, gamma-aminobutyric
acid; HH, Hodgkin-Huxley; Hz, hertz; HPA, hypothalamus-pituitary-adrenal; ISI, inter-
spike interval; LC, locus coeruleus; MDD, major depressive disorder; NA, noradrenaline
or noradrenergic; OCD, obsessive-compulsive disorder; PTSD, post-traumatic stress dis-
order; PVN, paraventricular nucleus (of hypothalamus); REM, rapid eye movement; SE,
serotonin or serotonergic.
1
Introduction
Neurons which exhibit (approximately) periodic spiking in the presumed absence of synap-
tic input are called autonomous pacemakers, and include neurons found in the subthalamic
nucleus, nucleus basalis, globus pallidus, raphe nuclei, cerebellum, locus ceruleus, ventral
tegmental area, and substantia nigra (Ramirez et al., 2011). In order to sustain spiking,
some pacemaking cells may require small amounts of depolarizing inputs, natural or lab-
oratory. Thus, for example Pan et al. (1994) found that all of 42 rat LC neurons fired
spontaneously, whereas in Williams et al. (1982) and Ishimatsu et al. (1996) it was re-
ported that most cells did not require excitatory input to fire regularly. These latter three
sets of results were obtained in vitro. Sanchez-Padilla et al. (2014) reported that spike
rate in mouse LC neurons was not affected by blockers of glutamatergic or GABAergic
synaptic input, supporting the idea that these cells were autonomous pacemakers.
In
Grace and Onn (1989), it was found that in vitro, rat midbrain dopamine neurons fired
periodically with a frequency which could be increased or decreased by the application
of depolarizing or hyperpolarizing currents without changing the pattern of spikes - see
Figure 1 herein. The mechanisms of pacemaking activity of DA neurons in the rat ventral
tegmental area (slice) was analyzed in Khaliq and Bean (2008).
The neurons with which we are mainly concerned are serotonergic neurons of the (dor-
sal) raphe nucleus, noradrenergic neurons of the locus coeruleus and midbrain dopamin-
ergic neurons. The electrophysiological properties of these cells have been much investi-
gated over the last several decades (Aghajanian and Vandermaelen, 1982; Aghajanian et
al., 1983; Alreja and Aghajanian, 1991; Williams et al., 1984; Grace and Bunney, 1983,
1984; Harris et al., 1989). The roles of serotonergic and noradrenergic neurons in stress
related disorders such as MDD and PTSD by means of reciprocal interactions with, inter
alia, the HPA axis (especially through the PVN), hippocampus and prefrontal cortex are
well documented (Lopez et al., 1999; Carrasco and Van de Kar, 2003; Lanfumey et al.,
2008; Mahar et al., 2014). For example, CRH neurons of the PVN project directly to SE
neurons of the DRN (Lowry et al., 2000) and NA neurons of the LC (Valentino et al.,
1993) and stress, via an upregulation of the cAMP pathway, leads to an increase in the
excitability of LC neurons (Nestler et al., 1999).
In addition many of these brainstem
neurons are endowed with glucocorticoid receptors which are activated by high levels of
corticosterone (cortisol) (Joels et al., 2007).
In most common experimentally employed animals except cat, the locus coeruleus is
almost completely homogeneous, consisting of noradrenergic neurons which in rat number
about 1500 (Swanson, 1976; Berridge and Waterhouse, 2003). The number of neurons
in the rat DRN is between about 12000 and 15000 (Jacobs and Azmitia, 1992; Vertes
and Crane, 1997) of which up to 50% are principal serotonergic cells (Vasudeva and
3
Waterhouse, 2014) but there are also present 1000 dopaminergic cells (Lowry et al., 2008)
and GABAergic cells, whose density varies throughout the divisions of the nucleus, as well
as several other types of neuron. In rat, the number of dopamine neurons (bilateral counts)
in the midbrain groups is approximately 71,000, with 40,000 in the VTA and 25,000 in
substantia nigra (German and Manaye, 1993; Nair-Roberts et al., 2008). In the latter two
cell groups there are also about 21,000 and 41,000 GABA-ergic cells, respectively.
The SE neurons of the DRN and NA neurons of the LC, often exhibit a slow regular
pattern of firing with frequencies of order 0.5 to 2 Hz in slice (Milnar et al., 2016) and
sometimes higher in vivo. Midbrain dopaminergic neurons exhibit similar firing patterns in
vitro. The origins of pacemaker firing differ amongst various neuronal types. Thus, brain-
stem dopaminergic neurons may fire regularly without excitatory synaptic input (Grace
and Bunney 1983; Harris et al., 1989). Underlying the rhythmic activity are subthreshold
oscillations which were demonstrated with a mathematical model to possibly reflect an
interplay between an L-type calcium current and a calcium-activated potassium current
(Amini et al, 1999) - but see Kuznetsova et al. (2010). The mechanisms of pacemaker
firing in LC neurons are not fully understood, although there is the possibility that it
is sustained by a TTX-insensitive persistent sodium current (De Carvalho et al., 2000;
Alvarez et al., 2002). For serotonergic neurons of the DRN, there have been no reports
of a persistent sodium current and L-type calcium currents are relatively small or absent
(Penington et al., 1991) so the main candidate for depolarization underlying pacemaking
is a combination of T-type calcium current and the classical fast TTX-sensitive sodium
current which dominates the pre-spike interval (Tuckwell and Penington, 2014). In some
cells the hyperpolarization-activated cation current may also play a role.
In Figure 1 are shown portions of spike trains in NA neurons of mouse and rat LC, SE
neurons in rat DRN and CRN and DA neurons in rat midbrain. It is noteworthy that the
frequency of firing of LC and DRN principal neurons depends on sleep stage. Thus for
example, in rats, waking, slow-wave sleep and REM sleep are accompanied by LC firing
rates of about 2.2 Hz, 0.7 Hz and 0.02 Hz, respectively (Foote et al., 1980; Aston-Jones
and Bloom, 1981; Luppi et al., 2012).
4
Figure 1: Some representative spikes from rat and mouse raphe nuclei and LC neurons.
A. Part of a train of spikes in a rat LC neuron in slice. Markers 20 mV and 200 ms.
(Andrade and Aghajanian, 1984). B. Detail of the course of the average membrane
potential in a mouse LC neuron during an interspike interval. (De Oliveira et al., 2010).
C. Action potentials in infant (7 to 12 days) and adult (8 to 12 weeks) mice. (De Oliveira
et al., 2011). D. Whole-cell current-clamp recording of spikes in a mouse (21 to 32 days)
LC neuron. (Sanchez-Padilla et al., 2014). E. A few spikes from rat dorsal raphe nucleus
(slice), (Vandermaelen and Aghajanian, 1983). F. Portion of a spike train from rat caudal
raphe nucleus. (Bayliss et al., 1997). G. Train of spontaneous spikes at a mean frequency
of 0.85 Hz for a rat LC neuron in vitro. (Jedema and Grace, 2004). H. In vitro spiking
in a rat dopaminergic neuron with two levels of applied current. (Grace and Onn, 1989).
Figure 2 shows computed spikes for the detailed model of rat DRN SE neurons of
Tuckwell and Penington (2014) with 4 different parameter sets. In this model the main
variables are membrane potential and intracellular calcium ion concentration which satisfy
ordinary differential equations. There are, however, 11 membrane currents which drive the
model, which results in a system with 18 components and over 120 parameters. In order
to study quantities like interspike interval distributions with various sources of random
synaptic input, it is helpful to have a simpler system of differential equations which might
yield insight into the properties of the complex model whose execution with random inputs
over hundreds of trials would be overly time consuming. It is also pointed out that LC and
DRN principal neurons are each responsive to activation of about 20 different receptor
types, making computational tasks even more cumbersome with an 18-component model
(Kubista and Boehm, 2006; Maejima et al., 2013). A simplified model of spike generation
5
would be useful in modeling the dynamics of serotonin release and uptake (Flower and
Wong-Lin, 2014) and in simplified models of brainstem neural networks (Joshi et al.,
2017). This approach has been proven useful by several authors, some of whom reduced
the number of component currents (for example Fitzhugh, 1961; Nagumo et al., 1962;
Destexhe et al., 2001) whereas others have simplified the geometry of the dendritic tree
(Rall, 1962 ; Walsh and Tuckwell, 1985; Bush and Sejnowski, 1993).
Figure 2: Examples of computed spikes in a model for serotonergic neurons of the rat
dorsal raphe nucleus from the model of Tuckwell and Penington (2014). Illustrated are
the prolonged afterhyperpolarizations after a spike. The subsequent climb to threshold is
plateau-like, sometimes being almost horizontal. Membrane potentials in mV are plotted
against time in ms.
2 A reduced physiological model with two component cur-
rents
With the multi-component neuronal model for DRN SE neurons (Tuckwell and Penington,
2014) there are 11 component currents and a variety of solution behaviors for different
choices of parameters.
It is often difficult to see which parameters are responsible for
various spiking properties. Some parameter sets lead to satisfactory spike trains with
properties similar to experiment, but sometimes spike durations are unacceptably long.
Other sets give rise to spontaneous activity but with spikes which have uncharacteristic
large notches on the repolarization phase. Doublets and triplets are also often observed,
and although these are sometimes found in experimental spike trains, it is desirable to find
solutions depicting the regular singlet spiking and relatively smooth voltage trajectories
usually observed.
Here we describe a simplified two-component model whose solutions exhibit membrane
potential trajectories resembling those for pacemaker spiking in many brainstem neurons,
as depicted in Figure 1. The several currents which act to mainly depolarize the neurons
are lumped together in a single current denoted by Ie, and similarly the currents which
6
act to repolarize the cells are lumped into Ii. Additionally there is an applied current
Ia which may be required to maintain pacemaker activity and which may also contain
excitatory and inhibitory synaptic input. Ie and Ii also subsume a leak current. The
basic differential equation for the voltage V (t) at time t is written
C
dV
dt
= −[Ie + Ii + Ia]
(1)
where C is capacitance. The initial value of V is usually taken to be the resting poten-
tial, VR. All potentials are in mV, all times are in ms, all conductances are in µS and
capacitances are in nF.
The form of the current Ie is based on that of the usual fast sodium transient currrent
IN a in the 11-component model for DRN SE cells and in the original Hodgkin-Huxley
(1952) model. The current Ii is based on the form of the delayed rectifier potassium
current IKDR in the 11-component model. The inclusion of just the currents Ie, Ii, and
if necessary a small constant driving applied current, is found, with appropriate sets of
parameters, to be sufficient to give spikes whose trajectories resemble many found in
pacemaker spiking of brainstem neurons as exemplified by those illustrated in Figures 1
and 2. These solutions have properties contrasting with the usual ones for the Hodgkin-
Huxley system - see for example Hodgkin and Huxley (1952), Tuckwell and Ditlevsen
(2016) and Figure 10 herein.
The depolarizing current, Ie is given by the classical form
Ie = ge,maxm3
ehe(V − Ve)
(2)
with activation variable me, inactivation variable he, maximal conductance ge,max and
reversal potential Ve. For the steady state activation we put
me,∞ =
1
1 + e−(V −Ve1 )/ke1
.
If the corresponding time constant is assumed to be voltage-dependent, we write it as
τm,e = ae + bee−(cid:0)(V −Ve2 )/ke2(cid:1)2
.
The steady state inactivation is given by
he,∞ =
1
1 + e(V −Ve3 )/ke3
,
and if its time constant is voltage dependent, we put
τh,e = ce + dee−(cid:0)(V −Ve4 )/ke4(cid:1)2
.
We take the hyperpolarizing current to be
Ii = gi,maxnnk(V − Vi)
(3)
(4)
(5)
(6)
(7)
where n (the traditional symbol) is the activation variable, nk is a usually positive integer-
valued index, gi,max is the maximal conductance and Vi is the reversal potential. The
steady state activation for Ii is written
n∞ =
1
1 + e−(V −Vi1 )/ki1
(8)
7
and the time constant is, if voltage dependent,
τn = ai +
bi
cosh((V − Vi2)/ki2)
(9)
Activation and inactivation variables for Ie are determined by the first order equations
dme
dt
=
me,∞ − me
τme
,
dhe
dt
=
he,∞ − he
τh,e
.
(10)
(11)
The activation variable n for Ii satisfies an equation like (10).
In some instances, the time constants are chosen to be constant, independent of V, in
which case they are denoted by τm,ec, τh,ec and τn,c for τm,e, τh,e and τn, respectively.
2.1 Examples of pacemaker-like firing
We illustrate the pacemaker-like solutions with two choices of parameter sets. Set 1 is
based on the parameters used for fast sodium and delayed rectifier potassium in the 11-
current component model (Tuckwell and Penington, 2014). Set 2 is adopted from the
parameters in a model of a rat sympathetic neuron with five conductances (Belluzzi and
Sacchi ,1991). For the second set we also use the resting potential and the cell capacitance
from Kirby et al. (2003) for rat DRN SE cells. The two sets of parameters are summarized
in Table 1.
Table 1: Two basic parameter sets
Parameter
Ve1
ke1
Ve3
ke3
VR
τm,ec
τh,ec
C
Vi1
ki1
nk
τn,c
ai
bi
Vi2
ki2
ge,max
gi,max
Ve
Vi
Set 1
-33.1
8
-50.3
6.5
-60
0.2
1.0
0.04
-15
7.0
1
-
1
4
-20
7
2.00
0.5
45
-93
Set 2
-36
7.2
-53.2
6.5
-67.8
0.1
2.0
0.08861
-6.1
8.0
1
3.5
-
-
-
-
1.5
0.5
45
-93
Both of these parameter sets led to repetitive spiking with the addition of a small
depolarizing current (see Table 2). Typical spike trains are shown in Figures 3 and 4, and
8
V
(mV)
I
(nA)
20
0
−20
−40
−60
−80
15
10
5
0
−5
−10
Set 1
600
700
800
900
1000
1100
1200
1300
I
e
I
i
t (ms)
0.2
0
F∞
−0.2
−0.4
−15
612
614
616
618
620
622
−80
−60
−40
−20
0
t (ms)
V (mV)
Figure 3: Top. Repetitive spiking in the two-current (Ie − Ii)) model for the parameter
set 1 at the approximate threshold for spiking. Bottom left. The currents Ie and Ii during
spikes. Bottom right. The function F∞(V ) defined in Equ. (12) without added current
(thin blue curve), whereby spiking does not occur, and with sufficient depolarizing current
of 0.0342 nA (red curve) to give rise to pacemaker activity.
Table 2 contains lists of some of the details of the spike and spike train properties. Spiking
for the second parameter set has a lower threshold for (repetitive) spiking, a longer ISI at
threshold, a longer spike duration and a larger spike amplitude. For both parameter sets,
most of these spike properties are in the ranges observed for DRN SE neurons. Figures 3
and 4 show well defined spikes with abruptly falling repolarization phases to a pronounced
level of hyperpolarization followed by a steady increase in depolarization until an apparent
spike threshold is reached. In the lower left-hand panels of Figures 3 and 4 are shown,
V
(mV)
0
−50
−100
15
10
5
0
−5
−10
−15
2700
I
(nA)
Set 2
1800
2000
2200
2400
2600
2800
t (ms)
3000
3200
3400
3600
3800
F∞
0.4
0.2
0
−0.2
I
e
I
i
2705
2710
t (ms)
−80
−60
−40
−20
0
V (mV)
Figure 4: Top. Repetitive spiking in the two-component model for the parameter set 2 at
the approximate threshold for spiking. Bottom left. The currents Ie and Ii during spikes.
Bottom right. The function F∞(V ) defined in Equ. (12) without added current (thin blue
curve), whereby spiking does not occur, and with sufficient depolarizing current of 0.018
nA (red curve) to give rise to pacemaker activity
9
on an expanded time scale, the excitatory current Ie and the inhibitory current Ii during
spikes. In the lower right-hand panels are shown plots (thin blue curves) of the function
F∞(V ) defined as the sum of the steady state (t → ∞) values of the quantities of Equ.
(2) and Equ. (7),
F∞(V ) = −[ge,maxm3
e,∞(V )he,∞(V )(V − Ve) + gi,maxnnk
∞ (V )(V − Vi)].
(12)
The behavior of this function near the resting potential has been found to provide a
heuristic indicator for the occurrence of spiking (Tuckwell, 2013; Tuckwell and Penington,
2014). It can be seen in both Figures 3 and 4 that, for both parameter sets, the function
F∞(V ) is negative for V in an interval of considerable size around the resting potential VR,
which is indicated by the vertical at -60 mV in Figure 3 and at - 67.8 mV in Figure 4. This
means that around V = VR the derivative of V with respect to time tends to be negative so
that spontaneous spiking is unlikely. The magnitude of the smallest depolarizing current
required to enable spiking is approximately the amount −µc which must be added to make
F∞(V ) − µ positive at and around VR. The resulting curves, shown in red in Figures 3
and 4, are approximately tangential to the V-axis, being obtained with µc = −0.0342 for
set 1 and µc = −0.018 for set 2. Hence these values of µ are estimates of the threshold
depolarizing current required for (pacemaker) spiking.
Many brainstem (particularly DRN SE neurons) often have characteristically long
plateau-like phases in the latter part of the ISI and this is apparent in Figures 3 and 4 for
spikes elicited near the threshold for spiking for both sets of parameters. The plateau for
the second set is nearly three times as long as that for the first set. For both parameter
sets, (not shown) at a particular value of µ the frequency jumps from zero to a positive
value, being 3.0 Hz for set 1 and 1.1 Hz for set 2, so that this model with the chosen
parameters would be classified as one with Hodgkin (1948) type 2 neuron properties.
Figure 5 shows, for both parameter sets, the computed spike trajectories and ISIs for
levels of excitation not much above threshold, µ being from 1 to 1.05 times the threshold
values of µc. In each case the spike trajectory displays a typical pronounced and prolonged
post-spike afterhyperpolarization followed by a plateau-like phase before the next spike.
The spiking properties for the two-component model with the parameter sets in Table
1 have several features in common with the experimental spike trains of many brainstem
neurons including DRN SE and LC NA cells. The frequency of action potentials is in good
agreement with experimental values near threshold. However, as the level of excitation
is increased to much greater values, the frequency becomes somewhat high (not shown)
relative to the most commonly reported values for sustained firing in these brainstem
neurons, although there are exceptions; for example, in rat LC neurons, Korf et al. (1974)
found in unanesthetized in vivo preparations, frequencies up to 30 Hz, and Sugiyama
et al.
In midbrain serotonergic raphe
neurons, Kocsis et al. (2006) obtained in vivo rates with mean 5.4 Hz. For experiments
with depolarizing current injection, Li and Bayliss (1998) obtained initial firing rates of
around 8 Hz for caudal raphe with an injected current of 60 pA; Li et al. (2001) reported
firing of rat DRN SE cells at frequencies as high as 35 Hz with current injection, and
Ohliger-Frerking et al. (2003) found rates as high as 8 Hz and 11 Hz with 100 pA current
injection in lean and obese Zucker rats, respectively. See also subsection 4.1 on firing
rates.
(2012) reported in vivo frequencies over 7 Hz.
The simplified model is expected to be useful in predicting approximate responses
to random synaptic inputs in contradistinction to a sustained depolarizing current as
employed in some experiments, which can lead to very large firing rates. However, with
other parameter sets the model may not exhibit such high frequencies as depolarization
level increases.
10
SET 1
µ=1.01 x µ
c
ISI=200
1000
1100
1200
0
−50
−100
900
0
−50
V
(mV)
µ=1.02 x µ
c
SET 2
µ=1.01 x µ
c
ISI=731
2000
2200
2400
2600
2800
µ=1.02 x µ
c
0
−50
−100
0
−50
ISI=161
ISI=625
−100
850
900
950 1000 1050 1100 1150
−100
2200
2400
2600
2800
3000
3200
0
−50
−100
1.05 x µ
c
ISI=118
850
900
950 1000 1050 1100
1.05 x µ
c
0
−50
ISI=479
−100
2000
2200
2400
2600
2800
3000
t (ms)
Figure 5: Spike trajectories and ISIs obtained in the reduced model for the parameter sets
of Table 1. Shown are results for Set 1 (left part) and Set 2 (right part), when applied
currents are increased by relatively small amounts of 1%. 2% and 5% above threshold
(top, middle, bottom curves).
Table 2: Spike train properties
Property
Spike threshold µc
ISI at threshold
Spike duration (-40 mV)
Max V
Min V
Set 1
-0.0342
331 ms
1.6 ms
+8
-90.0
Basic set 2
-0.018
948 ms
2.9 ms
+19.4
-91.2
As Figure 1 shows, not all regular spiking in brainstem neurons entails a long afterhy-
perpolarization and a long plateau. Figure 6 shows that spikes similar to those in Figures
1B, 1D and 1F are generated in the second model with increased depolarizing current.
11
20
0
−20
−40
−60
−80
−100
0
40
20
0
−20
−40
−60
−80
−100
0
50
100
150
200
250
300
50
100
150
200
250
300
Figure 6: Spiking in the reduced model for the parameter sets of Table 3 where in both
cases µ = −0.05, a value considerably above threshold and which results in short ISIs,
characterized by relatively short plateaux and short-lasting AHPs. Top part, set 1; bottom
part, set 2.
3 Effects of varying parameters
The simplified model has 19 parameters for set 1, which compares with over 120 for
the original model. The dependence of spiking on changes in the 19 parameters was
investigated by computing solutions with each parameter above and below its standard
value as given in Table 1. The focus in this subsection is the effect of parameter changes
on the ISI, athough 6 other properties were routinely computed on each run, being spike
duration, upslope on leading edge of spike, duration of AHP, amplitude of AHP, duration
of plateau phase and approximate voltage at the half-way point of the plateau. The
MATLAB computer programs employed to determine these properties are available from
the authors on request.
The % change in the ISI for increases and decreases in parameters by ± 0.1 % is
shown for 19 parameters in Table 3. The entries are ranked according to the degree
to which they change the ISI, with rank 1 for the parameter with the greatest effect.
Examination of these results reveals that small changes in 6 parameters, ranks 1 to 6,
can result in a very large change in the ISI, making this quantity infinite as spiking
is abolished altogether. Five of these 6 parameters are involved in the activation and
inactivation of Ie and the activation of Ii, the other one being the reversal potential for Ii.
The inhibitory conductance, gi,max is the 7th ranked parameter making changes of 47.0
% and -17.8 % for increases and decreases of 0.1 %, respectively. Small changes in the
parameters, ranked 8 to 11, make moderate changes in the ISI of order 10 to 20 %, being
the excitatory conductance, the depolarizing drive current µ and the excitatory reversal
potential. Changing the remaining parameters by ± 0.1% has minor to insignificant effects
on the ISI. The parameter whose such relatively small changes have the least effect on the
ISI is the resting potential, VR.
More details of the effects of changing the parameters are given in Table 4. Here the
parameters are grouped to reflect their different strengths in changing the ISI. The results
point to the extreme sensitivity of the ISI to small relative changes in the following 8
parameters.
12
Table 3: Percentage changes in ISI for ±0.1% changes in parameter
-48
∞
∞
-2.4
∞
-44.4
0.0
∞
∞
-40.2
-26.6
2.7
-28.7
Rank Parameter +0.1 % -0.1 %
1
3
6
11
5
2
19
4
15
16
13
18
17
14
8
7
9
12
10
Ve1
ke1
Ve3
ke3
Vi1
ki1
VR
Vi
τm,ec
τh,ec
ai
bi
Vi2
ki2
ge,max
gi,max
µ
C
Ve
∞
0.0
-30.3
0.009
-0.007
-0.03
-0.002
0.007
-0.01
-12.2
47.0
12.1
0.14
-6.3
-0.009
0.010
0.03
0.002
-0.007
0.01
21.1
-17.8
-8.6
-0.14
8.0
• gi,max, the maximal hyperpolarizing conductance. Increases of as little as 0.005%
are sufficient to prevent spiking. Decreases of 0.05% lead to a 43% decrease in the
ISI, or about a 50% increase in firing rate.
• Ve1, the half-activation potential for the depolarizing current, Ie. When this param-
eter is increased by only 0.05%, spiking is extinguished and when it is decreased by
0.05% the ISI is shorter by 37%.
• ke1, the slope factor for the activation function of Ie. Decreasing this quantity by
0.05% was sufficient to extinguish spiking, whereas increasing it by the same amount
led to a 28% decrease in the ISI.
• ki1, the slope factor for the activation of Ii. Decreasing this parameter by 0.05%
also extinguished spiking and an increase by 0.05% gave a 33% reduction in ISI, or
a 50% increase in firing frequency.
• ge,max, the maximal depolarizing conductance. A reduction of as little as 0.05% ln
this quantity was sufficient to abolish spiking, whereas a 0.05% increase resulted in
a 72% decrease in the ISI.
• Vi, the reversal potential for the hyperpolarizing current, A decrease of 0.05% of
this parameter gave a 65% increase in ISI, equivalent to a substantial reduction in
firing rate, whereas a 0.05% increase led to a 20% decrease in the ISI.
• C, the whole cell capacitance. Reducing this quantity by the relatively small amount
of 0.05% resulted in about a 50% increase in firing rate whereas a 0.5% increase in
C led to a 50% decrease in firing rate.
• µ, the added current. Increasing this by 0.05% (that is making it less negative and
hence less depolarizing) made the ISI infinite, whereas decreasing µ by 0.05% gave
a 27% reduction in the ISI.
13
Table 4: % changes in ISI for various % changes in set 1 parameters
0.01
∞ 25.7
-4.2
-16.9
-4.8
-18.6
9.1
-28.6
1.3
0.3
∞ 18.1
0.0
0.0
-5.3
-19.8
25
2.5
-2.1
20
-9.3
-0.5
1.0
-3.1
50
4.7
-4.2
50
% Changes → 0.05
Ve1
Ve3
Vi1
ke1
ke3
ki1
VR
Vi
% Changes
τm,ec
τh,ec
% Changes
ai
bi
Vi2
ki2
% Changes
ge,max
gi,max
% Changes
µ
% Changes
C
% Changes
Ve
-17.2
-1.3
1.6
-9.7
10
-72.4
5
-66.7
∞
∞
0.1
0.5
∞ 12.1
25
50
33.3
66.2
0.5
-22.3
-33.1
1
0.005
10.4
-2.3
-2.5
-4.9
0.1
7.8
0.0
-2.8
10
1.0
-0.8
10
-3.4
-0.2
0.6
-1.4
-0.005
-7.6
2.4
2.8
5.8
-0.1
-6.1
0.0
3.1
-10
-1.0
0.7
-10
3.3
0.3
-0.8
1.1
-1
8.9
∞ -10.4
0.05
-0.05
-4.6
5.5
-10
10
-13.5
13.4
0.1
-0.1
8.0
-6.3
-46.1
1
-0.05
-0.01
-37.0
-13.6
41.5
5.1
6.4
53.7
13.0 ∞
-1.2
-0.3
-11.1
-32.8
0.0
0.0
65.1
6.5
-25
-50
-5.3
-2.6
1.8
1.8
-50
-20
6.6
15.4
1.3
0.5
-7.8
1.9
1.8
2.5
-5
-10
21.1 ∞
-17.8
-0.1
-8.6
-25
-33.7
-0.5
-1
111.8 ∞
-43.1
-0.5
-27.9
-50
-67.4
Changes in the remaining 11 parameters have moderate to minor effects on the ISI -
see also Table 3.
3.1 More detailed analysis for Ve1 and µ
In this subsection the transition from non-spiking to spiking for two of the parameters is
examined more closely near the critical values.
3.1.1 Half-activation potential for Ie
The parameter ranked 1 in Table 3 is Ve1 which is the half-activation potential for the
activation variable of the excitatory current.
Its standard value was -33.1 mV which
gave rise to pacemaker-type spiking with an ISI of 331 ms. Since only a small increase
in this parameter abolished spiking, it is of interest to study the behavior of solutions
near a critical value above -33.1 at which no spiking occurs. To this end, solutions were
computed with set 1 parameters with varying values of Ve1. Results are shown in Figures
7 and 8. In Figure 7 are shown computed spike trains, i.e., V versus t, whereas in Figure
8 are shown plots of V versus n for the indicated time periods.
In panel A, the value of Ve1 is the standard value (-33.1) and the spike train is as
described above with a constant ISI (Figure 7A). The trajectories of V versus n for each
spike are practically identical, motion being clockwise on these orbits (Figure 8A).
In the panels B, the value of Ve1 is just 0.024800 % above the standard value and
14
A
C
V
V
20
0
−20
−40
−60
−80
−100
0
−50
−52
−54
−56
−58
−60
500
1000
T=30000
1
2
3
x 104
B
D
20
0
−20
−40
−60
−80
−100
−120
0
−50
−52
−54
−56
−58
−60
0
100
200
300
0
100
t
200
300
t
Figure 7: Voltage versus time plots for various values of the half-activation potential of
the excitatory current. A. The standard set of parameters including the value of Ve1 at
-33.1 mV. Pacemaker activity is sustained with an ISI of 331 ms. B. The value of Ve1 is
slightly above the critical value for spiking at 0.02480 % above the standard value. The
ISI is about 13000 ms. C. With the value of Ve1 slightly higher at 0.02481 % above the
standard value, there is no spiking and V approaches a limiting value of about -53 mV
after about 200 ms. D. An even larger value of Ve1, 2.5% above -33.1 mV, results in an
asymptotically smaller departure from rest to about -56 mV which is achieved in about
20 ms.
this is very close to the critical value (call it Ve1,c) beyond which spiking does not occur
because it is too far from the resting potential. As seen in Figure 7B, the first spike occurs
at 12980 ms and the second at 26010 ms, but somewhat remarkably, the orbits (for the
two spikes)as seen in Figure 8B are visually indistinguishable from those in Figure 8A
where the ISI is only about 1/40 as long at 331 ms.
The V versus t curve for a value of Ve1 which is slightly larger at 0.024810 % above the
standard value, and presumably greater than the critical value Ve1,c, is shown in Figure
7C. Here there are no spikes (up to 30000 ms) and the values of (V, n) have limiting values
(V ∗, n∗) of about (-53.0038, 0.004368) as seen in Figure 8C. Again, somewhat remarkably,
the trajectory from rest to (V ∗, n∗) lies almost exactly on the spiking orbit as it approaches
the limiting value.
With the considerably higher value of Ve1, at 2.5 % above the standard value, there
are, as expected, no spikes as seen in Figure 7D. The equilibrium point is closer to rest
at (V ∗, n∗)=(-56.4600, 0.002670). The trajectory in (n, V ) space still adheres to the early
part of the spike orbit.
3.1.2 Magnitude µ of the added current: comparison with usual Hodgkin-
Huxley parameters
The parameter ranked 9 in Table 3 is µ which is the magnitude of the added depolarizing
current which may be required to induce pacemaker-type spiking.
Its set 1 value was
µ = −0.0342 which was near the threshold value for spiking when all other parameters
had their set 1 values. The solutions are not as sensitive to changes in this parameter as
15
A
20
0
−20
−40
−60
−80
−100
B
20
0
−20
−40
−60
−80
−100
−0.1
0
0.1
0.2
0.3
−0.1
0
0.1
0.2
0.3
C
(n*, V*)
V
−50
−52
−54
−56
−58
−60
D
−50
−52
−54
−56
−58
−60
(n*, V*)
0
1
2
3
4
5
x 10−3
n
0
1
2
3
4
5
x 10−3
Figure 8: Plots of V versus n for the corresponding plots A-D of V versus t in Figure 7.
Note the similarity of the trajectories in A and B, despite enormously different ISIs. In C
and D, corresponding to only subthreshold responses, trajectories are shown approaching
asymptotic values at (V ∗, n∗).
to those in Ve1. Figure 9 shows the plots of V (t) versus n(t) for four values of µ.
In the left panel orbits are shown for the the standard value -0.0342 of µ, which gives
an ISI of 331 (red curves), and for a value 0.44 % above (more positive and hence more
hyperpolarizing) the standard value (green curves) which yields an extremely small spike
rate with the first spike not occurring until 13,700 ms. The orbits for these two values of
µ are almost coincident, but one can on close inspection distinguish parts of the red and
green curves. The critical value of µ, called µc, for spiking is apparently not far above the
second value.
In the right panel of Figure 9 are shown orbits, effectively from t = 0 to t = ∞, for
two values of µ above µc, at which there is no evidence of spiking. The red curve is for
a value of µ just 0.5 % above the standard value. The asymptotic value of (V, n) is with
V ∗ = −53.2313 and n∗ = 0.004229. Again the curve closely adheres to the early part
of the spike trajectory. This is also the case for the much higher value of µ > µc at 20
% above the standard value. The curve for this case is blue and terminates a shorter
distance from the rest point, with V ∗ ≈ −57.21 and n∗ = 0.0024.
Usual HH parameters
The manner in which the periodic orbit emerges as µ passes through a critical value
in the above calculations is seemingly quite different from that in which periodic spiking
arises in the HH model with the usual parameters as given, for example, in Tuckwell and
Ditlevsen (2016). The latter case is illustrated in Figure 10 where there are shown in A,
B, C, respectively the responses to input currents well below threshold for periodic spiking
(µ = 3.96), not far below (µ = 5.61) and above threshold (µ = 6.6). Here the subthreshold
responses are oscillatory and decay is into the rest point. This is in contradistinction to
the behavior of solutions with the parameters of the present model - see Figure 9. The
nature of the bifurcation to periodic solutions that occurs at the critical values for the
original Hodgkin-Huxley parameters was discussed in Tuckwell et al. (2009). Further
16
analysis is required to see if the bifurcation to periodic solutions in the present model has
a different mathematical structure from that in the original model.
SPIKING PATHS FOR
2 VALUES OF µ
V
20
0
−20
−40
−60
−80
−50
−51
−52
−53
−54
−55
−56
−57
−58
−59
−60
V
NON−SPIKING PATHS
FOR 2 VALUES OF µ
(n*, V*)
(n*, V*)
REST POINT
−100
−0.1
0
0.1
0.2
0.3
−61
0
n
1
2
3
n
4
5
x 10−3
Figure 9: Plots of V versus n as the depolarizing current is changed above and below
the critical value for spiking. In the left panel are shown spiking orbits for the standard
value of µ =-0.0342 (red curves) which gives an ISI of 331 ms, and a value which is 0.44
% above the standard value (green curve) which results in an ISI of about 14,000 ms.
The phase diagrams for these widely disparate values of the ISI are almost identical. In
the right hand panel are orbits for values of µ which are too large (hyperpolarizing) to
induce spiking. The orbits again asymptote to subthreshold values at (V ∗, n∗). These
subthreshold orbits coincide with the early parts of the spiking orbits shown in the left
panel.
3.2 Changing plateau level and the minimum of the AHP
Examination of the membrane potential trajectories during pacemaker-type firing often
shows a relatively long plateau between successive spikes during which the membrane
potential increases slowly, sometimes very slowly before climbing past threshold to give
rise to the subsequent spike. Such plateaux are exemplified by the recordings shown in
Figure 11, where the dopamine neuron spikes are from Grace and Bunney (1995), those
for noradrenaline from Andrade and Aghajanian (1984), and those for serotonin from
Vandermaelen and Aghajanian (1983).
In the case of the in vitro recording shown for
the DA cell, the plateau phase is not so flat, but an example with a flat plateau phase is
shown in Figure 1H (from Grace and Onn, 1989) for a cell with an applied hyperpolarizing
current of 0.07 nA. The resting potential for the DA cell shown in Figure 11 (middle dotted
line) is -50 mV, for rat NA cells the average resting potential is -58.2 mV based on 11 sets
17
8
6
4
2
0
0
100
50
0
0
100
50
0
0
µ=3.96
A
50
100
µ=5.61
B
50
µ=6.6
100
C
8
6
4
2
0
−2
100
50
0
100
50
0
0.34
0.36
0.38
0.4
0.3
0.4
0.5
0.6
0.7
50
100
0.3
0.4
0.5
0.6
0.7
Figure 10: Solutions V (t) versus t (left column) and phase diagrams V versus n (right
column) for the Hodgkin-Huxley equations with the original parameters. In A, the input
current is well below the spiking threshold, in B it is not far below and in C, it is above
the threshold for repetitive spiking. The subthreshold behaviors are quite different from
those illustrated in Figure 9, right hand diagram.
of experimental results, and the average resting potential for SE neurons of the DRN is
about -60 mV.
The DA and NA recordings are in vitro recordings from rat substantia nigra and locus
coeruleus, respectively. It is likely that regular pacemaker-type spiking in these two types
of cells usually only occurs in vitro. Grace and Onn (1989) stated that DA cells recorded
in vitro fired exclusively in a highly regular pacemaker-like pattern. This is supported by
results reported by Grenhoff et al. (1988) for DA neurons and by Svensson et al. (1989)
on NA neurons. In these two sets of in vivo experiments, application of the excitatory
amino acid antagonist kynurenate (or kynurenic acid) changed the irregular (bursting)
nature of spiking of DA and NA neurons to regular pacemaker type activity.
Since in vitro recordings usually have diminished and possibly no synaptic input, the
findings in these latter two references offer a plausible explanation why regular pacemaking
often occurs in vitro but rarely in vivo (see for example Sugiyama et al., 2011). Findings
related to that of Grenhoff et al. (1988) were that the GABAB agonist baclofen made
firing in DA neurons slower and more regular (Engberg et al.,1993), and that similar
effects resulted from the application of NMDA receptor blockers (Overton and Clark,
1992; Cherugi et al., 1993).
In another more recent study, Blythe et al. (2009) found
that in DA neurons of the substantia nigra, bursting could be induced by the application
of glutamate at somatic or dendritic sites. In a related interesting investigation in DA
neurons, Putzier et al. (2009) found that the voltage characteristics of L-type Cav1.3
18
Figure 11: Pacemaker-type spiking in a DA neuron (top) from Grace and Bunney (1995)
(also in Grace and Onn, 1989), from an NA neuron (middle) from Andrade and Agha-
janian (1984), and an SE neuron (bottom) from Vandermaelen and Aghajanian, (1983).
In each trajectory there is evidence of an extended plateau phase during which the mem-
brane potential undergoes relatively minor changes before ascending at the end of the ISI
to give rise to a subsequent spike.
channels rather than Ca2+ selectivity were important factors in pacemaking and bursting
activity.
For SE neurons, regular pacemaking has been found in both in vivo and in vitro
experiments, suggesting less fine tuning by excitatory synaptic inputs with a more robust
pacemaker mechanism. Levine and Jacobs (1992) applied kynurenic acid to cat DRN
serotonergic neurons and found that their spontaneous firing rates were only reduced by
about 4 %.
It can be seen that the voltage levels of the plateaux between the NA and SE spikes
in Figure 11 are often not much above resting potential. However for the model with
19
set 1 parameters, as shown in Figures 3 and 5, the plateau level is about 6.8 mV above
resting value. It turns out that the plateau level and minimum can be easily adjusted
with a change of parameters which are voltages, being in particular Ve, Vi, Ve1, Ve3, Vi1, Vi2.
These were all changed together by an amount ∆V . With ∆V = −5 mV applied to set 1
parameters no spikes emerged.
With all set 1 parameters unchanged with the exception that the reversal potential
for the inhibitory current was at a more depolarized level of Vi = −88 mV, spiking was
very rapid with an ISI of only 67.8 ms. If the maximal inhibitory conductance was then
multiplied by 1.1427, the ISI was back to 338.9 ms, close to the value for set 1 parameters.
Calling this set 3, with altered Vi and altered gi,max, the plateau level was 7 mV above
resting level. When the set Ve, Vi, Ve1, Ve3, Vi1, Vi2 was uniformly decreased by 5mV, (set
3'), the ISI was virtually unchanged, but the plateau level was reduced to only 2mV above
resting potential. The spike trains for sets 3 and 3' are shown in Figure 12, with red and
blue curves, respectively. The minimum value of V during the AHP for set 3 is -85.7 mV
and -90.7 for set 3'. Some of the remaining spike train properties for set 3 parameters
(with corresponding properties for set 3' in brackets) are: spike duration 1.6 ms (1.27),
upslope of spike 100 V/s (111), duration of AHP 38.8 ms (64), AHP amplitude 25.7 mV
(30.7) and plateau duration 270 ms (247). Several of these properties are quite different
for sets 3 and 3' and it is remarkable that the ISIs are almost identical. However, it has
been shown that the plateau level can be easily adjusted to various values above resting
level.
20
0
−20
V
−40
−60
−80
−100
0
500
1000
1500
Figure 12: Adjustment of plateau level and voltage minimum during AHP in pacemaker-
type spiking in the 2-component model with parameters given in text (red curve) as well
as with changes of ∆V = −5 mV in the 6 key voltages Ve, Vi, Ve1, Ve3, Vi1, Vi2 (blue curve).
4 Firing frequency and bursting
4.1 Firing rates
Examination of the experimental literature on firing rates and their responses to applied
currents for the three main types of neuron, DA, NA, and SE, considered above, reveals a
wide range of often disparate results. The origins of such variability lie in several factors
including animal types and ages and experimental conditions and techniques. Generally, in
pacemaking mode, firing rates of locus coeruleus (NA) neurons (Tuckwell, 2017, contains
20
a summary) are less than 5 Hz with a few in vivo experiments reporting higher or much
higher rates. For serotonergic neurons of the DRN (summarized in Tuckwell, 2013), firing
rates are usually less than a few Hz, rarely around 4-5 Hz and with applied current up to
10 Hz (Ohliger-Frerking et al., 2003). Rates as high as 20 Hz in cat DRN were reported
under glutamate application by Levine and Jacobs (1992). Bursting is also sometimes
reported in these cells (H´aj´os et al., 1996; H´aj´os et al., 2007). As remarked above, DA
cells in vivo generally exhibit burst firing and in vitro pacemaking was found to have
an average value of 5.4 Hz in one experiment (Grenhoff et al., 1988). Firing rates for
set 1 parameters in the simplified model were computed for both more hyperpolarizing
applied currents (Figure 13A) and more depolarizing currents (Figure 13B) relative to the
standard value of 34.2 pA. Frequencies ranged from 0 to 3 Hz in Figure 9A and from 3
to about 8 Hz in Figure 9B, the rather high frequencies in the latter resulting from large
applied depolarizing currents. The frequencies for set 2 parameters were considerably less
(see Figures 4 and 5).
f (Hz)
3
2.5
2
1.5
1
0.5
A
f (Hz)
B
9
8
7
6
5
4
more hyperpolarizing
0
−34.2
−34.15
−34.1
µ (pA)
−34.05
−34
−33.95
3
34.2
34.4
34.6
34.8
35
35.2
35.4
35.6
35.8
−µ (pA)
more depolarizing
Figure 13: A. Frequency of firing for set 1 parameters with values of µ which are more
hyperpolarizing than the standard value of −0.0342 nA. B. As in part A but that applied
currents are more depolarizing than −0.0342 nA.
4.2 Bursting
The results on the mechanisms of bursting in DA and LC neurons mentioned above
(Grenhoff et al.,1988; Svensson et al., 1989; Overton and Clark, 1992; Cherugi et al.,
1993; Engberg et al.,1993; Blythe et al., 2009) indicate that in several instances regular
pacemaking in these cells may be replaced by bursting if a (synaptic) excitatory drive is
activated or an inhibitory drive inactivated. This was explored in the simplified model by
briefly perturbing the constant excitatory conditions that led to regular pacemaking. The
perturbations mimic the effects of either increased excitatory input or decreased inhibitory
input. The results are shown in Figure 14. In all cases the perturbations lead to rapid
spiking which mimics bursting over the duration of the perturbation with a subsequent
return to pacemaker spiking when the perturbation is terminated. In Figures 14A and
14B, the perturbations consist of increases in the maximal excitatory conductance ge,max
for 200 ms (14A) or decreased inhibitory maximal conductance (14B). In Figures 14C and
14D, the depolarizing current parameter µ is increased for either 200 ms (14C) or 100
ms (14D). In all four cases the evidence is convincing that such perturbations may lead
to a disturbance of pacemaker firing. Thus, a steady but intermittent flow of excitatory
synaptic input or a reduction of inhibitory synaptic input could lead to non-pacemaker
firing and occasional or frequent bursting.
21
A
20
0
−20
V (mV)
B
20
0
−20
V (mV)
100
200
300
400
500
600
700
800
900
1000
t (ms)
−40
−60
−80
−100
0
20
0
−20
V (mV)
−40
−60
−80
100
200
300
400
500
600
700
800
900
1000
t (ms)
C
−40
−60
−80
−100
0
D
20
0
−20
V (mV)
−40
−60
−80
−100
0
100
200
300
400
500
t (ms)
600
700
800
900
1000
−100
0
200
400
600
800
1200
1400
1600
1800
1000
t (ms)
Figure 14: Induction of bursting-type spiking by means of brief changes in excitatory
or inhibitory parameters. Parameters are as in set 1 except with variations in a single
parameter as described. A. Increasing the maximal excitatory conductance ge,max from
the standard value of 2 to 2.5 for 450 < t < 650 gives rise to a burst of three spikes, followed
by a return to pacemaker activity. B. Reducing the maximal inhibitory conductance gi,max
from the standard value of 0.5 to 0.3 for the same time period as in A produces a burst of
4 spikes with a subsequent return to pacemaker-type spiking. C. A 10% increase in the
depolarizing current for the same time-period as in A and B gives rise to rapid spiking
followed by the resumption of regular pacemaking activity. D. As in C but now a 50%
increase in depolarizing current applied for the shorter time period 450 < t < 550 gives
rise to more rapid bursting-type spiking followed regular pacemaking activity.
5 Discussion
Principal brainstem neurons, particularly serotonergic cells of the dorsal raphe nucleus
and noradrenergic cells of the locus coeruleus are of great importance in the functioning
of many neuronal populations throughout cortical and subcortical structures. Of note is
the modulatory role the firing of neurons in these brainstem nuclei have on neurons of
the prefrontal cortex, including the orbitofrontal cortex, and hippocampus. These latter
structures have been strongly implicated in various pathologies, including depression, and
OCD. Lanfumey et al. (2008) contains a comprehensive summary of many of the biological
processes which are influenced by serotonin including those originating from serotonergic
neurons of the DRN. Modeling networks involving both serotonergic and noradrenergic
afferents requires plausible models for the spiking activity of the principal SE and NA
cells. Whereas detailed models of such activity are now available, their application to
22
many thousands of cells has the disadvantage of leading to very large computation time
and large memory requirements, so that the simplified models described in the present
article may provide useful approximating components for such complex computing tasks.
Accurate models of brainstem neurons (see below) involve many component currents and
very large numbers of parameters, several of which are possibly uncertain. Hence realistic
simplified mathematical models of brainstem neurons, beyond that provided by extremely
simplified models such as the leaky integrate and fire (or Lapicque) model (Tuckwell,
1988), are useful in order to investigate approximately the responses of these cells to their
complex array of synaptic and other input and to construct and analyze complex networks
involving these cells and those in other centers such as hippocampus, frontal cortex and
hypothalamus.
The simplified model we have used is a modified Hodgkin-Huxley model with currents
relabeled generically. With certain parameter sets such a model predicts spike trajectories
which are similar to those of brainstem neurons in pacemaker mode. The sensitivity of
solutions to changes in parameters was investigated in detail and very small changes
in some parameters were found to produce dramatic changes in the spiking behavior.
Details of such features as plateaux and AHPs could be adjusted to match experimental
results with judicious choices of voltage parameters. A preliminary investigation was
able to convert regular pacemaking to bursting behavior in accordance with experimental
findings on the application of kynurenic acid to DA and NA neurons. We end with a brief
discussion of previous computational modeling of DA, NA and SE brainstem neurons
5.1 LC NA neurons
Thus far there have been several mathematical models of locus coeruleus neurons per se,
which include a few ionic channels (Putnam et al., 2014; Contreras et al., 2015) or many
ionic channels including the usual sodium and potassium, high and low threshold calcium
currents, transient potassium IA, persistent sodium, leak and hyperpolarization activated
cation current Ih (De Carvalho et al., 2000; Alvarez et al. 2002; Carter et al., 2012).
Noteworthy is the omission of IA in the model of Alvarez et al. (2002) and its inclusion
in De Carvalho et al. (2000) and Carter et al. (2012). Also, a persistent sodium current
is included in De Carvalho et al. (2000) and Alvarez et al. (2002) but not in Carter et al.
(2012).
Despite such uncertainties in the mechanisms involved in pacemaker activity in LC
neurons, some of these works have included synaptic input and gap-junction inputs from
neighboring LC neurons. The pioneering article of De Carvalho et al. (2000) addressed the
mechanisms of morphine addiction and included several biochemical reactions involving
cAMP, µ-opioid receptors, morphine, G-protein, AC, CREB and Fos. Tuckwell (2017)
contains a summary of previous LC modeling as well as a review of LC neuron anatomy
and physiology. Brown et al. (2004) did employ Rose-Hindmarsh model neurons to study
a network of LC neurons but there have not appeared any plausible simplified models
of these cells per se. Thus the two-component model considered in this article provides
a good starting point for investigating, for example, the effects of synaptic input on LC
firing which will be performed in future articles.
5.2 DRN SE neurons
For serotonergic neurons of the dorsal raphe there has been only one detailed model
as described in the introduction (Tuckwell and Penington, 2014). Some authors have
addressed quantitatively serotonin release and included the effects of antidepressants but
without an explicit model for SE cell spiking (Geldof et al., 2008). Wong-Lin et al. (2012)
use a quadratic integrate and fire model for spiking DRN SE neurons in a network of such
23
cells along with inhibitory neurons. The model has a reset mechanism after spikes which
is artificial. In a related work, Cano-Colino et al. (2013, 2014) have modeled the influence
of serotonin on networks of excitatory and inhibitory cells in spatial working memory.
More recently, in a similar vein, Maia and Cano-Colino (2015) have made an interesting
study of serotonergic modulation of the strength of attractors in orbitofrontal cortex and
related this to the occurrence of OCD.
5.3 Midbrain DA neurons
Midbrain dopaminergic neuons have been implicated, inter alia, in reward pathways and
in several pathological conditions including Parkinson's disease (Kalia and Lang, 2015),
psychiatric conditions such as schizophrenia (Weinstein et al., 2017) and addiction to
drugs and natural rewards (Nestler, 2004). There have been several comprehensive com-
putational modeling studies of these neurons, many of which were carried out by Canavier
and coworkers since 1999. These include the model of Amini et al. (1999) with four cal-
cium currents, two potassium currents, a hyperpolarization activated current and various
pump currents which was used to explore the nature and role of a slow oscillatory potential
which was observed when spiking was blocked by tetrodotoxin. Compartmental models
were considered subsequently by Komendantov et al. (2004) and Kuznetsova et al. (2010),
with soma and two dendritic compartments. In the second of these works it was found
that spontaneous firing rates were determined mainly by an L-type calcium current and
the A-type potassium current. In a more recent experimental and computational study
using a compartmental model, Tucker et al. (2012) found that pacemaker frequency was
dependent on the density of Nav channels in the soma and their spatial distribution over
the soma-dendritic regions.
6 Acknowledgements
This research was partially supported by a Mathematical Biosciences Institute post-
doctoral fellowship to YZ from the National Science Foundation under Agreement No.
0931642.
7 References
Aghajanian, G.K., Vandermaelen, C.P., 1982. Intracellular recordings from serotonergic
dorsal raphe neurons: pacemaker potentials and the effect of LSD. Brain Res. 238,
463-469.
Aghajanian, G.K., Vandermaelen, C.P., Andrade, R., 1983. Intracellular studies on the
role of calcium in regulating the activity and reactivity of locus coeruleus neurons
in vivo. Brain Res. 273, 237-243.
Alreja, M., Aghajanian, G.K., 1991. Pacemaker activity of locus coeruleus neurons:
whole-cell recordings in brain slices show dependence on cAMP and protein kinase
A. Brain Res. 556, 339-343.
Alvarez, V.A., Chow, C.C., Van Bockstaele, E.J., Williams, J.T., 2002. Frequency-
dependent synchrony in locus ceruleus: role of electrotonic coupling. PNAS 99,
4032-4036.
Amini, B., Clark, J.W. Jr, Canavier, C.C., 1999. Calcium dynamics underlying pacemaker-
like and burst firing oscillations in midbrain dopaminergic neurons: A computational
study. J. Neurophysiol. 82, 2249-2261.
Andrade. R., Aghajanian, G.K., 1984. Locus coeruleus activity in vitro: intrinsic regu-
lation by a calcium-dependent potassium conductance but not α2-adrenoceptors J.
24
Neurosci 4, .161-170.
Aston-Jones, G., Bloom, F.E., 1981. Activity of norepinephrine-containing locus coeruleus
neurons in behaving rats anticipates fluctuations in the sleep-waking cycle. J. Neu-
rosci. 1, 876-886.
Bayliss, D.A., Li, Y.-W., Talley, E.M., 1997a. Effects of serotonin on caudal raphe neu-
rons: activation of an inwardly rectifying potassium conductance. J. Neurophysiol.
77, 13491361.
Belluzzi, O., Sacchi, O.,1991. A five-conductance model of the action potential in the rat
sympathetic neurone. Prog. Biophys. Molec. Biol. 55: 1-30
Berridge, C.W., Waterhouse, B.D., 2003. The locus coeruleusnoradrenergic system: mod-
ulation of behavioral state and state-dependent cognitive processes. Brain Res.
Rev. 42, 33-84.
Blythe, S.N., Wokosin, D., Atherton, J.F., Bevan, M.D., 2009. Cellular mechanisms
underlying burst firing in substantia nigra dopamine neurons. J. Neurosci. 29,
15531-15541.
Brown, E., Moehlis, J., Holmes, P. et al., 2004. The influence of spike rate and stimulus
duration on noradrenergic neurons. J. Comp. Neurosci. 17, 13-29.
Bush, P.C., Sejnowski, T.J., 1993. Reduced compartmental models of neocortical pyra-
midal cells. Journal of Neuroscience Methods 46,159-166.
Cano-Colino, M., Almeida, R., Compte, A., 2013. Serotonergic modulation of spatial
working memory: predictions from a computational network model. Frontiers in
Integrative Neuroscience 7, 71.
Cano-Colino, M., Almeida, R., Gomez-Cabrero, D. et al., 2014. Serotonin regulates per-
formance nonmonotonically in a spatial working memory network. Cerebral Cortex
24, 2449-2463.
Carrasco, G.A., Van de Kar, L.D., 2003. Neuroendocrine pharmacology of stress. Eur. J.
Pharm. 463, 235-272.
Carter, M.E., Brill, J., Bonnavion, P. et al., 2012. Mechanism for hypocretin-mediated
sleep-to-wake transitions. PNAS 109, E2635E2644.
Cherugi, K., Charl´ety, P.J., Akaoka, H. et al., 1993. Tonic activation of NMDA receptors
causes spontaneous burst discharge of rat midbrain dopamine neurons in vivo. Eur.
J. Neurosci. 5,137-144.
Contreras, S., Quintero, M., Putnam, R., Santin, J., Hartzler, L. and Cordovez, J., 2015.
A Computational Model of Temperature-Dependent Intracellular pH Regulation.
The FASEB Journal, 29(1 Supplement), pp.860-10
De Carvalho, L.A.V., De Azevedo, L.O., 2000. A model for the cellular mechanisms of
morphine tolerance and dependence. Math. Comp. Mod. 32, 933-953.
De Oliveira, R.B., Howlett, M.C.H., Gravina, F.S. et al., 2010. Pacemaker currents in
mouse locus coeruleus neurons. Neurosci 170, 166-177.
De Oliveira, R.B., Gravina, F.S., Lim, R. et al., 2011. Developmental changes in pace-
maker currents in mouse locus coeruleus neurons. Brain Res. 1425, 27-36.
Engberg, G., Kling-Peterson, T., Nissbrandt, H., 1993. GABAB-receptor activation alters
the firing pattern of dopamine neurons in the rat substantia nigra. Synapse 15, 229-
238.
Fitzhugh, R.,1961. Impulses and physiological states in theoretical models of nerve mem-
brane. Biophys. J. 1, 445-466.
Flower, G., Wong-Lin, K., 2014. Reduced computational models of serotonin synthesis,
release and reuptake. IEEE Trans. Biomed. Eng. 61, 1054-1061.
Foote, S.L., Aston-Jones, G., Bloom, F.E., 1980.
Impulse activity of locus coeruleus
neurons in awake rats and monkeys is a function of sensory stimulation and arousal.
Proc. Nati. Acad. Sci. USA 77, 3033-3037.
25
Geldof, M., Freijer, J.I., Peletier, L.A. et al., 2008. Mechanistic model for the acute effect
of fluvoxamine on 5-HT and 5-HIAA concentrations in rat frontal cortex. Eur. J.
Pharm. Sci. 33, 217-219.
German, D.C. and Manaye, K.F., 1993. Midbrain dopaminergic neurons (nuclei A8, A9,
and A10): Three-dimensional reconstruction in the rat. J. Comp. Neurol. 331,
297-309.
Grace, A. A., Bunney, B.S., 1983.
Intracellular and extracellular electrophysiology of
nigral dopaminergic neurons. I. Identification and characterization. Neurosci. 10:
301-315.
Grace, A., Bunney, B., 1984. The control of firing pattern in nigral dopamine neurons:
burst firing. J. Neurosci. 4, 2877-2890.
Grace, A.A., Onn, S-P., 1989. Morphology and electrophysiological properties of immuno-
cytochemically identified rat dopamine neurons recorded in vitro. J. Neurosci. 9,
3463-3481.
Grace, AA., Bunney, B.S., 1995. Electrophysiological Properties of Midbrain Dopamine
Neurons. Neuropsychopharmacology - 4th Generation of Progress. Eds. Bloom,
F.E., Kupfer, D.J. Raven Press, New York.
Grenhoff, J., Tung, C-S., Svensson, T.H., 1988. The excitatory amino acid antagonist
kynurenate induces pacemaker-like firing of dopamine neurons in rat ventral tegmen-
tal area in vivo. Acta Physiol. Scand. 134, 567-568.
H´aj´os, M., Sharp, T., Newberry, N.R., 1996.
Intracellular recordings from burst-firing
presumed serotonergic neurones the rat dorsal raphe nucleus in vivo. Brain Res.
737, 308-312.
H´aj´os, M., Allers, K.A., Jennings, K. et al., 2007. Neurochemical identification of stereo-
typic burst-firing neurons in the rat dorsal raphe nucleus using juxtacellular labelling
methods. Eur. J. Neurosci. 25, 119-126.
Harris, N.C., Webb, C., Greenfield, S.A., 1989. A possible pacemaker mechanism in pars
compacta neurons of the guinea-pig substantia nigra revealed by various ion channel
blocking agents. Neurosci. 31, 355-362.
Hodgkin, A.L., 1948. The local changes associated with repetitive action in a non-
medullated axon. J. Physiol. 107, 165-181.
Hodgkin, A.L., Huxley, A.F., 1952. A quantitative description of membrane current and
its application to conduction and excitation in nerve. J. Physiol. 117, 500-544.
Jedema, H.P., Grace, A.A., 2004. Corticotropin-releasing hormone directly activates nora-
drenergic neurons of the locus ceruleus recorded in vitro. J. Neurosci. 24, 9703-9713.
Ishimatsu, M., Williams, J.T., 1996. Synchronous activity in locus coeruleus results from
dendritic interactions in pericoerulear regions. J. Neurosci. 16, 5196-5204.
Jacobs, B.L., Azmitia, E.C., 1992. Structure and function of the brain serotonin system.
Physiol. Rev. 72, 165-229.
Joels, M, Karst, H., Krugers, H.J., Lucassen, P.J., 2007. Chronic stress: Implications for
neuronal morphology, function and neurogenesis. Front. Neuroendocrin. 28, 72-96.
Joshi, A., Youssofzadeh, V., Vemana, V., McGinnity, T.M., Prasad, G., Wong-Lin, K.,
2017. An integrated modelling framework for neural circuits with multiple neuro-
modulators. Journal of The Royal Society Interface 14, 20160902.
Kalia, L.V., Lang, A.E., 2015. Parkinsons disease. Lancet 386, 896-912.
Khaliq, Z.M., Bean, B.P., 2008. Dynamic, nonlinear feedback regulation of slow pacemak-
ing by A-type potassium current in ventral tegmental area neurons. J. Neurosci.
28, 10905-10917.
Kirby, L.G., Pernar, L., Valentino, R.J. et al., 2003. Distinguishing characteristics of
serotonin and nonserotonin- containing cells in the dorsal raphe nucleus: electro-
physiological and immunohistochemical studies. Neurosci. 116, 669-683.
26
Kocsis B, Varga V, Dahan L, Sik A (2006) Serotonergic neuron diversity: Identification
of raphe neurons with discharges time-locked to the hippocampal theta rhythm.
PNAS 103: 1059-1064.
Komendantov, A.O., Komendantova, O.G., Johnson, S.W., Canavier, C.C., 2004. A
modeling study suggests complementary roles for GABAA and NMDA receptors
and the SK channel in regulating the firing pattern in midbrain dopamine neurons.
J. Neurophysiol. 91, 346-357.
Korf, J., Bunney, B.S., Aghajanian, G.K., 1974. Noradrenergic neurons: morphine inhi-
bition of spontaneous activity. Eur.J. Pharmacol. 25, 165-169.
Kubista, H., Boehm, S., 2006. Molecular mechanisms underlying the modulation of exo-
cytotic noradrenaline release via presynaptic receptors. Pharmacol. Therap. 112,
213-242.
Kuznetsova, A.Y., Huertas, M.A., Kuznetsov, A.S., Paladini, C.A., Canavier, C.C., 2010.
Regulation of firing frequency in a computational model of a midbrain dopaminergic
neuron. Journal of computational neuroscience 28, 389-403.
Lanfumey ,L., Mongeau, R., Cohen-Salmon, C., Hamon, M., 2008. Corticosteroid-serotonin
interactions in the neurobiological mechanisms of stress-related disorders. Neurosci.
Biobehav. Rev. 32, 1174-1184.
Levine, E.S., Jacobs, B.L., 1992. Neurochemical afferents controlling the activity of sero-
tonergic neurons in the dorsal raphe nucleus: microiontophoretic studies in the
awake cat. J. Neurosci. 12, 4037-4044.
Li, Y-Q., Li, H., Kaneko, T., Mizuno, N., 2001. Morphological features and electrophys-
iological properties of serotonergic and non-serotonergic projection neurons in the
dorsal raphe nucleus An intracellular recording and labeling study in rat brain slices.
Brain Res. 900, 110-118.
Li, Y-W., Bayliss, D.A., 1998. Electrophysiological properties, synaptic transmission and
neuromodulation in serotonergic caudal raphe neurons. Clin. Exp. Pharm. Physiol.
25, 468-473.
Lopez. J.F., Akil, H., Watson, S.J., 1999. Neural circuits mediating stress. Biol. Psychi-
atry 46:1461-1471.
Lowry, C.A., Evans, A.K., Gasser, P.J. et al., 2008. Topographic organization and
chemoarchitecture of the dorsal raphe nucleus and the median raphe nucleus. In:
Serotonin and sleep: molecular, functional and clinical aspects, p 25-68, Monti, J.M.
et al., Eds. Basel: Birkhauser Verlag AG.
Lowry, C.A., Rodda, J.E., Lightman, S.L., Ingram, C.D., 2000. Corticotropin- releasing
factor increases in vitro firing rates of serotonergic neurons in the rat dorsal raphe
nucleus: evidence for activation of a topographically organized mesolimbocortical
serotonergic system. J. Neurosci. 20, 7728-7736.
Luppi. P-H., Clement, O., Sapin, E. et al., 2012. Brainstem mechanisms of paradoxical
(REM) sleep generation. Eur. J. Physiol. 463:43-52.
Maejima, T., Masseck, O.A., Mark, M.D., Herlitze, S., 2013. Modulation of firing and
synaptic transmission of serotonergic neurons by intrinsic G protein-coupled recep-
tors and ion channels. Frontiers in Integrative Neuroscience 7, 40.
Mahar, I., Bambico, F.R., Mechawar, N., Nobrega J.N., 2014. Stress, serotonin, and hip-
pocampal neurogenesis in relation to depression and antidepressant effects. Neu-
rosci. Biobehav. Rev. 38, 173-192.
Maia, T.V., Cano-Colino, M., 2015. The role of serotonin in orbitofrontal function
and obsessive-compulsive disorder. Clinical Psychological Science 3, 460-482, and
Supplemental-data.
Milnar, B., Montalbano, A., Piszczek et al., 2016. Firing properties of genetically iden-
tified dorsal raphe serotonergic neurons in brain slices. Front. Cell. Neurosci. 10,
27
195.
Nagumo, J. S., Arimoto, S., Yoshizawa, S., 1962. An active pulse transmission line
simulating nerve axon. Proc. I.R.E. 50, 2061-2070.
Nair-Roberts, R.G., Chatelain-Badie, S.D., Benson, E., White-Cooper, H., Bolam, J.P.,
Ungless, M.A., 2008. Stereological estimates of dopaminergic, GABAergic and glu-
tamatergic neurons in the ventral tegmental area, substantia nigra and retrorubral
field in the rat. Neurosci. 152, , 1024-1031.
Nestler, E.J., 2004. Molecular mechanisms of drug addiction. Neuropharmacol. 47, 24-32.
Nestler, E. J., Alreja, M., Aghajanian, G.K., 1999. Molecular control of locus coeruleus
neurotransmission. Biol. Psychiatry 46, 1131-1139.
Ohliger-Frerking, P., Horwitz, B.A., Horowitz, J.M., 2003. Serotonergic dorsal raphe
neurons from obese zucker rats are hyperexcitable. Neurosci. 120, 627-634.
Overton, P., Clark., D.,1992.
Iontophoretically administered drugs acting at the N-
methyl-D-aspartate receptor modulate burst firing in A9 dopamine neurons in the
rat. Synapse 10,131-140.
Pan, W.J., Osmanovi´c, S.S., Shefner, S.A., 1994. Adenosine decreases action potential
duration by modulation of A-current in rat locus coeruleus neurons. J. Neurosci.
14, 1114-1122.
Penington, N.J., Kelly, J.S., Fox, A.P., 1991. A study of the mechanism of Ca2+ current
I7,
inhibition produced by serotonin in rat dorsal raphe neurons. J. Neurosci.
3594-3609.
Putnam, R., Quintero, M., Santin, J. et al., 2014. Computational modeling of the effects
of temperature on chemosensitive locus coeruleus neurons from bullfrogs. Faseb J.
28, Supp. 1128.3. (Abstract only).
Putzier, I., Kullmann, P.H.M., Horn, J.P., Levitan, E.S., 2009. Cav1.3 channel voltage
dependence, not Ca2+ selectivity, drives pacemaker activity and amplifies bursts in
nigral dopamine neurons. J. Neurosci. 29, 15414-15419.
Rall, W., 1962. Theory of physiological properties of dendrites. Ann. NY Acad. Sci. 96:
1071-1092.
Ramirez, J-M., Koch, H., Garcia, A.J. III, et al., 2011. The role of spiking and bursting
pacemakers in the neuronal control of breathing. J Biol. Phys. 37, 241-261.
Sanchez-Padilla, J., Guzman, J.N., Ilijic, E. et al., 2014. Mitochondrial oxidant stress in
locus coeruleus is regulated by activity and nitric oxide synthase. Nat. Neurosci.
17, 832-842.
Svensson, T.H., Engberg, C-S., Tung, C-S., Grenhoff, J., 1989. Pacemaker-like firing of
noradrenergic locus coeruleus neurons in vivo induced by the excitatory amino acid
antagonist kynurenate in the rat. Acta Physiol. Scand. 135, 421-422.
Sugiyama, D., Hur, S.W., Pickering, A:E. et al. 2012. In vivo patch-clamp recording from
locus coeruleus neurones in the rat brainstem. J Physiol. 590, 2225-2231.
Swanson, L.W., 1976. The locus coeruleus: a cytoarchitectonic, golgi and immunohisto-
chemical study in the albino rat. Brain Res. 110, 39-56.
Tucker, K.R., Huertas, M.A., Horn, J.P., Canavier, C.C., Levitan, E.S., 2012. Pacemaker
rate and depolarization block in nigral dopamine neurons: a somatic sodium channel
balancing act. J. Neurosci 32, 14519-14531.
Tuckwell, H.C., 1988. Introduction to Theoretical Neurobiology. Cambridge University
Press, Cambridge UK.
Tuckwell, H.C., 2013. Biophysical properties and computational modeling of calcium
spikes in serotonergic neurons of the dorsal raphe nucleus. BioSystems 112, 204-
213.
Tuckwell, H.C., 2017. Computational modeling of spike generation in locus coeruleus
noradrenergic neurons. Preprint.
28
Tuckwell, H.C., Ditlevsen, S., 2016. The space-clamped Hodgkin-Huxley system with ran-
dom synaptic input: inhibition of spiking by weak noise and analysis with moment
equations. Neural Computation 28, 2129-2161.
Tuckwell, H.C., Jost, J., GutKin, B.S., 2009.
Inhibition and modulation of rhythmic
neuronal spiking by noise. Phys. Rev. E. 80, 031907.
Tuckwell, H.C., Penington, N.J., 2014. Computational modeling of spike generation in
serotonergic neurons of the dorsal raphe nucleus. Prog. Neurobiol. 118, 59-101.
Valentino, R.J., Foote, S.L., Page, M.E., 1993. The locus coeruleus as a site for integrating
corticotropin-releasing factor and noradrenergic mediation of stress responses. Ann.
NY.Acad. Sci 697, 173-188.
Vandermaelen, C.P., Aghajanian, G.K., 1983. Electrophysiological and pharmacological
characterization of serotonergic dorsal raphe neurons recorded extracellularly and
intracellularly in rat brain slices. Brain Res. 289, 109119.
Vasudeva, R.K., Waterhouse, B.D., 2014. Cellular profile of the dorsal raphe lateral
wing sub-region: relationship to the lateral dorsal tegmental nucleus. J. Chem.
Neuroanat. 57-58, 15-23.
Vertes, R.P., Crane, A.M., 1997. Distribution, quantification, and morphological charac-
teristics of serotonin-immunoreactive cells of the supralemniscal nucleus (b9) and
pontomesencephalic reticular formation in the rat. J. Comp. Neurol. 378, 411-424.
Walsh, J.B., Tuckwell, H.C., 1985. Determination of the electrical potential over dendritic
trees by mapping onto a nerve cylinder. J. Theor. Neurobiol. 4, 27-46.
Weinstein, J.J., Chohan, M.O., Slifstein, M., Kegeles, L.S., Moore, H., Abi-Dargham,
A., 2017. Pathway-specific dopamine abnormalities in schizophrenia. Biological
psychiatry, 81(1), pp.31-42.
Williams, J.T., Egan, T.M., North, R.A., 1982. Enkephalin opens potassium channels on
mammalian central neurons. Nature 299, 74-77.
Williams, J.T., North, R.A., Shefner S.A. et al., 1984. Membrane properties of rat locus
coeruleus neurones. Neuroscience 13, 137-156.
Wong-Lin, K-F., Joshi, A., Prasad, G., McGinnity, T.M., 2012. Network properties of a
computational model of the dorsal raphe nucleus. Neural Netw. 32, 15-25.
29
|
1311.4035 | 1 | 1311 | 2013-11-16T08:18:02 | Analog and digital codes in the brain | [
"q-bio.NC",
"physics.data-an"
] | It has long been debated whether information in the brain is coded at the rate of neuronal spiking or at the precise timing of single spikes. Although this issue is essential to the understanding of neural signal processing, it is not easily resolved because the two mechanisms are not mutually exclusive. We suggest revising this coding issue so that one hypothesis is uniquely selected for a given spike train. To this end, we decide whether the spike train is likely to transmit a continuously varying analog signal or switching between active and inactive states. The coding hypothesis is selected by comparing the likelihood estimates yielded by empirical Bayes and hidden Markov models on individual data. The analysis method is applicable to generic event sequences, such as earthquakes, machine noises, human communications, and enhances the gain in decoding signals and infers underlying activities. | q-bio.NC | q-bio |
Analog and digital codes in the brain
Yasuhiro Mochizuki∗ and Shigeru Shinomoto†
Department of Physics, Kyoto University, Kyoto 606-8502, Japan
(Dated: November 19, 2013)
It has long been debated whether information in the brain is coded at the rate of neuronal spiking
or at the precise timing of single spikes. Although this issue is essential to the understanding of
neural signal processing, it is not easily resolved because the two mechanisms are not mutually
exclusive. We suggest revising this coding issue so that one hypothesis is uniquely selected for a
given spike train. To this end, we decide whether the spike train is likely to transmit a continu-
ously varying analog signal or switching between active and inactive states. The coding hypothesis
is selected by comparing the likelihood estimates yielded by empirical Bayes and hidden Markov
models on individual data. The analysis method is applicable to generic event sequences, such as
earthquakes, machine noises, human communications, and enhances the gain in decoding signals
and infers underlying activities.
I.
INTRODUCTION
Sensation and motion are represented and processed in
the brain as series of neuronal discharges called firings or
spikes [1]. In the early 1900s, the number of neuronal dis-
charges in a given time interval was found to be related to
the tension in the associated muscle [2]. Since then, cor-
relating the rate of neuronal firings with animal behavior
has become standard protocol. Other coding hypothe-
ses have also been studied both experimentally [3–5] and
theoretically [6–9]. Among these alternatives is temporal
coding, which emphasizes the importance of precise spike
timings [5].
Coding hypotheses have retained researchers’ inter-
est, less for unproductive taxonomy purposes, but be-
cause they assist our understanding of neuronal informa-
tion processing in the brain. Theoretically, it has been
demonstrated that a discrete (as opposed to continuous)
firing rate increases the rate of information transmis-
sion, depending on the width of the time window [10–
14]. Thus, assuming that neural systems have evolved to
maximize their information transmission rate, different
areas of the brain may process signals in different ways.
The coding problem has become the sub ject of theoret-
ical modeling. For instance, in attractor network the-
ory, neuronal activity undergoes transitions among qua-
sistationary states [15–17]. Attractor states may mani-
fest as distinct changes in the firing condition. Informa-
tion processing can feasibly be represented by jumping
among quasi-stationary states. In particular, the change-
point detection of neurons approaches the theoretical op-
timum [18]. Of more practical interest, coding identifica-
tion may lead to improved information decoding, which
would benefit real-time applications such as brain ma-
chine interfaces.
Nevertheless, rate coding and temporal coding are not
clearly delineated because they are not mutually exclu-
sive. For instance, spike timing may be reinterpreted as
∗ [email protected]
† [email protected]
high firing rate in a small time window. A clearer distinc-
tion between the coding types has been suggested, such
that any spike train is classifiable into rate or temporal
coding, depending on whether the underlying rate varies
slowly or rapidly with time, respectively [19, 20]. How-
ever, this principle is not directly applicable to data, be-
cause the firing rate cannot be uniquely determined from
a single spike train. Rather, spiking is generally irregular
and sparse, and the underlying rate can be obtained only
from multiple spike train analyses in repeated trials. The
fine details of original rate fluctuations are easily erased
by inter-trial jittering [21]. For this reason, the coding
hypothesis should be identified on a single trial basis.
Here, we suggest a method that selects a unique
hypothesis for any given single spike train.
The
conventional timing-based classification is replaced by
?
FIG. 1.
Selecting a coding hypothesis for a single
spike train. A spike train is examined to determine whether
it is likely transmitting a continuously varying analog signal
or discontinuously switching binary signals.
(a)
(b)
2
]
z
H
[
e
t
a
R
]
z
H
[
e
t
a
R
]
z
H
[
e
t
a
R
Ornstein-Uhlenbeck Process (OUP)
60
30
0
0
Spike train
0
5
5
10
10
Hidden Markov Model (HMM)
60
30
0
0
5
10
Empirical Bayes Model (EBM)
60
30
0
0
5
Time [s]
10
]
z
H
[
e
t
a
R
]
z
H
[
e
t
a
R
]
z
H
[
e
t
a
R
Switching State Process (SSP)
60
30
0
0
Spike train
0
5
5
10
10
Hidden Markov Model (HMM)
60
30
0
0
5
10
Empirical Bayes Model (EBM)
60
30
0
0
5
Time [s]
10
FIG. 2. Continuous and discontinuous rate processes and their decoding.
(a) A spike train is derived from the
continuously modulated rate given by the Ornstein–Uhlenbeck process (OUP) (blue). The continuous empirical Bayes model
(EBM) (green) estimates the original rate better than the discontinuous hidden Markov model (HMM) (orange). (b) A spike
train is derived from the discontinuous rate given by the switching state process (SSP) (red). The HMM better estimates the
original rate than the EBM. Simulation parameters are µ = 25 [Hz], σ = 17 [Hz], and τ = 1 [s] for both the OUP and SSP.
an analog–digital classification criterion that inquires
whether the spike train is likely transmitting a contin-
uously varying analog signal or discontinuously switch-
ing binary signals (Fig. 1). The proposed classification
scheme is similar to the timing-based scheme because
analog and digital signals may be represented by the fir-
ing rate and timing of spike bursts, respectively. Thus,
rather than dispense with the rate-coding hypothesis, we
select the best interpretation of a single spike train from
alternative rate estimators based on rivalry principles.
Specifically, we select a single stochastic model, either
the empirical Bayes model (EBM) or the hidden Markov
model (HMM), by comparing their likelihood estimates
for a given spike train. The EBM and HMM represent the
analog and digital codes, respectively. The effectiveness
of the inference method is tested on synthetic data de-
rived from inhomogeneous Poisson processes whose con-
tinuous and discrete rates are given by the Ornstein–
Uhlenbeck process (OUP) and the switching state pro-
cess (SSP), respectively. Finally, to determine whether
different areas of the brain encode signals continuously or
discretely, the suggested analysis is applied to biological
data.
II. MODELING SPIKING PROCESSES
To examine how efficiently the rate estimators infer the
underlying rate, we first consider the idealized inhomo-
geneous Poisson processes, in which spikes are randomly
drawn from a given rate function of time. OUP and SSP
represent rate functions that fluctuate continuously and
discontinuously in time, respectively.
The Ornstein–Uhlenbeck process (OUP): A typ-
ical continuously fluctuating process is illustrated in
Fig. 2(a). We represent this process by the OUP, origi-
nally introduced to describe the fluctuating velocities of
Brownian particles. OUP is modeled by the following
stochastic differential equation:
1
2
dλ(t)
dt
λ(t) − µ
τ
= −
+
σ
√τ
ξ (t),
(1)
where ξ (t) is the Gaussian white noise characterized by
the ensemble average hξ (t)i = 0 and hξ (t)ξ (t′ )i = δ(t−t′ ).
Due to random fluctuations in the OUP, λ(t) can be neg-
ative even if µ exceeds the typical fluctuation amplitude
σ . Interpreting λ(t) as the temporally fluctuating rate,
the firing rate is regarded as zero if λ(t) < 0.
Switching state process (SSP): A typical discon-
tinuously fluctuating process is illustrated in Fig. 2(b).
For this process, we adopt random telegraph state switch-
ing, in which the on/off states stochastically alternate
under the random telegraph process. Here we consider a
symmetric case in which the average inter-transition in-
terval is the same for both states. This process can be re-
alized by repeating the Bernoulli trials for making a tran-
sition to another state with the rate 1/τ ; equivalently,
by drawing inter-transition intervals from the exponen-
tial distribution with mean τ , p(t) = τ −1 exp (−t/τ ). For
two states we assign the firing rates λ(t) = µ − σ or µ + σ
(µ > σ > 0).
1st and 2nd order statistics: The sample rate pro-
cesses generated by the OUP and SSP are apparently
different (Fig. 2). However, given the same parameter
set, {µ, σ, τ }, both processes deliver identical first and
second order statistics, (mean and correlation function
of the rate, respectively), given by
(2)
λ(t) = µ,
δλ(t + s)δλ(t) = σ2 exp (cid:18)−
τ (cid:19),
2s
where the overline represents the time-average and
δλ(t) ≡ λ(t) − µ.
Inhomogeneous Poisson processes: Given a time-
dependent rate process λ(t), a Poisson spike train can be
derived by subdividing the time axis into small bins of
width δ t and repeating the Bernoulli trials for generating
a spike in every time bin with a probability of λ(t)δ t.
(3)
III. RATE ESTIMATORS
Here we introduce EBM and HMM as stochastic mod-
els of continuous and discontinuous rate processes, re-
spectively. These models are used as rate estimators.
Empirical Bayes model (EBM): We assume that
spikes are independently drawn from the underlying fir-
ing rate λ(t). In this scenario, termed the inhomogeneous
Poisson process, the probability of spike occurrences at
times {tj } ≡ {t1 , t2 , . . . , tn} in the period t ∈ [0, T ] is
analytically given as [22],
P ({tj }λ(t)) =
λ(tj )
λ(t)dt! .
exp − Z T
n
Yj=1
0
The underlying rate is inferred from the spike train using
Bayes’ theorem,
(4)
.
(5)
P (λ(t){tj }) =
P ({tj }λ(t))P (λ(t))
P ({tj })
Here we give a prior distribution functional, by assuming
the rate to be generally flat:
Pγ (λ(t)) ∝ exp −
dt (cid:19)2
2γ 2 Z T
0 (cid:18) dλ
1
dt! ,
(6)
3
(7)
where γ is a hyperparameter representing the flatness of
the rate process. The hyperparameter can be selected by
maximizing the marginal likelihood, or “evidence,”
Pγ ({tj }) ≡ Z P ({tj }λ(t))Pγ (λ(t))D{λ(t)},
where D{λ(t)} denotes that the functional is integrated
over all possible rate processes.
If data are provided,
this marginalization integral can be maximized with re-
spect to γ by the expectation and maximization (EM)
algorithm [23]. The negative of the logarithm of the
marginal likelihood corresponds to the free energy [24–
26]. With the hyperparameter γ that maximizes the
marginal likelihood or minimizes the free energy, we ob-
tain the maximum a posteriori (MAP) estimate of the
rate process, λ(t), that maximizes the posterior distribu-
tion functional, pγ (λ(t){tj }) [27]. An application pro-
gram for estimating the MAP rate for a given spike train
is accessible by Ref. [28].
Hidden Markov model (HMM): When estimating
the firing rate by HMM [29, 30], the spike train is de-
rived from the rate of transition between different states
according to the Markov process. Here we adopt a two-
state HMM model, in which the rate takes one of two
values. Model parameters are the two rates λ1 and λ2 ,
the elements of the transition matrix, and the proba-
bilities of the initial states. These parameters are esti-
mated by the Baum–Welch algorithm, and then the most
likely sequence of hidden states is then obtained from the
Viterbi algorithm. The estimated rate λ(t) is regarded as
the sequence of alternating rates assigned as the selected
hidden states. An application program for performing
HMM rate estimation is accessible by Ref. [31].
IV. TESTING THE RATE ESTIMATORS USING
SYNTHETIC SPIKE TRAINS
In this section, we compare the accuracy of EBM and
HMM in estimating the underlying rates from OUP- and
SSP-derived spike trains, representing continuous and
discontinuous rate processes, respectively.
A. The Kullback–Leibler (KL) divergence
Apparently the firing rate of the spike train derived
from OUP is better estimated by EBM than by HMM,
while the opposite is true for the SSP-derived spike train
(Fig. 2).
If the underlying rates of synthetic data are
known, the estimation accuracy can be evaluated by
directly measuring the deviation of the estimated rate
λ(t) from the underlying rate λ(t). Furthermore,
if
spikes are derived independently from the underlying rate
λ(t) and the resulting spike train follows a Poisson pro-
cess, the goodness of the rate estimator is the deviation
of the normalized density of individual spikes, p(t) ≡
4
Switching State Process (SSP)
(a)
0.1
Ornstein-Uhlenbeck Process (OUP)
e
c
n
e
g
r
e
v
i
d
-
L
K
0.05
(b)
0.1
0.05
0
0
5
15
20
10
σ
0
0
5
15
20
10
σ
FIG. 3. Kullbac–Leibler (KL) divergence of the estimated distribution p(t) from the underlying distribution
p(t). Numerically estimated EBM and HMM values are depicted in green and orange, respectively. (a) The continuous OUP
rate process. (b) The discontinuous SSP rate process. The rates are estimated from spike trains of n=1,000 spikes. The black
solid line is the analytical result obtained by the path integral, Eq. (23). The edges of the light green and yellow regions
represent the upper/lower quartiles of KL divergence estimated from 1,000 samples. Other parameters are µ=25 [Hz], τ =1 [s],
giving a theoretical detection limit of σc = pµ/τ =5 [Hz].
p(t) log (p(t)/ p(t))dt ≥ 0,
λ(t)/ R T
λ(t′ )dt′ from the normalized underlying density,
0
p(t) ≡ λ(t)/ R T
0 λ(t′ )dt′ , assuming that the average rate
λ(t′ )dt′ = R T
is correctly captured; i.e., R T
0 λ(t′ )dt′ . The
0
deviation of distribution functions may be represented by
the Kullback–Leibler (KL) divergence [32, 33], defined as
D(p p) ≡ Z T
0
where the equality holds if the two distribution functions
are equal. The KL divergence is represented as the sur-
plus of the cross entropy [34],
H (p, p) ≡ − Z T
0
over the entropy of the underlying distribution p(t),
H (p) ≡ − Z T
0
p(t) log p(t)dt,
p(t) log p(t)dt,
(10)
(8)
(9)
p(t) log(1/T )dt = log T .
The initial quadratic increase with σ may be interpreted
as follows. If the rate fluctuation in a spike train is small,
it will not be detected by any rate estimator. Rather than
sampling a fluctuation, principled estimators such as the
EBM will draw a fixed rate of p(t) = 1/T , whereby the
cross entropy is given as
H (p, p) = − Z T
0
The entropy can be approximated by expanding p(t) in
terms of the deviation of the normalized distribution from
the mean 1/T ,
H (p) = − Z T
0 (cid:18) 1 + δp(t)
(cid:19) log (cid:18) 1 + δp(t)
T
T
σ2
≈ log T − δp2/2 = log T −
2µ2 .
Thus, in both the OUP and SSP, the KL divergence of
the constant probability p = 1/T is approximated as
(cid:19)dt
(12)
(13)
that is;
D(p p) = H (p, p) − H (p).
Rate estimation using the EBM: In EBM, the KL
divergences of the OUP- and SSP-derived spike trains de-
pend similarly on σ (the green lines in Figs. 3(a) and (b)).
(11)
D0 = H (p, p) − H (p) ≈
σ2
2µ2 .
(14)
Theoretical estimate of the EBM using the path
integral: As the rate fluctuation σ increases further,
the KL divergence departs downward from this quadratic
(a)
y
g
r
e
n
e
e
e
r
F
3
2
1
0
σ <σc
σ =σc
σ >σc
0
5
10 15
γ
(b)
Ornstein-Uhlenbeck Process
(OUP)
10
γ
5
0
0
5
10
Switching State Process
(SSP)
10
γ
5
0
0
5
σ
10
FIG. 4. Detectable-undetectable phase transitions of
the empirical Bayes model. (a) Free energy defined by the
log-likelihood Eq. (19) for the cases of σ < σc , = σc , and > σc .
(b) An optimal hyperparameter γ representing the degree of
flatness. The conditions of phase transition for the OUP and
SSP are given by Eq. (20). The error bars of numerical data
represent the average and upper/lower quartiles of the hyper-
parameter γ determined by applying practical optimization
algorithms to 1,000 synthetic data numerically generated by
simulating the inhomogeneous Poisson processes. Parameters
of synthetic data used for numerical analysis are the same as
those of Fig.3.
line, implying that rate fluctuations have begun to be
appropriately detected by EBM. The EBM marginal-
ization integral has been shown to be analytically solv-
able [35, 36]. Equation (7) can be transformed into a
path integral format as [37]
1
Z (γ ) Z D{λ(t)}e− R T
0 L(λ, λ,t)dt ,
where the “Lagrangian” L(λ, λ, t) is
Pγ ({tj }) =
(15)
(16)
1
2γ 2
L(λ, λ, t) =
λ2 + λ −
δ(t − ti ) log λ.
n
Xi=1
The “classical path” corresponding to the MAP esti-
mate of the rate process λ(t) is obtained by the Euler–
Lagrange equation:
∂ λ (cid:19) −
dt (cid:18) ∂L
d
Here, the log marginal likelihood is averaged over possible
realizations of spike trains derived from a rate function.
When deriving a spike train from an underlying rate λ(t),
the fluctuations in spike counts within a given time in-
terval are Poisson-distributed. In this case, the variance
in the spike count equals the mean, and the rate of an
individual spike train is described by a stochastic func-
tion:
(17)
∂L
∂ λ
= 0.
λ(t) + pλ(t)ξ (t),
(18)
5
where ξ (t) denotes Gaussian white noise. The path inte-
gral can be evaluated by expanding the “action integral”
to quadratic order in the deviation of the rate from the
mean [35, 36]. The free energy is analytically obtained
as
(19)
(20)
F (γ ) = − hlog Pγ ({tj })i
2τ σ2
4√µ (cid:18)1 −
2µ + γ τ √µ (cid:19) + const.,
T γ
≈
where the angle brackets represent the averaging opera-
tion with respect to the ξ (t) ensemble. The hyperparam-
eter γ may be selected by minimizing the free energy:
F (γ ) = (cid:26) 0,
if σ < σc ,
2(σ − σc )/√τ , otherwise,
γ = arg min
γ
where σc = pµ/τ . Thus, γ vanishes, or equivalently, the
flatness of the rate diverges, if the fluctuation amplitude
of the underlying rate is below the critical value σc , im-
plying that the rate fluctuation is undetectable (Fig. 4).
The MAP estimate of the rate, or solution of the Euler–
Lagrange equation, is given as
2õ Z T
γ
(δλ(s) + √µξ (s))e−
γ t−s√µ ds,
λ(t) ≈ µ +
(21)
0
where pλ(t)ξ (s) is approximated by √µξ (s). The cross
entropy can be obtained by expanding p(t) in terms of
the deviation of the normalized distribution, and by av-
eraging over all possible realizations of spike trains ξ (s)
as
(23)
(22)
(cid:19)(cid:29) dt
H (p, p) = − Z T
(cid:28)log (cid:18) 1 + δ p(t)
1 + δp(t)
T
T
0
≈ log T − hδpδ p − δ p2/2i.
By calculating the ensemble average, the KL divergence
is obtained as
D(p p) ≈ (cid:26) σ2 /(2µ2),
for σ < σc ,
σσc /(2µ2), otherwise.
The phase transition at σ = σc , above which the fluctu-
ations become detectable, is discernible in the analytical
KL divergence curves (the black solid lines in Fig. 3).
Though this solution assumes that σ ≪ µ, it reasonably
agrees with the numerical solutions (the green lines in
Fig. 3) even when σ is comparable to µ.
Rate estimation using the HMM: To enable com-
parison between the rate-detection performances of EBM
and HMM, the KL divergence of HMM is also plotted in
Fig. 3 (the orange lines).
When σ < σc , the KL divergences of the HMM are
higher than D0 = σ2 /(2µ2) for both OUP- and SSP-
derived spike trains. From this result, we infer that HMM
needlessly tracked the fluctuations of individual data; ac-
cordingly, the rate estimation is inferior to that obtained
by simply indicating the constant mean rate.
When rate fluctuations are large, σ > σc , the KL di-
vergences obtained by HMM differ widely between the
OUP and SSP data. For the SSP, the KL divergence is
lower than that of EBM, implying that HMM more accu-
rately estimates the underlying transition rate between
the two states. Contrariwise, for the OUP data, which
realize continuous rate changes, the performance of EBM
is always superior to that of HMM.
B. Validating rate estimators
The KL divergence is an impractical measure because
the original rate is not known in real applications. Here
we suggest a random subsampling validation method that
evaluates the rate estimators in terms of their goodness
of estimation.
Given a spike train of n spikes, we randomly remove
m(≪ n) spikes and estimate the rate profile λn−m (t)
from the remaining n − m spikes. Because every spike
is independently derived from the underlying rate, the
likelihood of the rate profile of the unused m spikes is the
product of the likelihoods of normalized densities p(t) =
Ornstein-Uhlenbeck Process Switching State Process
ching State Process
ching State Process
(OUP)
(SSP
)
(a)
-3.9
(b)
-3.4
M
B
E
l
-3.65
-3.65
l
-3.4
-3.6
-3.35
l
-3.1
-3.1
-3.35
-3.9
-3.9
(c)
-3.65
lHMM
-3.6
-3.6
-3.4
-3.35
lHMM
-3.1
-0.15
0
HMM - l
l
EBM
0.15
-0.15
0
HMM - l
l
EBM
0.15
FIG. 5. Validating the likelihoods of analog and digital
coding hypotheses. (a) Likelihood distributions of EBM
and HMM, (the blue and red distributions, respectively), cal-
culated by Eq. (24), for trains of 1,000 spikes derived from
continuous OUP and discontinuous SSP. (b) Scatter plots of
the likelihoods on the (lHMM , lEBM )-plane (c) Distribution of
the likelihood difference, lHMM − lEBM . Model parameters:
(µ, σ , τ ) = (25 [Hz], 10 [Hz], 1 [s]) for the OUP and (25
[Hz], 20 [Hz], 1 [s]) for the SSP. Number of spikes n=1,000;
subsampling spikes m=10; sampling trials k=100.
6
(24)
1
m
λn−m (t)/ R T
λn−m (t′ )dt′ . Thus, the log-likelihood per a
0
single spike is estimated as
log λn−m (ti )
λn−m (t′ )dt′ ! .
m
Xi=1
l =
R T
0
Repeating this procedure k times, we compute the mean
and the standard error of the log-likelihood of a given
spike train. The cross-validated log-likelihood should ap-
proximate the negative cross entropy, −H (p, p) (Eq. (9)).
When the distributions of the likelihoods for EBM and
HMM are directly compared, their relative superiority or
inferiority is not evident (Fig. 5(a)). This occurs be-
cause the entropy H (p) of individual rate processes fluc-
tuates among samples, and the estimated log-likelihood
and negative cross entropy −H (p, p) alone do not reflect
the goodness of the rate estimation. Thus, we suggest
comparing the likelihoods of EBM and HMM for individ-
ual spike trains (Fig. 5(b)) or computing their difference
lHMM − lEBM (Fig. 5(c)). The conformity of the data to
HMM and EBM is now clearly detected even from se-
quences of O(1,000) spikes, which are typically acquired
from experiments.
V. ANALYSIS OF BIOLOGICAL DATA
Finally, we apply our method to real data. To this
end, we analyze biological neuronal spike trains by the
cross-validation method. The test is conducted on pub-
licly available spike data. All data were collected from
the visual cortical areas, primary visual cortex (V1) and
middle temporal area (MT), and from the thalamus and
lateral geniculate nucleus (LGN) of monkeys (Macaca
fascicularis) repeatedly presented with a drifting sinu-
soidal grating [38, 39]. In each trial, the recording times
of a single run were 6,000 or 3,000 ms for V1, 1,280 ms
for MT, and 5,138 ms for the LGN. Only the trials with
mean firing rate greater than 10 Hz were accepted, and
spike trains recorded in different trials were concatenated
into a final spike train of 1,000 spikes. The numbers of
accepted neurons were 39, 40, and 49, respectively for
V1, MT, and LGN.
The results of the cross-validation analysis are shown
in Fig. 6. Fractions of neurons exhibiting analog and
digital coding patterns differ between the three brain re-
gions. In particular, more discontinuous firing patterns
were observed in LGN neurons than in V1 and MT neu-
rons (15/49 in the LGN, versus 3/39 and 7/40 in the
V1 and MT, respectively). Several spike trains, together
with their rates estimated by continuous EBM and dis-
continuous HMM rate estimators, are presented in Fig. 7.
VI. DISCUSSION
In this paper, we selected alternative coding hypothe-
ses for individual spike trains. For this purpose, we com-
(a)
: V1
-2.75
M
B
E
l
-3
(b)
-3.25
-3.25
t
n
u
o
c
-0.2
7
-2
-3
: MT
-2
-3
: LGN
-2
-3
-4
-4
-3
-2
-4
-4
-3
-2
-4
-4
-3
-2
-2.75
-3
-2.75
-3
-3
l
HMM
-2.75
-3.25
-3.25
-3
l
HMM
-2.75
-3.25
-3.25
-3
l
HMM
-2.75
0
HMM - l
l
EBM
0.2
-0.2
0
HMM - l
l
EBM
0.2
-0.2
0
HMM - l
EBM
l
0.2
FIG. 6. Analysis of biological neuronal spike trains. (a): Scatter plots of the log-likelihood of EBM against the log-
likelihood of HMM. Each cross represents the standard error of the mean likelihoods obtained for a neuron in V1 (green), MT
(pink) and LGN (orange). (b): Histograms of differences between the two log-likelihoods shown in Fig. 5(c). Number of spikes
n=1,000; subsampling spikes m=10; sampling trials k=100.
pared rate estimators of the continuous EBM and dis-
continuous HMM, respectively representing analog and
digital neuronal codes. We first determined whether the
class of rate process could be identified from synthetic
spike trains derived from OUP and SSP. Next, we ap-
plied our analytical method to biological data obtained
from the visual cortical areas V1 and MT, and the tha-
lamus LGN, and found significant differences among the
firing patterns of different brain areas.
Here we assumed two hypotheses; that information
transmitted by neurons is coded in an analog or digital
manner. If the purpose of selecting coding hypotheses is
to best estimate the unknown underlying rate, more cod-
ing hypotheses can be accommodated by adopting a suit-
able model selection principle. For instance, we suggested
that two-state HMM represents discontinuous rate pro-
cesses, but numerous variants are possible. For example,
the number of states can exceed two; the number of firing
rates is not necessarily fixed in advance but may change
arbitrarily in every switching; the firing rate may fluctu-
ate during the inter-transition interval. Such possibilities
could be examined using multistate HMMs, the infinite
HMM [40] and the switching state-space model [41].
Throughout this study, we have assumed the inhomo-
geneous Poisson process, in which individual spikes are
independently derived from a given rate function of time.
However, it should be noted that spiking events are sig-
nificantly influenced by their predecessors. Consequently,
real neuronal firings are not precisely modeled by Poisson
processes [42, 43]. Thus, one may extend our analysis to
contend with deviation from Poisson firing, as has been
done in Refs [36, 44].
Nevertheless, assuming the simple Poisson process is
suitable for diverse problems due to its general applica-
bility.
In random point processes such as earthquakes,
machine noises, and human communications, it would be
worthwhile to examine whether the underlying condition
is better interpreted as active/inactive, or continuously
fluctuating.
ACKNOWLEDGMENTS
This study was supported in part by Grants-in-Aid
for Scientific Research to SS from the MEXT Japan
V1: 472l002
100
]
z
50
H
[
0
e
100
t
a
R
50
0
0
V1: 532l057
]
z
H
[
e
t
a
R
60
30
0
60
30
0
0
MT: 524l017
]
z
H
[
e
t
a
R
80
40
0
80
40
0
0
MT: 519l017
120
]
z
60
H
[
0
e
120
t
a
R
60
0
0
LGN: 474l017
]
z
H
[
e
t
a
R
60
30
0
60
30
0
0
LGN: 474l008
100
]
z
50
H
[
0
e
100
t
a
R
50
0
0
8
HMM - l
l
EBM
= - 0.11 ± 0.02
HMM - l
l
EBM
= - 0.08 ± 0.02
HMM - l
l
EBM
= - 0.09 ± 0.03
HMM - l
l
EBM
= + 0.03 ± 0.02
HMM - l
l
EBM
= + 0.05 ± 0.03
HMM - l
l
EBM
= + 0.06 ± 0.03
10
10
10
10
10
10
5
5
Time [s]
5
5
Time [s]
5
5
Time [s]
FIG. 7. Sample of biological spike trains and their estimated rates. Spike trains are recorded from V1, MT, and LGN
and the rates are estimated from the continuous and discontinuous estimators, EBM (green) and HMM (orange), respectively.
The title of each set of plots indicates the neuron ID in Refs [39], and the rate estimation selected according to the likelihood
are shown with the tick marks.
(25115718, 25240021), and by JST, CREST.
9
[1] F. Rieke, Spikes: exploring the neural code (MIT Press,
Cambridge, MA, 1999).
[2] E. D. Adrian and Y. Zotterman, “The impulses produced
by sensory nerve-endings part II. the response of a single
end-organ,” J. Physiol. 61, 151–171 (1926).
[3] D. H. Hubel and T. N. Wiesel, “Receptive fields, binoc-
ular interaction and functional architecture in the cat’s
visual cortex,” J. Physiol. 160, 106 (1962).
[4] J. H. Maunsell and D. C. Van Essen, “Functional prop-
erties of neurons in middle temporal visual area of the
macaque monkey. I. selectivity for stimulus direction,
speed, and orientation,” J. Neurophysiol. 49, 1127–1147
(1983).
[5] B. J. Richmond, L. M. Optican, M. Podell and H. Spitzer,
“Temporal encoding of two-dimensional patterns by sin-
gle units in primate inferior temporal cortex. I. response
characteristics,” J. Neurophysiol. 57, 132–146 (1987).
[6] J.-P. Nadal and N. Parga, “Nonlinear neurons in the low-
noise limit: a factorial code maximizes information trans-
fer,” Network 5, 565–581 (1994).
[7] R. Ben-Yishai, R. Lev Bar-Or and H. Sompolinsky, “The-
ory of orientation tuning in visual cortex,” Proc. Natl.
Acad. Sci. 92, 3844–3848 (1995).
[8] L. Bonnasse-Gahot and J.-P. Nadal, “Neural coding of
categories: information efficiency and optimal population
codes,” J. Comput. Neurosci. 25, 169–187 (2008).
[9] R. Rubin, R. Monasson and H. Sompolinsky, “Theory of
spike timing-based neural classifiers,” Phys. Rev. Lett.
105, 218102 (2010).
[10] J. G. Smith, “The information capacity of amplitude-and
variance-constrained scalar gaussian channels,” Inform.
Control 18, 203–219 (1971).
[11] G. Lewen, W. Bialek and R. Steveninck, “Neural cod-
ing of naturalistic motion stimuli,” Network 12, 317–329
(2001).
[12] M. Bethge, D. Rotermund and K. Pawelzik, “Second or-
der phase transition in neural rate coding: binary encod-
ing is optimal for rapid signal transmission,” Phys. Rev.
Lett. 90, 088104 (2003).
[13] S. Ikeda and J. H. Manton, “Capacity of a single spiking
neuron channel,” Neural Comput. 21, 1714–1748 (2009).
[14] A. P. Nikitin, N. G. Stocks, R. P. Morse and M. D.
McDonnell, “Neural population coding is optimized by
discrete tuning curves,” Phys. Rev. Lett. 103, 138101
(2009).
[15] D. J. Amit, Modeling Brain Function: The World of At-
tractor Neural Networks (Cambridge University Press,
Cambridge, 1989).
[16] M. Abeles, H. Bergman, I. Gat, I. Meilijson, E. Seide-
mann, N. Tishby and E. Vaadia, “Cortical activity flips
among quasi-stationary states,” Proc. Natl. Acad. Sci.
92, 8616–8620 (1995).
[17] R. Monasson and S. Rosay, “Crosstalk and transitions
between multiple spatial maps in an attractor neural net-
work model of the hippocampus: Phase diagram,” Phys.
Rev. E 87, 062813 (2013).
[18] H. Kim, B. J. Richmond and S. Shinomoto, “Neurons as
ideal change-point detectors,” J. Comput. Neurosci. 32,
137–146 (2012).
[19] F. Theunissen and J. P. Miller, “Temporal encoding in
nervous systems: a rigorous definition,” J. Comput. Neu-
rosci. 2, 149–162 (1995).
[20] P. Dayan and L. F. Abbott, Theoretical Neuroscience:
Computational and Mathematical Modeling of Neural
Systems (MIT Press, Cambridge, MA, 2001).
[21] C. D. Brody, “Correlations without synchrony,” Neural
Comput. 11, 1537–1551 (1999).
[22] D. J. Daley and D. Vere-Jones, An introduction to the
theory of point processes: volume II: general theory and
structure (Springer, New York, 2003).
[23] A. C. Smith and E. N. Brown, “Estimating a state-space
model from point process observations,” Neural Comput.
15, 965–991 (2003).
[24] D. J. C. MacKay, “Bayesian interpolation,” Neural Com-
put. 4, 415–447 (1992).
[25] A. D. Bruce and D. Saad, “Statistical mechanics of hy-
pothesis evaluation,” J. Phys. A 27, 3355 (1994).
[26] B. P. Carlin and T. A. Louis, Bayes and Empirical Bayes
Methods for Data Analysis (Chapman and Hall, London,
1996).
[27] S. Koyama and L. Paninski, “Efficient computation of the
maximum a posteriori path and parameter estimation in
integrate-and-fire and more general state-space models,”
J. Comput. Neurosci. 29, 89–105 (2010).
[28] T. Shimokawa, “Web application for the Bayesian estima-
tion of the firing rate,” [http://www.ton.scphys.kyoto-
u.ac.jp/shino/toolbox/ssBayes/bayes.html] (2010).
[29] C. M. Bishop, Pattern Recognition and Machine Learning
(Springer, New York, 2006).
[30] T. Shintani and S. Shinomoto, “Detection limit for rate
fluctuations in inhomogeneous poisson processes,” Phys.
Rev. E 85, 041139 (2012).
application
es-
rate
for
“Web
[31] Y. Mochizuki,
hidden Markov
two-state
timation
using
a
model,”
[http://www.ton.scphys.kyoto-
u.ac.jp/shino/toolbox/msHMM/HMM.html] (2013).
[32] T. M. Cover and J. A. Thomas, Elements of Information
Theory (Wiley, New York, 1991).
[33] M. M´ezard and A. Montanari, Information, Physics, and
Computation (Oxford University Press, Oxford, 2009).
[34] K. P. Murphy, Machine Learning: A Probabilistic Per-
spective (MIT Press, Cambridge, MA, 2012).
[35] S. Koyama, T. Shimokawa and S. Shinomoto, “Phase
transitions in the estimation of event rate: A path in-
tegral analysis,” J. Phys. A 40, F383 (2007).
[36] S. Koyama, T. Omi, R. E. Kass and S. Shinomoto, “In-
formation transmission using non-poisson regular firing,”
Neural Comput. 25, 854–876 (2013).
[37] W. Bialek, C. G. Callan and S. P. Strong, “Field theories
for learning probability distributions,” Phys. Rev. Lett.
77, 4693–4697 (1996).
[38] W. Bair, J. R. Cavanaugh, M. A. Smith, and J. A.
Movshon, “The timing of response onset and offset in
macaque visual neurons,” J. Neurosci. 22, 3189-3205
(2002); W. Bair and J. A. Movshon, “Adaptive tempo-
ral integration of motion in direction-selective neurons in
macaque visual cortex,” ibid. 24, 7305-7323 (2004); A.
Kohn, J. A. Movshon, et al., “Neuronal adaptation to
visual motion in area MT of the macaque,” Neuron 39,
681-692 (2003).
J. R. Cavanaugh, M. A. Smith,
[39] W. Bair,
and
nsa2004.3,”
sig.
J. A. Movshon,
“Neur.
arch.,
[http://www.neuralsignal.org]
(2002); A. Kohn,
J.
A. Movshon, et al., “Neur. sig. arch., nsa2004.6,” ibid.
(2003); W. Bair and J. A. Movshon, “Neur. sig. arch.,
nsa2004.4,” ibid. (2004).
[40] M. J. Beal, Z. Ghahramani and C. E. Rasmussen, “The
infinite hidden Markov model,” Adv. Neural Inf. Process.
Syst. 14, 577–584 (2002).
10
[41] Z. Ghahramani and G. E. Hinton, “Variational learning
for switching state-space models,” Neural Comput. 12,
831–864 (2000).
[42] E. D. Gershon, M. C. Wiener, P. E. Latham and B. J.
Richmond, “Coding strategies in monkey V1 and infe-
rior temporal cortices,” J. Neurophysiol. 79, 1135–1144
(1998).
[43] S. Shinomoto, H. Kim, T. Shimokawa, N. Matsuno, S.
Funahashi, et al., “Relating neuronal firing patterns to
functional differentiation of cerebral cortex,” PLoS Com-
put. Biol. 5, e1000433 (2009).
[44] S. Tokdar, P. Xi, R. C. Kelly and R. E. Kass, “Detection
of bursts in extracellular spike trains using hidden semi-
Markov point process models,” J. Comput. Neurosci. 29,
203–212 (2010).
|
1210.5348 | 1 | 1210 | 2012-10-19T09:07:33 | Operational Design Considerations for Retinal Prostheses | [
"q-bio.NC"
] | Three critical improvements for present day and future retinal vision implants are proposed and discussed: (1) A time profile for the stimulation current that leads predominantly to transverse stimulation of nerve cells; (2) auxiliary electric currents for electric field shaping with a time profile chosen such that these currents have small probability to cause stimulation; and (3) a local area scanning procedure that results in high pixel density for image/percept formation (except for losses at the boundary of an electrode array). | q-bio.NC | q-bio | !"#$%&'()%*+,#-'.)+/()-'0#$%&'()-+
#$%&'!()!*&'+%,"-.-/!01,!(2345015!6%17.-/-8!
1($+2#&')%*+3$(-&4#-#-+
!"91:;%;<;=! 24!>'=2$=;%&03! ?'@:%&:-!>AB%15=1!C1%D=$:%;@-!E<4! ,=$!F2$5=1:;=33=! "8-! G.HGI! >AJ
B%15=1-!K=$+01@!
.L%:<03! 01,! E<;212+2<:! #MN32$0;%21! *@:;=+:! O=:=0$&'! P0B2$0;2$@-! Q%D%:%21! 24! ?'@:%&:-!
F0;'=+0;%&:!01,!E:;$212+@-!R03%42$1%0!91:;%;<;=!24!>=&'12325@-!?0:0,=10-!RE!S"".T-!C*E!
/L%:<03! 01,! E<;212+2<:! #MN32$0;%21! *@:;=+:! O=:=0$&'! P0B2$0;2$@-! Q=N0$;+=1;:! 24! #3=&J
;$%&03!U!R2+N<;=$!#15%1==$%15-!V%2+=,%&03!#15%1==$%15-! *@:;=+:!U! 91,<:;$%03!#15%1==$%15-!
01,!WN';'03+2325@!U!L%:%21!*&%=1&=-!C1%D=$:%;@!24!E$%X210-!><&:21-!EY!ZTG."-!C*E!
8Q2'=1@! #@=! 91:;%;<;=-! [=&7! *&'223! 24!F=,%&%1=-! C1%D=$:%;@! 24! *2<;'=$1! R03%42$1%0-! P2:! E1J
5=3=:-!RE!SHH//-!C*E!
>'$==! &$%;%&03! %+N$2D=+=1;:! 42$! N$=:=1;! ,0@! 01,! 4<;<$=! $=;%103! D%:%21! %+N301;:! 0$=! N$2J
+
N2:=,!01,!,%:&<::=,\!]"^!E!;%+=!N$24%3=!42$!;'=!:;%+<30;%21!&<$$=1;!;'0;!3=0,:!N$=,2+%101;J
3@! ;2! ;$01:D=$:03! :;%+<30;%21! 24! 1=$D=! &=33:_! ].^! 0<M%3%0$@! =3=&;$%&! &<$$=1;:! 42$! =3=&;$%&! 4%=3,!
56-&$%7&+
:'0N%15!`%;'!0!;%+=!N$24%3=!&'2:=1!:<&'!;'0;!;'=:=!&<$$=1;:!'0D=!:+033!N$2B0B%3%;@!;2!&0<:=!
:;%+<30;%21_! 01,! ]/^! 0! 32&03! 0$=0! :&011%15! N$2&=,<$=! ;'0;! $=:<3;:! %1! '%5'! N%M=3! ,=1:%;@! 42$!
%+05=aN=$&=N;!42$+0;%21!]=M&=N;!42$!32::=:!0;!;'=!B2<1,0$@!24!01!=3=&;$2,=!0$$0@^)!
O=:=0$&'! 21! $=;%103! %+N301;:! '0:! B==1! 2152%15! 42$! 0B2<;! ;`2! ,=&0,=:\! =)5)-! bR'! "SS/-! c<!
+
"SSI-! Y$! "SSG-! Y$! .HH.-! O%! .HH/-! c<! .HH/-! R'! .HH8-! P%! .HH8-! ?0! .HHT-!F&! .HHG-! c2! .HHZ-!
c<! .HHS-! Y$! .H"Hd)! >'=! %,=0! %:! ;2! $=:;2$=! 0! :+033! N0$;! 24! D%:%21! ;2! N=2N3=! :<44=$%15! 4$2+!
89+:)&$(0;7&'()+
B3%1,1=::! ,<=! ;2! $=;%1%;%:! N%5+=1;2:0! 2$! ,<=! ;2! 05=! $=30;=,! +0&<30! ,=5=1=$0;%21)! E! &'%N!
`%;'!01!0$$0@!24!=3=&;$2,=:!%:!N30&=,!%1;2!01!=N%J$=;%103!2$!:<BJ$=;%103!N2:%;%21)!#3=&;$%&!&<$J
$=1;:!=+=$5%15!4$2+!;'=!=3=&;$2,=:!0$=!:==1!B@!;'=!B3%1,!N=$:21!0:!:+033!N'2:N'=1=:)!>'=@!
:=$D=! 0:! N%M=3:! 42$! N$=:=1;%15! 01! %+05=)! Q=:N%;=! :2+=! =1&2<$05%15! $=:<3;:! ;'=! 5203! 24! N$=J
:=1;%15!0!5$0@:&03=,!N%&;<$=!`%;'!0!;'2<:01,!2$!+2$=!N%M=3:!'0:!12;!@=;!B==1!$=0&'=,)!
91! 0! N$=&=,%15! N0N=$! b*&! .H"Hd! &2+N<;0;%2103! ;223:! '0D=! B==1! N$=:=1;=,-! `'%&'! +%5';! B=!
<:=4<3! %1! ;'=! ,=D=32N+=1;! 24! 0! $=;%103! N$2:;'=:%:)! >'=! 0NN3%&0B%3%;@! 24! ;'=! ;223:! '0:! B==1!
,=+21:;$0;=,!`%;'!:=D=$03!=M0+N3=:)!>'=!=M0+N3=:!:'=,!:2+=! 3%5';!21;2! ;'=!,%44%&<3;%=:!24!
%+N3=+=1;%15!;'=!%,=0!24!0!$=;%103!N$2:;'=:%:)!91!=::=1&=!;'=:=!,%44%&<3;%=:!0$=\!
") >'=! <:=! 24! ;22! :%+N3=! time profiles of the electric stimulation signal:)! >'=! +2:;! &2+J
+21!N$24%3=! %:!;'=!+212N'0:%&!:e<0$=!D23;05=!N<3:=-!`'%&'!@%=3,:!0!B%N'0:%&!&<$$=1;!
N<3:=)! W$! ;'=! B%N'0:%&! :e<0$=! D23;05=! N<3:=-! `'%&'! @%=3,:! 0! ;$%N'0:%&! &<$$=1;! N<3:=)!
*<&'!N<3:=:! 3=0,! ;2! ;'=!,=N230$%X0;%21! ]2$!'@N=$N230$%X0;%21^!24! 0! &=33!+=+B$01=! %1!
"!
!
!
;'=!4%$:;!N'0:=!24!;'=!&<$$=1;-!01,!;2!0!N230$%X0;%21!24!2NN2:%;=!:%51!%1!;'=!1=M;!N'0:=!
24!;'=!&<$$=1;)!>'<:!0!f@=:g!%:!42332`=,!B@!0!f12)g!
2. >'=! %,=0! 24! <:%15! 213@! 21=! =3=&;$2,=! N=$! N%M=3)! *=D=$03! $=:=0$&'! 5$2<N:! 0$=! <:%15! 0!
&2++21!&2<1;=$!=3=&;$2,=!0;! %14%1%;@)!W1=!5$2<N! %1!E<:;$03%0! %:!<:%15!I!counter elec-
/) Too little effort for shaping the electric field (or current). As has been discussed earlier
trodes around a center electrode on a hexagonal grid and, by current splitting, a common
counter electrode at infinity as well [Lo 2005]. Counter electrodes at infinity, however,
lead to cross-talk [Sc 2010].
trodes. But it also means undesired cross-talk between neighboring electrodes. >'=! E<:J
;$03%01! 5$2<N! %:! ;'=! 4%$:;! 21=! ;2! N=$42$+! ;$<=! 4%=3,! :'0N%15! B@! :%M! :2J&033=,! f5<0$,!
[Sc 2010], simultaneous firing of neighboring electrodes leads to bunching of field lines.
=3=&;$2,=:g!bP2!.HHTd)!
This bunching means field shaping, i.e., increased density of field lines above the elec-
91!b*&!.H"Hd!;'=:=!,%44%&<3;%=:!'0D=!B==1!%33<:;$0;=,)!c2`!;2!2D=$&2+=!;'=!,%44%&<3;%=:!%:!:<BJ
h=&;!24!;'=!N$=:=1;!N0N=$)!
91!*=&;%21!.!`=!:;<,@!;'=!;%+=!N$24%3=!24!;'=!=3=&;$%&!4%=3,!]2$!&<$$=1;^)!9;!'0:!B==1!:'2`1!%1!
b*&! .H"Hd! ;'0;! 0! :'2$;=$! N<3:=! `%;'! 0! +212N'0:%&! &<$$=1;-! N3<:! 0! 30;=1&@! N=$%2,! 04;=$! ;'=!
N<3:=-!'0:!01!0,D01;05=!2D=$!;'=!&2++213@!<:=,!N<3:=!;@N=:)!(=!`%33!$=D%:%;!;'%:!4%1,%15! %1!
*=&;%21!.)")! 91!*=&;%21!.).!01!=1;%$=3@!,%44=$=1;! ;@N=!24! :;%+<30;%21-!10+=3@!21=! ;'0;!=M&%;=:!
N$=,2+%101;3@!B@!;'=!,%=3=&;$%&!]2$!;$01:%=1;^!N0$;!24!;'=!&<$$=1;!`%33!B=!,%:&<::=,)!
*=&;%21! /! `%33! ,=03! `%;'! ;'=! N$2B3=+! 24! 4%=3,! :'0N%15)! W1=! `2<3,! 3%7=! ;2! '0D=! 01! =3=&;$%&!
:;%+<30;%21! 4%=3,! ;'0;!B=&2+=:!0&;%D=! %1!0!:+033! ;0$5=;!D23<+=! %1! ;'=!$=;%10-!`%;'2<;!'0D%15!
01@! +0h2$! =44=&;! =3:=`'=$=)! C142$;<10;=3@-! 01! =3=&;$%&! &<$$=1;-! 01,! ;'=! =3=&;$%&! 4%=3,! ;'0;!
,$%D=:! ;'=! &<$$=1;-! &0112;! B=! 42&<:=,! 3%7=! 0! B=0+! 24! 3%5';)!('0;! 21=! &01! ,2-! '2`=D=$-! %:! ;2!
<:=! :=N0$0;$%&=:! 42$! 5<%,%15! ;'=! &<$$=1;)! >'%:!`%33! B=! ,%:&<::=,! %1! *=&;%21! /)! >'=! B0:%&! =3=J
+=1;! 42$! 4%=3,!:'0N%15!`%33!B=!;'=!,%N23=!$%,5=-!0:!,%:&<::=,! %1!*=&;%21!/)")!F<3;%N23=!N%M=3:-!
`%;'! 0! &=1;=$! =3=&;$2,=! 01,! 4%=3,!5<%,%15! =3=&;$2,=:! 0$2<1,! ;'=+-!`%33!B=!,%:&<::=,! %1! *=&J
;%21! /).)! >'=! <:=! 24! :=D=$03! =3=&;$2,=:! 42$! 21=! N%M=3! 3=0,:! ;2! 0! $=,<&=,! N%M=3! ,=1:%;@)! P2&03!
0$=0!:&011%15!:;%+<30;%21!`%33!0332`!<:!;2!$=50%1!;'=! %1%;%03!'%5'!N%M=3!,=1:%;@-!`%;'2<;!'0DJ
%15!;2!%1&$=0:=!;'=!N'@:%&03!1<+B=$!24!=3=&;$2,=:-!0:!,%:&<::=,!%1!*=&;%21!/)/)!91!*=&;%21!/)8!
0! :;2&'0:;%&! 2N;%+%X0;%21! 4$0+=`2$7! %:! ,%:&<::=,! ;'0;! &01! B=! =+N32@=,! ]0^! ;2! 0&'%=D=! 2NJ
;%+03!4%=3,!:'0N%15!`%;'!0!5%D=1!=3=&;$2,=!0$$0@!]&'%N^_!]B^!;2!2N;%+033@!,=:%51!;'=!=3=&;$2,=!
0$$0@! %;:=34-! %)=)-! ;2! ,=;=$+%1=! ;'=! :N0;%03! 0$$015=+=1;! 01,! :%X=a,%0+=;=$! 24! ;'=! =3=&;$2,=:!
21! ;'=!=3=&;$2,=!0$$0@_!01,!]&^! ;2!2N;%+033@!,$%D=! ;'=!:;%+<30;%21! ;'$2<5'!0!5%D=1-! %+N301;J
=,!=3=&;$2,=!0$$0@! %1!$=03!;%+=!,<$%15!<:05=!24!;'=!$=;%103! %+N301;)!R21&3<,%15!$=+0$7:!0$=!
N$=:=1;=,!%1!*=&;%21!8)!
!
O=:=0$&'!21!=3=&;$%&03!:;%+<30;%21!24!1=$D=!&=33:!:;0$;=,! %1! ;'=! 30;=!"Z;'!&=1;<$@!`'=1!P<%5%!
K03D01%! :0`! 4$25! 3=5:! ;`%;&'! 01,! 42<1,! 2<;! ;'0;! ;'=! ;`%;&'%15! `0:! 01! =3=&;$%&03! =44=&;! bK0!
<9+!"&'='>').+&4#+&'=#+"$(1'*#+(1+%+-&'=;*%&'()+-'.)%*+
"GS"d)!
!
.!
!
E&;<033@! K03D01%! 42<1,! ;`2! 7%1,:! 24! =3=&;$%&03! :;%+<30;%21! ;'0;! `=$=! ,%:;%1&;3@! ,%44=$=1;! %1!
;=$+:!24!N'@:%&:)!*%1&=!%;!%:!%+N2$;01;!42$!;'=!N$=:=1;!N0N=$-!`=!B$%=43@!,%:&<::!;'=!;`2!'%:J
;2$%&03!=MN=$%+=1;:)!
9;! :;0$;=,! `%;'! P<%5%! K03D01%! 7%33%15! 4$25:! 01,! N$=N0$%15! ;'=%$! 3=5:! 42$! ;'=! 7%;&'=1)! c=! ,%,!
;'%:! %1! '%:! 30B2$0;2$@)! E;! ;'=! :0+=! ;%+=! '%:! 0::%:;01;! `0:! `2$7%15! 1=0$B@! `%;'! 01! =3=&;$2J
:;0;%&! 5=1=$0;2$)! 9;! '0NN=1=,! ;'0;! K03D01%! `0:! ;2<&'%15! ;'=! :N%1=! 24! 0! 4$25! `%;'! '%:! 71%4=!
`'=1-!0;! ;'=! :0+=!+2+=1;-! ;'=!=3=&;$%&!5=1=$0;2$! :N0$7=,)!E+0X%153@-! ;'=! 3=5:!24! ;'=!,=0,!
4$25! ;`%;&'=,i! K03D01%! 01,! '%:! 0::%:;01;! $=N=0;=,! ;'=! =MN=$%+=1;)! >'=! 4$25! 3=5:! ;`%;&'=,)!
>'=$=! `0:! 12! ,2<B;\! ;'=! ;`%;&'%15! `0:! &0<:=,! B@! ;'=! :N0$7! ,%:&'0$5=! 24! ;'=! =3=&;$2:;0;%&!
5=1=$0;2$)! !91!2$,=$!;2!5%D=!&$=,%;!;2!K03D01%!42$!'%:!'%:;2$%&!,%:&2D=$@!`=!,=12;=!;'%:!=D=1;!
0:!0!!"#$"%&'()*+&,-#"+&.%)!
K03D01%!B<%3;!01!01;=110!01,!42<1,!2<;!;'0;!03:2!;'<1,=$:;2$+!3%5';1%15!&0<:=,!;'%:!7%1,!24!
:;%+<30;%21)!E;!;'=!;%+=-!03+2:;!0!&=1;<$@!B=42$=!;'=!40+2<:!=MN=$%+=1;!B@!c=%1$%&'!c=$;X-!
P<%5%! K03D01%! '0,! 12! &'01&=! 24! 4%1,%15! 2<;! %1! ,=;0%3! `'0;! `0:! 52%15! 21! %1! ;'%:! 503D01%J"!
:;%+<30;%21)!
*221!04;=$!;'=!4%$:;!,%:&2D=$@!24!=3=&;$%&03!:;%+<30;%21!;'=$=!&0+=!0!:=&21,!21=)!>'%:!;%+=! %;!
'0NN=1=,!`%;'2<;!01@! :N0$7%15!=3=&;$2:;0;%&!5=1=$0;2$)!P<%5%!K03D01%!'0,!'<15!<N! &3=01=,!
4$25! 3=5:! 0;! 01! %$21! 5$%33)! ('=1! '=! ;2<&'=,! 0! 3=5! `%;'! 0! N%=&=! 24! &2NN=$! ;'=! 3=5! ;`%;&'=,i!
E50%1-! K03D01%! $=N=0;=,! ;'=! =MN=$%+=1;! 01,! +2,%4%=,! ;'=! :=;;%15)! >'=! :;%+<30;%21! =44=&;!
`0:! $=N$2,<&%B3=)! c=! 42<1,! 2<;! ;'0;! ;`2! ,%44=$=1;! 7%1,:! 24!+=;03! 01,! 0! &32:=,! =3=&;$%&! &%$J
&<%;!`=$=!=::=1;%03)!E:!`=!712`-! ;'%:! ;%+=!'%:! $=:=0$&'!`0:!D=$@! 4$<%;4<3!01,! 3=,! ;2! ;'=!,%:J
&2D=$@!24!;'=!503D01%&!&=33)!E50%1!`=!`01;!;2!5%D=!&$=,%;!;2!K03D01%!42$!'%:!'%:;2$%&!,%:&2D=$@!
01,!&033!;'=!:=&21,!7%1,!24!=3=&;$%&03!:;%+<30;%21!0!!"#$"%&'/)*+&,-#"+&.%)!
>'=:=! ;`2! 7%1,:! 24! :;%+<30;%21! 42<1,! B@! P<%5%! K03D01%! 0$=! ,%:;%1&;3@! ,%44=$=1;)!('0;! %:! ;'=!
B0:%&! ,%44=$=1&=j! >'=! =3=&;$%&03! N$2&=::! %1! ;'=! 503D01%J"! =MN=$%+=1;! %:! B@! :=D=$03! 2$,=$:! 24!
+051%;<,=! 40:;=$! ;'01! ;'=! 21=! %1! ;'=! 503D01%J.! =MN=$%+=1;)! >'=! D23;05=! :<$5=! +0@! '0D=!
B==1! :;$215=$! %1! 503D01%J"! B<;-! ,<=! ;2! ;'=! =M;$=+=3@! :'2$;! N<3:=! ,<$0;%21! ]%1! ;'=! 1012:=&J
21,! $015=i^-! ;'=! ;$01:4=$$=,! =3=&;$%&! &'0$5=! '0:! B==1! B@! 2$,=$:! 24!+051%;<,=! :+033=$! ;'01!
%1!;'=!503D01%J.!:;%+<30;%21)!
>2,0@-!+2$=!;'01!;`2!'<1,$=,!@=0$:!30;=$-!=3=&;$%&03!:;%+<30;%21!%:!:;%33!'0$,!;2!<1,=$:;01,)!
>'=! $=0:21! %:! ;'0;! B%2325%&03! ;%::<=! %:! 0! D=$@! &2+N3%&0;=,! =3=&;$%&! &21,<&;2$)! 9;! %:! N0$;3@! 01!
=3=&;$23@;=-! N0$;3@! 01! %1:<30;2$-! N0$;3@! 0! &2332%,)! 9;! &21;0%1:! +0&$2+23=&<3=:! `%;'! D0$%2<:!
N230$%X0B%3%;%=:!01,! %;!'0:!0!&2+N3%&0;=,!:;$<&;<$=! %1!:N0&=-!`%;'!&=33:!01,!&3=4;:! %1!B=;`==1!
;'=!&=33:)!O=:=0$&'!21! %;:!=3=&;$%&03!N$2N=$;%=:!'0:!B==1!2152%15!=D=$!:%1&=!K03D01%k:!=MN=$J
%+=1;:! 01,! %:! 12;! &2+N3=;=)! >'=! '%:;2$@! 24! ;'%:! $=:=0$&'! %:! `=33! ,2&<+=1;=,! %1! 0! $=D%=`!
N0N=$!B@!62:;=$!01,!*&'`01!b62!"SZSd)!
>'=$=! %:! 012;'=$! ,%44%&<3;@! %1! 0! ;'=2$=;%&03! ,=:&$%N;%21! 24! :;%+<30;%21)! >'=! :;%+<30;%21! N$2J
&=::! 24! 0! 1=$D=! &=33! %:! 0! 121J3%1=0$! 2N=$0;2$! %1! +0;'=+0;%&03! 3015<05=)! 62$! ;'=! 0103@:%:! 24!
=MN=$%+=1;03! $=:<3;:-! 62<$%=$! ;$01:42$+0;%21:! 0$=! &2++213@! <:=,! b62! "SZSd)! E! 121J3%1=0$!
2N=$0;2$! ,2=:! 12;! &2++<;=! `%;'! ;'=:=! ;$01:42$+0;%21:)! >'%:! +=01:! ;'0;! :;%+<30;%21! &01J
12;! B=! :;<,%=,! %1! 62<$%=$! ,=&2+N2:%;%21i! F2:;! =MN=$%+=1;:! 21! ;'=! =3=&;$%&! N$2N=$;%=:! 24!
B%2325%&03! ;%::<=:-! '2`=D=$-! 0$=! ,21=! `%;'! 03;=$10;%15! &<$$=1;:! 24! D0$%2<:! 4$=e<=1&%=:-! %)=)-!
%1!62<$%=$!,=&2+N2:%;%21!b62!"SZSd)!
!
/!
!
>'=2$=;%&03! ,=:&$%N;%21:! 24! =3=&;$%&! &<$$=1;:! %1! B%2325%&03! ;%::<=:! 01,! ;'=2$=;%&03! ,=:&$%NJ
;%21:!24! ;'=! :;%+<30;%21!N$2&=::!24!1=$D=! &=33:! 0$=! 03`0@:!B0:=,!21! :%+N3%4%=,!+2,=3:-! :<&'!
0:!&21;%1<<+!+2,=3:-!:;0;%:;%&03!+2,=3:-!:<:N=1:%21!24!:N'=$%&03!&=33:! %1!01!=3=&;$23@;=-!=;&)!
>'=!N$=,%&;%D=!N2`=$!24!:<&'!+2,=3!,=:&$%N;%21:!%:!12;!D=$@!'%5')!l=D=$;'=3=::-!;'=!+2,=3:!
0$=! <:=4<3! 0:! 0! B0:%:! 42$! %1;<%;%D=! ,=:&$%N;%21:-! 01,! ;'=@! 0$=! <:=4<3! 42$! ,=:%51%15! =MN=$%J
+=1;:)!>'=!4%103!01:`=$-!24!&2<$:=-!`%33!03`0@:!B=!5%D=1!B@!=MN=$%+=1;!01,!12;!B@!;'=2$@)!
91!;'=!42332`%15!`=!`%33!,=:&$%B=!;'$==!D=$@!,%44=$=1;!`0@:!24!=3=&;$%&03!:;%+<30;%21!24!1=$D=!
&=33:\! ]"^! ;'=!=3=&;$%&03! :;%+<30;%21!B@!0! :<:;0%1=,!2'+%&! &<$$=1;-! ].^! ;'=!=3=&;$%&03! :;%+<30J
;%21! B@! 0! ,=3;0J4<1&;%21J3%7=! &'0$5=! %1h=&;%21-! 01,! ]/^! ;'=! :;2&'0:;%&! :;%+<30;%21! B@! 0! :+033!
2'+%&!&<$$=1;)!
!
>'=! :;01,0$,!+2,=3! 24! =3=&;$%&03! :;%+<30;%21! %:! B0:=,! 21! ;'=! 4%1,%15:! 24!c2,57%1! 01,!c<MJ
3=@!0B2<;!;'=!=3=&;$%&!N$2N=$;%=:!24!1=$D=!&=33:!bc2!"ST.d!01,!21!c=0D%:%,=k:!&0B3=!=e<0;%21)!
>'=!2$%5%103! &0B3=! =e<0;%21!,=:&$%B=:! ;'=! 5$0,<03! 32::!24! 0! ;=3=5$0N'! :%5103! %1! 01!2&=01! &0J
<989++?().'&;0')%*+-&'=;*%&'()+
B3=)!9;!03:2!,=:&$%B=:!;'=!:%;<0;%21!`'=1!;'=!:%5103!;$0D=3:!%1!;'=!2&=01!01,!;'=!&0B3=!N%&7:!%;!
<N!3%7=!01!01;=110_!213@!;'=!,$%D%15!;=$+!24!;'=!=e<0;%21!%:!,%44=$=1;!%1!;'=!;`2!&0:=:)!('=1!
0! ,=1,$%;=! 2$! 01! 0M21! '0:! ;'=! +0;'=+0;%&03! N$2N=$;%=:! 24! 0! &0B3=-! c=0D%:%,=k:! =e<0;%21! %:!
03:2!D03%,!%1!0!&21;%1<<+!+2,=3!24!0!1=$D=!;%::<=!:<&'!0:!;'=!$=;%10)!
91! b*&! .H"Hd! ;'=! 01;=110JD=$:%21! 24! ;'=! &0B3=! =e<0;%21! '0:! B==1! ,=$%D=,-! ;`2! +=;'2,:! 24!
:23<;%21!'0D=!B==1!N$=:=1;=,-!01,!N$0&;%&03!=M0+N3=:!'0D=!B==1!,%:&<::=,)!
?$=:=1;3@-! 3215%;<,%103!:;%+<30;%21! %:!12;!%1!;'=!42&<:!24!2<$! %1;=$=:;)!(=!0$=!+2$=!%1;=$=:;J
=,! %1! ;$01:D=$:=! :;%+<30;%21-! %)=)-! N=$N=1,%&<30$! ;2! ;'=! 0M21! 2$! ,=1,$%;=! 24! 0! 1=$D=! &=33)!
>'=$=42$=!`=!213@!`01;!;2!:'2$;3@!$=D%=`!`'0;!'0:!B==1!,%:&<::=,!%1!b*&!.H"Hd!21!3215%;<J
,%103!:;%+<30;%21)!
E! :<:;0%1=,! 2'+%&! &<$$=1;! %:! N$2,<&=,! B=;`==1! 01! =3=&;$2,=! 01,! 0! &2<1;=$! =3=&;$2,=! B@! 0!
:2J&033=,! D23;05=J&21;$233=,! 5=1=$0;2$! 2$! 0! &<$$=1;J&21;$233=,! 5=1=$0;2$)! E! D23;05=J! &21J
;$233=,! 5=1=$0;2$! ;@N%&033@! 0NN3%=:! 0! &21:;01;! D23;05=! 24! 0! 4=`! D23;:! 42$! 0! ;%+=! %1;=$D03!
:2+=`'=$=! B=;`==1! H)/! 01,! /!+%33%:=&21,:)! >'=! &<$$=1;-! 04;=$! $=0&'%15! 0!+0M%+<+-! 52=:!
,2`1! `'%3=! ;'=! c=3+'23;XJ30@=$:! 24! ;'=! =3=&;$2,=:! 0$=! &'0$5%15! <N)! E! &<$$=1;J&21;$233=,!
5=1=$0;2$! 7==N:! ;'=! &<$$=1;! 0;! 0! &21:;01;! D03<=! `'%3=! ;'=! 0NN3%=,! D23;05=! %:! 0,h<:;=,)! 91!
B2;'! &0:=:-! &0$=! %:! ;07=1! ;2! 0D2%,! %$$=D=$:%B3=! &'=+%&03! $=0&;%21:)! 91! ;'=! 4%$:;! &0:=-! ;'=! 0NJ
N3%=,! D23;05=! %:! :=;! 0&&2$,%153@! 32`-! %1! ;'=! :=&21,! &0:=-! ;'=! %1h=&;=,! &'0$5=! %:! 3%+%;=,! B@!
&'22:%15!0!:+033!&<$$=1;!01,!0!:'2$;!=12<5'!;%+=!%1;=$D03)!
#M&=N;! 42$! ;'=! D=$@! B=5%11%15! 24! &'0$5=! %1h=&;%21-! `=! `%33! '0D=! 0! :<:;0%1=,! 2'+%&! &<$$=1;!
432`%15!03215!;'=!&3=4;:!B=;`==1!;'=!&=33:!24!;'=!$=;%10)!91!;'=!<1,=$3@%15!&21;%1<<+!+2,=3-!
;'=!&3=4;:!01,!&=33:!24!;'=!$=;%10!0$=!:+=0$=,!2<;-!=M&=N;!42$!;'=!1=$D=!&=33!<1,=$!&21:%,=$0J
;%21)!('=1! ;'=!&<$$=1;!'0:!0! 3215%;<,%103!]%)=)-!N0$033=3^!&2+N21=1;!`%;'!$=:N=&;! ;2! ;'=!0M%:!
24! ;'=! 0M21! 2$! ;'=! ,=1,$%;=! <1,=$! &21:%,=$0;%21-! %;:! =3=&;$%&! 4%=3,! `%33! =1;=$! %1;2! ;'=! 1=$D=!
&=33)!>'=!N$2&=::! %:!,=:&$%B=,!B@! ;'=! 01;=110JD=$:%21!24!c=0D%:%,=k:! &0B3=! =e<0;%21)!(=! $=J
4=$!;2!b*&!.H"Hd!01,!:'2`!0!N%&;<$=!24!;'0;!N<B3%&0;%21)!
!
8!
!
!
@'.9+ 89+ A&'=;*%&'()+ -'.)%*+ #)&#$').+ ')&(+ %+ -#7&'()+ (1+ %)+ %B()+ ($+ 0#)0$'&#+ (1+ !""+µ#+ *#).&49+ C%D+ &4#+ 7%E
91!;'%:!=M0+N3=-!;'=!5=1=$0;2$! %:!D23;05=J&21;$233=,!01,!0NN3%=:!0!D23;05=!24!J/)0!42$!0!,<$0J
&4(0'7+F(*&%.#+";*-#+(1+ $%&+'+%)0+0;$%&'()+()*""%µ+%%""*'#0+&(+&4#+#*#7&$(0#G+C6D+&4#+&'=#+"$(1'*#+(1+&4#+
;%21! 24! 122!+%&$2:=&21,:-! 0:! :'2`1! %1! 6%5)! "0)! 6%5)! "B! :'2`:! ;'=! &<$$=1;! N$24%3=! ;@N%&03! 42$!
$#-;*&').+7;$$#)&+C')+%$6'&$%$H+;)'&-DG+C7D+&4#+&$%)-=#=6$%)#+"(&#)&'%*+')+&4#+%B()+($+0#)0$'&#+%1&#$+*%
01! =3=&;$2,=! 24! :+033-! B<;! :<44%&%=1;! &0N0&%;@)! 6%5)! "&! :'2`:! ;'=! :23<;%21! 24! ;'=! &0B3=! =e<0J
µ+,%!"%µ+,%&"%µ++%)0+*"%µ+I+&4#+$#-&').+"(&#)&'%*+'-+-#&+&(+$"-".%'9++
;%21-! 42$! ;@N%&03! &=33! N0$0+=;=$:! 01,! 0! :2J&033=,! &32:=,! =1,! B2<1,0$@! &21,%;%21)! 9;! %:! :==1!
'2`! ;'=! =3=&;$%&! 4%=3,! 24! ;'=! 2'+%&! &<$$=1;! =1;=$:! %1;2! ;'=! &=33)! E! '@N=$N230$%X0;%21! %:! N$2J
,<&=,!0;!21=!=1,!01,!,=N230$%X0;%21! %:!N$2,<&=,!0;! ;'=!2;'=$!=1,!24! ;'=!&=33!:=&;%21)!('=1!
;'=!,=N230$%X0;%21!'0:!$=0&'=,!%;:!&$%;%&03!D03<=!;'=!N0::%D=!&0B3=!`%33!B=&2+=!01!0&;%D=!21=!
0&&2$,%15! ;2!c2,57%1! 01,!c<M3=@! bc2! "ST.d-! %)=)-! ;'=! :;%+<30;%21! N$2&=::! &21;%1<=:! 01,! 01!
0&;%21!N2;=1;%03!0$%:=:)!
9;!%:!%1;=$=:;%15!;2!'0D=!0!&32:=$!3227!0;!;'=!+0;'=+0;%&03!+2,=3!N$=:=1;=,!%1!b*&!.H"Hd)!W1=!
&01! :==! ;'0;! 213@! ;'=! 3215%;<,%103! ]%)=)-! N0$033=3^! &2+N21=1;! 24! 01! =3=&;$%&! 4%=3,-! $=30;%D=! ;2!
;'=! 0M%:!24! ;'=! 0M21!2$!,=1,$%;=-!`%33!'0D=! 01! =44=&;!21! ;'=! :23<;%21!24! ;'=! &0B3=! =e<0;%21_! 0!
;$01:D=$:=!]%)=)-!N=$N=1,%&<30$^!&2+N21=1;!`%33!12;!'0D=!01@!=44=&;!0;!033)!
!
E:! '0:! B==1! :;0;=,! %1! ;'=! N$=&=,%15! :=&;%21! ;'=! &0B3=! +2,=3! 1==,:! 0! 3215%;<,%103! &<$$=1;!
42$! ;'=! :;%+<30;%21! 24! 0! 121J+@=3%10;=,! 0M21! 2$! 0! ,=1,$%;=)! E! ;$01:D=$:=! ]%)=)-! N=$N=1,%&<J
<9<9+J$%)-F#$-#+-&'=;*%&'()+
30$^! =3=&;$%&! &<$$=1;! 2$! 4%=3,! &0112;! :;%+<30;=! %1! ;'=! &0B3=!+2,=3)! 6$2+! =MN=$%+=1;:-! '2`J
=D=$-!`=!712`!;'0;!%;!&01i!
91! 0! B=0<;%4<3! =MN=$%+=1;-! *'=33=@! 6$%=,! b6$! .HHSd! :=0$&'=,! 42$! :=&;%21:! 24! 01! 0M21! =+=$5J
%15! 4$2+! 0! $=;%103! 50153%21! &=33! ;'0;! 0$=! =:N=&%033@! :=1:%;%D=! ;2! =3=&;$%&03! :;%+<30;%21)! c=!
N30&=,! 0! &21%&03! =3=&;$2,=! %1;2! ;'=! $=;%10-! :%,=`0@:! 4$2+! 01! 0M21! 01,! N2%1;%15! ;2`0$,! ;'=!
0M21! ]%)=)-! N=$N=1,%&<30$! ;2! ;'=! 0M21^-! 01,! %1h=&;=,! 0! &21:;01;! &<$$=1;! 42$! ;'=! $0;'=$! :'2$;!
;%+=!24!"HHJ.HH!+%&$2:=&21,:)!>'=!;$01:D=$:=!&<$$=1;!:;%+<30;=,!;'=!0M21i!c2`!&2<3,!;'0;!
'0NN=1j!
!
T!
!
P=;!<:!4%$:;!;$@!;2!<1,=$:;01,!`'0;!'0NN=1=,!%1!K03D01%k:!4%$:;!=MN=$%+=1;)!E:!'0:!B==1!:0%,-!
01! 0::%:;01;! 24! K03D01%! 2N=$0;=,! 01! =3=&;$2:;0;%&! 5=1=$0;2$-! &'0$5=,! <N! ;'=! +=;03! :N'=$=:!
<1;%3! ;'=@! ,%:&'0$5=,!`%;'! 0! 32<,! :N0$7)!E;! ;'%:!+2+=1;!K03D01%!`0:! :;%33! &3=01%15! ;'=! 4$25!
3=5:)!c%:! %$21!71%4=!`0:! %1!&21;0&;!`%;'! ;'=!:N%103!&2$,!24! ;'=!,=0,! 4$25)!>'=!71%4=-! ;25=;'=$!
`%;'! ;'=!0$+!24!K03D01%-!0&&%,=1;033@!B=&0+=!01!01;=110!01,!0!D=$@!:+033!&'0$5=!0NN=0$=,!
0;! ;'=! ;%N!24! ;'=!71%4=)! 91! ;'%:!`0@! ;'=! ;%N!24! ;'=!71%4=!B=&0+=!0!:;%+<30;%15!=3=&;$2,=)! 91!2$J
,=$! ;2! <1,=$:;01,! ;'=! =3=&;$%&! N$2&=::! 3=0,%15! ;2! ;'%:! :;%+<30;%21! 3=;! <:! ;$@! ;2! D%:<03%X=!
;'%15:!%1!0!:32`J+2;%21!f+2D%=)g!
!
@'.9+ <9+ /$(--E-#7&'()+ (1+ %+ 0#)0$'&#G+ ($+ )()E=H#*')%�+ %B()G+ %&+ &4#+ &'=#+ (1+=%B'=;=+ "(*%$'>%&'()9+ 2#E
P=;!;'=!0NN=0$01&=!24!0!:+033!=3=&;$%&!&'0$5=!0;!;'=!;%N!24!;'=!71%4=!B=!;'=!4%$:;!;%+=!4$0+=!24!
*%B%&'()+=#%)-+ &4%&+ '()-+ %$#+ 0$'F#)+ -'0#K%H-G+ %-+ -4(K)+ 6H+ %$$(K-G+ %)0+='B+K'&4+ '()-+ 1$(=+ &4#+ (&4#$+
2<$! :32`! +2;%21! +2D%=)! W4! &2<$:=-! ;'=! 5=1=$0;2$! &0112;! ,=N2:%;! ;'=! &'0$5=! %1! X=$2! ;%+=)!
-'0#+(1+ &4#+7'$7*#G+&4;-+ 1($=').+%+)#;&$%*+-%*')#+%.%')9+L4'*#+$#*%B%&'()+ '-+.(').+()G+&4#+7#**+=#=6$%)#+
V<;!%;!&01!,2!:2!%1!D=$@!:'2$;!;%+=-!%)=)-!21!;'=!2$,=$!24!0!+%&$2:=&21,!2$!3=::)!
'-+0#"(*%$'>#0+%&+()#+-'0#+(1+&4#+7'$7*#+%)0+4H"#$"(*%$'>#0+%&+&4#+(""(-'&#+-'0#9+
91! 0! :=&21,! ;%+=! 4$0+=! `=! &21:%,=$! ;'=! R2<32+B! 4%=3,! 24! ;'=! &'0$5=! ;'0;! '0:! 0NN=0$=,! 0;!
;'=! ;%N! 24! ;'=! 71%4=)! >'%:! R2<32+B! 4%=3,! N230$%X=:! ;'=! $=;%10-! 2$! 01@! 2;'=$! B%2325%&03! ;%::<=-!
;'0;! 4%33:! ;'=!:N0&=!B=;`==1! ;'=! ;%N!24! ;'=!71%4=!01,!0!$=+2;=!&2<1;=$!=3=&;$2,=)!>'=! %21:!24!
;'=! :03%1=-! ;25=;'=$! `%;'! ;'=%$! &32<,:! 24! `0;=$! +23=&<3=:-! `%33! +2D=! 0! 3%;;3=-! ,$%D=1! B@! ;'=!
R2<32+B! 4%=3,)!>'=@!`%33! 42$+! ;'=!c=3+'23;X! 30@=$:! 0;! ;'=! =3=&;$2,=! 01,! &2<1;=$! =3=&;$2,=)!
E1,! ;'=@!`%33! 42$+!,2<B3=! 30@=$:!0;!&=33!+=+B$01=:-!2$!0;!01@!2;'=$!121J&21,<&;%15!2B:;0J
&3=-!0:!:&'=+0;%&033@!,=N%&;=,!%1!6%5)!.)!P=;!;'=!:;%+<30;%15!=3=&;$2,=!`%;'!N2:%;%D=!&'0$5=!B=!
0;! ;'=!B2;;2+!24! ;'=!N%&;<$=! ]6%5)!.^!01,! ;'=! $=+2;=! &2<1;=$!=3=&;$2,=!0;! ;'=! ;2N!01,! 3=;!01!
0M21!2$!,=1,$%;=!&<;!;'=! %+05=!N301=! %1!:<&'!0!`0@!;'0;!;'=!&$2::!:=&;%21! %:!:==1!0:!;'=!&%$J
&3=!,=N%&;=,! %1!6%5)!.)!>'=!,2<B3=! 30@=$!24!&'0$5=:!`%33!B=!:;$215=:;!0;!;'=!B2;;2+!01,!]`%;'!
2NN2:%;=! :%51^! 0;! ;'=! ;2N! 24! ;'=! &%$&3=)! 91! ;'=!+%,,3=! B=;`==1! ;2N! 01,! B2;;2+! ;'=$=!`%33! B=!
12!N230$%X0;%21)!
!
I!
!
P=;! <:! 3227! 12`! 0;! 4<$;'=$! ;%+=! 4$0+=:! 24! ;'=! :32`! +2;%21! +2D%=)! >'=! %1:<30;%21! 24! ;'=!
+=+B$01=! %:!12;!N=$4=&;-!B<;!522,!=12<5'! ;2!7==N! ;'=! %21:!01,! &2<1;=$! %21:!24! ;'=!,2<B3=!
30@=$! 0N0$;! 42$! :2+=! ;%+=)!>'=! 213@!,=5$==! 24! 4$==,2+! 24! ;'=:=!N0%$:! 24! &'0$5=:! %:! ;2!+2D=!
:%,=`0@:)!>'=!,$%D%15!42$&=!42$!;'%:!+2D=+=1;!%:!12;!D=$@!:;$215-!B<;!:<44%&%=1;!;2!+07=!;'=!
&'0$5=:!:N$=0,!2<;!2D=$!;'=!:<$40&=!24!;'=!+=+B$01=_!0;!;'=!+=,%01!24!;'=!&%$&3=!&'0$5=:!24!
2NN2:%;=! :%51!+%M! 01,! 42$+! 0! 1=<;$03! :03%1=! 050%1)! >'=! $=:<3;%15! &<$$=1;! %:! &033=,! 0! $=30M0J
;%21!&<$$=1;)! 91! ;'=! ;'=2$@!24!=3=&;$%&!&<$$=1;:! %1!B%2325%&03! ;%::<=:!:<&'!$=30M0;%21!&<$$=1;:!
N30@! 01! %+N2$;01;! $23=)! 91! 40&;-! ;'=$=! =M%:;! :=D=$03! $=30M0;%21! &<$$=1;:! `%;'! ,%44=$=1;! 3250J
$%;'+%&! ,=&0@! ;%+=:! b62! "SZSd)! c=$=! `=! 0$=! 213@! %1;=$=:;=,! %1! ;'=! $=30M0;%21! N$2&=::! ;'0;!
N30@:!;'=!,2+%101;!$23=!%1!;$01:D=$:=!:;%+<30;%21!24!1=$D=!&=33:)!
P=;! <:! $=&033! ;'=! 4$0+=:! 24! 2<$! :32`!+2;%21!+2D%=\! 34"5!6) &%763+&.%)8)9.#"5&:"+&.%).;),6,'
<5"%6*)8)56#"="+&.%)3-556%+>)9;!%:!0!:%+N3%4%=,!+2,=3)!(=!<:=!%;!%1!2$,=$!;2!%33<:;$0;=!'@N=$N2J
30$%X0;%21!01,!,=N230$%X0;%21!24! ;'=!+=+B$01=!24!0!1=$D=!&=33!B@!0! ;$01:D=$:=!=3=&;$%&! 4%=3,)!
91! $=03%;@-! ;'=! ;%+=! 4$0+=:! 2D=$30N-! 01,! 0! :+033! 2'+%&! &<$$=1;! `%33! 0&&2+N01@! ;'=! ,%:J
N30&=+=1;!&<$$=1;)!
l=D=$;'=3=::-! ;'=!:%+N3%4%=,!+2,=3!24!2<$!:32`!+2;%21!+2D%=! %:!<:=4<3)! 9;!5%D=:!<:!01!=MN30J
10;%21! 24! `'0;! P<%5%! K03D01%! '0:! 3%7=3@! :==1! %1! '%:! D=$@! 4%$:;! =MN=$%+=1;)! E:! '0:! B==1! :0%,!
0B2D=-!;'=!'01,!24!K03D01%!01,!;'=!71%4=!%1!'%:!'01,!0&;=,!3%7=!01!01;=110!01,-!%1!3=::!;'01!0!
+%&$2:=&21,-! 0! D=$@! :+033! =3=&;$%&! &'0$5=! 0NN=0$=,! 0;! ;'=! ;%N! 24! ;'=! 71%4=)! E! ,%:N30&=+=1;!
&<$$=1;! $01! ;'$2<5'! ;'=! ;%::<=! 24! ;'=! 4$25! 01,! N230$%X=,! ;'=! +=+B$01=! 24! 50153%21! &=33:)!
>'=$=! `0:! 12! &2<1;=$! N<3:=-! 3%7=! %1! 01! ER! &<$$=1;)! >'=! N230$%X0;%21! ,=&0@=,! 0&&2$,%15! ;2!
$=30M0;%21)!ENN0$=1;3@-!,=N230$%X0;%21! 24!N0$;:! 24! ;'=! &=33!+=+B$01=!,<$%15! $=30M0;%21-! 423J
32`=,!B@!0! 30;=1&@!N=$%2,-!`0:!:<44%&%=1;! ;2! ;$%55=$!01!0&;%21!N2;=1;%03! ;'0;! 3=;! ;'=! 4$25! 3=5:!
;`%;&')!
(=!&2+=!B0&7!12`!;2!;'=!=MN=$%+=1;!B@!*'=33=@!6$%=,!01,!'%:!&2330B2$0;2$:!b6$!.HHSd)!>'=!
&0B3=!+2,=3! ,2=:! 12;! =MN30%1! ;'=%$! 4%1,%15:)! >'=$=! '0:! B==1! 01! 2'+%&! &<$$=1;! 42$! "HHJ.HH!
+%&$2:=&21,:)! V<;! ;'%:! &<$$=1;! `0:! 0%+%15! 0;! ;'=! :N2;! %1! `'%&'! ;'=! %1D=:;%50;2$:! `=$=! %1J
;=$=:;=,-! %)=)-! '0,! 0! ,%$=&;%21! N=$N=1,%&<30$! ;2! ;'=! 0M21)! 60$;'=$! 0`0@! 4$2+! ;'0;! :N2;! %;! ,%,!
'0D=! 0! 3215%;<,%103! &2+N21=1;-! B<;! ;'=! &'2:=1! N230$%;@! @%=3,=,! '@N=$N230$%X0;%21! %1:;=0,!
24!,=N230$%X0;%21)!>'%:!N$=&3<,=:! 3215%;<,%103! :;%+<30;%21)!>'=! :'0$N!21:=;!24! &'0$5=! %1h=&J
;%21! `%;'! ;'=! $=&;015<30$! ;%+=! N$24%3=! 24! ;'=! &<$$=1;-! '2`=D=$-! N$2,<&=,! 0! :<44%&%=1;3@!
:;$215! D23;05=! :<$5=! 42$! 0! 503D01%J"! 2$! ;$01:D=$:=! :;%+<30;%21)! >'=! $=:;! N=$%2,! 1==,=,! 42$!
;'=!503D01%J"!:;%+<30;%21!'0,!B==1!:=;!;2!"H!+%33%:=&21,:!01,!`0:!N0$;!24!;'=!;%+=!N$24%3=)!
>'<:-!42$!;'=!2N=$0;%21!24!0!$=;%103!%+N301;-!;$01:D=$:=!:;%+<30;%21!:==+:!;2!B=!01!=M&=33=1;!
03;=$10;%D=! ;2! 3215%;<,%103! :;%+<30;%21)! E1! %1h=&;=,! &'0$5=J&21;$233=,! 5=1=$0;2$-! %1:;=0,! 24!
0! D23;05=J&21;$233=,! 2$! &<$$=1;J&21;$233=,! 21=-! `2<3,! B=! %,=03)! 9;! %:! %+N2$;01;! ;'0;! ;'=!
&'0$5=!%:!%1h=&;=,!D=$@!40:;!m!;=$+=,!430:'!2$!:'2&7!:;%+<30;%21!m!40:;=$!;'01!$=30M0;%21-!01,!
`%;'2<;!0!D23;05=!B2<1,)!>'%:!+=01:! ;'0;! ;'=!=1=$5@!1==,=,!'0:! ;2!B=! :;2$=,!B=42$=! %1h=&J
;%21! %1! ;'=! &%$&<%;$@! 24! ;'=! 5=1=$0;2$! ]=)5)-! %1! 0! &0N0&%;2$^)! 91h=&;%21! +0@! B=! $=N=0;=,! 04;=$!
;'=! $=30M0;%21! &<$$=1;! '0:! ,=&$=0:=,! 4$2+! %;:! +0M%+<+! D03<=! ;2! :2+=! 5%D=1! 32`=$! D03<=-!
B=42$=!,%:&'0$5%15!;'=!=3=&;$2,=)!#3=&;$%&!4%=3,!:'0N%15!+0@!B=!1==,=,!;2!'%;!;'=!;0$5=;!&=33:!
]:==!B=32`^)!
!
G!
+
!
*;2&'0:;%&! :;%+<30;%21!24! ;'=! $=;%10! ;07=:!N30&=!`'=1! ;'=! :;%+<30;%21! &<$$=1;! %1! =M;$0&=33<J
30$! :N0&=! %1;=$4=$=:! `%;'! ;'=! 2+1%N$=:=1;! :@10N;%&! 12%:=! 24! 0! 1=<$03! 1=;`2$7)! *@10N;%&!
<9M9++A&(74%-&'7+-&'=;*%&'()+
12%:=!%:!712`1!4$2+!;'=!&2$;=M!24!;'=!B$0%1_!%;!%:!=MN=&;=,!;2!B=!N$=:=1;!03:2!%1!;'=!50153%21!
&=33! 30@=$! 24! ;'=! $=;%10)! >'=! `2$,! 12%:=! +=01:! f121J,%$=&;%2103g! 2$! f12! ;$01:N2$;! 24! %142$J
+0;%21g)!E!:+033!=3=&;$%&! 4%=3,!+0@!B=!:<44%&%=1;! ;2!+07=!:<&'!12%:=!f,%$=&;%2103g-! %)=)-! ;$01:J
N2$;%15!%142$+0;%21)!91!;'=!$=;%10!:<&'!N$%+%;%D=!%142$+0;%21!+%5';!B=!N=$&=%D=,!0:!0!N'2:J
N'=1=)! >'=$=! %:! 0! :%+%30$%;@! ;2! ;'=! F0M`=33JV23;X+011! D=32&%;@! ,%:;$%B<;%21! 24! 50:! +23=J
&<3=:\! 0!D=$@! :+033!,=D%0;%21! 4$2+! 0! :N'=$%&03-! %:2;$2N%&!,%:;$%B<;%21! %:!+0&$2:&2N%&033@!2BJ
:=$D=,!0:! f`%1,g-! %)=)-!,%$=&;%2103)!>'=!+051%;<,=!24!01!=3=&;$%&!&<$$=1;!:<44%&%=1;! ;2!&0<:=!0!
N'2:N'=1=! %:!=MN=&;=,! ;2!B=!:+033-!:+033=$! ;'01! ;'=!&<$$=1;!1==,=,! 42$! 3215%;<,%103!:;%+<J
30;%21)!(=!=MN=&;!:<&'!N'2:N'=1=:!;2!0NN=0$-!42$!%1:;01&=-!%1!0$=0!R!24!6%5)!/)!
91!+2:;! =MN=$%+=1;03! 0NN3%&0;%21:!21=!'0:! 0!+%M;<$=!24! 3215%;<,%103! 01,! ;$01:D=$:=! :;%+<J
+
30;%21)! >'=! ;%+=! N$24%3=! :'2`1! %1! 6%5)! "0-! 42$! %1:;01&=-! :;0$;:! `%;'! 0! &'0$5=! %1h=&;%21! 0;! ;'=!
B=5%11%15! 01,! 52=:! 2D=$! %1;2! 01! =MN21=1;%033@! ,=&0@%15! 2'+%&! &<$$=1;-! B=42$=! ,%:&'0$5%15!
<9N9+O'B#0+*().'&;0')%*+%)0+&$%)-F#$-#+-&'=;*%&'()+
:;0$;:! 0;! ;n>)! 94!`=! ;07=! %;! 0:! 0! :<N=$N2:%;%21! 24! 0! :%5103! 42$! 3215%;<,%103! :;%+<30;%21! N3<:! 0!
:%5103!42$!;$01:D=$:=!:;%+<30;%21-!;'=!2'+%&!&<$$=1;!N'0:=!&01!B=!D%=`=,!0:!;'=! 30;=1&@!N=J
$%2,!24!;'=!:%5103!42$!;$01:D=$:=!:;%+<30;%21)!
94!`=!`01;!;2!+=0:<$=!N<$=!]2$!03+2:;!N<$=^!3215%;<,%103!:;%+<30;%21!`=!'0D=!;2!:;0$;!`%;'!
0!D23;05=!$0+N-!%1:;=0,!24!0!D23;05=!:;=N-!;2!0D2%,!;$01:D=$:=!:;%+<30;%21)!#D=1!%1!;'0;!&0:=-!
:2+=!;$01:D=$:=!:;%+<30;%21!+%5';!2&&<$!`'%3=!`=!0$=!52%15!<N!;'=!$0+N!B=&0<:=!24!0!121J
X=$2! ;%+=!,=$%D0;%D=!24! ;'=! D23;05=)!>'%:! %:!`'@!N<$=! 3215%;<,%103! :;%+<30;%21! %:!'0$,! ;2! %1J
D=:;%50;=)!
94!`=!%1&$=0:=!;'=!D03<=!24!>!%1!6%5)!"0!01,!5=;!0:!0!$=:<3;!01!%1&$=0:=,!:;%+<30;%21!N$2B0B%3J
%;@-! 42$! %1:;01&=-! %;! %:!'0$,! ;2! ;=33!`'=;'=$! ;'%:! %1&$=0:=! %:!,<=! ;2! 3215=$! 30;=1&@!24!503D01%J"!
:;%+<30;%21-!2$!,<=!;2!;'=!3215=$!2'+%&!&<$$=1;!N'0:=!%1!0!503D01%J.!:;%+<30;%21)!
9;! %:! =0:%=$! ;2! ,=:%51! 01! =MN=$%+=1;! 42$! N<$=! 503D01%J"! 2$! ;$01:D=$:=! :;%+<30;%21)!(=! 213@!
'0D=!;2!0NN3@!0!&'0$5=!%1h=&;%21!;'0;!%:!:'2$;=$!;'01!;'=!;%+=!&21:;01;!24!;'=!$=30M0;%21!&<$J
$=1;)!F2,=$1! =3=&;$21%&:! &01! ,2! ;'0;!`%;'2<;! K03D01%k:! =3=&;$2:;0;%&! 5=1=$0;2$)! 9;!+%5';! B=!
'=3N4<3! ;2!+=0:<$=! 033!N$2N=$;%=:! 24! N<$=! ;$01:D=$:=! :;%+<30;%21! 4%$:;-! B=42$=! ;$@%15! ;2! 010J
3@X=!;'=!:;%+<30;%21!=44=&;!24!;'=!+%M=,!:%5103)!
('=1!`=! ;037! 0B2<;! =3=&;$%&03! :;%+<30;%21! 24! ;'=! $=;%10!`=! ;0&%;3@! 0::<+=! ;'0;! :;%+<30;%21!
+
;07=:!N30&=!%1!0!N$=:&$%B=,!;0$5=;!D23<+=)!91!b*&!.H"Hd!=M0+N3=:!24!:<&'!;0$5=;!D23<+=:!0$=!
:'2`1! 42$! ;'=! &0:=!24!21=! 0&;%D0;=,! =3=&;$2,=! %1! 0! :<BJ$=;%103!N2:%;%21-! 0! &2<1;=$! =3=&;$2,=!
M9+!"&'='>').+&4#+-"%&'%*+"$(1'*#+(1+%+-&'=;*%&'()+-'.)%*+
0;!%14%1%;@-!01,!;'=!D%;$=2<:!$=N30&=,!B@!:%3%&21!2%3)!6%5)!").!%1!b*&!.H"Hd!%:!N$=:=1;=,!'=$=!0:!
6%5)!/)!
!
Z!
!
L%;$=2<:!&0D%;@!4%33=,!`%;'!:%3%&21=!2%3!
!
@'.9+ M9+ 5+ &H"'7%*+ -&'=;*%&'()+ 1'#*0+ "$(0;7#0+ 6H+ ()#+ %7&'F%�+ #*#7&$(0#+ ')+ -;6E$#&')%*+ "(-'&'()9+ J4#+
7(;)&#$+ #*#7&$(0#+ '-+ %&+ ')1')'&HG+ &4#+ F'&$#(;-+ '-+ $#"*%7#0+ 6H+ -'*'7()+ ('*9+ J4#+ #*#7&$'7+ 7;$$#)&+ 1'#*0+ '-+
-4(K)+ ')+ -(*'0+ $#0+ *')#-G+ &4#+ #P;'"(&#)&'%*+ #*#7&$'7+ 1'#*0+ '-+ -4(K)+ ')+ 6$(Q#)+ 6*;#+ *')#-I+ &4#+ 0#)-'&H+ (1+
>'=!0$=0!0$2<1,!E! %:! ;'=! %,=03! ;0$5=;!D23<+=! 42$! ;'=! :;%+<30;%21!24!B%N230$! &=33:!B@!0! 3215%J
6*;#+*')#-+'-+"$("($&'()%*+&(+&4#+-&$#).&4+(1+&4#+7;$$#)&9+J4#+*#&&#$-+5G+R+%)0+/+=%$Q+&H"'7%*+&%$.#&+F(*E
;<,%103! 4%=3,)! 91! ;'=! ;0$5=;! D23<+=! V! 21=! 4%1,:! B%N230$! &=33:! 42$! ;$01:D=$:=! :;%+<30;%21! 2$!
;=#-+ 1($+ &4#+ -&'=;*%&'()+ (1+ 6'"(*%$+ 7#**-G+ 0#)0$'&'7+ 7())#7&'()-G+ ($+ )#;$%*+ )#&K($Q-+ ')+ &4#+ .%).*'()+
,=1,$%;%&! &211=&;%21:)! 91! ;'=! ;0$5=;! D23<+=! R! 21=! =MN=&;:! ;2! 4%1,! 1=<$03! 1=;`2$7:! 24! ;'=!
7#**+*%H#$G+$#-"#7&'F#*H9+
50153%21! &=33! 30@=$)! Q=N=1,%15! 21! ;'=! ;@N=! 24! :;%+<30;%21! 21=! `2<3,! 3%7=! ;2! f0%+g! 0;! :<&'!
;0$5=;! D23<+=:-! %)=)-! 21=!`2<3,! 3%7=! ;2! '0D=! 0! :;$215! :;%+<30;%21! 4%=3,! %1! ;'0;! D23<+=! 01,! 0!
3=::!:;$215!4%=3,!=3:=`'=$=)!
C142$;<10;=3@-!01!=3=&;$%&! &<$$=1;! &0112;!B=! 42&<:=,! 3%7=!0! 30:=$!B=0+)!>'=!=3=&;$%&! &<$$=1;!
4%=3,!'0:!;2!:0;%:4@!;'=!?2%::21!=e<0;%21-!01,!;'=$=!%:!0!121J&$2::%15!$<3=!24!4%=3,!3%1=:_!`'0;!
%:! ;$<=! 42$!=3=&;$%&! &<$$=1;:! %:!03:2! ;$<=! 42$!=3=&;$%&! 4%=3,:)! 94!`=!`01;! ;2!,2! :2+=! 4%=3,! :'0NJ
%15!`=!'0D=!;2!3227!42$!+=01:!;'0;!0$=!&21:%:;=1;!`%;'!;'=:=!$<3=:)!
('=1! 01! =3=&;$%&! 4%=3,! 2$! 0! &<$$=1;! 4%=3,! %:! N$2,<&=,! B@! 01! 0$$0@! 24! =3=&;$2,=:-! =D=$@! 4%=3,!
3%1=!52=:!4$2+!=M0&;3@!21=!24!;'=!012,=:!;2!21=!24!;'=!&0;'2,=:)!E!1=%5'B2$%15!3%1=!0;!%14%1%J
+
;=:%+03!,%:;01&=!52=:! 4$2+! ;'=!:0+=!012,=! ;2! ;'=!:0+=!&0;'2,=)!>'=$=!0$=!0$=0:-!'2`=D=$-!
M98+@'#*0+-4%"').+;-').+-#"%$%&$'7#-+
%1!`'%&'!4%=3,!3%1=:!=+=$5=!4$2+!012;'=$!012,=!01,a2$!52!;2!012;'=$!&0;'2,=)!#D=$@!;`2!24!
:<&'!0$=0:!0$=!:=N0$0;=,!4$2+!=0&'!2;'=$!B@!0!+0;'=+0;%&03!:<$40&=!&033=,!f:=N0$0;$%Mg!bP=!
"SSHd)!>'=!N2:%;%21!01,!:'0N=!24!:<&'!:=N0$0;$%&=:!0$=!,=;=$+%1=,!B@!;'=!5=2+=;$@!24!=3=&J
;$2,=!N2:%;%21:!01,!B@!;'=!0&;%D0;%21!N2;=1;%03:)!91!;'=!42332`%15!`=!`01;!;2!<:=!:<&'!:=N0J
$0;$%&=:!42$!:'0N%15!;'=!=3=&;$%&!&<$$=1;!24!2<$!:;%+<30;%21!:%5103:)!
!
S!
!
E:!01!=M0+N3=!`=!&21:;$<&;!0!:=N0$0;$%M! ;'0;!'0:! ;'=! 42$+!24!0!$%,5=)! 9;! %:! 42$+=,!B@!01!0$J
$015=+=1;! 24! =3=&;$%&! ,%N23=:-! :==! 6%5)! 80)! >'=! ,%$=&;%21! 24! ;'=! ,%N23=:! %:! N=$N=1,%&<30$! ;2!
M9898+A#"%$%&$'B+1($=#0+6H+%)+%$$%).#=#)&+(1+0'"(*#-+
;'=!$%,5=)!>'=!$%,5=!'0:!N=07:!01,!:0,,3=:)!>'=!:;%+<30;%21!=3=&;$2,=! %:!:%;<0;=,! %1! 4$21;!24!
;'=! $%,5=-! %;:! &2<1;=$! =3=&;$2,=! %:! :%;<0;=,! B='%1,! ;'=! $%,5=)! >'=! :<$40&=! 24! ;'=! $%,5=! %:! ;'=!
:=N0$0;$%M!42$!;'=!:;%+<30;%21!&<$$=1;)!>'=!4%=3,!B=;`==1!:;%+<30;%21!=3=&;$2,=!01,!&2<1;=$!
=3=&;$2,=!&0112;!N=1=;$0;=!%1;2!;'=!$%,5=\!%;!'0:!;2!&3%+B!2D=$!%;_!;'=!&<$$=1;!,=1:%;@!`%33!B=!
'%5'! 1=0$! 0! :0,,3=)! E1! =M0+N3=! 24! 4%=3,! 3%1=:! %:! :'2`1! %1! 6%5)! 8B)! >'=! '%5'=:;! N2%1;! 24! ;'=!
:=N0$0;$%M! %:! 0! :0,,3=! %1! ;'%:! 4%5<$=)! >'=! 4%=3,! 3%1=:! %1! &32:=! D%&%1%;@! 24! ;'=! :;%+<30;%15! =3=&J
;$2,=! 0$=! 12;! :'2`1! 42$! N32;;%15! $=0:21:)! >'=! 4%=3,! 3%1=:! B=;`==1! ;'=! =3=&;$2,=:! 24! ;'=! ,%J
N23=:!]:'2`1!%1!B3<=!%1!6%5)!80^!4%33!;'=!:N0&=!B=;`==1!;'=!:=N0$0;$%M!01,!;'=!MJ0M%:-!%)=)-!<1J
,=$1=0;'! ;'=! :=N0$0;$%M)!>'=@!0$=! &21:%,=$=,!B=%15!01!0<M%3%0$@! 4%=3,!01,!0$=!12;! :'2`1! %1!
6%5)!8B)!
!
@'.9+N+5+-#"%$%&'B+&4%&+4%-+&4#+-4%"#+(1+%+$'0.#+1($+&4#+-&'=;*%&'()+7;$$#)&+'-+6;'*&+6H+N+0'"(*#-9+:)+C%D+%+
&("+F'#K+(1+&4#+$#*#F%)&+#*#7&$(0#-+()+&4#+%$$%H+ '-+-##)9+J4#+6*;#+#*#7&$(0#-+%$#+&4#+N+0'"(*#-+"$(0;7E
').+ &4#+ $'0.#G+ &4#+ $#0+ #*#7&$(0#-+ %$#+ &4#+ -&'=;*%&'()+ #*#7&$(0#-9+ :)+ C6D+ &4#+ -&'=;*%&'()+ 1'#*0+ '-+ -##)+ ')+
$#0+ ()+ %+ "*(&+ -7$##)+ F#$&'7%*+ &(+ &4#+ %$$%H+ K'&4+ %+ 6%-#+ *')#+ ')7*;0').+ &4#+ -&'=;*%&'()+ #*#7&$(0#-9+ J4#+
>'=! 0<M%3%0$@! 4%=3,! %:! 12;! :<NN2:=,! ;2! %1;=$4=$=! `%;'! ;'=! :;%+<30;%21)! >'%:! %:! 0&'%=D=,! B@!
6$(Q#)+ *')#+ '-+ &4#+ -#"%$%&$'B9+ J4#+ -#='7'$7*#-+ %$(;)0+ &4#+ -&'=;*%&').+ #*#7&$(0#-+ 4%F#+ )(+ "4H-'7%*+
5%D%15! %;! 012;'=$! ;%+=! N$24%3=)! 62$! N$=,2+%101;3@! 503D01%J"! :;%+<30;%21-! 42$! %1:;01&=-! ;'%:!
=#%)').I+ &4#H+%$#+ ')&$(0;7#0+ 1($+"*(&&').+ $#%-()-9+J4#+ 1'#*0+ *')#-+(1+ &4#+#*#7&$(0#-+ -4(K)+ ')+6*;#+%$#+
&01!B=!,21=!B@!&'22:%15!0! ;%+=!N$24%3=!24! ;'=!0<M%3%0$@!&<$$=1;! ;'0;!@%=3,:!0!N$=,2+%101;3@!
)(&+"*(&	+
503D01%J.! :;%+<30;%21)! 62$! ,%,0&;%&03! :%+N3%&%;@! 3=;! <:! ;07=! 0! 10$$2`! 50<::%01! :'0N=! 42$! ;'=!
:;%+<30;%21! 4%=3,!01,!0!`%,=$!50<::%01!:'0N=! 42$!;'=!0<M%3%0$@! 4%=3,)!>'=!`%,;'!24!;'=!:;%+<J
30;%15! 50<::%01! +%5';! B=! ;'=! 3250$%;'+%&! ,=&$=+=1;! 24! $=30M0;%21_! &2+N0$=! ;2! ;'=! ,%:&<:J
:%21! 24! 6%5)! .)! >'=! `%,;'! 24! ;'=! 50<::%01! N$24%3=! 24! ;'=! 0<M%3%0$@! 4%=3,! +%5';! B=! ;'$==! ;%+=:!
30$5=$)! >'=! &'0$5=! %1h=&;%21! 24! ;'=! 503D01%J"! :;%+<30;%21! %:! D=$@! :+033-! 01,-! 03;'2<5'! ;'=!
!
"H!
!
&'0$5=! %1h=&;%21!24! ;'=!503D01%J.!:;%+<30;%21! %:!:=D=$03! ;%+=:! 30$5=$-! ;'=!0<M%3%0$@!&<$$=1;! %:!
:;%33! ;22!:+033! ;2!&0<:=!'=0;!,0+05=!24! ;'=! ;%::<=)!>'=!$%,5=!B=&2+=:!'%5'=$!`'=1! ;'=!=3=&J
;$2,=:! 42$+%15!;'=!$%,5=!=+%;!+2$=!&<$$=1;)!9;!B=&2+=:! 32`=$!`'=1!;'=@!=+%;! 3=::!&<$$=1;)!
91! 6%5)! 8B! ;'=! :%;<0;%21! %:! :'2`1! 42$! ;'=! &0:=! ;'0;! 033! =3=&;$2,=:! =+%;a$=&=%D=! ;'=! :0+=!
0+2<1;!24!&<$$=1;)+
+
E1! =M0+N3=! 24! <:%15! :=D=$03! =3=&;$2,=:! 24! 01! 0$$0@! 42$! 5<%,%15! 0! :;%+<30;%15! &<$$=1;! ;2! 0!
M989<+O;*&'"(*#-+1($=#0+6H+-&'=;*%&').+#*#7&$(0#-+%)0+1'#*0+-4%"').+#*#7E
N$=:&$%B=,! ;0$5=;! D23<+=! %:! :==1! 03$=0,@! %1! 6%5)! 8)! (=! 1==,! 0;! 3=0:;! .! ,%N23=:-! %)=)-! 8! =3=&J
&$(0#-+
;$2,=:-! ;2! 42$+! 0! 5<%,%15! 4%=3,!N3<:!.! =3=&;$2,=:! 42$! :;%+<30;%21)!>'=! &214%5<$0;%21! %:! :<%;0J
B3=!42$!:;%+<30;%21!%1!;0$5=;!0$=0!V!24!6%5)!/)!>'=!,%:;01&=!24!;'=!;0$5=;!D23<+=!4$2+!;'=!&'%N!
&01! B=! %1&$=0:=,! =%;'=$! B@! %1&$=0:%15! ;'=! :;$=15;'! 24! ;'=! 5<%,%15! 4%=3,! 2$! B@! <:%15! =3=&J
;$2,=:!;'0;!0$=!40$;'=$!0N0$;!4$2+!=0&'!2;'=$)!!
91! 2$,=$! ;2! :;%+<30;=! %1! ;0$5=;! 0$=0! E! 24! 6%5)! /-! `%;'! 0! D=$;%&03! &<$$=1;-! `=! 1==,! 0! &=1;=$!
=3=&;$2,=!=1&%$&3=,!B@!$%,5=!42$+%15!,%N23=:)!(=!:==!:<&'!0!&214%5<$0;%21!%1!6%5)!T\!>'=!&'%N!
&0$$%=:! 0! '=M052103! 5$%,! 24! =3=&;$2,=:! 0:! <:=,! %1! =MN=$%+=1;:! B@! ;'=! E<:;$03%01! 5$2<N! bP2!
.HHTd)! E50%1-! ;'=! =3=&;$2,=:! 42$+%15! ;'=! :=N0$0;$%M! 42$! ;'=! :;%+<30;%21! &<$$=1;! 0$=! :'2`1!
%1!B3<=)!>'=!:;%+<30;%15!=3=&;$2,=!%1!;'=!&=1;=$!01,!;'=!$=;<$1!=3=&;$2,=:!0$=!:'2`1!%1!$=,)!!
!
@'.9+ S9+ 5+ =;*&'"(*#+ 7()-'-&').+ (1+ 8T+ #*#7&$(0#-G+ ')+ %+ 4#B%.()%*+ %$$%H9+ J4#+ =;*&'"(*#+ $#"$#-#)&-+ ()#+
!
"'B#*9+ J4#+ -&'=;*%&').+ 7;$$#)&+ #=#$.#-+ 1$(=+ &4#+ 7#)&#$+ #*#7&$(0#+ %)0+ .(#-+ &(+ &4#+ -'B+ 7(;)&#$+ #*#7E
&$(0#-+ -4(K)+ ')+ $#09+ J4#+ #*#7&$(0#-+ -4(K)+ ')+ 6*;#+ 6;'*0+ ;"+ %+ -#"%$%&$'B+ 1($+ .;'0').+ &4#+ -&'=;*%&'()+
7;$$#)&9+ J4#+ -'.)-+ -4(K)+ %$#+ .'F#)+ 1($+ %)(0'7+ -&'=;*%&'()I+ 1($+ 7%&4(0'7+ -&'=;*%&'()+ &4#+ -'.)-+ 4%F#+ &(+
6#+$#F#$-#09+
!
""!
!
>'=! :;%+<30;%21! &<$$=1;! &01! B=! D%=`=,! 0:! 0! 42<1;0%1-! $%:%15! %1! ;'=!+%,,3=-! ,%D%,%15! <N! 01,!
4033%15! ,2`1! ;2! ;'=! $=;<$1! =3=&;$2,=:_! %1! ;'=! 1<+=$%&03! :%+<30;%21! ]6%5)! TB^-! ;'=! =3=&;$2,=!
'0:!B==1!:<B:;%;<;=,!B@!01!0,=e<0;=!B2<1,0$@!&21,%;%21!21!0!'=+%:N'=$=)!
('0;! '0:! B==1! ,21=! 42$! 0! '=M052103! 5$%,! 24! =3=&;$2,=:! &01! 03:2! B=! ,21=! 42$! 0! &<B%&! 21=! 0:!
:'2`1! %1!6%5)!I)!E50%1-!,%N23=:!0$2<1,!0!&=1;=$!=3=&;$2,=!0$=! 42$+%15!0!$%,5=! 42$!;'=!:;%+<J
30;%15!&<$$=1;)!
!
!
@'.9+U9+5+=;*&'"(*#+7()-'-&').+(1+8T+#*#7&$(0#-G+ ')+%+P;%0$%&'7+%$$%H9+J4#+=;*&'"(*#+$#"$#-#)&-+()#+"'BE
#*9+ J4#+ -&'=;*%&').+ 7;$$#)&+ #=#$.#-+ 1$(=+ &4#+ 7#)&#$+ #*#7&$(0#+ %)0+ .(#-+ &(+ &4#+ -'B+ 7(;)&#$+ #*#7&$(0#-+
-4(K)+ ')+ $#09+J4#+#*#7&$(0#-+ -4(K)+ ')+6*;#+6;'*0+;"+%+ -#"%$%&$'B+ 1($+.;'0').+ &4#+ -&'=;*%&'()+7;$$#)&9+
J4#+ -'.)-+ -4(K)+ %$#+ .'F#)+ 1($+ %)(0'7+ -&'=;*%&'()G+ 1($+ 7%&4(0'7+ -&'=;*%&'()+ &4#+ -'.)-+ 4%F#+ &(+ 6#+ $#E
>'=!%,=0!24!%+05%15!`%;'!01!0$$0@!24!=3=&;$2,=:!%:!;2!<;%3%X=!0:!+01@!N%M=3:!0:!;'=$=!0$=!=3=&J
F#$-#09+
;$2,=:-! 01,! 12;! ;2! <:=! +01@! =3=&;$2,=:! 42$! :'0N%15! 01! =3=&;$%&! &<$$=1;! 4%=3,)! >'=! 30;;=$! %:!
+
M9<+?(7%*+%$#%+-7%))').+-&'=;*%&'()+
,21=! %1!6%5)!8-!01,!+2$=!:2! %1!6%5:)!T!01,!I)! 91! ;'=!N$=:=1;!:<B:=&;%21! %;!`%33!B=!:'2`1!'2`!
;2!5=;!B0&7!;2!;'=!B0:%&!%,=0)!
O=:=0$&'!21! $=;%103! %+N301;:! %:!&0$$%=,!B@! ;'=!'2N=! ;'0;! ;'=!N'2:N'=1=:!N$2,<&=,!B@!=3=&J
;$%&03! :;%+<30;%21! 0$=! :+033! 01,! ;'<:! &01! B=! <:=,! 0:! N%&;<$=! =3=+=1;:! &033=,! fN%M=3:g)! 9;!
`2<3,!B=!%,=03!;2!'0D=!0:!+01@!N%M=3:!0:!;'=$=!0$=!=3=&;$2,=:!21!;'=!&'%N)!
91! 6%5)! T0! 01,! %1! 6%5)! I!`=! 0$=! <:%15! "S! 01,! .T! =3=&;$2,=:-! $=:N=&;%D=3@-! 42$! N$2,<&%15! 213@!
21=! N%M=3)! >'=! &=1;=$! =3=&;$2,=! %:! <:=,! 42$! ;$01:+%;;%15! ;'=! :;%+<30;%21! :%5103! 01,! .8! =3=&J
;$2,=:!0$=!<:=,!N0$;3@!0:!&2<1;=$!=3=&;$2,=:!01,!N0$;3@!42$!:'0N%15!;'=!:;%+<30;%21!4%=3,)!E;!
4%$:;!:%5';!`=!:==+!;2!B=!<:%15!;22!+01@!=3=&;$2,=:!42$!N$2,<&%15!21=!:%153=!N%M=3)!
Q<$%15!;'=!2N=$0;%21!24!0!$=;%103! %+N301;!`=!,2!12;!`01;!;2!;$01:+%;!213@!21=!:%153=! %+05=!
B<;! 2BD%2<:3@! 0! D%,=2! :=e<=1&=! 24! %+05=:)! (=! '0D=! ;2! 3227! 0;! ;'=! ;%+=! :&03=)! >@N%&033@-! 0!
D%,=2! :=e<=1&=! '0:! .T! 4$0+=:! N=$! :=&21,)! >'%:!+=01:! ;'0;! ;'=$=! 0$=! 8H!+%33%:=&21,:! ;%+=!
42$! ;$01:+%;;%15! 21=! 4$0+=)! >'%:! %:! +2$=! ;%+=! ;'01! 1==,=,! 42$! &2+N3=;%15! 01! =3=&;$%&03!
:;%+<30;%21!N$2&=::)!
!
".!
!
62$! 3215%;<,%103! :;%+<30;%21! ;'=! ;@N%&03! 3=15;'! 24! 0! +212N'0:%&! D23;05=! N<3:=! %:! H)T! +%33%J
:=&21,:-! 0:! %1! 6%5)! "0)! >'=! $=:<3;%15! B%N'0:%&! &<$$=1;! N<3:=! ,=&0@:! ;2! 0! &<$$=1;! ,=1:%;@! B=J
32`! :;%+<30;%21! ;'$=:'23,! 04;=$! 0B2<;! H)Z! +%33%:=&21,:-! :==! 6%5)! "B)! >'=! +=+B$01=! 1==,:!
:2+=! ;%+=! ;2!$=:N21,)! 94!`=!5%D=! ;'=!:0+=!0+2<1;!24! ;%+=! 42$! ;'0;-! ;'=!:;%+<30;%21!N$2&=::!
:'2<3,! B=! &2+N3=;=,! 04;=$! 0B2<;! ")I! +%33%:=&21,:)! 62$! ;$01:D=$:=! :;%+<30;%21! ;'=! :;%+<30J
;%21! N$2&=::! 1==,:! 0B2<;! 0! ,&35.*63.%?! 42$! &'0$5=! %1h=&;%21! 01,! 0! $=:;! N=$%2,! 24! <N! ;2! "!
+%33%:=&21,)! >'<:! `=! 1==,! 0! ;%+=! `%1,2`! 24! ")I! +%33%:=&21,:! 42$! &2+N3=;%15! 0! N$2&=::! 24!
3215%;<,%103! :;%+<30;%21! 01,! 0! 3%;;3=! 3=::! 42$! ;$01:D=$:=! :;%+<30;%21)! >'%:! +=01:! ;'0;! `=!
'0D=!0;!3=0:;!.T!;%+=!:3%&=:!42$!=D=$@!21=!24!;'=!8H!+%33%:=&21,!4$0+=:!24!;'=!D%,=2!:=e<=1&=)!
E4;=$! =0&'! 21=! 24! ;'=:=! .T! ;%+=! :3%&=:! ;'=! 0332&0;%21! 24! =3=&;$2,=:! ;2! +<3;%N23=:! &01! B=!
&'015=,)! #M&=N;! 42$! ;'=! B2<1,0$@! 24! ;'=! 0$$0@-! =D=$@! =3=&;$2,=! &01! B=&2+=! ;'=! &=1;=$! 24! 0!
+<3;%N23=!3%7=!;'=!21=:!:'2`1!%1!6%5:)!T0!01,!I)!>'=!&'015=!24!0332&0;%21!%:!%33<:;$0;=,!%1!6%5)!
G-! 0:! 01! =M0+N3=)! 91! ;'%:! `0@! `=! $=&2D=$! ;'=! 4<33! $=:23<;%21! 24! ;'=! =3=&;$2,=! 0$$0@-! =M&=N;!
42$!0!&=$;0%1! 32::!0;!;'=!B2<1,0$@!24!;'=!0$$0@)!>'%:! 32::!03215!;'=!&%$&<+4=$=1&=!24!01!=3=&J
;$2,=! 0$$0@! `2<3,! 40D2$! &%$&<30$! 2D=$! $=&;015<30$! =3=&;$2,=! 0$$015=+=1;:-! `'%&'! `2<3,!
03:2! &2142$+! B=;;=$! ;2! ;'=! 4<1,<:)! 91! &0:=! 24! ;'=!+<3;%N23=! :'2`1! %1! 6%5)! I! ;'=! h<+N%15! :=J
e<=1&=!+0@!B=!$2`!B@!$2`-!07%1!;2!0!;=3=D%:%21!:&$==1)!
!
!
@'.9+V9+ WB%="*#+ (1+ *(7%*+ %$#%+ -7%))').+K'&4+=;*&'"(*#-+ (1+ &4#+ &H"#+ -4(K)+ ')+ @'.9+ S%9+J4#+ 7#)&#$-+ (1+ &4#+
=;*&'"(*#-+ X;="+%1&#$+#%74+&'=#+-*'7#+&(+%+)#K+"(-'&'()+()+&4#+4#B%.()%*+.$'0G+%-+-4(K)+6H+%$$(K-+1($+
&4#+ =;*&'"(*#+ ')+ &4#+ 7#)&#$9+ J4#+ -#P;#)7#+ (1+ X;="').+ '-+ '$$#*#F%)&+ %-+ *().+ %-+ '&+ 7(F#$-+ %**+ #*#7&$(0#-+
*2! 40$! `=! ,%,! 12;! &21:%,=$! ;'=! &0:=! %1! `'%&'! ;'=! ,=:%$=,! :;%+<30;%21! D23<+=! %:! :%;<0;=,!
K'&4')+&4#+&'=#+6#&K##)+&K(+7()-#7;&'F#+F'0#(+1$%=#-9+
$%5';!21!;2N!24!;'=!=3=&;$2,=-!'0:!0!,%0+=;=$!12;!30$5=$!;'01!;'=!=3=&;$2,=!01,!0!;'%&71=::!24!
+
M9M+J$%)-F#$-#G+=()("(*%$+-&'=;*%&'()+K'&4+%+$#=(&#+7(;)&#$+#*#7&$(0#+
3=::!;'01!"HH!µ+)! 91! ;'%:!&0:=-!<1%N230$!:;%+<30;%21!`%;'!21=! 30$5=!$=+2;=!$=;<$1!=3=&;$2,=!
!
"/!
!
:==+:! ;2! B=! 4=0:%B3=-! N$2D%,=,! ;'0;! 503D01%J"! :;%+<30;%21! 0&&2$,%15! ;2! *=&;%21! .).! %:! =+J
N32@=,)!
E50%1! ;'=! &$%;=$%0! 42$! 0! 503D01%J"! :;%+<30;%21! 0$=! 0:! :;0;=,! 0B2D=\! ]"^! 01! =M;$=+=3@! :'2$;!
&'0$5=! %1h=&;%21! ]%1! ;'=! $015=! 24! 1012:=&21,:-! <N! ;2! 0! 4=`!+%&$2:=&21,:^-! ].^! 0! D=$@! :+033!
0+2<1;! 24! %1h=&;=,! &'0$5=! ]24! 3=::! ;'01! "! 1R^! 01,! ]/^! 0! $0;'=$! 3215! $=:;! N=$%2,! ]24! 0B2<;! "!
+%33%:=&21,^!04;=$!=0&'!&'0$5=! %1h=&;%21)!>'=!2&&<$$=1&=!24!&$2::J;037!,<$%15!;'=!:'2$;!;%+=!
24! &'0$5=! %1h=&;%21! &01! =0:%3@! B=! 0D2%,=,)! *2+=! &$2::J;037! +0@! 0$%:=! `'=1! 0! :=&21,! =3=&J
;$2,=! %1!&32:=!1=%5'B2$'22,! 4%$=:!`'%3=!;'=! 4%$:;!=3=&;$2,=! %:!:;%33! %1! %;:!$=:;!N=$%2,)!>'%:!&01!
B=! 0D2%,=,! B@! $01,2+%X%15! ;'=! 4%$%15! 2$,=$-! 2$-! %1! &2+N<;=$J&21;$233=,! 0NN3%&0;%21:-! %;! &01!
B=!0D2%,=,!B@!0!&2+N<;=$!0352$%;'+!;'0;!52D=$1:!;'=!:;%+<30;%21!:=e<=1&=)!
!
91! ;'%:! *=&;%21! /)8! 0! :;2&'0:;%&! 2N;%+%X0;%21! 4$0+=`2$7! %:! ,%:&<::=,! ;'0;! &01! B=! =+N32@=,!
M9N+ A"%&'(E&#="($%*+ ("&'='>%&'()+ (1+ #*#7&$(0#+ %$$%H+ 0#-'.)+ %)0+ #*#7&$'7+
42$! ;'=! :N0;%2J;=+N2$03! 2N;%+%X0;%21! 24! ;'=! =3=&;$2,=! 0$$0@! ,=:%51! ]%)=)-! ;'=! :N0;%03! 0$J
1'#*0+-4%"').+F'%+-&(74%-&'7+("&'='>%&'()+1$%=#K($Q+
$015=+=1;:! 24! ;'=! =3=&;$2,=:! 21! ;'=! &'%N^! 01,! ;'=! N$2&=::! 24! =3=&;$%&! 4%=3,! :'0N%15)! >'=!
5203:!0$=\!]0^!;2!0&'%=D=!2N;%+03!4%=3,!:'0N%15!`%;'!0!5%D=1!=3=&;$2,=!0$$0@!]%)=)-!&'%N^-!]B^!;2!
2N;%+033@! ,=:%51! ;'=! =3=&;$2,=! 0$$0@! %;:=34-! %)=)-! ;2! ,=;=$+%1=! ;'=! :N0;%03! 0$$015=+=1;! 01,!
:%X=a,%0+=;=$!24!;'=!=3=&;$2,=:!21!;'=!=3=&;$2,=!0$$0@-!01,!]&^!;2!2N;%+033@!,$%D=!;'=!:;%+<J
30;%21! ;'$2<5'! 0! 5%D=1-! %+N301;=,! =3=&;$2,=! 0$$0@! %1! $=03! ;%+=! ,<$%15! <:05=! 24! ;'=! $=;%103!
%+N301;)!
91! .HHZ! 21=! 24! <:! ](6^! '0:! %1;$2,<&=,! 0! *;2&'0:;%&! WN;%+%X0;%21! 6$0+=`2$7! ]*W6^! b6%!
.HHZ0d! ;'0;! 0332`:! 2N;%+%X%15! 0! :@:;=+! 2$! N$2&=::! ;'0;! %:! 52D=$1=,! B@! 1<+=$2<:! 0,h<:;0J
B3=! N0$0+=;=$:)! >'=! <1,=$3@%15! N$%1&%N3=! %:! ;'=! +%1%+%X0;%21! 24! 0! 4%;1=::! 4<1&;%21! ;'0;!
+=0:<$=:!;'=!,%44=$=1&=!B=;`==1!0!,=:%$=,!2<;&2+=!01,!01!0&;<03!2<;&2+=!`'%3=!2N=$0;%15!
;'=! :@:;=+! 2$! N=$42$+%15! ;'=! N$2&=::)! >'=! +%1%+%X0;%21! 24! ;'%:! 4%;1=::! 4<1&;%21! %:! 0&&2+J
N3%:'=,! D%0! +<3;%J,%+=1:%2103! :;2&'0:;%&! 2N;%+%X0;%21! 0352$%;'+:-! :<&'! 0:! *%+<30;=,! E1J
1=03%15! bF=! "ST/_! [%! "SZ/d-! K=1=;%&! E352$%;'+:! bK2! "SZSd-! 01,! 2;'=$! #D23<;%210$@! E352J
$%;'+:)!>'=!&2++21!&'0$0&;=$%:;%&!24!;'=:=!0352$%;'+:!%:!;'=!&0N0B%3%;@!;2!=:&0N=!32&03!+%1J
%+0! 24! 0! 4%;1=::! 4<1&;%21! `'%3=! 0NN$2M%+0;%15! ]01,! %,=033@! $=0&'%15^! ;'=! 532B03! +%1%+<+)!
>'%:!&'0$0&;=$%:;%&! %:! %1!:;0$7!&21;$0:;! ;2!,=;=$+%1%:;%&-!5$0,%=1;J,=:&=1;JB0:=,!0352$%;'+:-!
:<&'! 0:! ;'=! F0$e<0$,;JP=D=1B=$5! 0352$%;'+! b?$! "SS"d-! `'%&'! ;=1,! ;2! B=! N22$! N=$42$+=$:!
%1!+<3;%J,%+=1:%2103! 301,:&0N=:! ;'0;! =M'%B%;!+<3;%N3=! 32&03!+%1%+0! 01,! 0$=! 32&033@! $<55=,)!
>'=!:;2&'0:;%&!2N;%+%X0;%21!0352$%;'+:!<:=!0:!;'=%$!%1N<;!;'=!4%;1=::!4<1&;%21!D03<=!42$!=0&'!
%;=$0;%21! 01,! 5=1=$0;=! 0! 1=`-!+2,%4%=,! :=;! 24! N0$0+=;=$! D03<=:! 0:! 0! $=:<3;-! <3;%+0;=3@! &21J
D=$5%15! ;2!0!:=;!24!N0$0+=;=$!D03<=:! ;'0;!@%=3,:! ;'=!,=:%$=,! 4%;1=::! 4<1&;%21!D03<=-! %)=)-!&32:=!
;2!X=$2)!
>'=!*;2&'0:;%&!WN;%+%X0;%21!6$0+=`2$7!'0:!B==1!0NN3%=,! ;2! ;'=!2N;%+%X0;%21!24!N$2:;'=;%&!
D%:%21-! %1! N0$;%&<30$! N$2D%,=,! B@! =N%J$=;%103! %+N301;:! b6%! .HHZB-! 6%! .H"H0d-! ;2! +01%N<30;=!
=3=&;$%&! :;%+<30;%21!N0$0+=;=$:!24!01! %+N301;=,!=N%J$=;%103!=3=&;$2,=!0$$0@! ;2!2N;%+%X=! ;'=!
$=:<3;%15!D%:<03!N=$&=N;%21!24!;'=!B3%1,!:<Bh=&;)!
!
!
"8!
!
91!0!:%+%30$!+011=$-!;'=!*W6!&01!B=!0NN3%=,!;2!;'=!42332`%15!2N;%+%X0;%21!:&=10$%2:\!
E!5%D=1!=3=&;$2,=!0$$0@!+=01:! ;'0;! ;'=!:N0;%03! 32&0;%21!01,!:N=&%4%&0;%21:! 42$!=0&'!=3=&J
;$2,=!]=)5)-!,%0+=;=$-!=3=&;$2,=!+0;=$%03^!0$=!4%M=,)!>'=!*W6!&01!12`!B=!=+N32@=,!;2!2NJ
;%+%X=! ;'=! &<$$=1;! ;'$2<5'! ;'=! D0$%2<:! =3=&;$2,=:! :<&'! ;'0;! ;'=! $=:<3;%15! /Q! :'0N=! 24!
!- /012#34%52647%+83029:%;218%3%:2<69%646=1>?76%3>>3@%A2-6-,%=820BC%
;'=! =3=&;$%&! 4%=3,! 0NN$2M%+0;=:! `%;'%1! <:=$J,=4%1=,! ;23=$01&=:! ;'=! ,=:%$=,! 21=)! >'=! /Q!
:'0N=! 24! ;'=! =3=&;$%&! 4%=3,! &01! B=! 0&&<$0;=3@! :%+<30;=,! 0:! ,=:&$%B=,! %1! b*&! .H"Hd)! >'=!
<1,=$3@%15! =3=&;$2:;0;%&! +2,=3! 24! ;'=! =3=&;$2,=! 0$$0@! `%33! ,=3%D=$! ;'=! 4%;1=::! 4<1&;%21!
1=&=::0$@!42$!;'=!*W6)!
>'%:!:&=10$%2!%:!01!=M;=1:%21!24!;'=!4%$:;!21=!0B2D=\!c=$=!;'=!&21:;$0%1;!24!'0D%15!0!:N0J
;%033@! 4%M=,! =3=&;$2,=! 0$$015=+=1;! %:! $=30M=,-! 01,! ;'=! *W6! %:! =+N32@=,! ;`%&=\! ]"^! %1! 01!
2<;=$! 2N;%+%X0;%21! 322N-! 2N;%+%X%15! ;'=! :N0;%03! 32&0;%21a0$$015=+=1;! 01,! =D=1! ;'=!
&- /012#34%646=1>?76%3>>3@%76+2:9C%
:N=&%4%&0;%21:!42$!=0&'!=3=&;$2,=!]=)5)-!=3=&;$2,=!,%0+=;=$^-!01,!].^!%1!01!%11=$!2N;%+%X0J
;%21! 322N-! N=$42$+%15! =M0&;3@! ;'=! =3=&;$%&! &<$$=1;! 2N;%+%X0;%21! 0:! ,=:&$%B=,! %1! :&=10$%2!
")!>'=!$=:<3;!`%33!B=!01!2N;%+%X=,!=3=&;$2,=!0$$0@!,=:%51)!
C:%15! 0! :<44%&%=1;3@! &0N0B3=! +%1%0;<$%X=,! &2+N<;%15! :@:;=+! ]=)5)-! b6%! .H"HBd^! 21=! &01!
N=$42$+!;'=!*W6JB0:=,!2N;%+%X0;%21!,=:&$%B=,!%1!:&=10$%2!"!%1!$=03!;%+=!]%)=)-!,<$%15!0&J
;<03! <:05=! 24! ;'=! D%:%21! N$2:;'=:%:^! 5%D=1! ;'=! 4%M=,! 5=2+=;$@! 01,! :N=&%4%&0;%21:! 24! ;'=!
D- E634$12#6%+12#F4312?9%?012#2G312?9C%
%+N301;=,!=3=&;$2,=!0$$0@!01,!;'=!=3=&;$2:;0;%&!+2,=3!24!;'=!=3=&;$2,=!0$$0@!;2!5=1=$0;=!
;'=! 4%;1=::! 4<1&;%21! 42$! ;'=!*W6)!(%;'!N$%2$! %+05=!N$2&=::%15!24! ;'=!&0+=$0! %+05=:! ;'0;!
4==,! %1;2! ;'=! 0$;%4%&%03! D%:%21! %+N301;:! b6%! .H"HBd! ]=:N=&%033@! $=;%103! %+N301;:^! 21=! &01!
=:;%+0;=! `'0;! ;'=! ,=:%$=,! /Q! :'0N=! 24! ;'=! =3=&;$%&! 4%=3,! 42$! =0&'! %1,%D%,<03! N$2&=::=,!
&0+=$0!4$0+=!:'2<3,!B=)!>'%:!+07=:!0!&2+N0$%:21!24!;'=!$=:<3;%15!/Q!:'0N=!24!;'=!=3=&J
;$%&! 4%=3,!0&$2::! ;'=!=3=&;$2,=! 0$$0@!,<$%15! ;'=!*W6JB0:=,!2N;%+%X0;%21!N$2&=::!01,! ;'=!
,=:%$=,!/Q!:'0N=!N2::%B3=-!;'<:!,=3%D=$%15!;'=!4%;1=::!4<1&;%21)!l2;=-!;'0;!033!24!;'%:!'0NJ
N=1:! N$%2$! ;2! ;'=! 0&;<03! =3=&;$%&! :;%+<30;%21! 24! =0&'! &0+=$0! 4$0+=! D%0! ;'=! =3=&;$2,=! 0$J
$0@\! 213@! ;'=! $=:N=&;%D=! &<$$=1;! N$24%3=:! 24! 033! =3=&;$2,=:! <1,=$3@%15! ;'=! :<44%&%=1;3@!
&21D=$5=,!/Q!:'0N=!24!;'=!$=:<3;%15!=3=&;$%&!4%=3,!&2$$=:N21,%15!;2!;'=!%+05=JN$2&=::=,!
&0+=$0!4$0+=!`%33!B=!:;%+<30;=,!B=42$=!;'=!1=M;!&0+=$0!4$0+=!%:!N$2&=::=,!0&&2$,%153@)!
E:! ;'=! ,=;0%3:! 24! ;'=:=! 2N;%+%X0;%21! :&=10$%2:! 0$=! B=@21,! ;'=! :&2N=! 24! ;'%:!`2$7-! ;'=@!`%33!
B=!:<Bh=&;!;2!0!42$;'&2+%15!N0N=$!%1&3<,%15!2N;%+%X0;%21!=M0+N3=:)!
!
>'$==!,%44=$=1;!+=&'01%:+:!24!=3=&;$%&03!:;%+<30;%21!24!;'=!$=;%10!'0D=!B==1!,%:&<::=,)!
>'=! ;&5*+! 21=! %:! &033=,! 3215%;<,%103! :;%+<30;%21! 2$! 503D01%J.! :;%+<30;%21)! 9;! 0$%:=:! `'=1! 01!
N9+A;==%$H+%)0+0'-7;--'()++
2'+%&!&<$$=1;!N$2,<&=:!0!D23;05=!,$2N!03215!%;:!&2<$:=!;'$2<5'!;'=!;%::<=!24!;'=!$=;%10)!>'=!
&2+N21=1;! 24! ;'=! D23;05=! ,$2N! N0$033=3! ;2! 0! 121J+@=3%10;=,! 0M21! 2$! ,=1,$%;=! =1;=$:!
;'$2<5'! ;'=!&=33!+=+B$01=! %1;2! ;'=!&=33!0:!,=:&$%B=,!B@! ;'=!&0B3=!=e<0;%21)!>'%:!N$2&=::! %:!
`=33! 712`1! 01,! <:=,! %1! +01@! 0NN3%&0;%21:)! 9;! &2$$=:N21,:! ;2! ;'=! :=&21,! 4$25! 3=5! =MN=$%J
+=1;!B@!P<%5%!K03D01%)!
!
"T!
!
>'=! *63.%?! +=&'01%:+! %:! &033=,! ;$01:D=$:=! :;%+<30;%21! 2$! 503D01%J"! :;%+<30;%21)! 9;! 0$%:=:!
`'=1! 0! D=$@! :+033! =3=&;$%&03! &'0$5=! 0NN=0$:! :<,,=13@! 0;! ;'=! :<$40&=! 24! 01! =3=&;$2,=! 01,!
:=1,:! 0! ,%:N30&=+=1;! &<$$=1;! ;'$2<5'! ;'=! ;%::<=)! ?230$%X%15! &'0$5=:! 0NN=0$! 0;! ;'=! :<$40&=!
24!=D=$@!%1:<30;2$-!01,!03:2!0;!;'=!:<$40&=!24!&=33!+=+B$01=:)!*%1&=!;'=!+=+B$01=:!0$=!:<$J
$2<1,=,! B@! 01! =3=&;$23@;=! ;'=! :<$40&=! &'0$5=:! ,%:0NN=0$! B@! $=30M0;%21)! Q<$%15! ;'=! ;%+=! 24!
$=30M0;%21! ;'=!+=+B$01=! 24! 0! 1=$D=! &=33! &01! $=0&;! B@! 2N=1%15! %21! &'011=3:! &0<:%15! 0&;%21!
N2;=1;%03:)! (=! &033! ;'%:! N$2&=::! ;$01:D=$:=! :;%+<30;%21! B=&0<:=! ;'=! 4%=3,! &2+N21=1;! N=$J
N=1,%&<30$!;2!;'=!&=33!+=+B$01=!%:!&0<:%15!;'=!N230$%X0;%21_!B=&0<:=!24!;'=!:'2$;1=::!24!;'=!
=3=&;$%&! N<3:=!`=! $=4=$! ;2! ;'%:! N$2&=::! 03:2! 0:! 430:'! 2$! :'2&7! :;%+<30;%21)! 9;! &2$$=:N21,:! ;2!
;'=!4%$:;!4$25!3=5!=MN=$%+=1;!B@!P<%5%!K03D01%)!
>'=! +4&5?! +=&'01%:+! %:! :;2&'0:;%&! :;%+<30;%21! B@! 0! D=$@! :+033! =3=&;$%&! &<$$=1;)! *@10N;%&!
12%:=! %1! 0! 1=<$03! 1=;`2$7! ,2=:! 12;! &0$$@! %142$+0;%21)! *%5103:! ;$0D=3! :;0;%:;%&033@! B=;`==1!
1=<$21:! `%;'2<;! ,%$=&;%2103! N$=4=$=1&=)! E! :+033! =3=&;$%&! &<$$=1;! &01! %+N2:=! 0! ,%$=&;%2103!
N$=4=$=1&=! 01,! 3=0,! ;2! 0! 1=;! :%5103)! 9;! +0@! &0<:=! ;'=! N=$&=N;%21! 24! N'2:N'=1=:)! W1=! ,2=:!
12;! =MN=&;! :<&'! N'2:N'=1=:! ;2! B=! :+033! %1! :%X=! 01,! :<%;0B3=! 42$! %+05%15)! >'=$=42$=! 21=!
:'2<3,!;$@!;2!0D2%,!;'%:!+=&'01%:+)!
9;!`2<3,! B=! %,=03! ;2! f42&<:g! ;'=! :;%+<30;%21! &<$$=1;! ;2! 0! :+033! ;0$5=;! D23<+=! 01,! ;2! :;%+<J
30;=! 1=<$21:! 24! ;'=! $=;%10! 213@! %1! ;'%:! D23<+=)! C142$;<10;=3@-! 01! =3=&;$%&! &<$$=1;! '0:! 0! f42J
&<:g! 213@! 0;! :2<$&=:! 2$! :%17:)!c2`=D=$-! :2J&033=,! :=N0$0;$%&=:! &01! B=! <:=,! %1! 2$,=$! ;2! &21J
,<&;!:2+=!4%=3,!:'0N%15)!9;!'0:!B==1!,%:&<::=,!'2`!0!:=N0$0;$%M!&01!B=!<:=,!;2!&$=0;=!:2+=J
;'%15! 3%7=!0!+2<1;0%1!N0::\!E1!0<M%3%0$@! 4%=3,!&$=0;=:!0!$%,5=!`%;'!:<++%;:!01,!N0::=:)!>'=!
:;%+<30;%21! &<$$=1;! &0112;! N=1=;$0;=! %1;2! ;'=! 0<M%3%0$@! 4%=3,-! B=&0<:=! 24! ;'=! :=N0$0;$%M! B=J
;`==1! ;'=! ;`2! 4%=3,:)! 9;! '0:! ;2! &3%+B! 2D=$! ;'=! $%,5=-! N$=4=$0B3@! 2D=$! 0! N0::)! E;! 0! N0::! ;'=!
4%=3,!3%1=:!0$=!+2$=!,=1:=3@!N0&7=,)!>'%:!+0@!B=!&21:%,=$=,!0!:<B:;%;<;=!42$!f42&<:%15g)!>'=!
,$0`B0&7! %:! ;'0;! 21=! 1==,:! :=D=$03! =3=&;$2,=:! 24! ;'=! %+N301;=,! =3=&;$2,=! 0$$0@! ]&'%N^-! %1!
0,,%;%21! ;2! ;'=! :;%+<30;%15! =3=&;$2,=-! 42$! ;'%:!7%1,!24! 4%=3,! :'0N%15)!E;! 4%$:;! :%5';-!21=! :0&$%J
4%&=:!+01@!=3=&;$2,=:!42$!&$=0;%15!%+05%15!N%M=3:-!`'%&'!+=01:!322:%15!%+05=!$=:23<;%21)!
9;! '0:! B==1! ,%:&<::=,! '2`! %+05=! 0&<%;@! &01! B=! $=&2D=$=,! D%0! 32&03! 0$=0! :&011%15)! W1=!
712`:!'2`!+<&'! ;%+=! %:!1==,=,! 42$! &2++<1%&0;%15! 21=! 4$0+=! 24! ;'=! D%,=2! :=e<=1&=-! 01,!
21=! 712`:! '2`! +<&'! ;%+=! %:! 1==,=,! 42$! :;%+<30;%15! 24! 0! 1=$D=! &=33)! 9;! ;<$1:! 2<;! ;'0;! ;'=!
;%+=! 1==,=,! 42$! 21=! D%,=2! 4$0+=! &01! B=! ,%D%,=,! <N! %1;2! ;%+=! :3%&=:)! 91! =D=$@! 21=! 24! ;'=:=!
;%+=! :3%&=:! ;'=! +<3;%N23=! &01! &'015=! N2:%;%21! 21! ;'=! &'%N)! 91! 21=! &0:=! ;'=$=! 0$=! .T! =3=&J
;$2,=:!1==,=,! 42$! &$=0;%15!0! :;%+<30;%15! &<$$=1;!N3<:!5<%,%15! &<$$=1;:)!(%;'!.T! ;%+=! :3%&=:!
;'=! &=1;=$! 24! ;'=! .TJN23=! &01!+2D=! 2D=$! .T! N2:%;%21:! 01,! N$2,<&=! .T! N%M=3:-! ;'<:! $=&2D=$J
%15! ;'=! 4<33! %+05=! $=:23<;%21_! :2+=! 32::! 0$%:=:! 213@! 0;! ;'=! &%$&<+4=$=1&=! 24! ;'=! =3=&;$2,=!
0$$0@)!
>'=!2N=$0;%2103!,=:%51!24!0!$=;%103!D%:%21!%+N301;!B=&2+=:!:%+N3=!`'=1!;'=!;0$5=;!D23<+=:!
0$=! 32&0;=,!,%$=&;3@!0B2D=!=0&'!=3=&;$2,=!01,!`'=1!503D01%J"!:;%+<30;%21! %:!=+N32@=,)!>'=!
213@! ,%44%&<3;@! %1! ;'%:! &0:=! %:! 0! ;=&'1%&03! 21=\! W1=! '0:! ;2! $01,2+%X=! ;'=! ;%+=! :=e<=1&=! 24!
:;%+<30;%21! 01,! ;'=! &%$&<%;$@! '0:! ;2! B=! &0N0B3=! 24! N$2,<&%15! ;'=! D=$@! :'2$;! D23;05=! :<$5=!
1==,=,!42$!;'=!503D01%J"!:;%+<30;%21)!
9;! :'2<3,! B=! =+N'0:%X=,! ;'0;! ;'=! ,=:&$%B=,! :;%+<30;%21! +2,03%;%=:! ]%)=)-! 503D01%J"! 01,! 530J
D01%J.!:;%+<30;%21:^!01,! ;'=!=3=&;$%&! 4%=3,!:'0N%15! ;=&'1%e<=!D%0!:=N0$0;%&=:!0$=!N2;=1;%033@!
!
"I!
!
0NN3%&0B3=! ;2! ;'=! :;%+<30;%21!24!2;'=$! ;%::<=a1=$D=! $=5%21:-!=)5)-!2N;%&!1=$D=-! 30;=$03!5=1%&<J
30;=!1<&3=<:!]PKl^-!D%:<03!&2$;=M-!,==N!B$0%1-!01,!N0$03@X=,!3%+B:)!
57Q)(K*#0.#=#)&Y+6$<%;4<3!,%:&<::%21:!`%;'!O2B=$;!(%37=!0$=!5$0;=4<33@!0&712`3=,5=,)!
!
,'-7*(-;$#-Y+ E<;'2$:! #(*! 01,! (6! +0@! '0D=! N$2N$%=;0$@! %1;=$=:;! %1! ;'=! :;%+<30;%21! +2J
+
,03%;%=:!N$=:=1;=,!%1!;'%:!+01<:&$%N;!0:!N$2D%:%2103!N0;=1;:!'0D=!B==1!%::<=,)!
!
bR'! "SS/d! E)! o)! R'2`-! #3=&;$%&03! :;%+<30;%21! 24! ;'=! $0BB%;! $=;%10!`%;'! :<B$=;%103! =3=&;$2,=:!
%+N301;:)! 91D=:;! WN';03+23! L%:! *&%!
01,! '%5'! ,=1:%;@! +%&$2N'2;2,%2,=! 0$$0@!
2#1#$#)7#-+
]*<NN3^"SS/_!/8\!Z/T!
!bR'! .HH8d! E)! o)! R'2`-! L)! o)! R'2`-! [)c)! ?0&72-! p)*)! ?2330&7-! K)E)! ?=@+01-! O)! *&'<&'0$,! .HH8!
>'=! 0$;%4%&%03! :%3%&21! $=;%10!+%&$2&'%N! 42$! ;'=! ;$=0;+=1;!24!D%:%21! 32::! 4$2+! $=;%1%;%:!N%5J
+=1;2:0)!E$&')!WN';'03+23)!8<<]8^\!8IHJ8IS!
!b6%! .HHZ0d! ()! 6%17_! *;2&'0:;%&! WN;%+%X0;%21! 6$0+=`2$7! ]*W6^! 42$! R2+N<;=$JWN;%+%X=,!
Q=:%51-! #15%1==$%15-! 01,! ?=$42$+01&=! 24! F<3;%JQ%+=1:%2103! *@:;=+:! 01,! ?$2&=::=:_!
?$2&)!*?9#-!L23)!ISIH-!ISIHHl!].HHZ^_!QW9\"H)"""Ga".)GZ888H!]%1D%;=,!N0N=$^!
!b6%!.HHZBd!6%17!(-!>0$B=33!FE_!*;2&'0:;%&!WN;%+%X0;%21!6$0+=`2$7!42$!;'=!WN;%+%X0;%21!24!
?$2:;'=;%&! L%:%21_! EOLW! ]E::2&%0;%21! 42$! O=:=0$&'! %1! L%:%21! 01,! WN';'03+2325@^! .HHZ!
R214=$=1&=-! 6;)! P0<,=$,03=-! 632$%,0-! 91D=:;)! WN';'03+23)! L%:)! *&%)! .HHZ! 8S\! #JEB:;$0&;!
"GGS-!0B:;$0&;!01,!N2:;=$!
!b6%! .H"H0d! 6%17!(-!>0$B=33!FE_!?0;%=1;J%1J;'=J322N!WN;%+%X0;%21! 24!?$2:;'=;%&!L%:%21_!l=<J
$03!91;=$40&=:!R214=$=1&=-!P215!V=0&'-!RE-!p<1=!."J./-!.H"H!
!b6%!.H"HBd!6%17!(-!o2<!Rq-!>0$B=33!FE_!rEL*.\!F%&$2&2+N<;=$JB0:=,!E$;%4%&%03!L%:%21!*<NJ
N2$;!*@:;=+!42$!O=03J>%+=!9+05=!?$2&=::%15!42$!R0+=$0JQ$%D=1!L%:<03!?$2:;'=:=:_!p)!V%J
2+=,)!WN;)-!L23)!"T-!H"IH"/!].H"H^_!,2%\"H)"""Ga")/.S.H".!
!b6$!.HHSd!*)! 9)!6$%=,-!E)!R)!()!P0:7=$-!l)! p)!Q=:0%-!Q)![)!#,,%15;21-! p)!6)!O%XX2!/$,_!EM2103!*2,%J
<+JR'011=3!V01,:!*'0N=!;'=!O=:N21:=!;2!#3=&;$%&!*;%+<30;%21!%1!O=;%103!K0153%21!R=33:_!
p!l=<$2N'@:%23!"H"\!"SG.J"SZG-!.HHS!
!b62! "SZSd![)!O)! 62:;=$-!c)! ?)! *&'`01_!Q%=3=&;$%&! N$2N=$;%=:! 24! ;%::<=:! 01,! B%2325%&03!+0;=$%J
03:\!0!&$%;%&03!$=D%=`_!R$%;!O=D!V%2+=,!#15)!"SZS_!8V]"^\+.TJ"H8!
!bK0!"GS"d!E32@:%%!K03D01%_!Q=!D%$%B<:!=3=&;$%&%;0;%:! %1!+2;<!+<:&<30$%!&2++=1;0$%<:)!WN<:J
&<30-!N)!/I/J8"Z-!V232510!"GS"!
!
!
"G!
!
bK2!"SZSd!Q)#)!K23,B=$5_!K=1=;%&!E352$%;'+:!%1!*=0$&'-!WN;%+%X0;%21!01,!F0&'%1=!P=0$1%15)!
E,,%:21J(=:3=@-!"SZS!
!bc2!"ST.d!E)!P)!c2,57%1-!E)!6)!c<M3=@_!E!e<01;%;0;%D=!,=:&$%N;%21!24!+=+B$01=!&<$$=1;!01,!
%;:!0NN3%&0;%21!;2!&21,<&;%21!01,!=M&%;0;%21!%1!1=$D=_!p)!?'@:%23)!"ST._!""G\!THHJT88!
!bc<! "SSId!F)*)! c<+0@<1-! #)! ,=! p<01-! p$)-! K)! Q051=3%=-! O)p)! K$==1B=$5-! O)c)! ?$2N:;-! Q)c)! ?'%3J
3%N:)!L%:<03!N=$&=N;%21!=3%&%;=,!B@!=3=&;$%&03! :;%+<30;%21!24! $=;%10! %1!B3%1,!'<+01:)!E$&')!
WN';'03+23-!""8]"^\8HJ8I-!"SSI!
!bc<!.HH/d!F)*)!c<+0@<1-!p)!(=%301,-!K)!6<h%%-!O)p)!K$==1B=$5-!O)!(%33%0+:21-!p)!P%;;3=-!V)!F=&'-!
L)! R%++0$<:;%-! K)! D01! V2=+=3-! K)! Q051=3%=-! #)! ,=! p<01! p$-! fL%:<03! N=$&=N;%21! %1! 0! B3%1,!
:<Bh=&;!`%;'! 0! &'$21%&!+%&$2=3=&;$21%&! $=;%103!N$2:;'=:%:-! f!L%:%21!O=:=0$&'-! D23)!8/-!12)!
.8-!NN!.TG/J.TZ"-!.HH/!
!bc<!.HHSd!F)!c<+0@<1!.HHS!?$=3%+%10$@! $=:<3;:! 4$2+!E$5<:! 99! 4=0:%B%3%;@! :;<,@\!0!IH!=3=&J
;$2,=!=N%$=;%103!N$2:;'=:%:)!91D=:;)!WN';'03+23)!L%:)!*&%)!SZ-!#JEB:;$0&;!8G88!
!b[%! "SZ/d! *)! [%$7N0;$%&7-! R)Q)! K=30;-!F)?)! L=&&'%_! WN;%+%X0;%21! B@! *%+<30;=,! E11=03%15-! *&%J
=1&=-!..H-!IG"!m!IZH-!"SZ/!
!bP=!"SSHd!K)!P='1=$-!#3=7;$2+051=;%:&'=!6=3,;'=2$%=!4A$!915=1%=<$=!<1,!?'@:%7=$-!*N$%15=$!
"SSH-!9*Vl!/JT8HJT./"SJG!
!bP%!.HH8d!()!P%<-!F)*)!c<+0@<1-!fO=;%103!?$2:;'=:%:-g!9###!91;=$10;%2103!*23%,J*;0;=!R%$&<%;:!
R214=$=1&=!Q%5=:;!24!>=&'1%&03!?0N=$:-!NN!."Z!m!."S-!6=B$<0$@!.HH8!
!bP2!.HHTd!l)!c)!P2D=33-!*)!Q272:-!*)P)!R32'=$;@-!?)p)!?$=:;21-!K)p)!*<01%15_!R<$$=1;!,%:;$%B<;%21!
,<$%15!N0$033=3!:;%+<30;%21\! %+N3%&0;%21:! 42$!01!=N%$=;%103!1=<$2N$2:;'=:%:_!#15%1==$%15!
%1! F=,%&%1=! 01,! V%2325@! *2&%=;@-! .HHT)! .G;'! E11<03! 91;=$10;%2103! R214=$=1&=! 24! ;'=!
9###J#FV*!
.HHT)-!
D23)-!
12)-!
NN)T.8.JT.8T-!
"GJ"Z!
p01)!
.HHI_!
,2%\!
"H)""HSa9#FV*).HHT)"I"TII"!
!bF&! .HHGd! F)! p)! F&F0'21-! E)! R0:N%-! p)Q)! Q2$1-! =;! 03)! ].HHG^! *N0;%03! D%:%21! %1! B3%1,! :<Bh=&;:!
%+N301;=,! `%;'! ;'=! *=&21,! *%5';! $=;%103! N$2:;'=:%:)! 91D=:;! WN';'03+23! L%:! *&%! 8Z\! #J
EB:;$0&;!888/!
!bF=!"ST/d!l)!F=;$2N23%:-!E)()!O2:=1B3<;'-!F)l)!O2:=1B3<;'-!E)c)!>=33=$-!#)!>=33=$_!#e<0;%21!
24! *;0;=! R03&<30;%21! B@! 60:;! R2+N<;%15! F0&'%1=:-! p)! 24! R'=+)! ?'@:)-! ."-! "HZG! m! "HS"-!
"ST/!
!b?0!.HHTd!Q)!?03017=$-!E)!L0172D-!?)!c<%=-! *)!V0&&<:_!Q=:%51!24! 0!c%5'!O=:23<;%21!WN;2=3=&J
;$21%&!O=;%103!?$2:;'=:%:_!p)!l=<$03!#15)!<Y+*"HTm*".H!].HHT^!
!
!
"Z!
!
b?$!"SS"d!()c)!?$=::-!V)?)!63011=$@-!*)E)!>=<723:7@-!01,!()>)!L=;;=$3%15-!l<+=$%&03!O=&%N=:!
%1! R\! >'=! E$;! 24! *&%=1;%4%&! R2+N<;%15-! NN)! .ZIJ.ZS-! R0+B$%,5=! C1%D=$:%;@! ?$=::-! R0+J
B$%,5=-!lo!]"SS"^!
!bO%! .HH/d! p)6)! O%XX2! 999-! p)!(@0;;-! p)! P2=`=1:;=%1-! *)! [=33@-! Q)! *'%$=)! F=;'2,:! 01,! N=$&=N;<03!
;'$=:'23,:! 42$! :'2$;J;=$+! =3=&;$%&03! :;%+<30;%21! 24! '<+01! $=;%10! `%;'! +%&$2=3=&;$2,=!
0$$0@:)!91D=:;)!WN';'03+23)!L%:)!*&%)!88]".^\T/TTJT/I"-!.HH/!
!b*&! .H"Hd! #)! ()! *&'+%,-! O)! (%37=_! #3=&;$%&! *;%+<30;%21! 24! ;'=! O=;%10_! .H"H-!
0$q%D\"H".)TSTZD"!beJB%2)lRd!
!bY$! "SSGd! #)! Y$=11=$-! [)JQ)! F%3%&X=7-! L)?)! K0B=3-! c)K)! K$04-! #)! K<=1;'=$-! c)! c0=++=$3=-! V)!
c2=443%15=$-! [)! [2'3=$-! ()! l%:&'-! F)! *&'<B=$;-! E)! *;=;;-! *)! (=%::-! f>'=! Q=D=32N+=1;! 24!
*<B$=;%103! F%&$2N'2;2,%2,=:! 42$! O=N30&=+=1;! 24! Q=5=1=$0;=,! ?'2;2$=&=N;2$:-g! WN'J
;'03+%&!O=:=0$&'_!.S\.ISJ.Z-!"SSG!
!bY$! .HH.d! #)! Y$=11=$-! f(%33! O=;%103! 9+N301;:! O=:;2$=! L%:%21j-g! *R9#lR#-! D23! .ST-! 12! TTTG-!
NN!"H..!m!"H.T-!.HH.!
!bY$!.H"Hd!#)!Y$=11=$-![)!C)!V0$;XJ*&'+%,;-!c)!V=10D-!Q)!V=:&'-!E)!V$<&7+011-!L)J?)!K0B=3-!6)!
K=7=3=$-!C)!K$=NN+0%=$-!E)!c0$:&'=$-!*)![%BB=3-! p)![2&'-!E)![<:1@=$%7-!>)!?=;=$:-![)!*;%153-!
c)! *0&':-! E)! *;=;;-! ?)! *X<$+01-! V)! (%3'=3+-! O)! (%37=_! *<B$=;%103! =3=&;$21%&! &'%N:! 0332`!
B3%1,! N0;%=1;:! ;2! $=0,! 3=;;=$:! 01,! &2+B%1=! ;'=+! ;2!`2$,:_! ?$2&! V%23! *&%)! .H""!F0@! .._!
.GZ]"G""^\!"8ZSJ"8SG!
!
!
"S!
|
1208.2720 | 1 | 1208 | 2012-08-13T22:20:05 | High fidelity optogenetic control of individual prefrontal cortical pyramidal neurons in vivo | [
"q-bio.NC"
] | Precise spatial and temporal manipulation of neural activity in specific genetically defined cell populations is now possible with the advent of optogenetics. The emerging field of optogenetics consists of a set of naturally-occurring and engineered light-sensitive membrane proteins that are able to activate (e.g., channelrhodopsin-2, ChR2) or silence (e.g., halorhodopsin, NpHR) neural activity. Here we demonstrate the technique and the feasibility of using novel adeno-associated viral (AAV) tools to activate (AAV-CaMKll{\alpha}-ChR2-eYFP) or silence (AAV-CaMKll{\alpha}-eNpHR3.0-eYFP) neural activity of rat prefrontal cortical prelimbic (PL) pyramidal neurons in vivo. In vivo single unit extracellular recording of ChR2-transduced pyramidal neurons showed that delivery of brief (10 ms) blue (473 nm) light-pulse trains up to 20 Hz via a custom fiber optic-coupled recording electrode (optrode) induced spiking with high fidelity at 20 Hz for the duration of recording (up to two hours in some cases). To silence spontaneously active neurons we transduced them with the NpHR construct and administered continuous green (532 nm) light to completely inhibit action potential activity for up to 10 seconds with 100% fidelity in most cases. These versatile photosensitive tools combined with optrode recording methods provide experimental control over activity of genetically defined neurons and can be used to investigate the functional relationship between neural activity and complex cognitive behavior. | q-bio.NC | q-bio | F1000 Research 2012, 1:7 (doi: 10.3410/f1000research.1-7.v1). © Usage Licensed by Creative Commons CC-BY 3.0
High fidelity optogenetic control of individual prefrontal
cortical pyramidal neurons in vivo
Shinya Nakamura, Michael V Baratta, Matthew B Pomrenze, Samuel D Dolzani,
Donald C Cooper*
Department of Psychology and Neuroscience, Institute for Behavioral Genetics, University of Colorado, Boulder, CO 80303 USA
*Corresponding author: Donald C Cooper, email: [email protected]
Abstract Precise spatial and temporal manipulation of neural activity in specific genetically defined
cell populations is now possible with the advent of optogenetics. The emerging field of optogenetics
consists of a set of naturally-occurring and engineered light-sensitive membrane proteins that are able to
activate (e.g., channelrhodopsin-2, ChR2) or silence (e.g., halorhodopsin, NpHR) neural activity. Here we
demonstrate the technique and the feasibility of using novel adeno-associated viral (AAV) tools to activate
(AAV-CaMKllα-ChR2-eYFP) or silence (AAV-CaMKllα-eNpHR3.0-eYFP) neural activity of rat prefrontal
cortical prelimbic (PL) pyramidal neurons in vivo. In vivo single unit extracellular recording of ChR2-
transduced pyramidal neurons showed that delivery of brief (10 ms) blue (473 nm) light-pulse trains up
to 20 Hz via a custom fiber optic-coupled recording electrode (optrode) induced spiking with high fidelity
at 20 Hz for the duration of recording (up to two hours in some cases). To silence spontaneously active
neurons we transduced them with the NpHR construct and administered continuous green (532 nm) light
to completely inhibit action potential activity for up to 10 seconds with 100% fidelity in most cases. These
versatile photosensitive tools combined with optrode recording methods provide experimental control over
activity of genetically defined neurons and can be used to investigate the functional relationship between
neural activity and complex cognitive behavior.
Introduction
A method for selective and rapid reversible manipulation of
neuronal activity is important for parsing out the relationship
between prefrontal cortical (PFC) neuronal activity and cogni-
tion. Recent optogenetic technologies, which allow neurons
to respond to specific wavelengths of light with action poten-
tial output, are now providing a significant advance in our abil-
ity to control the activity of select cell populations and neural
circuits in behaving animals1. These tools allow for bidirectional
control over the neuronal activity2. To date, the majority of op-
togenetic experiments have used transgenic, virally-mediated, or a
combination of the two approaches in mice3,4,5. Here we use virally-
mediated gene delivery of the light-responsive proteins ChR26 or
halorhodopsin7 in Sprague-Dawley rat PFC PL pyramidal neurons.
Expression of ChR2 and NpHR enables neurons to be depolar-
ized and silenced by pulses of blue and green light, respectively. We
tested how well PL pyramidal neurons expressing ChR2 followed
pulses (10 ms) of blue light delivered at 20 Hz (fidelity). Given the
potential importance of long duration light delivery for a variety of
protocols we determined if stable responses could be elicited from
individual neurons for at least 2 hrs. Lastly, we tested green light-
induced silencing of spontaneously active NpHR expressing PL
neurons to determine if network-driven action potential activity
could be silenced continuously for 10 seconds, thus establishing
the feasibility of long duration silencing for behavioral testing.
Materials and methods
Subjects. Male Sprague-Dawley rats (6–8 weeks) were housed in
pairs on a 12-h light/dark cycle. Rats were allowed to acclimate
to colony conditions for 7–10 days prior to surgery. All animal
procedures were approved by the Institutional Animal Care and
Use Committee of University of Colorado at Boulder.
Virus injection. An AAV vector carrying the opsin gene encod-
ing the light-gated nonselective cation channel ChR2 or the
light-driven third generation chloride pump NpHR under the
control of the excitatory neuron-specific promoter CaMKllα
(AAV-CaMKllα-ChR2-eYFP or AAV-CaMKllα-eNpHR3.0-
eYFP) was injected into the PL (A/P: +2.7 mm; M/L: ±0.5;
D/V: -2.2 mm) using a 10 μl syringe and a thin 31 gauge metal
needle with a beveled tip (Hamilton Company). The total in-
jection volume (1 μl) and rate (0.1 μl/min) were controlled
with a microinjection pump (UMP3-1, World Precision In-
struments). The virus titer was 3 x 1012 particles/ml.
Optrode preparation
For simultaneous extracellular recording and light delivery, we
developed a custom-made optrode consisting of a tungsten
electrode (1~1.5 MΩ, MicroProbes) attached to an optical
fiber (200 μm core diameter, 0.48 NA, Thorlabs). First, a tung-
sten electrode was attached to a glass capillary tube. An optical
Page 1 of 4
F1000 Research 2012, 1:7 (doi: 10.3410/f1000research.1-7.v1). © Usage Licensed by Creative Commons CC-BY 3.0
fiber was then inserted into the capillary tube and fixed loosely
to the electrode using suture thread. Position of the fiber tip
was adjusted so that the center-to-center distance between the
electrode tip and the fiber tip was ~300 μm.
In vivo extracellular recording
Each rat transduced in the PL with ChR2 or NpHR was an-
esthetized with urethane (1.5 g/kg, i.p.), and a small hole was
made on the skull above the PL. An optrode was then carefully
lowered through the hole until emergence of optically evoked
or inhibited signals. Recorded signals were band-pass filtered
(0.3–8 kHz), amplified (ExAmp-20K, Kation Scientific), digi-
tized at 20 kHz (NI USB-6009, National Instruments), and
stored in personal computer using custom software (LabVIEW,
National Instruments). The custom software was also used for
controlling a blue (473 nm) or green (532 nm) single diode
laser (Shanghai Laser & Optics Century Co.).
Results
To test the functionality of ChR2-transduced PL, we performed
in vivo extracellular recordings from PL neurons during deliv-
ery of blue light. Indeed, PL neurons showed spiking with per-
fect fidelity by the blue-light pulse trains at various frequencies
(1–20 Hz) in ChR2-transduced PL (n=8, Figure 1a).
Photoactivation of a ChR2-transduced PL neuron at various
frequencies
157 Data Files
(http://dx.doi.org/10.6084/m9.figshare.93270)
b.
20 Hz, 10 s
Figures 1b and 1c show the results of 10 ms blue-light pulse
delivery (<250mW/mm2) at 20 Hz for 5 seconds (100 pulses
total). Even when using this longer duration protocol, each
light pulse was able to evoke an action potential with high re-
producibility. We then recorded 120 minutes (2 hrs) of light-
evoked responses in PL neurons (n=2).
Repeated photoactivation of a ChR2-transduced PL neuron
133 Data Files
(http://dx.doi.org/10.6084/m9.figshare.93271)
The average spontaneous firing rate during the baseline period
(5 seconds) during the first 10 minutes of recording was 0.68
+/- 0.10 (Cell 1) and 1.13 +/- 0.19 (Cell 2). Five seconds after
delivery of 100 10 ms blue light pulses at 20 Hz the average
spontaneous firing rate was 0.56 +/- 0.13 (Cell 1) and 0.067
+/- 0.067 (Cell 2). The average spontaneous firing rate dur-
ing the baseline period during the last 10 minutes of the 120
minute recording was 0.12 +/- 0.08 (Cell 1) and 1.83 +/- 0.23
(Cell 2) and after blue light delivery the average firing rate was
0.16 +/- 0.074 (Cell 1) and 0.4 +/- 0.15 (Cell 2).
After a baseline period of 5 seconds, trains of 10 ms light pulses
(100 pulses total) were repeatedly delivered at 20 Hz with an
inter-train interval of 2 min for 120 minutes (Figure 2). Figure 2b
shows the high fidelity responsiveness to the light train across
the 2 hr recording. Mean spike probabilities (± standard error of
the mean) in response to each of the 100 light pulses delivered at
20 Hz for first and last 10 min recording session were as follows:
1.04 ± 0.020 (Cell 1, first) and 0.91 ± 0.034 (Cell 1, last); 2.16 ±
0.055 (Cell 2, first) and 2.19 ± 0.056 (Cell 2, last, Figure 2b).
a.
0 min
b.
a.
1 Hz
5 Hz
10 Hz
20 Hz
1 s
1 s
1 s
1 s
c.
y
t
i
l
i
b
a
b
o
r
p
g
n
i
r
i
F
0.8
0.6
0.4
0.2
0
–10
1 ms
10
0
Time (ms)
20
60 min
120 min
1 s
c.
)
z
H
(
.
q
e
r
F
40
20
0
0
5
10
Time (s)
15
Figure 1. ChR2 evokes temporally precise unit activity in rat PL
cortex in vivo. a) Photoactivation of ChR2-transduced PL neurons
at various frequencies (1–20 Hz). b) Voltage trace showing blue
(473 nm) light-evoked spiking of PL neuron to 20 Hz delivery of
10 ms blue light pulses. Inset: representative light-evoked single-
unit response. c) Averaged response of this neuron to 10 ms light
pulse calculated from all events in above trace (bin width of 1 ms).
Figure 2. Repeated 20 Hz blue light (10 ms) stimulation of a
ChR2-transduced PL neuron. a) Voltage traces of the light-evoked
spiking acquired at the time point of 0, 60 and 120 min after the
beginning of recordings. b) Raster plot showing all 61 repetitions
(2 min interval, 120 min total) of the light-induced activation.
c) Average firing rate (thick line) with standard error of the mean (SEM,
thin line) calculated from raster plot above (bin width of 100 ms).
Page 2 of 4
F1000 Research 2012, 1:7 (doi: 10.3410/f1000research.1-7.v1). © Usage Licensed by Creative Commons CC-BY 3.0
We then tested in vivo photoinhibition of spontaneous activity
in NpHR-transduced PL.
Photoinhibition of NpHR-transduced PL neuron activity
61 Data Files
(http://dx.doi.org/10.6084/m9.figshare.93272)
Figure 3 shows the results of NpHR-induced inhibition of PL
neuronal activity using 10 seconds of continuous green light
exposure (<250 mW/mm2). In contrast to the ChR2 results,
we observed silencing of spontaneous activity in the PL during
light delivery (n=7) compared to pre- and post-light periods.
Only four (13%, from two cells) out of a total of 30 trials had
break through action potentials during the period of 10 second
laser illumination (e.g. Figure 3b). To functionally character-
ize the neuronal activity we constructed autocorrelation histo-
grams (ACH) and calculated the average firing rate during pre-
and post-laser period for each neuron (Figure 4). The ACH
revealed two types of neurons that were functionally classified
into either phasic or non-phasic firing. Phasic neurons showed
correlated activity marked by peaks in their ACH (Figure 4a),
indicating their periodic network-driven firing pattern, but
non-phasic neurons showed no correlated activity (Figure 4b).
Two out of three phasic neurons showed a rebound increase in
excitability after the light illumination (Figure 4c), but overall
a.
b.
c.
)
z
H
(
.
q
e
r
F
8
4
0
0
10
20
Time (s)
30
Figure 3. NpHR rapidly and reversibly silences spontaneous ac-
tivity of rat PL neurons in vivo. a) Representative trace of spontane-
ous activity of a NpHR-transduced PL neuron that was inhibited by
the continuous green (532 nm) laser delivery. b) Raster plot showing
five repetitions of the light-induced silencing in this neuron. Each unit
activity is plotted as a dot. c) Average firing rate (thick line) with SEM
(thin line) calculated from these five repetitions (bin width of 1000 ms).
b.
Non-phasic
Pre
0
1000
2000
3000
4000
5000
0
1000
2000
3000
4000
5000
40
30
20
10
0
40
30
20
10
0
20
15
10
5
1000
2000
3000
4000
5000
40
30
20
10
0
0
40
30
20
10
0
0
20
15
10
5
0
0
Post
1000
2000
3000
1000
2000
3000
1000
2000
3000
Pre
Phasic
20
15
10
Post
0
1000
2000
3000
4000
5000
5
0
0
1000
2000
3000
4000
5000
5
0
0
1000
2000
3000
4000
20
15
10
5
0
5000 0
20
15
10
1000
2000
3000
4000
5000
0
1000
2000
3000
Time (ms)
4000
5000
5
0
0
1000
2000
3000
4000
5000
a.
20
15
10
5
0
20
15
10
t
n
e
v
e
#
20
15
10
5
0
c.
)
z
H
(
.
q
e
r
F
3
2
1
0
Pre
Post
0
0
d.
6
4
2
0
Figure 4. Comparison of spontaneous activities during pre- and post-laser period for phasic (left) and non-phasic (right) PL
neurons. a, b) Auto-correlation histograms (ACH) for each recorded neuron showing periodic firing pattern in phasic neurons, but not in
non-phasic neurons. ACH were created by summing the number of intervals between non-consecutive events as well as intervals between
consecutive events (bin width of 20 ms). c, d) Average firing rate of each neuron during pre- and post-laser period. Note that the result of
one neuron is not shown because of the low number of spikes.
Page 3 of 4
Pre
Post
F1000 Research 2012, 1:7 (doi: 10.3410/f1000research.1-7.v1). © Usage Licensed by Creative Commons CC-BY 3.0
there appeared to be no change (Figures 4, c and d). Further
recordings are necessary to determine if continuous activation
of NpHR with green light leads to some form of plasticity de-
pending on their type of neuronal inputs. Control recordings
using blue and green light (light intensities of 250 mW/mm2)
did not alter spontaneous firing rates in PL neurons that did
not express either ChR2 or NpHR as inferred from the lack
of time-locked light responses (data not shown), which sug-
gests no unexpected off-target effects of light alone on neuronal
activity.
Discussion
In this study we demonstrated high fidelity in vivo recording
of individual PL pyramidal neurons transduced with ChR2
or NpHR. Single-unit responses were recorded in response to
trains of 10 ms blue light pulses up to 20 Hz with a mode of
activation onset time of 7 ms. Light delivered through a cus-
tom optrode (<250 mW/mm2) induced stable and robust ac-
tivation that persisted for the duration of recording (2 hrs in
some cases).
To understand how neuronal activity influences behavior it is
necessary to not only activate, but also silence neuronal activ-
ity with precision. We used 5–10 seconds of green light (<250
mW/mm2) to silence spontaneously active phasic and nonpha-
sic firing NpHR transduced PL pyramidal neurons. Both pha-
sic and nonphasic pyramidal neurons were potently inhibited
for the duration of light delivery (up to 10 seconds). In some
phasic firing neurons we observed a slight rebound increase
in excitability upon termination of light. More recordings are
necessary to determine the mechanism for rebound excitabil-
ity or if it is particular for phasic firing neurons. A transient
increase in activity shortly after termination of light-induced
hyperpolarization (~100 ms) as seen in Figure 3C would be
consistent with activation of a hyperpolarization activated (I h )
inward current, while more persistent increases in activity may
indicate indirect effects resulting from activation of the NpHR
Cl– pumps and the resulting change in extracellular Cl–. Re-
cently, Raimondo et al. (2012) reported that in vitro synap-
tically-evoked spike probability significantly increases shortly
after termination of photoinhibition in NpHR-expressing
hippocampal neurons8. Our results are the first to suggest this
phenomenon may happen in vivo to a subpopulation of phasic
firing cells in the PL. Future experiments are necessary to un-
derstand the relationship between neuronal activation/silenc-
ing on behavior and the possible off target effects of ChR2 or
NpHR activation.
The technique of in vivo light delivery and simultaneous re-
cording of neuronal responses holds the promise of establishing
direct causal relationships between the onset/offset and pattern
of activity in specific genetically-determined neuronal popu-
lations and corresponding time-locked behavioral events. Our
results establish the feasibility of long duration high fidelity
activation or continuous silencing of individual neurons for a
variety of behavioral experiments.
Author contributions
S.N., M.V.B, and D.C.C. designed experiments, analyzed data
and wrote the paper. S.N., M.V.B., M.B.P., and S.D.D. carried
out experiments. M.V.B. and M.B.P. assisted with molecular
biology and virus preparation.
Competing Interests
No relevant competing interests declared.
Grant information
This work was supported by National Institute on Drug
Abuse grant R01-DA24040 (to D.C.C.), NIDA K award
K-01DA017750 (to D.C.C.)
References
1. Bernstein, J., G. & Boyden, E. S. Optogenetic tools for analyzing
the neural circuits of behavior. Trends. Cogn. Sci. 15, 592–600
(2011).
2. Baratta, M. V., Nakamura, S., Dobelis, P., Pomrenze, M. B.,
Dolzani, S. D. & Cooper, D. C. Optogenetic control of genetically-
targeted pyramidal neuron activity in prefrontal cortex. Nat.
Precedings (posted Apr 2, 2012), doi:10.1038/npre.2012.7102.1.
3. Tye, K. M., Prakash, R., Kim, S. Y., Fenno, L. E., Grosenick, L.,
Zarabi, H., Thompson, K. R., Gradinaru, V., Ramakrishnan, C.
& Deisseroth, K. Amygdala circuitry mediating reversible and
bidirectional control of anxiety. Nature 471, 358–362 (2011).
4. Zhao, S., Ting, J. T., Atallah, H. E., Qiu, L., Tan, J., Gloss, B.,
Augustine, G. J., Deisseroth, K., Luo, M., Graybiel, A. M. & Feng G.
Cell type-specific channelrhodopsin-2 transgenic mice for optogenetic
dissection of neural circuitry function. Nat. Methods 8, 745–752 (2011).
5. Witten, I. B., Lin, S. C., Brodsky, M., Prakash, R., Diester, I.,
Anikeeva, P., Gradinaru, V., Ramakrishnan, C. & Deisseroth, K.
Cholinergic interneurons control local circuit activity and cocaine
conditioning. Science 330, 1677–1681 (2010).
6. Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G. & Deisseroth, K.
Millisecond-timescale, genetically targeted optical control of neural
activity. Nat. Neurosci. 8, 1263–1268 (2005).
7. Zhang, F., Wang, L. P., Brauner, M., Liewald, J. F., Kay, K.,
Watzke, N., Wood, P. G., Bamberg, E., Nagel, G., Gottschalk, A.
& Deisseroth, K. Multimodal fast optical interrogation of neural
circuitry. Nature 446, 633–639 (2007).
8. Raimondo, J. V., Kay, L., Ellender, T. J. & Akerman, C. J.
Optogenetic silencing strategies differ in their effects on
inhibitory synaptic transmission. Nat. Neurosci. (2012),
doi:10.1038/nn.3143.
Page 4 of 4
|
1912.00270 | 1 | 1912 | 2019-11-30T21:51:28 | Challenges for automated spike sorting: beware of pharmacological manipulations | [
"q-bio.NC"
] | The advent of large-scale and high-density extracellular recording devices allows simultaneous recording from thousands of neurons. However, the complexity and size of the data makes it mandatory to develop robust algorithms for fully automated spike sorting. Here it is shown that limitations imposed by biological constraints such as changes in spike waveforms induced under different drug regimes should be carefully taken into consideration in future developments. | q-bio.NC | q-bio | Challenges for automated spike sorting:
beware of pharmacological manipulations
Gerrit Hilgen
Biosciences Institute, Faculty of Medical Sciences, Newcastle University,
Newcastle, NE2 4HH, United Kingdom
The advent of large-scale and high-density extracellular recording devices allows simultaneous
recording from thousands of neurons. However, the complexity and size of the data makes it
mandatory to develop robust algorithms for fully automated spike sorting. Here it is shown that
limitations imposed by biological constraints such as changes in spike waveforms induced under
different drug regimes should be carefully taken into consideration in future developments.
Extracellular probes provide an excellent tool for long-term in vitro and in vivo recordings of neurons.
The variety of these devices spans from single electrodes (1 channel) over tetrodes (4 channels) to
multielectrode arrays (MEAs) and probes (up to thousands of channels). These devices usually pick
up the activity of multiple neurons on each channel, which necessitates spike sorting to isolate the
responses of single-units. For devices with few and well-separated channels, it is standard to detect
and extract the spike waveforms and combine dimensionality reduction of the waveforms with
posthoc clustering (for review, Rey, Pedreira, and Quian Quiroga 2015), either in a supervised or
semisupervised manner. For the newest generation of CMOS (complementary metal-oxide-
semiconductor-based) MEA devices with thousands of closely spaced channels, however, the sheer
volume combined with the intricacy of the recordings makes it mandatory to minimize manual
intervention and to develop new fully-unsupervised spike sorting workflows (for review, Hennig,
Hurwitz, and Sorbaro 2018). These new algorithms not only need to tackle the computational
challenges but they also need to consider acute changes in the spike waveform caused by overlapping
spikes, drifting electrodes or bursting activity (for review, see Rey, Pedreira, and Quian Quiroga 2015).
In addition, acute changes can occur in spike waveforms in the presence of pharmacological agents
that affect ionic conductances. Such variability, if not taken into consideration, can impair the quality
of automated spike sorting, and consequently, bias the biological interpretation of the results.
Recently developed unsupervised spike sorters (see Hennig, Hurwitz, and Sorbaro 2018 for a complete
list) use either template-matching or density-based approaches to overcome these challenges. For
template-matching the cluster size is set lower than estimated, ending up with multiple clusters from
the same unit. These clusters are then iteratively compared to each other and merged if they
represent the same neuron (for review, see Lefebvre, Yger, and Marre 2016). This approach works
very well for very short-term spike waveform changes caused by changes in activity levels. However,
for longer-term changes (over the course of minutes to hours) induced by direct or indirect
modulation of ion channel properties under different pharmacological and chemogenetic regimes, the
template information provided by these affected waveforms would be falsely interpreted as spikes
originating from different neurons, requiring manual intervention to reassign the waveforms to the
correct neurons. On the other hand, density-based algorithms (Chung et al. 2017; Hilgen et al. 2017;
Jun et al. 2017) are able to produce reliable results in such conditions, without necessitating manual
intervention.
This review summarizes the different ionic components of the spike waveform, lists conditions which
can affect these components either acutely or chronically, and gives a brief overview of the existing
unsupervised density-based spike sorters.
The spike waveform components: Considering the classical Hodgkin Huxley model of axonal spikes, a
spike waveform is composed of the initial resting membrane potential phase (Fig. 1A, 1) followed by
a small change of the membrane potential to the point where it crosses the threshold potential. At
that point, voltage-gated Na+ channels open, resulting in a massive Na+ influx into the neuron
(depolarization, Fig. 1A, 2). After about a very short period of threshold depolarization, these same+
channels inactivate, resulting in the repolarization phase of the action potential towards resting levels
(determined by the high K+ leakage permeability). At that point, the delayed-rectifier voltage-gated K+
channels open, increasing K+ efflux and rapid repolarization (Fig. 1A, 3). Until the delayed rectifier
channels close again, the permeability to K+ is higher than in resting conditions, causing the membrane
potential to move closer to the equilibrium potential for K+(undershoot or afterhyperpolarization, Fig.
1A, 4). Somatic spikes, however, have more than a dozen distinct voltage-dependent conductances,
and ionic channel contributions to the spike waveform are much more complex than in the axon (for
review, Bean 2007).
Pharmacology-induced changes in spike waveforms: Obviously, applying various voltage-gated ion
channel agonists and antagonists will directly modify the spike waveform. Drugs such as tetrodotoxin
(TTX, Na+ channels), 4-aminopyridine (4-AP, K+ channels), tetraethylammonium (TEA, K+ channels) and
cadmium chloride (CdCl2, Ca2+ channels) for example, have dramatic effects on various components
of the spike waveform. The Na+ channel opener veratridine leads to persistent activation, therefore
affecting the spike waveform as well (Akanda et al. 2009). Such compounds are routinely used in
intracellular, but rarely in extracellular recordings. However, many different drugs are used to probe
neural function and connectivity on a larger scale in extracellular recordings. Many of these
compounds can change the spike waveform. For example, strychnine, an antagonist of the inhibitory
neurotransmitter glycine, is known to affect voltage-gated Na+ channels as well (Reiser, Günther, and
Hamprecht 1982), thereby profoundly changing the spike waveform. The antimalarial drug quinine is
often used to block specific gap junctions (Cx36), but it is also a K+ channel modulator, and at high
concentrations (mM range) it even modulates Na+ channels (Akanda et al. 2009). Indeed, we recorded
spikes with an HD MEA from the mouse ganglion cell layer and found the effect of quinine on the
average spike waveform to be severe, even at 100 M concentration (Fig. 1A, blue vs beige). The
GABAA receptor antagonist bicuculline also blocks Ca+-activated K+ channels (Bruening-wright,
Adelman, and Maylie 1999) which are responsible for the slow afterhyperpolarization phase in some
types of neurons (Fig. 1A, 4). Mecamylamine is an antagonist of nicotinic cholinergic receptors with
no known potential effects on spike waveforms. In the presence of a cocktail of 20 M mecamylamine
and 20 M bicuculline, we observed a decrease in the amplitude of both phases of the average spike
waveform amplitude (Fig. 1D, blue vs beige).
Chemogenetic-induced changes of spike waveforms: Neuroscientist now commonly use chemogenetic
tools that alter various ionic conductances over periods of several hours, but little attention has been
given so far to the underlying implications this might have for spike waveforms. For example, PSAM
(pharmacologically selective actuator module, Sternson and Roth 2014) is a chimeric chloride-
permeable ligand-gated channel for neuron activation and silencing that causes changes in membrane
potential, thereby affecting spike waveforms. Further, inhibitory hM4Di (human M4 muscarinic
receptor) DREADDs (designer receptors exclusively activated by designer drugs, Sternson and Roth
2014) activated by clozapine N-oxide induce hyperpolarization by activation of GIRK (G-protein
inwardly rectifying K+ ) channels and excitatory hM3Dq (human M4 muscarinic receptor) DREADDs
induce tonic depolarization of the cell membrane by intracellular calcium release. The extent of
membrane potential changes has a direct effect on the amplitude of the spike waveform, with smaller
spikes when the membrane is depolarized, and larger spikes when the membrane is hyperpolarized
(Nádasdy Z et al. 1998).
Density-based algorithms: Nádasdy et al. (1998) described a silicon probe with a densely-spaced and
diamond-shaped recording site to map the isopotential contours of a spontaneously spiking rat
pyramidal cell. By advancing the probe in small steps in the vicinity of the cell while measuring the
amplitudes of the successively recorded spikes, it was possible to interpolate the barycenter (center
of mass) of the spike amplitude. They stated that it is feasible to determine the location of the action
potential generation site by using three or more recording sites with less than 50 m spacing. Further,
they also estimated the spike origins of rat cortical recordings made with a high-density 8 x 16 shank
probe array on a larger scale. Briefly, spikes were detected for each channel and a weighting factor,
determined by the interelectrode distance, was assigned to each measured spike amplitude. The
mean of the weighted amplitudes is then plotted as a function of the sum of the amplitudes as a two-
dimensional density map which can be very efficiently clustered. Indeed, Nádasdy et al. used K-Means
clustering to isolate potential single units. This study laid the foundation of how to use spatial spike
information for spike sorting, an approach that has recently been revived and massively improved by
density-based unsupervised spike sorters (MountainSort: Chung et al. 2017; Herding Spikes: Hilgen et
al. 2017; JRCLUST/IRONCLUST: Jun et al. 2017). In high-density arrays and probes with less than 60 m
spacing, spike events from the same neuron are recorded on multiple channels, a fact that is used to
estimate the barycentre from amplitudes in adjacent channels. As mentioned above, the two-
dimensional spacing of spike localisations can be very efficiently clustered but it is mandatory to
combine that information with the information from extracted spike waveform features obtained
through dimensionality reduction for successful isolation of single-units. The current density-based
sorters use different clustering strategies to solve this task: JRCLUST/IRONCLUST uses a newly
developed density-based clustering algorithm named DPCLUS and identifies local density peaks by
calculating pairwise distances between neighboring points. MountainSort is also using a newly
developed non-parametric density-based algorithm named ISO-SPLIT, which generates unimodal
clusters by using isotonic regression. Herding Spikes is using the mean-shift algorithm to "herd" data
points towards high-density areas.
What these density-based sorters have in common is the fact that they use the spatial origin of the
spike waveform as a foundation for their sorting. This is a very important point because the spike
localization remains unchanged in control and drug conditions, even when the extracted spike
waveform principal components change under different drug regimes. To further illustrate this crucial
issue, we plotted the first two whitened principal components of control (beige) and drug condition
(blue) from the experiments mentioned above in their feature space (Fig. 1, C & F). We did not cluster
these points but instead colored them according to their condition origin. It is already visible by eye
that the two conditions form separate clusters, even though originating from the same neuron, and
conventional clustering based on these principal components alone would likely fail. In contrast, we
plotted the interpolated spike localization for all these waveforms (Fig. 1 B & E) and, again, colored
them according to their condition origin. The spatial origin of the spike waveform remains unchanged
in control and drug conditions and these example units were sorted correctly, unsupervised and in
near real-time, with Herding Spikes.
The examples in Figure 1 are rather extreme and for most conditions, template-matching and density-
based sorters perform equally well (https://spikeforest.flatironinstitute.org/). However, both sorting
strategies have their limitations. Template-matching sorters often require manual curation and do not
perform well on recordings with low firing rates, whereas density-based sorters require relatively
uniform and dense spacing of electrodes. For data that is suitable for both types of sorters, it is
advisable to use more than one sorter algorithm. A very convenient approach is to use SpikeInterface
(http://homepages.inf.ed.ac.uk/mhennig/news/spikeinterface/), a newly developed spike sorting
workflow to sort the same data with different sorters and compare the results in a comprehensive
and standardized way in a GUI.
Acknowledgments: I thank Evelyne Sernagor (Newcastle University, UK) and Matthias Hennig
(University of Edinburgh, UK) for helpful discussions and critical readings of this manuscript.
Experiments were performed in Evelyne Sernagor's lab and funding was provided by the Leverhulme
Trust grant RPG-2016-315 (ES).
Akanda, Nesar, Peter Molnar, Maria Stancescu, and James J. Hickman. 2009. "Analysis of Toxin-
Induced Changes in Action Potential Shape for Drug Development." Journal of Biomolecular
Screening 14(10): 1228 -- 35.
Bean, Bruce P. 2007. "The Action Potential in Mammalian Central Neurons." Nature reviews.
Neuroscience 8(6): 451 -- 65.
Bruening-wright, Radwan Khawaled Andrew, John P Adelman, and James Maylie. 1999. "Bicuculline
Block of Small-Conductance Calcium-Activated Potassium Channels." : 314 -- 21.
Chung, Jason E. et al. 2017. "A Fully Automated Approach to Spike Sorting." Neuron 95(6): 1381-
1394.e6. https://doi.org/10.1016/j.neuron.2017.08.030.
Hennig, Matthias H., Cole Hurwitz, and Martino Sorbaro. 2018. "Scaling Spike Detection and Sorting
for Next Generation Electrophysiology." In Eprint ArXiv:1809.01051, , 1 -- 14.
http://arxiv.org/abs/1809.01051.
Hilgen, G. et al. 2017. "Unsupervised Spike Sorting for Large-Scale, High-Density Multielectrode
Arrays." Cell Reports 18(10).
Jun, James et al. 2017. "Real-Time Spike Sorting Platform for High-Density Extracellular Probes with
Ground-Truth Validation and Drift Correction." bioRxiv: 1 -- 29.
Lefebvre, Baptiste, Pierre Yger, and Olivier Marre. 2016. "Recent Progress in Multi-Electrode Spike
Sorting Methods." Journal of Physiology Paris 110(4): 327 -- 35.
http://dx.doi.org/10.1016/j.jphysparis.2017.02.005.
Nádasdy Z et al. 1998. "Extracellular Recording and Analysis of Neuronal Activity : From Single Cells
to Ensembles." In Neuronal Ensembles: Strategies for Recording and Decoding., eds.
Eichenbaum HB and Davis JL. , 17 -- 55.
Reiser, Georg, Andrea Günther, and Bernd Hamprecht. 1982. "Strychnine and Local Anesthetics
Block Ion Channels Activated by Veratridine in Neuroblastoma × Glioma Hybrid Cells." FEBS
Letters 143(2): 306 -- 10.
Rey, Hernan Gonzalo, Carlos Pedreira, and Rodrigo Quian Quiroga. 2015. "Past, Present and Future
of Spike Sorting Techniques." Brain Research Bulletin 119: 106 -- 17.
http://dx.doi.org/10.1016/j.brainresbull.2015.04.007.
Sternson, Scott M., and Bryan L. Roth. 2014. "Chemogenetic Tools to Interrogate Brain Functions."
Annual Review of Neuroscience 37(1): 387 -- 407.
Figures
Figure 1: Drug-induced changes of the spike waveforms from mouse retinal ganglion cells recorded
with an HD MEA (3Brain, Switzerland). A) The average spike waveform of a representative selected
RGC in control (beige) and after application of 100M quinine (blue). The average spike waveform is
profoundly changed in the presence of quinine. Numbers indicate the different stages of a spike
waveform as explained in the main text. As a convention, extracellular spikes such as those illustrated
here have opposite polarity to intracellular spikes. B) The spatial origin of the spike source, the spike
localization remains similar in both conditions. C) Plotting the first two principal components (Python
scikit-learn module, sklearn.decomposition.PCA, whiten=true) from control and 100M quinine spike
waveforms shows two very well-separated clusters. It is difficult to assign one cluster in principal
component space for this single RGC unit. D-F) Another example of drug-induced changes of spike
waveforms is shown here by using 20M bicuculline in combination with 20M mecamylamine. D)
The average waveform changes after application of this drug cocktail but the spike localization remains
the same (E) while the feature space shows two clusters.
|
1612.05660 | 3 | 1612 | 2018-04-20T20:27:30 | Gr\"obner Bases of Neural Ideals | [
"q-bio.NC",
"math.AC",
"math.AG"
] | The brain processes information about the environment via neural codes. The neural ideal was introduced recently as an algebraic object that can be used to better understand the combinatorial structure of neural codes. Every neural ideal has a particular generating set, called the canonical form, that directly encodes a minimal description of the receptive field structure intrinsic to the neural code. On the other hand, for a given monomial order, any polynomial ideal is also generated by its unique (reduced) Gr\"obner basis with respect to that monomial order. How are these two types of generating sets -- canonical forms and Gr\"obner bases -- related? Our main result states that if the canonical form of a neural ideal is a Gr\"obner basis, then it is the universal Gr\"obner basis (that is, the union of all reduced Gr\"obner bases). Furthermore, we prove that this situation -- when the canonical form is a Gr\"obner basis -- occurs precisely when the universal Gr\"obner basis contains only pseudo-monomials (certain generalizations of monomials). Our results motivate two questions: (1)~When is the canonical form a Gr\"obner basis? (2)~When the universal Gr\"obner basis of a neural ideal is {\em not} a canonical form, what can the non-pseudo-monomial elements in the basis tell us about the receptive fields of the code? We give partial answers to both questions. Along the way, we develop a representation of pseudo-monomials as hypercubes in a Boolean lattice. | q-bio.NC | q-bio |
GR OBNER BASES OF NEURAL IDEALS
REBECCA GARCIA, LUIS DAVID GARC´IA PUENTE, RYAN KRUSE, JESSICA LIU, DANE MIYATA,
ETHAN PETERSEN, KAITLYN PHILLIPSON, AND ANNE SHIU
Abstract. The brain processes information about the environment via neural codes. The neural
ideal was introduced recently as an algebraic object that can be used to better understand the
combinatorial structure of neural codes. Every neural ideal has a particular generating set, called
the canonical form, that directly encodes a minimal description of the receptive field structure
intrinsic to the neural code. On the other hand, for a given monomial order, any polynomial ideal
is also generated by its unique (reduced) Grobner basis with respect to that monomial order. How
are these two types of generating sets – canonical forms and Grobner bases – related? Our main
result states that if the canonical form of a neural ideal is a Grobner basis, then it is the universal
Grobner basis (that is, the union of all reduced Grobner bases). Furthermore, we prove that this
situation – when the canonical form is a Grobner basis – occurs precisely when the universal Grobner
basis contains only pseudo-monomials (certain generalizations of monomials). Our results motivate
two questions: (1) When is the canonical form a Grobner basis? (2) When the universal Grobner
basis of a neural ideal is not a canonical form, what can the non-pseudo-monomial elements in the
basis tell us about the receptive fields of the code? We give partial answers to both questions. Along
the way, we develop a representation of pseudo-monomials as hypercubes in a Boolean lattice.
Keywords: neural code, receptive field, canonical form, Grobner basis, Boolean lattice
MSC classes: 92-04 (Primary), 13P25, 68W30 (Secondary)
1. Introduction
The brain is tasked with many important functions, but one of the least understood is how it
builds an understanding of the world. Stimuli in one's environment are not experienced in isolation,
but in relation to other stimuli. How does the brain represent this organization? Or, to quote from
Curto, Itskov, Veliz-Cuba, and Youngs, "What can be inferred about the underlying stimulus space
from neural activity alone?" [7].
Curto et al. pursued this question for codes where each neuron has a region of stimulus space,
called its receptive field, in which it fires at a high rate. They introduced algebraic objects that
summarize neural-activity data, which are in the form of neural codes (0/1-vectors where 1 means
the corresponding neuron is active, and 0 means silence) [7]. The neural ideal of a neural code is an
ideal that contains the full combinatorial data of the code. The canonical form of a neural ideal is
a generating set that is a minimal description of the receptive-field structure. Hence, the questions
posed above have been investigated via the neural ideal or the canonical form [6, 7, 8, 10]. As
a complement to algebraic approaches, combinatorial and topological arguments are employed in
related works [5, 11, 13].
The aim of our work is to investigate, for the first time, how the canonical form is related to
other generating sets of the neural ideal, namely, its Grobner bases. This is a natural mathematical
question, and additionally the answer could improve algorithms for computing the canonical form.
Currently, there are two distinct methods to compute the canonical form of a neural ideal: the
original method proposed in [7] and an iterative method introduced in [16]. The former method
requires the computation of primary decomposition of pseudo-monomial ideals. As a result, this
method is rather inefficient. Even in dimension 5, one can find codes for which this algorithm
takes hundreds or even thousands of seconds to terminate or halts due to lack of memory. The
Date: April 24, 2018.
1
2
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
more recent iterative method relies entirely on basic polynomial arithmetic. This algorithm can
efficiently compute canonical forms for codes in up to 10 dimensions; see [16]. On the other hand,
Grobner basis computations are generally computationally expensive. Nevertheless, we take full
advantage of tailored methods for Grobner basis over Boolean rings [3]. As we show later in Table I,
for small dimensions less than or equal to 8, Grobner basis computations are faster than canonical
form ones. For larger dimensions, we have observed that in general Grobner basis computations
are faster but the standard deviation on computational time is much larger. In dimension 9, the
average time to compute a Grobner basis is around 3 seconds, but there are codes for which that
computation takes close to 10 hours to finish.
Nevertheless, we believe that a thorough study of Grobner basis of neural ideals is not only of
theoretical interest, but it can lead to better procedures able to perform computations in larger
dimensions.
Indeed, among small codes, surprisingly many have canonical forms that are also
Grobner bases. Moreover, the iterative nature of the newer canonical form algorithm hints towards
the ability to compute canonical forms and Grobner bases of neural codes in large dimensions by
'gluing' those of codes on small dimensions. Such decomposition results are a common theme in
other areas of applied algebraic geometry [1, 9].
The outline of this paper is as follows. Section 2 provides background on neural ideals, canonical
forms, and Grobner bases. In Section 3, we prove our main result: if the canonical form of a neural
ideal is a Grobner basis, then it is the universal Grobner basis (Theorem 3.1). We also prove a
partial converse:
if the universal Grobner basis of a neural ideal contains only so-called pseudo-
monomials, then it is the canonical form (Theorem 3.12). Our results motivate other questions:
(1) When is the canonical form a Grobner basis?
(2) If the universal Grobner basis of a neural ideal is not a canonical form, what can the
non-pseudo-monomial elements in the basis tell us about the receptive fields of the code?
Sections 4 and 5 provide some partial answers these questions. Finally, a discussion is in Section 6.
2. Background
This section introduces neural ideals and related topics, which were first defined by Curto, Itskov,
Veliz-Cuba, and Youngs [7], and recalls some basics about Grobner bases. We use the notation
[n] := {1, 2, . . . , n}.
2.1. Neural codes and receptive fields. A neural code (also known as a combinatorial
code) on n neurons is a set of binary firing patterns C ⊂ {0, 1}n, that is, a set of binary strings of
neural activity. Note that neither timing nor rate of neural activity are recorded in a neural code.
An element c ∈ C of a neural code is a codeword. Equivalently, a codeword is determined by
the set of neurons that fire:
supp(c) := {i ∈ [n] ci = 1} ⊆ [n] .
Thus, the entire code is identified with a set of subsets of co-firing neurons: supp(C) = {supp(c)
c ∈ C} ⊆ 2[n].
In many areas of the brain, neurons are associated with receptive fields in a stimulus space. Of
particular interest are the receptive fields of place cells, which are neurons that fire in response to an
animal's location. More specifically, each place cell is associated with a place field, a convex region
of the animal's physical environment where the place cell has a high firing rate [15]. The discovery
of place cells and related neurons (grid cells and head direction cells) won neuroscientists John
O'Keefe, May Britt Moser, and Edvard Moser the 2014 Nobel Prize in Physiology and Medicine.
GR OBNER BASES OF NEURAL IDEALS
3
X
U1
U2
U3
Figure 1. Receptive fields Ui for which the code is C(U ) = {∅, 1, 123, 13, 3}.
Given a collection of sets U = {U1, ..., Un} in a stimulus space X (here Ui is the receptive field
of neuron i), the receptive field code, denoted by C(U ), is:
C(U ) :=
c ∈ {0, 1}n :
\
i∈supp(c)
Ui
\
[
j /∈supp(c)
Uj
6= ∅
.
As mentioned earlier, we often identify this code with the corresponding set of subsets of [n]. Also,
we use the following convention for the empty intersection: Ti∈∅ Ui := X.
Example 2.1. Consider the sets Ui in a stimulus space X depicted in Figure 1. The corresponding
receptive field code is C(U ) = {∅, 1, 123, 13, 3}.
2.2. The neural ideal and its canonical form. A pseudo-monomial in F2[x1, . . . , xn] is a
polynomial of the form
f = Y
i∈σ
xiY
j∈τ
(1 + xj) ,
where σ, τ ⊆ [n] with σ ∩ τ = ∅. Every term in a pseudo-monomial f = Qi∈σ xiQj∈τ (1 + xj)
divides its highest-degree term, Qi∈σ∪τ xi. We will use this fact several times in this work.
Each v ∈ {0, 1}n defines a pseudo-monomial ρv as follows:
ρv :=
nY
(1 − vi − xi) = Y
i=1
{ivi=1}
xi Y
{jvj=0}
(1 + xj) = Y
{i∈supp(v)}
xi Y
{j6∈supp(v)}
(1 − xj) .
Notice that ρv is the characteristic function for v, that is, ρv(x) = 1 if and only if x = v.
Definition 2.2. Let C ⊆ {0, 1}n be a neural code. The neural ideal JC is the ideal in F2[x1, . . . , xn]
generated by all ρv for v 6∈ C:
JC := h{ρvv 6∈ C}i .
It follows that the variety of the neural ideal is the code itself: V (JC ) = C. The following lemma
provides the algebraic version of the previous statement:
Lemma 2.3 (Curto, Itskov, Veliz-Cuba, and Youngs [7, Lemma 3.2]). Let C ⊆ {0, 1}n be a neural
code. Then
where I(C) is the ideal of the subset C ⊆ {0, 1}n.
I(C) = JC + hxi(1 + xi) i ∈ [n]i ,
Note that the ideal generated by the Boolean relations hxi(1 + xi) : i ∈ [n]i is contained in I(C),
regardless of the structure of C.
A pseudo-monomial f in an ideal J in F2[x1, . . . , xn] is minimal if there does not exist another
pseudo-monomial g ∈ J, with g 6= f , such that f = gh for some h ∈ F2[x1, . . . , xn].
4
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
Definition 2.4. The canonical form of a neural ideal JC , denoted by CF(JC ), is the set of all
minimal pseudo-monomials of JC.
Algorithms for computing the canonical form were given in [7, 8, 16]. In particular, [16] describes
an iterative method to compute the canonical form that is significantly more efficient than the
original method presented in [7].
The canonical form CF(JC ) is a particular generating set for the neural ideal JC [7]. The main
goal in this work is to compare CF(JC ) to other generating sets of JC , namely, its Grobner bases.
Example 2.5. Returning to Example 2.1, the codewords v that are not in C(U ) = {∅, 1, 123, 13, 3}
are 2, 12, and 23, so the neural ideal is JC = h{x2(1 + x1)(1 + x3), x1x2(1 + x3), x2x3(1 + x1)}i.
The canonical form is CF(JC(U )) = {x2(1+ x1), x2(1+ x3)}. We will interpret these canonical-form
polynomials in Example 2.7 below.
2.3. Receptive-field relationships. It turns out that we can interpret pseudo-monomials in JC
(and thus in the canonical form) in terms of relationships among receptive fields. First we need the
following notation: for any σ ⊆ [n], define:
xσ := Y
i∈σ
xi
and Uσ := \
Ui ,
i∈σ
where, by convention, the empty intersection is the entire space X.
Lemma 2.6 (Curto, Itskov, Veliz-Cuba, and Youngs [7, Lemma 4.2]). Let X be a stimulus space,
let U = {Ui}n
i=1 be a collection of sets in X, and consider the receptive field code C = C(U ). Then
for any pair of subsets σ, τ ⊆ [n],
xσY
(1 + xi) ∈ JC ⇐⇒ Uσ ⊆ [
Ui .
i∈τ
i∈τ
Thus, three types of receptive-field relationships (RF relationships) can be read off from pseudo-
monomials in a neural ideal (e.g., those in the canonical form) [7]:
Type 1: xσ ∈ JC ⇐⇒ Uσ = ∅ (where σ 6= ∅).
Type 2: xσQi∈τ (1 + xi) ∈ JC ⇐⇒ Uσ ⊆ Si∈τ Ui (where σ, τ 6= ∅).
Type 3: Qi∈τ (1 + xi) ∈ JC ⇐⇒ X ⊆ Si∈τ Ui (where τ 6= ∅), and thus X = Si∈τ Ui.
Example 2.7. The canonical form in Example 2.5, which is {x2(1 + x1), x2(1 + x3)}, encodes two
Type 2 relationships: U2 ⊆ U1 and U2 ⊆ U3. Indeed, we can verify this in Figure 1.
In this work, we reveal more types of RF relationships, which arise from non-pseudo-monomials.
They often appear in Grobner bases of neural ideals (see Section 5).
2.4. Grobner bases. Here we recall some basics about Grobner bases [2, 4, 12].
Fix a monomial ordering < of a polynomial ring R = k[x1, . . . , xn] over a field k, and let I be
an ideal in R. Let LT<(I) denote the ideal generated by all leading terms, with respect to the
monomial ordering <, of elements in I.
Definition 2.8. A Grobner basis of I, with respect to <, is a finite subset of I whose leading
terms generate LT<(I).
One useful property of a Grobner basis is that given a polynomial f and a Grobner basis G, the
remainder of f when divided by the set of elements in G is uniquely determined.
A Grobner basis is reduced if (1) every f ∈ G has leading coefficient 1, and (2) no term of
any f ∈ G is divisible by the leading term of any g ∈ G for which g 6= f . For a given monomial
ordering, the reduced Grobner basis of an ideal is unique.
GR OBNER BASES OF NEURAL IDEALS
5
Definition 2.9. A universal Grobner basis of an ideal I is a Grobner basis that is a Grobner
basis with respect to every monomial ordering. The universal Grobner basis of an ideal I is
the union of all the reduced Grobner bases of I.
The universal Grobner basis is an instance of a universal Grobner basis, given that the set of
all distinct reduced Grobner bases of an ideal I is finite [2, pg. 515]. This fact is actually the main
result of the theory of Grobner fans first introduced in [14].
3. Main Result
In this section, we give the main result of our paper: if the canonical form is a Grobner basis,
then it is the universal Grobner basis (Theorem 3.1). Beyond being a natural expansion of some
of Curto et al.'s results [7], our theorem is also of mathematical interest since there are few classes
of ideals whose universal Grobner bases are known. Indeed, such characterizations in general are
known to be computationally difficult.
Theorem 3.1. If the canonical form of a neural ideal JC is a Grobner basis of JC with respect to
some monomial ordering, then it is the universal Grobner basis of JC.
The proof of Theorem 3.1, which appears in Section 3.3, requires the following related results:
Lemma 3.2. For a pseudo-monomial f = xσQj∈τ (1 + xj) in F2[x1, . . . , xn], the leading term of f
with respect to any monomial ordering is its highest-degree term, xσ∪τ .
Proof. This follows from the fact that every term of f divides xσ∪τ , and two properties of a mono-
mial ordering [4]: it is a well-ordering (so, 1 < xi), and xα < xβ implies xα∪γ < xβ∪γ.
(cid:3)
Proposition 3.3. If the canonical form of a neural ideal JC is a Grobner basis of JC with respect
to some monomial ordering, then it is a universal Grobner basis of JC.
Proof. Let G denote the canonical form, and assume that G is a Grobner basis with respect to
some monomial ordering <1. Let <2 denote another monomial ordering. As always, we have the
containment LT<2(G) ⊆ LT<2(JC ), which we must prove is an equality. Accordingly, let f ∈ JC .
We must show that LT<2(f ) ∈ LT<2(G). With respect to <1, the reduction of f by G is 0, so we
can write f as a polynomial combination of some of the gi ∈ G in the following form:
(1)
f =
LT<1(f )
LT(g1)
g1 +
LT<1(r1)
LT(g2)
g2 + · · · +
LT<1(rt−1)
LT(gt)
gt = h1 + · · · + ht ,
LT(gi)
where (for i = 1, . . . , t) we have gi ∈ G, hi := LT<1 (ri−1)
gi, r0 := f , and ri = f − h1 − · · · − hi is
the remainder after the i-th division of f by G. Note that in equation (1), the polynomial gi may
appear multiple times, but this does not affect our arguments. By Lemma 3.2, the leading term of
gi does not depend on the monomial ordering. Moreover, each hi is the product of a monomial and
a pseudo-monomial, gi, so by a straightforward generalization of Lemma 3.2, the leading term of hi
with respect to any monomial ordering is LT<1(hi). Also note that when dividing by the Grobner
basis G, LT<1(ri) <1 LT<1(ri−1) so the LT<1(ri) are distinct. This implies that the LT<1(hi) are
distinct since LT<1(hi) = LT<1(ri−1).
Hence, among the list of monomials {LT(hi)}t
i=1, there is a unique largest monomial with respect
to <2, which we denote by LT(hi∗ ). Next, by examining the sum in (1), and noting that every term
of hi divides the leading term of hi, we see that LT<2(f ) = LT(hi∗ ). Thus, because gi∗ divides hi∗ ,
it follows that LT(gi∗ ) divides LT<2(f ), and so, LT<2(f ) ∈ LT<2(G).
Thus, if the canonical form is a Grobner basis with respect to some monomial ordering, then it
(cid:3)
is a Grobner basis with respect to every monomial ordering.
6
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
3.1. Pseudo-monomials and hypercubes. To prove our main result (Theorem 3.1), we need
to develop the connection between pseudo-monomials and hypercubes in the Boolean lattice. The
Boolean lattice on [n] is the power set P ([n]) := 2[n], partially ordered by inclusion. Also, for
σ ⊆ [n], we let P (σ) denote the power set of σ.
The support of a monomial Qn
Definition 3.4. Let f = xσQj∈τ (1+xj) be a pseudo-monomial in F2[x1, . . . , xn]. The hypercube
of f , denoted by H(f ), is the sublattice of the Boolean lattice on [n] formed by the support of each
term of f .
is the set {i ∈ [n] ai > 0}.
i=1 xai
i
Remark 3.5. The hypercube of f is the interval of the Boolean lattice from σ to σ ∪ τ :
H(f ) = {ω σ ⊆ ω ⊆ σ ∪ τ } ⊆ P ([n]) ,
and thus its Hasse diagram is a hypercube (this justifies its name). This is because:
f = xσ Y
(1 + xj) = X
xσ∪θ .
j∈τ
{θθ⊆τ }
Example 3.6. Let f = x1x2(1 + x3)(1 + x4) = x1x2x3x4 + x1x2x3 + x1x2x4 + x1x2. Figure 2 shows
part of the Hasse diagram of P ([4]), with the hypercube of f indicated by circles and solid lines.
1234
123
124
134
234
12
13
14
1
2
23
3
∅
24
34
4
Figure 2. Displayed is part of the Hasse diagram of the Boolean lattice P ([4]).
The hypercube of f = x1x2(1 + x3)(1 + x4) is indicated by circles and solid lines,
and P ([2]) is marked by dotted lines. If g is a pseudo-monomial that divides f , then
its hypercube is contained in either the hypercube of f or one of the dashed-line
squares "parallel" to the hypercube of f (see Example 3.8).
Via hypercubes, divisibility of pseudo-monomials has a nice geometric interpretation:
Lemma 3.7. For pseudo-monomials f = xσQj∈τ (1 + xj) and g = xαQj∈β(1 + xj), the following
are equivalent:
(1) gf ,
(2) α ⊆ σ and β ⊆ τ ,
(3) H(g) ⊆ P (σ ∪ τ ) and H(g) ∩ P (σ) = {α}, and
(4) H(g) ⊆ P (σ ∪ τ ) and H(g) ∩ P (σ) = 1.
Proof. The implication (1) ⇐ (2) is clear, and (1) ⇒ (2) follows from the fact that F2[x1, . . . , xn]
is a unique factorization domain. For (2) ⇒ (3), assume that α ⊆ σ and β ⊆ τ . Then H(g) ⊆
P (α ∪ β) ⊆ P (σ ∪ τ ). So, we need only show that H(g) ∩ P (σ) = {α}. To see this, we first recall:
(2)
H(g) = {α ∪ θ θ ⊆ β}
GR OBNER BASES OF NEURAL IDEALS
7
from Remark 3.5. Thus,
H(g) ∩ P (σ) = {α ∪ θ θ ⊆ β and θ ⊆ σ} = {α} ,
where the second equality follows from hypotheses: α ⊆ σ and σ ∩ β ⊆ σ ∩ τ = ∅ (because β ⊆ τ ).
(3) ⇒ (4) is clear, so we need only show (2) ⇐ (4). Accordingly, suppose H(g) ⊆ P (σ ∪ τ ) and
I := H(g) ∩ P (σ) consists of only one element. We claim that this element is α. Indeed, let ω ∈ I
(i.e., ω ∈ H(g) and ω ⊆ σ); then, α also is in I (because α ∈ H(g) and α ⊆ ω ⊆ σ). So, α = ω ⊆ σ.
To complete the proof, we must show that β ⊆ τ . To this end, let k ∈ β. Then α ∪ {k} is in
H(g), by equation (2), so it is not in P (σ) (because H(g) ∩ P (σ) = {α}). So, k ∈ (β \ σ). Finally,
(β \ σ) ⊆ τ , because α ∪ β ⊆ σ ∪ τ follows from the hypothesis H(g) ⊆ P (σ ∪ τ ). So, k ∈ τ .
(cid:3)
Example 3.8. We return to the pseudo-monomial f = x1x2(1 + x3)(1 + x4), which we rewrite as
f = xσQj∈τ (1 + xj), where σ = {1, 2} and τ = {3, 4}. In Figure 2, P (σ) = P ([2]) is marked by the
dotted line. According to Lemma 3.7, a pseudo-monomial h divides f if and only if the hypercube
of h satisfies two conditions: it includes a vertex from P (σ), and it is contained within either the
hypercube of f or one of the dashed-line squares "parallel" to the hypercube of f in Figure 2.
3.2. Multivariate division by pseudo-monomials. The following result concerns reducing a
given pseudo-monomial by a set of pseudo-monomials.
Theorem 3.9. Consider a pseudo-monomial f = xσQi∈τ (1 + xi) ∈ F2[x1, . . . , xn], and let G be a
finite set of pseudo-monomials in F2[x1, . . . , xn]. If some remainder upon division of f by G is 0
for some monomial ordering, then there exists g ∈ G such that g divides f .
Proof. Suppose that some remainder on division of f by G is 0:
(3)
f =
LT(f )
LT(g1)
g1 +
LT(r1)
LT(g2)
g2 + · · · +
LT(rt−1)
LT(gt)
gt = h1 + · · · + ht ,
where, as in the proof of Proposition 3.3, for i = 1, . . . , t, we have gi ∈ G, hi := LT(ri−1)
LT(gi) gi, and
ri = f − h1 − · · · − hi is the remainder after the i-th division (and r0 := f ). Also, each term of hi
divides the leading term of hi.
By construction, gihi. So, it suffices to show that there exists i such that hif .
We now claim that LT(hi)LT(f ) holds for all i. We prove this claim by induction on i. For the
i = 1 case, LT(h1) = LT(f ). If i ≥ 2, then LT(hi) is the leading term of:
(4)
ri−1 = f − h1 − · · · − hi−1 .
We now examine the summands in (4). As f is a pseudo-monomial, each term in f divides LT(f ),
and the same holds for each remaining summand hi: as noted above, its terms divide LT(hi), and
thus (by induction hypothesis) divide LT(f ). So, LT(hi) = LT(ri−1)LT(f ), proving our claim.
We now assert that hi is a pseudo-monomial. To see this, recall that hi is the product of a
monomial and a pseudo-monomial (namely, gi), so we just need to show that its leading term is
square-free. Indeed, this follows from two facts: LT(hi)LT(f ) and f is a pseudo-monomial.
Hence, H(hi) ⊆ P (σ ∪ τ ) for every i, because every term in hi divides LT(hi) which in turn
divides xσ∪τ = LT(f ). Thus, by Lemma 3.7, it is enough to show that H(hi) ∩ P (σ) = 1 for some
i (because this would imply that hif ).
The sum in (3) is over F2, so the polynomials f, h1, . . . , ht together must contain an even number
of each term. We focus now on only those terms with support in P (σ). The pseudo-monomial f
has only one such term (namely, xσ). Thus, some hi∗ has an odd number of terms in P (σ), i.e.,
H(hi∗ ) ∩ P (σ) is odd. On the other hand, both H(hi∗ ) and P (σ) are hypercubes in the Boolean
lattice, so their intersection, if nonempty, also is a hypercube and thus has size 2q for some q ≥ 0.
Hence, q = 0, so H(hi∗ ) ∩ P (σ) = 1. This completes our proof.
(cid:3)
8
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
3.3. Proof of Theorem 3.1. Theorem 3.9 allows us to prove that when a canonical form is a
Grobner basis, it is reduced:
Proposition 3.10. If the canonical form of a neural ideal JC is a Grobner basis of JC, then it is
a reduced Grobner basis of JC .
Proof. Suppose for contradiction that CF(JC ) is a Grobner basis, but not a reduced Grobner basis.
Then there exist f, g ∈ CF(JC ), with f 6= g, such that LT(g) divides some term of f . Thus, LT(g)
divides LT(f ) (because every term in a pseudo-monomial divides the leading term). Thus, CF(JC )
and CF(JC ) \ {f } both generate the same ideal of leading terms, and hence CF(JC ) \ {f } is also
a Grobner basis of JC . It follows that the remainder on division of f by CF(JC ) \ {f } is 0, so by
Theorem 3.9, there exists h ∈ CF(JC ) \ {f } such that hf . Hence, f is a non-minimal element of
the canonical form, which is a contradiction.
(cid:3)
Now we can prove Theorem 3.1, which states that a canonical form that is a Grobner basis is
the universal Grobner basis:
Proof of Theorem 3.1. Follows from Propositions 3.3 and 3.10.
(cid:3)
3.4. Every pseudo-monomial in a reduced Grobner basis is in the canonical form. In this
subsection, we prove the following partial converse of Theorem 3.1: if the universal Grobner basis of
a neural ideal consists of only pseudo-monomials, then it equals the canonical form (Theorem 3.12).
We first show that every pseudo-monomial in a reduced Grobner basis is in the canonical form.
Proposition 3.11. Let JC be a neural ideal.
(1) Let G be a reduced Grobner basis of JC. Then every pseudo-monomial in G is in the
canonical form of JC.
(2) Let bG be the universal Grobner basis of JC . Then every pseudo-monomial in bG is in the
canonical form of JC.
Proof. Let f be a pseudo-monomial in G. Suppose that f is not a minimal pseudo-monomial in
for some pseudo-monomial h ∈ JC such that deg(h)<deg(f ), hf . Then for some g ∈ G,
JC :
LT(g)LT(h). Hence, LT(g)LT(f ) (because LT(h)LT(f )) and also g 6= f (because deg(g) ≤
deg(h) < deg(f )). This is a contradiction: f and g cannot both be in a reduced Grobner basis.
Finally, (2) follows directly from (1).
(cid:3)
Theorem 3.12. Let JC be a neural ideal. The following are equivalent:
(1) the canonical form of JC is a Grobner basis of JC ,
(2) the canonical form of JC is the universal Grobner basis of JC, and
(3) the universal Grobner basis of JC consists of pseudo-monomials.
Proof. The implication (1)⇒(2) is Theorem 3.1, and both (1)⇐(2) and (2)⇒(3) are clear. For
(3)⇒(1), assume that the universal Grobner basis bG consists of pseudo-monomials. Then, by
Proposition 3.11(2), bG is contained in the canonical form of JC. Thus, the canonical form contains
a Grobner basis of JC (namely, bG) and hence is itself a Grobner basis.
Remark 3.13. Suppose we want to know whether a code's canonical form is a Grobner basis.
Theorem 3.12 tells us how to do so without computing the canonical form: compute the universal
Grobner basis, and then check whether it contains only pseudo-monomials. See Example 3.14.
(cid:3)
Under certain conditions, e.g. small number of neurons, computing the Grobner basis is more
efficient than computing the canonical form, but is there some way to avoid computations entirely
and yet still decide whether the canonical form is a Grobner basis? In the next section, we give
conditions under which we can resolve this decision problem quickly.
Example 3.14. Consider the neural code C = {0100, 0101, 0111}. The universal Grobner basis of
GR OBNER BASES OF NEURAL IDEALS
9
Example 3.15. Consider the neural code C = {0101, 1100, 1110}. The universal Grobner basis
JC is bG = {x3(x4 + 1), x2 + 1, x1}, so it contains only pseudo-monomials. Thus, by Theorem 3.12,
bG is the canonical form.
of JC is bG = {x4x3, x3(x1 + 1), x1 + x4 + 1, x2 + 1}, which contains the non-pseudo-monomial
x1 + x4 + 1. Thus, by Theorem 3.12, the canonical form is not a universal Grobner basis of JC .
Indeed, the canonical form is CF(JC) = {x3(x1 + 1), x2 + 1, (x4 + 1)(x1 + 1), x4x1, x4x3}, and,
for a monomial ordering where x4 > x1, the leading term of the non-pseudo-monomial x1 + x4 + 1
is x4, which is not divisible by any of the leading terms from the canonical form.
4. When is the canonical form a Grobner basis?
In this section we present some results that partially solve the question of when is the canonical
form a Grobner basis for the neural ideal. A complete answer to this question is not only of
theoretical interest but perhaps also of practical relevance. Extensive computations suggest that,
under certain conditions, Grobner bases of neural ideals can be computed more efficiently than
canonical forms. This is true for small neural codes. Moreover, the iterative nature of the newer
canonical form algorithm hints towards the ability to compute canonical forms and Grobner bases of
neural codes in large dimensions by 'gluing' those of codes on small dimensions. Such decomposition
results are a common theme in other areas of applied algebraic geometry such as algebraic statistics
and phylogenetic algebraic geometry [1, 9].
Table I displays a runtime comparison between the iterative canonical form algorithm described
in [16] and a specialized Grobner basis algorithm for Boolean rings implemented in SageMath based
on the work in [3]. We report the mean time (in seconds) of 100 randomly generated codes on n
neurons for n = 4, . . . , 8. More precisely, for each code, a number m was chosen uniformly at
random from {1, . . . , 2n − 1} and then m codewords were chosen at random. These computations
were performed on SageMath 7.2 running on a Macbook Pro with a 2.8 GHz Intel Core i7 processor
and 16 GB of memory.
Dimension
Canonical form 0.0016
Grobner basis
4
5
0.0076
6
0.108
7
0.621
8
1.964
0.00147 0.00202 0.00496 0.01604 0.16638
Table I. Runtime comparison of canonical form versus Grobner basis computations.
For codes on a larger number of neurons, our computations indicate that in general Grobner
bases computations are still more efficient than canonical form computations. However, even in the
case of n = 9 neurons we found codes whose Grobner bases took over 6 hours to be computed.
Proposition 4.1. Let C be a neural code on n neurons. If C = 1 or C = 2n − 1, then the
canonical form of JC is the universal Grobner basis of JC.
. . . , xn − cni. When
Proof. If C = {c}, then Lemma 2.3 implies that JC = hx1 − c1, x2 − c2,
C = 2n − 1, then by definition JC = hρvi for the unique v /∈ C. In either case, the indicated
generating set is both the canonical form and the universal Grobner basis of JC.
(cid:3)
A set of subsets ∆ ⊆ 2[n] is an (abstract) simplicial complex if σ ∈ ∆ and τ ⊆ σ implies
τ ∈ ∆. A neural code C is a simplicial complex if its support supp(C) is a simplicial complex.
Proposition 4.2. If C is a simplicial complex, then the canonical form of JC is the universal
Grobner basis of JC .
10
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
Proof. If C is a simplicial complex, then JC is a monomial ideal generated by the minimal Type 1
relationships (indeed, it is the Stanley-Reisner ideal of the simplicial complex supp(C)) [7, Lemma
4.4]. These minimal Type-1 relationships comprise the canonical form of JC, and also form the
universal Grobner basis of JC .
(cid:3)
The next result gives conditions that guarantee that the canonical form is not a Grobner basis.
Proposition 4.3. Let U = {Ui}n
denote the corresponding receptive field code.
canonical form of JC is not a Grobner basis of JC :
i=1 be a collection of sets in a stimulus space X, and let C = C(U )
If one of the following conditions hold, then the
(1) Two proper, nonempty receptive fields coincide: ∅ 6= Ui = Uj ( X for some i 6= j ∈ [n].
(2) Two nonempty receptive fields are complementary: Ui = X \ Uj for some i 6= j ∈ [n] with
Ui 6= ∅ and Uj 6= ∅.
Proof. (1) Suppose Ui, Uj ∈ U are two sets with ∅ 6= Ui = Uj ( X. By Lemma 2.6, both
f = xi(xj + 1) and g = xj(xi + 1) are in JC. In fact, f and g are minimal pseudo-monomials in JC
(because ∅ 6= Ui = Uj 6= X), so f, g ∈ CF(JC ). Under any monomial ordering, LT(f ) = LT(g) =
xixj (by Lemma 3.2), so the set CF(JC ) is not reduced and thus cannot be a reduced Grobner
basis. Hence, by Proposition 3.10, CF(JC ) cannot be a Grobner basis.
(2) Now assume that Ui = X \Uj for some i 6= j ∈ [n], with Ui 6= ∅ and Uj 6= ∅. Thus, Ui ∩Uj = ∅
and Ui ∪ Uj = X, so Lemma 2.6 implies that f = xixj and g = (xi + 1)(xj + 1) are in JC . Now
we proceed as in the previous paragraph: f and g are minimal pseudo-monomials in CF(JC ), and
LT(f ) = LT(g) = xixj, so, by Proposition 3.10, CF(JC ) cannot be a Grobner basis.
(cid:3)
The last result in this section concerns a class of codes that we call complement-complete.
Definition 4.4. The complement of c ∈ {0, 1}n is the codeword c ∈ {0, 1}n defined by ci = 1 if
and only if ci = 0. A neural code C is complement-complete if for all c ∈ C, then c is also in C.
Example 4.5. The complement of the codeword c1 = 1000 is c1 = 0111, and the complement of
c2 = 1010 is c2 = 0101. Thus, the code C = {1000, 0111, 1010, 0101} is complement-complete.
Definition 4.6. The complement of a pseudo-monomial f = xσQi∈τ (1 + xi) is the pseudo-
monomial f = xτ Qj∈σ(1 + xj).
Lemma 4.7. Consider pseudo-monomials f = xσQi∈τ (1 + xi) and g = xσ′ Qi∈τ ′(1 + xi). If f
divides g, then f divides g.
Proof. This follows from the fact that f g if and only if σ′ ⊆ σ and τ ′ ⊆ τ (Lemma 3.7).
(cid:3)
Proposition 4.8. Let C be a code on n neurons, with C ( {0, 1}n. If C is complement-complete,
then the canonical form of JC is not a Grobner basis of JC.
Proof. Note that since C 6= {0, 1}n, JC is not trivial. We make the following claim:
Claim: If h is a pseudo-monomial in JC , then h is also in JC.
To see this, let S be the set of all degree-n pseudo-monomials in F2[x1, . . . , xn] that are multiples
of h (so, S ⊆ JC). Degree-n pseudo-monomials in F2[x1, . . . , xn] are characteristic functions ρv, so,
every element of S is some ρv, where v /∈ C. Thus, every element of S := {f f ∈ S} has the form
ρv = ρv, where v /∈ C, which is equivalent to v /∈ C, as C is complement-complete. So, S ⊆ JC .
Next, let s ∈ S, that is, s = hq for some pseudo-monomial q. Then hq is also in S. Since
gcd(q, q) = 1, it follows that h = gcd(hq, hq), so h = gcd{S}. Thus, h = gcd{S}, so h ∈ JC
(because S ⊆ JC), which proves the claim.
Now let f ∈ CF (JC ). By the claim, f is in JC, and now we assert that, like f , the pseudo-
monomial f is in CF (JC ). Indeed, if a pseudo-monomial d in JC divides f , then by Lemma 4.7,
the pseudo-monomial d divides f . Also, d ∈ JC (by the claim), so d = f (because f is minimal),
GR OBNER BASES OF NEURAL IDEALS
11
and thus d = f . Hence, f is minimal, and so f is also in CF (JC). Thus, CF (JC) contains two
polynomials (f and f ) with the same leading term, and so is not a reduced Grobner basis, and thus
(by Proposition 3.10) is not a Grobner basis of JC.
(cid:3)
Example 4.9. Consider again the complement-complete code C = {1000, 0111, 1010, 0101} from
Example 4.5. The canonical form is CF (JC) = {(x1 + 1)(x2 + 1), (x1 + 1)(x4 + 1), x1x2, x2(x4 +
1), x1x4, x4(x2 +1)}. Note that CF (JC) is itself complement-complete; for example, f = x2(x4 +1)
and f = x4(x2 + 1) are both in CF (JC ). Also, we can show directly that CF (JC) is not a Grobner
basis, which is consistent with Proposition 4.8: with respect to any monomial ordering, the leading
term of f + f = x2 + x4 is not divisible by any of the leading terms in CF (JC ).
5. New receptive-field relationships
We saw earlier that if the universal Grobner basis of a neural ideal consists of only pseudo-
monomials, then it equals the canonical form (Theorem 3.12). When this is not the case, there are
non-pseudo-monomial elements in the universal Grobner basis, so it is natural to ask what they tell
us about the receptive fields of the code. In other words, what types of RF relationships, besides
those of Types 1–3 (Lemma 2.6), appear in Grobner bases? Here we give a partial answer:
Theorem 5.1. Let U = {Ui}n
i=1 be a collection of sets in a stimulus space X. Let C = C(U )
denote the corresponding receptive field code, and let JC denote the neural ideal. Then for any
subsets σ1, σ2, τ1, τ2 ⊆ [n], and m indices 1 ≤ i1 < i2 < · · · < im ≤ n, with m ≥ 2, we have RF
relationships as follows:
Type 4: xσ1 Qi∈τ1(1 + xi) + xσ2 Qj∈τ2(1 + xj) ∈ JC ⇒ Uσ1 ∩(cid:0)Ti∈τ1 U c
Type 5: xi1 + · · · + xim ∈ JC ⇒ Uik ⊆ Sj∈[m]\{k} Uij for all k = 1, . . . , m, and if, additionally, m is
Type 6: xi1 + · · · + xim + 1 ∈ JC ⇒ Sm
i(cid:1) = Uσ2 ∩(cid:16)Tj∈τ2 U c
j(cid:17).
odd, then Tm
k=1 Uik = X.
k=1 Uik = ∅.
Proof. Throughout the proof, for p ∈ X, we let c(p) denote the corresponding codeword in C.
Type 4. Let f1 := xσ1 Qi∈τ1(1 + xi), and let f2 := xσ2 Qj∈τ2(1 + xj). Also, let W1 := Uσ1 ∩
j(cid:17). By symmetry, we need only show that W1 ⊆ W2. To
(cid:0)Ti∈τ1 U c
i(cid:1), and let W2 := Uσ2 ∩(cid:16)Tj∈τ2 U c
this end, let p ∈ W1 (so, c(p) ∈ C). First, because f1 + f2 ∈ JC and V (JC ) = C, it follows that
f1(c(p)) = f2(c(p)). Next, for i = 1, 2, we have p ∈ Wi if and only if fi(c(p)) = 1. Thus, p ∈ W2.
Type 5. Let g := xi1 + · · · + xim. By symmetry, we need only show that Ui1 ⊆ Sm
end, let p ∈ Ui1 (so, c(p)i1 = 1). Then g ∈ JC implies the following equality in F2:
l=2 Uil. To this
(5)
0 = g(c(p)) = c(p)i1 + c(p)i2 + · · · + c(p)im = 1 + c(p)i2 + · · · + c(p)im .
Now assume, additionally, that m is odd. Suppose, for contradiction, that there exists q ∈
k=1 Uik . Then, like the sum (5) above, we have 0 = g(c(q)) = 1 + · · · + 1 = m, which contradicts
Thus, for some k ≥ 2, we have c(p)ik = 1, i.e., p ∈ Uik . Hence, p ∈ Sm
Tm
the hypothesis that m is odd. So, Tm
Type 6. Let h := xi1 +· · ·+xim +1. Let p ∈ X (so, c(p) ∈ C). We must show that p ∈ Sm
c(p)ik = 1, i.e., p ∈ Uik . Hence, p ∈ Sm
k=1 Uik .
Because h ∈ JC , we have 0 = h(c(p)) = c(p)i1 + · · · + c(p)im + 1. Thus, for some k ∈ [m], we have
(cid:3)
k=1 Uik = ∅.
k=1 Uik .
l=2 Uil.
Remark 5.2. Like the earlier RF relationships (those of Types 1–3 from Lemma 2.6), some of our
new ones (Types 4–6) are containments and some are equalities.
Example 5.3. Recall the code C = {0101, 1100, 1110}, from Example 3.15, for which the universal
Grobner basis of JC is bG = {x4x3, x3(x1 + 1), x1 + x4 + 1, x2 + 1}. The polynomial x1 + x4 + 1
encodes a Type 6 relationship: U1 ∪ U4 = X. Also, the polynomial x2 + 1 encodes a Type 3
12
GARCIA, GARC´IA PUENTE, KRUSE, LIU, MIYATA, PETERSEN, PHILLIPSON, AND SHIU
relationship: U2 = X, which together gives us U1 ∪ U4 = U2. The canonical form also contains the
polynomial x1x4, which encodes a Type 1 relationship: U1 ∩U4 = ∅. We conclude that U1 ∪U4 = U2.
Example 5.4. Consider the code C = {00, 11}. The universal Grobner basis of C is bG = {x1(1 +
x1), x1 + x2, x2(1 + x2)}. The polynomial x1 + x2 encodes a Type 4 relationship: U1 = U2. (The
polynomial x1 + x2 also encodes Type 5 relationships.) This points to one of the advantages of our
new RF relationships: we can read off some set equalities more quickly than from the canonical
Indeed, the canonical form is CF(JC ) = {x1(1 + x2), x2(1 + x1)}, in which the Type 2
form.
relationships are U1 ⊆ U2 and U2 ⊆ U1 – and only from there do we infer the equality U1 = U2.
6. Discussion
In this work, we proved that if a code's canonical form is a Grobner basis of the neural ideal,
then it is the universal Grobner basis. Additionally, we gave conditions that guarantee or preclude
this situation, and found three new types of receptive-field relationships that arise in neural ideals.
Going forward, there are natural extensions to pursue:
(1) Give a complete characterization of codes for which the canonical form is a Grobner basis.
(2) Beyond those of Types 1–6, what other receptive-field relationships can be read off from a
Grobner basis, and what do they tell us about a code?
Solutions to these problems would help us extract information about the receptive-field structure
directly from the neural code.
Finally, we expect that our results can be used to improve canonical-form algorithms. Indeed, our
experiments indicate that under certain conditions, Grobner bases can be computed more efficiently
than canonical forms. Moreover, every pseudo-monomial in the universal Grobner basis of a neural
ideal is in the canonical form – so, that subset of the canonical form can be obtained directly from
the universal Grobner basis. And, in the case when the universal Grobner basis contains only
pseudo-monomials, then we can conclude immediately that the basis is in fact the canonical form.
Moreover, we hope to develop decomposition results to build canonical forms and Grobner basis of
codes in large dimensions by 'gluing' those of codes in smaller dimensions.
Acknowledgments. DM, RK, and EP conducted this research as part of the 2015 Pacific Under-
graduate Research Experience in Mathematics Interns Program funded by the NSF (DMS-1045147
and DMS-1045082) and the NSA (H98230-14-1- 0131), in which RG and LG served as mentors and
KP was a GTA. JL conducted this research as part of the 2016 NSF-funded REU in the Department
of Mathematics at Texas A&M University (DMS-1460766), in which AS served as mentor and KP
was a GTA. The authors thank Ihmar Aldana, Carina Curto, Vladimir Itskov, and Ola Sobieska
for helpful discussions. LG was supported by the Simons Foundation Collaboration grant 282241.
AS was supported by the NSF (DMS-1312473/DMS-1513364). The authors thank an anonymous
referee for helpful comments which improved this work.
References
[1] Elizabeth S. Allman and John A. Rhodes. Phylogenetic ideals and varieties for the general Markov model. Adv.
in Appl. Math., 40(2):127–148, 2008.
[2] Thomas Becker and Volker Weispfenning. Grobner bases: a computational approach to commutative algebra.
Graduate texts in mathematics. Springer, New York, 1993.
[3] Michael Brickenstein and Alexander Dreyer. PolyBoRi: a framework for Grobner-basis computations with
Boolean polynomials. J. Symbolic Comput., 44(9):1326–1345, 2009.
[4] David A. Cox, John Little, and Donal O'Shea. Ideals, varieties, and algorithms. Undergraduate Texts in Math-
ematics. Springer, fourth edition, 2015. An introduction to computational algebraic geometry and commutative
algebra.
[5] Joshua Cruz, Chad Giusti, Vladimir Itskov, and William Kronholm. On open and closed convex codes. Available
at arxiv:1609.03502.
GR OBNER BASES OF NEURAL IDEALS
13
[6] Carina Curto, Elizabeth Gross, Jack Jeffries, Katherine Morrison, Mohamed Omar, Zvi Rosen, Anne Shiu, and
Nora Youngs. What makes a neural code convex? SIAM Journal on Applied Algebra and Geometry, 1(1):222–238,
2017.
[7] Carina Curto, Vladimir Itskov, Alan Veliz-Cuba, and Nora Youngs. The neural ring: an algebraic tool for
analyzing the intrinsic structure of neural codes. Bull. Math. Biol., 75(9):1571–1611, 2013.
[8] Carina Curto and Nora Youngs. Neural ring homomorphisms and maps between neural codes. Available at
arXiv:1511.00255.
[9] Alexander Engstrom, Thomas Kahle, and Seth Sullivant. Multigraded commutative algebra of graph decompo-
sitions. J. Algebraic Combin., 39(2):335–372, 2014.
[10] Elizabeth Gross, Nida Kazi Obatake, and Nora Youngs. Neural ideals and stimulus space visualization. Adv.
Appl. Math., 95:65–95, 2018.
[11] R Amzi Jeffs, Mohamed Omar, Natchanon Suaysom, Aleina Wachtel, and Nora Youngs. Sparse neural codes
and convexity. Available at arXiv:1511.00283, 2015.
[12] Martin Kreuzer and Lorenzo Robbiano. Computational commutative algebra 1. Springer-Verlag, Berlin, 2008.
Corrected reprint of the 2000 original.
[13] Caitlin Lienkaemper, Anne Shiu, and Zev Woodstock. Obstructions to convexity in neural codes. Adv. Appl.
Math., 85:31–59, 2017.
[14] Teo Mora and Lorenzo Robbiano. The Grobner fan of an ideal. J. Symbolic Comput., 6(2-3):183–208, 1988.
Computational aspects of commutative algebra.
[15] John O'Keefe and Jonathan Dostrovsky. The hippocampus as a spatial map. preliminary evidence from unit
activity in the freely-moving rat. Brain research, 34(1):171–175, 1971.
[16] Ethan Petersen, Nora Youngs, Ryan Kruse, Dane Miyata, Rebecca Garcia, and Luis David Garcia Puente.
Neural ideals in SageMath. Available at arXiv:1609.09602, 2016.
(R. Garcia and L. D. Garc´ıa Puente) Department of Mathematics and Statistics, Sam Houston State
University, Huntsville, TX 77341-2206
E-mail address: [email protected]
E-mail address: [email protected]
(R. Kruse) Mathematics Department, Central College, Pella, IA 50219
E-mail address: [email protected]
(J. Liu) Department of Mathematics, Bard College, Annandale, NY 12504
E-mail address: [email protected]
(D. Miyata) Department of Mathematics, Willamette University, Salem, OR 97301
E-mail address: [email protected]
(E. Petersen) Department of Mathematics, Rose-Hulman Institute of Technology, Terre Haute, IN
47803
E-mail address: [email protected]
(K. Phillipson) Department of Mathematics, St. Edwards University, Austin, Texas 78704-6489
E-mail address: [email protected]
(A. Shiu) Department of Mathematics, Texas A&M University, College Station, TX 77843
E-mail address: [email protected]
|
1902.04704 | 2 | 1902 | 2019-03-01T00:19:24 | Neural network models and deep learning - a primer for biologists | [
"q-bio.NC",
"cs.LG",
"cs.NE"
] | Originally inspired by neurobiology, deep neural network models have become a powerful tool of machine learning and artificial intelligence, where they are used to approximate functions and dynamics by learning from examples. Here we give a brief introduction to neural network models and deep learning for biologists. We introduce feedforward and recurrent networks and explain the expressive power of this modeling framework and the backpropagation algorithm for setting the parameters. Finally, we consider how deep neural networks might help us understand the brain's computations. | q-bio.NC | q-bio | Neural network models and deep learning
-- a primer for biologists
Nikolaus Kriegeskorte1,2,3,4 and Tal Golan4
1Department of Psychology, 2Department of Neuroscience
3Department of Electrical Engineering
4Zuckerman Mind Brain Behavior Institute, Columbia University
[email protected], [email protected]
Originally inspired by neurobiology, deep neural network models have become a
powerful tool of machine learning and artificial intelligence. They can approximate
functions and dynamics by learning from examples. Here we give a brief
introduction to neural network models and deep learning for biologists. We
introduce feedforward and recurrent networks and explain the expressive power of
this modeling framework and the backpropagation algorithm for setting the
parameters. Finally, we consider how deep neural network models might help us
understand brain computation.
Neural network models of brain function
Brain function can be modeled at many different levels of abstraction. At one extreme,
neuroscientists model single neurons and their dynamics in great biological detail. At the
other extreme, cognitive scientists model brain information processing with algorithms
that make no reference to biological components. In between these extremes lies a model
class that has come to be called artificial neural network (Rumelhart et al., 1987; LeCun
et al., 2015; Schmidhuber, 2015; Goodfellow et al., 2016).
A biological neuron receives multiple signals through the synapses contacting its
dendrites and sends a single stream of action potentials out through its axon. The
conversion of a complex pattern of inputs into a simple decision (to spike or not to spike)
suggested to early theorists that each neuron performs an elementary cognitive function:
it reduces complexity by categorizing its input patterns (McCulloch and Pitts, 1943;
Rosenblatt, 1958). Inspired by this intuition, artificial neural network models are
composed of units that combine multiple inputs and produce a single output.
The most common type of unit computes a weighted sum of the inputs and transforms
the result nonlinearly. The weighted sum can be interpreted as comparing the pattern of
inputs to a reference pattern of weights, with the weights corresponding to the strengths
of the incoming connections. The weighted sum is called the preactivation. The strength
of the preactivation reflects the overall strength of the inputs and, more importantly, the
match between the input pattern and the weight pattern. For a given input strength
(measured as the sum of squared intensities), the preactivation will be maximal if the input
pattern exactly matches the weight pattern (up to a scaling factor).
Kriegeskorte & Golan (2019) Neural network models and deep learning
The preactivation forms the input to the unit's nonlinear activation function. The activation
function can be a threshold function (0 for negative, 1 for positive preactivations),
indicating whether the match is sufficiently close for the unit to respond (Rosenblatt,
1958). More typically, the activation function is a monotonically increasing function, such
as the logistic function (Figure 1) or a rectifying nonlinearity, which outputs the
preactivation if it is positive and zero otherwise. These latter activation functions have
non-zero derivatives (at least over the positive range of preactivations). As we will see
below, non-zero derivatives make it easier to optimize the weights of a network.
The weights can be positive or negative. Inhibition, thus, need not be relayed through a
separate set of inhibitory units, and neural network models typically do not respect Dale's
law (which states that a neuron performs the same chemical action at all of its synaptic
connections to other neurons, regardless of the identity of the target cell). In addition to
the weights of the incoming connections, each unit has a bias parameter: the bias is
added to the preactivation, enabling the unit to shift its nonlinear activation function
horizontally, for example moving the threshold to the left or right. The bias can be
understood as a weight for an imaginary additional input that is constantly 1.
Figure 1 Function approximation by a feedforward neural network. A feedforward neural
network with two input units (bottom), three hidden units (middle), and two output units (top). The
input patterns form a two-dimensional space. The hidden and output units here use a sigmoid
(logistic) activation function. Surface plots on the left show the activation of each unit as a function
of the input pattern (horizontal plane spanned by inputs x1 and x2). For the output units, the
preactivations are shown below the output activations. For each unit, the weights (arrow
thickness) and signs (black, positive; red, negative) of the incoming connections control the
orientation and slope of the activation function. The output units combine the nonlinear ramps
computed by the hidden units. Given enough hidden units, a network of this type can approximate
any continuous function to arbitrary precision.
2
Kriegeskorte & Golan (2019) Neural network models and deep learning
Neural networks are universal approximators
Units can be assembled into networks in many different configurations. A single unit can
serve as a linear discriminant of its input patterns. A set of units connected to the same
set of inputs can detect multiple classes, with each unit implementing a different linear
discriminant. For a network to discriminate classes that are not linearly separable in the
input signals, we need an intermediate layer between input and output units, called a
hidden layer (Figure 1).
If the units were linear -- outputting the weighted sum directly, without passing it through
a nonlinear activation function -- then the output units reading out the hidden units would
compute weighted sums of weighted sums and would, thus, themselves be limited to
weighted sums of the inputs. With nonlinear activation functions, a hidden layer makes
the network more expressive, enabling it to approximate nonlinear functions of the input,
as illustrated in Figure 1.
A feedforward network with a single hidden layer (Figure 1) is a flexible approximator of
functions that link the inputs to the desired outputs. Typically, each hidden unit computes
a nonlinear ramp, for example sigmoid or rectified linear, over the input space. The ramp
rises in the direction in input space that is defined by the vector of incoming weights. By
adjusting the weights, we can rotate the ramp in the desired direction. By scaling the
weights vector, we can squeeze or stretch the ramp to make it rise more or less steeply.
By adjusting the bias, we can shift the ramp forward or backward. Each hidden unit can
be independently adjusted in this way.
One level up, in the output layer, we can linearly combine the outputs of the hidden units.
As shown in Figure 1, a weighted sum of several nonlinear ramps produces a qualitatively
different continuous function over the input space. This is how a hidden layer of linear --
nonlinear units enables the approximation of functions very different in shape from the
nonlinear activation function that provides the building blocks.
It turns out that we can approximate any continuous function to any desired level of
precision by allowing a sufficient number of units in a single hidden layer (Cybenko, 1989;
Hornik et al., 1989). To gain an intuition of why this is possible, consider the left output
unit (y1) of the network in Figure 1. By combining ramps overlapping in a single region of
the input space, this unit effectively selects a single compact patch. We could tile the
entire input space with sets of hidden units that select different patches in this way. In the
output layer, we could then map each patch to any desired output value. As we move
from one input region to another, the network would smoothly transition between the
different output values. The precision of such an approximation can always be increased
by using more hidden units to tile the input space more finely.
Deep networks can efficiently capture complex functions
A feedforward neural network is called 'deep' when it has more than one hidden layer.
The term is also used in a graded sense, in which the depth denotes the number of layers.
We have seen above that even shallow neural networks, with a single hidden layer, are
universal function approximators. What, then, is the advantage of deep neural networks?
Deep neural networks can re-use the features computed in a given hidden layer in higher
hidden layers. This enables a deep neural network to exploit compositional structure in a
3
Kriegeskorte & Golan (2019) Neural network models and deep learning
function, and to approximate many natural functions with fewer weights and units (Lin et
al., 2017; Rolnick and Tegmark, 2017). Whereas a shallow neural network must piece
together the function it approximates, like a lookup table (although the pieces overlap and
sum), a deep neural network can benefit from its hierarchical structure. A deeper
architecture can increase the precision with which a function can be approximated on a
fixed budget of parameters and can improve the generalization after learning to new
examples.
Deep learning refers to the automatic determination of parameters deep in a network on
the basis of experience (data). Neural networks with multiple hidden layers are an old
idea and were a popular topic in engineering and cognitive science in the 1980s
(Rumelhart et al., 1987). Although the advantages of deep architectures were understood
in theory, the method did not realize its potential in practice, mainly because of insufficient
computing power and data for learning. Shallow machine learning techniques, such as
support vector machines (Cortes and Vapnik, 1995; Schölkopf and Smola, 2002), worked
better in practice and also lent themselves to more rigorous mathematical analysis. The
recent success of deep learning has been driven by a rise in computing power -- in
particular the advent of graphics processing units, GPUs, specialized hardware for fast
matrix -- matrix multiplication -- and web-scale data sets to learn from. In addition,
improved techniques for pretraining, initialization, regularization, and normalization, along
with the introduction of rectified linear units, have all helped to boost performance. Recent
work has explored a wide variety of feedforward and recurrent network architectures,
improving the state-of-the-art in several domains of artificial intelligence and establishing
deep learning as a central strand of machine learning in the last few years.
The function that deep neural networks are trained to approximate is often a mapping
from input patterns to output patterns, for example classifying natural images according
to categories, translating sentences from English to French, or predicting tomorrow's
weather from today's measurements. When the cost minimized by training is a measure
of the mismatch between the network's outputs and desired outputs (that is, the 'error'),
for a training set of example cases, the training is called supervised. When the cost
minimized by training does not involve prespecified desired outputs for a set of example
inputs, the training is called unsupervised.
Two examples of unsupervised learning are autoencoders and generative adversarial
networks. Autoencoder networks (Rumelhart et al., 1986; Hinton and Salakhutdinov,
2006) learn to transform input patterns into a compressed latent representation by
exploiting inherent statistical structure. Generative adversarial networks (Goodfellow et
al., 2014) operate in the opposite direction, transforming random patterns in a latent
representation into novel, synthetic examples of a category, such as fake images of
bedrooms. The generator network is trained concurrently with a discriminator network that
learns to pick out the generator's fakes among natural examples of the category. The two
adversarial networks boost each other's performance by posing increasingly difficult
challenges of counterfeiting and detection to each other. Deep neural networks can also
be trained by reinforcement (deep reinforcement learning), which has led to impressive
performance at playing games and robotic control (Mnih et al., 2015).
4
Kriegeskorte & Golan (2019) Neural network models and deep learning
Deep learning by backpropagation
Say we want to train a deep neural network model with supervision. How can the
connection weights deep in the network be automatically learned? The weights are
randomly initialized and then adjusted in many small steps to bring the network closer to
the desired behavior. A simple approach would be to consider random perturbations of
the weights and to apply them when they improve the behavior. This evolutionary
approach is intuitive and has recently shown promise (Such et al., 2017; Stanley et al.,
2019), but it is not usually the most efficient solution. There may be millions of weights,
spanning a search space of equal dimension. It takes too long in practice to find directions
to move in such a space that improve performance. We could wiggle each weight
separately, and determine if behavior improves. Although this would enable us to make
progress, adjusting each weight would require running the entire network many times to
assess its behavior. Again, progress with this approach is too slow for many practical
applications.
In order to enable more efficient learning, neural network models are composed of
differentiable operations. How a small change to a particular weight affects performance
can then be computed as the partial derivative of the error with respect to the weight. For
different weights in the same model, the algebraic expressions corresponding to their
partial derivatives share many terms, enabling us to efficiently compute the partial
derivatives for all weights.
For each input, we first propagate the activation forward through the network, computing
the activation states of all the units, including the outputs. We then compare the network's
outputs with the desired outputs and compute the cost function to be minimized (for
example, the sum of squared errors across output units). For each unit, we then compute
how much the cost would drop if the activation changed slightly. This is the sensitivity of
the cost to a change of activation of each output unit. Mathematically, it is the partial
derivative of the cost with respect to the each activation. We then proceed backwards
through the network propagating the cost derivatives (sensitivities) from the activations to
the preactivations and through the weights to the activations of the layer below. The
sensitivity of the cost to each of these variables depends on the sensitivities of the cost
to the variables downstream in the network. Backpropagating the derivatives through the
network by applying the chain rule provides an efficient algorithm for computing all the
partial derivatives (Werbos, 1982).
The critical step is computing the partial derivative of the cost with respect to each weight.
Consider the weight of a particular connection (red arrow in Figure 2). The connection
links a source unit in one layer to a target unit in the next layer. The influence of the weight
on the cost for a given input pattern depends on how active the source unit is. If the source
unit is off for the present input pattern, then the connection has no signal to transmit and
its weight is irrelevant to the output the network produces for the current input. The
activation of the source unit is multiplied with the weight to determine its contribution to
the preactivation of the target unit, so the source activation is one factor determining the
influence of the weight on the cost. The other factor is the sensitivity of the cost to the
preactivation of the target unit. If the preactivation of the target unit had no influence on
the cost, then the weight would have no influence, either. The derivative of the cost with
respect to the weight is the product of its source unit's activation and its target unit's
influence on the cost.
5
Kriegeskorte & Golan (2019) Neural network models and deep learning
Figure 2 The backpropagation algorithm. Backpropagation is an efficient algorithm for
computing how small adjustments to the connection weights affect the cost function that the
network is meant to minimize. A feedforward network with two hidden layers is shown as an
example. First, the activations are propagated in the feedforward direction (upward). The
activation function (gray sigmoid) is shown in each unit (circle). In the context of a particular input
pattern (not shown), the network is in a particular activation state, indicated by the black dots in
the units (horizontal axis: preactivation, vertical axis: activation). Second, the derivatives of the
cost function (squared-error cost shown on the right) are propagated in reverse (downward). In
the context of the present input pattern, the network can be approximated as a linear network
(black lines indicating the slope of the activation function). The chain rule defines how the cost
(the overall error) is affected by small changes to the activations, preactivations, and weights. The
goal is to compute the partial derivative of the cost with respect to each weight (bottom right).
Each weight is then adjusted in proportion to how much its adjustment reduces the cost. The
notation roughly follows Nielsen (2015), but we use bold symbols for vectors and matrices. The
symbol ⨀ denotes element-wise multiplication (Hadamard product).
We adjust each weight in the direction that reduces the cost (the error) and by an amount
proportional to the derivative of the cost with respect to the weight. This process is called
gradient descent, because it amounts to moving in the direction in weight space in which
the cost declines most steeply. To help our intuition, let us consider two approaches we
might take. First, consider the approach of taking a step to reduce the cost for each
individual training example. Gradient descent will make a minimal and selective
adjustments to reduce the error, which makes sense as we do not want learning from the
current example to interfere with what we've learned from other examples. However, our
goal is to reduce the overall error, which is defined as the sum of the errors across all
examples. So second, consider the approach of summing up the error surfaces (or,
equivalently, the gradients) across all examples before taking a step. We can still only
take a small step, because the error surface is nonlinear and so the gradient will change
as we move away from the point about which we linearized the network.
In practice, the best solution is to use small batches of training examples to estimate the
gradient before taking a step. Compared to the single-example approach, this gives us a
more stable sense of direction. Compared to the full-training-set approach, it greatly
6
Kriegeskorte & Golan (2019) Neural network models and deep learning
reduces the computations required to take a step. Although the full-training-set approach
gives exact gradients for the training-set error, it still does not enable us to take large
steps, because of the nonlinearity of the error function. Using batches is a good
compromise between stability of the gradient estimate and computational cost. Because
the gradient estimate depends on the random sample of examples in the current batch,
the method is called stochastic gradient descent (SGD). Beyond the motivation just given,
the stochasticity is thought to contribute also to finding solutions that generalize well
beyond the training set (Poggio and Liao, 2017).
The cost is not a convex function of the weights, so we might be concerned about getting
stuck in local minima. However, the high dimensionality of weight space turns out to be a
blessing (not a curse) for gradient descent: there are many directions to escape in, making
it unlikely that we will ever find ourselves trapped, with the error surface rising in all
directions (Kawaguchi, 2016). In practice, it is saddle points (where the gradient vanishes)
that pose a greater challenge than local minima (Dauphin et al., 2014). Moreover, the
cost function typically has many symmetries, with any given set of weights having many
computationally equivalent twins (that is, the model computes the same overall function
for different parameter settings). As a result, although our solution may be one local
minimum among many, it may not be a poor local minimum: It may be one of many
similarly good solutions.
Recurrent neural networks are universal approximators of
dynamical systems
So far we have considered feedforward networks, whose directed connections do not
form cycles. Units can also be configured in recurrent neural networks (RNNs), where
activity is propagated in cycles, as is the case in brains (Dayan and Abbott, 2001;
Goodfellow et al., 2016). This enables a network to recycle its limited computational
resources over time and perform a deeper sequence of nonlinear transformations. As a
result, RNNs can perform more complex computations than would be possible with a
single feedforward sweep through the same number of units and connections.
For a given state space, a suitable RNN can map each state to any desired successor
state. RNNs, therefore, are universal approximators of dynamical systems (Schäfer and
Zimmermann, 2006). They provide a universal language for modeling dynamics, and one
whose components could plausibly be implemented with biological neurons.
Much like feedforward neural networks, RNNs can be trained by backpropagation.
However, backpropagation must proceed through the cycles in reverse. This process is
called backpropagation through time. An intuitive way to understand an RNN and
backpropagation through time is to 'unfold' the RNN into an equivalent feedforward
network (Figure 3). Each layer of the feedforward network represents a timestep of the
RNN. The units and weights of the RNN are replicated for each layer of the feedforward
network. The feedforward network, thus, shares the same set of weights across its layers
(the weights of the recurrent network).
7
Kriegeskorte & Golan (2019) Neural network models and deep learning
Figure 3 Recurrent neural networks. (a) A recurrent neural network models with two input
units (in blue box), three hidden units (green box), and two output units (pink box). The hidden
units here are fully recurrently connected: each sends its output to both other units. The arrows
represent scalar weights between particular units. (b) Equivalent feedforward network. Any
recurrent neural network can be unfolded along time as a feedforward network. To this end, the
units of the recurrent neural network (blue, green, pink sets) are replicated for each time step.
The arrows here represent weights matrices between sets of units in the colored boxes. For the
equivalence to hold, the feedforward network has to have a depth matching the number of time
steps that the recurrent network is meant to run for. Unfolding leads to a representation that is
less concise, but easier to understand and often useful in software implementations of recurrent
neural networks. Training of the recurrent model by backpropagation through time is equivalent
to training of the unfolded model by backpropagation.
For tasks that operate on independent observations (for example, classifying still images),
the recycling of weights can enable an RNN to perform better than a feedforward network
with the same number of parameters (Spoerer et al., 2017). However, RNNs really shine
in tasks that operate on streams of dependent observations. Because RNNs can maintain
an internal state (memory) over time and produce dynamics, they lend themselves to
tasks that require temporal patterns to be recognized or generated. These include the
perception speech and video, cognitive tasks that require maintaining representations of
hidden states of the agent (such as goals) or the environment (such as currently hidden
objects), linguistic tasks like the translation of text from one language into another, and
control tasks at the level of planning and selecting actions, as well as at the level of motor
control during execution of an action under feedback from the senses.
Deep neural networks provide abstract process models of
biological neural networks
Cognitive models capture aspects of brain information processing, but do not speak to its
biological implementation. Detailed biological models can capture the dynamics of action
potentials and the spatiotemporal dynamics of signal propagation in dendrites and axons.
However, they have only had limited success in explaining how these processes
contribute to cognition. Deep neural network models, as discussed here, strike a balance,
explaining feats of perception, cognition, and motor control in terms of networks of units
that are highly abstracted, but could plausibly be implemented with biological neurons.
For engineers, artificial deep neural networks are a powerful tool of machine learning. For
neuroscientists, these models offer a way of specifying mechanistic hypotheses on how
cognitive functions may be carried out by brains (Kriegeskorte and Douglas, 2018;
Kietzmann et al., 2019; Storrs and Kriegeskorte, in press). Deep neural networks provide
8
Kriegeskorte & Golan (2019) Neural network models and deep learning
a powerful language for expressing information-processing functions. In certain domains,
they already meet or surpass human-level performance (for example, visual object
recognition and board games) while relying exclusively on operations that are biologically
plausible.
Neural network models in engineering have taken inspiration from brains, far beyond the
general notion that computations involve a network of units, each of which nonlinearly
combines multiple inputs to compute a single output (Kriegeskorte, 2015; Yamins and
DiCarlo, 2016). For example, convolutional neural networks (Fukushima and Miyake,
1982; LeCun et al., 1989), the dominant technology in computer vision, use a deep
hierarchy of retinotopic layers whose units have restricted receptive fields. The networks
are convolutional in that weight templates are automatically shared across image
locations (rendering the computation of a feature map's preactivations equivalent to a
convolution of the input with the weight template). Although the convolutional aspect may
not capture an innate characteristic of the primate visual system, it does represent an
idealization of the final product of development and learning in primates, where
qualitatively similar features are extracted all over retinotopic maps at early stages of
processing. Across layers, these networks transform a visuospatial representation of the
image into a semantic representation of its contents, successively reducing the spatial
detail of the maps and increasing the number of semantic dimensions (Figure 4).
The fact that a neural network model was inspired by some abstract features of biology
and that it matches overall human or animal performance at a task, does not make it a
good model of how the human or animal brain performs the task. However, we can
compare neural network models to brains in terms of detailed patterns of behavior, such
as errors and reaction times for particular stimuli. Moreover, we can compare the internal
representations in neural networks to those in brains.
In the 'white-box' approach, we evaluate a model by looking at its internal representations.
Neural network models form the basis for predicting representations in different brain
regions for a particular set of stimuli (Diedrichsen and Kriegeskorte, 2017). One approach
is called encoding models. In encoding models, the brain activity pattern in some
functional region is predicted using a linear transformation of the representation in some
layer of the model (Kay et al., 2008; Mitchell et al., 2008). In another approach, called
representational similarity analysis (Kriegeskorte et al., 2008; Nili et al., 2014;
Kriegeskorte and Diedrichsen, 2016), each representation in brain and model is
characterized by a representational dissimilarity matrix. Models are evaluated according
to their ability to explain the representational dissimilarities across pairs of stimuli. A third
approach is pattern component modeling (Diedrichsen et al., 2011, 2017), where
representations are characterized by the second moment of the activity profiles.
Recent results from the domain of visual object recognition indicate that deep
convolutional neural networks are the best available model of how the primate brain
achieves rapid recognition at a glance, although they do not explain all of the explainable
variance in neuronal responses. (Cadieu et al., 2014; Khaligh-Razavi and Kriegeskorte,
2014; Yamins et al., 2014; Güçlü and van Gerven, 2015; Cichy et al., 2016; Eickenberg
et al., 2017; Wen et al., 2017; Nayebi et al., 2018), and yet multiple functional
incompatibilities have already been reported (Szegedy et al., 2013; Geirhos et al., 2017;
Jo and Bengio, 2017; Rajalingham et al., 2018).
9
Kriegeskorte & Golan (2019) Neural network models and deep learning
Figure 4 Deep convolutional feedforward neural networks. The general structure of Alexnet,
a convolutional deep neural network architecture which had a critical role in bringing deep neural
networks into the spotlight. Unlike the visualization in the original report on this model, here the
tensors' dimensions are drawn to scale, so it is easier to appreciate how the convolutional deep
neural network gradually transforms the input image from a spatial to semantic representation.
For sake of simplicity, we did not visualize the pooling operations, as well as the splitting of some
of these layers between two GPUs. The leftmost box is the input image, (a tensor of the
dimensions 227×227×3, where 227 is the length of the square input-image edges and three is the
number of color components). It is transformed by convolution into the first layer (second box from
the left), a tensor with smaller spatial dimensions (55×55) but a larger number of feature maps
(96). Each feature map in this tensor is produced by a convolution of the original image with a
particular 11×11×3 filter. Therefore, the preactivation of each unit in this layer is a linear
combination of one rectangular receptive field in the image. The boundaries of such a receptive
field are visualized as a small box within the image tensor. In the next, second layer, the
representation is even more spatially smaller (27×27) but richer with respect of the number of
feature maps (256). Note that from here and onwards, each feature is not a linear combination of
pixels but a linear combination of the previous layer's features. The sixth layer (see the small
overview inset at the top-right) combines all feature maps and locations of the fifth layer to yield
4096 different scalar units, each with its own unrestricted input weights vector. The final eighth
layer has 1000 units, one for each output class. The eight images on the bottom were produced
by gradually modifying random noise images so excite particular units in each of the eight layers
(Erhan et al., 2009). The rightmost image was optimized to activate the output neuron related to
the class 'Mosque'. Importantly, these are only local solutions to the activation-maximization
problem. Alternative activation-maximizing images may be produced by using different starting
conditions or optimization heuristics.
In the 'black-box' approach, we evaluate a model on the basis of its behavior. We can
reject models for failing to explain detailed patterns of behavior. This has already helped
reveal some limitations of convolutional neural networks, which appear to behave
differently from humans under noisy conditions (Geirhos et al., 2017) and to show
different patterns of failures across exemplars (Rajalingham et al., 2018).
Deep neural networks bridge the gap between neurobiology and cognitive function,
providing an exciting framework for modeling brain information processing. Theories of
how the brain computes can now be subjected to rigorous tests by simulation. Our
10
Kriegeskorte & Golan (2019) Neural network models and deep learning
theories, and the models that implement them, will evolve as we learn to explain the rich
measurements of brain activity and behavior provided by modern technologies in animals
and humans.
References
Cadieu CF, Hong H, Yamins DLK, Pinto N, Ardila D, Solomon EA, Majaj NJ, DiCarlo JJ
(2014) Deep neural networks rival the representation of primate IT cortex for core
visual object recognition. PLoS Comput Biol 10:e1003963.
Cichy RM, Khosla A, Pantazis D, Torralba A, Oliva A (2016) Comparison of deep neural
networks to spatio-temporal cortical dynamics of human visual object recognition
reveals hierarchical correspondence. Sci Rep 6:27755.
Cortes C, Vapnik V (1995) Support-vector networks. Mach Learn 20:273 -- 297.
Cybenko G (1989) Approximation by superpositions of a sigmoidal function. Math Control
Signals Systems 2:303 -- 314.
Dauphin YN, Pascanu R, Gulcehre C, Cho K, Ganguli S, Bengio Y (2014) Identifying and
Attacking the Saddle Point Problem in High-dimensional Non-convex Optimization.
In: Proceedings of the 27th International Conference on Neural Information
Processing Systems - Volume 2, pp 2933 -- 2941 NIPS'14. Cambridge, MA, USA: MIT
Press.
Dayan P, Abbott LF (2001) Chapter 7.4, Recurrent Neural networks. In: Theoretical
neuroscience. Cambridge, MA: MIT Press.
Diedrichsen J, Kriegeskorte N (2017) Representational models: A common framework for
representational-similarity
understanding encoding, pattern-component, and
analysis. PLoS Comput Biol 13:e1005508.
Diedrichsen J, Ridgway GR, Friston KJ, Wiestler T (2011) Comparing the similarity and
spatial structure of neural representations: a pattern-component model. Neuroimage
55:1665 -- 1678.
Diedrichsen J, Yokoi A, Arbuckle SA (2017) Pattern component modeling: A flexible
approach for understanding the representational structure of brain activity patterns.
Neuroimage Available at: http://dx.doi.org/10.1016/j.neuroimage.2017.08.051.
Eickenberg M, Gramfort A, Varoquaux G, Thirion B (2017) Seeing it all: Convolutional
network layers map the function of the human visual system. Neuroimage 152:184 --
194.
Erhan D, Bengio Y, Courville A, Vincent P (2009) Visualizing higher-layer features of a
deep network. University of Montreal 1341:1.
Fukushima K, Miyake S (1982) Neocognitron: A Self-Organizing Neural Network Model
for a Mechanism of Visual Pattern Recognition. In: Competition and Cooperation in
Neural Nets, pp 267 -- 285. Springer Berlin Heidelberg.
Geirhos R, Janssen DHJ, Schütt HH, Rauber J, Bethge M, Wichmann FA (2017)
Comparing deep neural networks against humans: object recognition when the signal
gets weaker. arXiv [csCV] Available at: http://arxiv.org/abs/1706.06969.
11
Kriegeskorte & Golan (2019) Neural network models and deep learning
Goodfellow I, Bengio Y, Courville A (2016) Deep Learning. MIT Press.
Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, Courville A,
Bengio Y (2014) Generative Adversarial Nets. In: Advances in Neural Information
Processing Systems 27 (Ghahramani Z, Welling M, Cortes C, Lawrence ND,
Weinberger KQ, eds), pp 2672 -- 2680. Curran Associates, Inc.
Güçlü U, van Gerven MAJ (2015) Deep Neural Networks Reveal a Gradient in the
Complexity of Neural Representations across the Ventral Stream. J Neurosci
35:10005 -- 10014.
Hinton GE, Salakhutdinov RR (2006) Reducing the dimensionality of data with neural
networks. Science 313:504 -- 507.
Hornik K, Stinchcombe M, White H (1989) Multilayer feedforward networks are universal
approximators. Neural Netw 2:359 -- 366.
Jo J, Bengio Y (2017) Measuring the tendency of CNNs to Learn Surface Statistical
Regularities. arXiv [csLG] Available at: http://arxiv.org/abs/1711.11561.
Kawaguchi K (2016) Deep Learning without Poor Local Minima. In: Advances in Neural
Information Processing Systems 29 (Lee DD, Sugiyama M, Luxburg UV, Guyon I,
Garnett R, eds), pp 586 -- 594. Curran Associates, Inc.
Kay KN, Naselaris T, Prenger RJ, Gallant JL (2008) Identifying natural images from
human brain activity. Nature 452:352 -- 355.
Khaligh-Razavi S-M, Kriegeskorte N (2014) Deep supervised, but not unsupervised,
models may explain IT cortical representation. PLoS Comput Biol 10:e1003915.
Kietzmann TC, McClure P, Kriegeskorte N (2019) Deep Neural Networks in
Computational Neuroscience. Available at:
http://oxfordre.com/neuroscience/view/10.1093/acrefore/9780190264086.001.0001/
acrefore-9780190264086-e-46.
Kriegeskorte N (2015) Deep Neural Networks: A New Framework for Modeling Biological
Vision and Brain Information Processing. Annu Rev Vis Sci 1:417 -- 446.
Kriegeskorte N, Diedrichsen J (2016) Inferring brain-computational mechanisms with
models of activity measurements. Philos Trans R Soc Lond B Biol Sci 371 Available
at: http://dx.doi.org/10.1098/rstb.2016.0278.
Kriegeskorte N, Douglas PK (2018) Cognitive computational neuroscience. Nat Neurosci
21:1148 -- 1160.
Kriegeskorte N, Mur M, Bandettini P (2008) Representational similarity analysis -
connecting the branches of systems neuroscience. Front Syst Neurosci 2:4.
Krizhevsky A, Sutskever I, Hinton GE (2012) ImageNet Classification with Deep
Convolutional Neural Networks. In: Advances in Neural Information Processing
Systems 25 (Pereira F, Burges CJC, Bottou L, Weinberger KQ, eds), pp 1097 -- 1105.
Curran Associates, Inc.
LeCun Y, Bengio Y, Hinton G (2015) Deep learning. Nature 521:436.
LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W, Jackel LD (1989)
Backpropagation Applied to Handwritten Zip Code Recognition. Neural Comput
12
Kriegeskorte & Golan (2019) Neural network models and deep learning
1:541 -- 551.
Lin HW, Tegmark M, Rolnick D (2017) Why Does Deep and Cheap Learning Work So
Well? J Stat Phys 168:1223 -- 1247.
McCulloch WS, Pitts W (1943) A logical calculus of the ideas immanent in nervous
activity. Bull Math Biophys 5:115 -- 133.
Mitchell TM, Shinkareva SV, Carlson A, Chang K-M, Malave VL, Mason RA, Just MA
(2008) Predicting human brain activity associated with the meanings of nouns.
Science 320:1191 -- 1195.
Mnih V, Kavukcuoglu K, Silver D, Rusu AA, Veness J, Bellemare MG, Graves A,
Riedmiller M, Fidjeland AK, Ostrovski G, Petersen S, Beattie C, Sadik A, Antonoglou
I, King H, Kumaran D, Wierstra D, Legg S, Hassabis D (2015) Human-level control
through deep reinforcement learning. Nature 518:529.
Nayebi A, Bear D, Kubilius J, Kar K, Ganguli S, Sussillo D, DiCarlo JJ, Yamins DL (2018)
Task-Driven convolutional recurrent models of the visual system. In: Advances in
Neural Information Processing Systems, pp 5295 -- 5306.
Nielsen MA (2015) Neural networks and deep learning. Determination Press.
Nili H, Wingfield C, Walther A, Su L, Marslen-Wilson W, Kriegeskorte N (2014) A toolbox
for representational similarity analysis. PLoS Comput Biol 10:e1003553.
Poggio T, Liao Q (2017) Theory ii: Landscape of the empirical risk in deep learning.
Rajalingham R, Issa EB, Bashivan P, Kar K, Schmidt K, DiCarlo JJ (2018) Large-scale,
high-resolution comparison of the core visual object recognition behavior of humans,
monkeys, and state-of-the-art deep artificial neural networks. bioRxiv:240614
Available at: https://www.biorxiv.org/content/early/2018/01/01/240614
[Accessed
April 16, 2018].
Rolnick D, Tegmark M (2017) The power of deeper networks for expressing natural
functions. arXiv [csLG] Available at: http://arxiv.org/abs/1705.05502.
Rosenblatt F (1958) The perceptron: a probabilistic model for information storage and
organization in the brain. Psychol Rev 65:386 -- 408.
Rumelhart DE, Hinton GE, Williams RJ (1986) Parallel Distributed Processing:
Explorations in the Microstructure of Cognition, Vol. 1. In (Rumelhart DE, McClelland
JL, PDP Research Group C, eds), pp 318 -- 362. Cambridge, MA, USA: MIT Press.
Rumelhart DE, McClelland JL, Group PR, Others (1987) Parallel distributed processing.
MIT press Cambridge, MA.
Schäfer AM, Zimmermann HG (2006) Recurrent Neural Networks Are Universal
Approximators. In: Artificial Neural Networks -- ICANN 2006 (Kollias SD, Stafylopatis
A, Duch W, Oja E, eds), pp 632 -- 640. Berlin, Heidelberg: Springer Berlin Heidelberg.
Schmidhuber J (2015) Deep learning in neural networks: An overview. Neural Netw
61:85 -- 117.
Schölkopf B, Smola AJ (2002) Learning with Kernels: Support Vector Machines,
Regularization, Optimization, and Beyond. MIT Press.
Spoerer CJ, McClure P, Kriegeskorte N (2017) Recurrent Convolutional Neural Networks:
13
Kriegeskorte & Golan (2019) Neural network models and deep learning
A Better Model of Biological Object Recognition. Front Psychol 8:1551.
Stanley KO, Clune J, Lehman J, Miikkulainen R (2019) Designing neural networks
through neuroevolution. Nature Machine Intelligence 1:24 -- 35.
Storrs KR, Kriegeskorte N (in press.) Deep learning for cognitive neuroscience. In: The
Cognitive Neurosciences (6th Edition). (Gazzaniga M, ed). Boston: MIT Press.
Such FP, Madhavan V, Conti E, Lehman J, Stanley KO, Clune J (2017) Deep
Neuroevolution: Genetic Algorithms Are a Competitive Alternative for Training Deep
Neural Networks
[csNE] Available at:
http://arxiv.org/abs/1712.06567.
for Reinforcement Learning. arXiv
Szegedy C, Zaremba W, Sutskever I, Bruna J, Erhan D, Goodfellow I, Fergus R (2013)
[csCV] Available at:
Intriguing properties of neural networks. arXiv
http://arxiv.org/abs/1312.6199.
Wen H, Shi J, Zhang Y, Lu K-H, Cao J, Liu Z (2017) Neural Encoding and Decoding with
Deep Learning for Dynamic Natural Vision. Cereb Cortex:1 -- 25.
Werbos PJ (1982) Applications of advances in nonlinear sensitivity analysis. In: System
Modeling and Optimization, pp 762 -- 770. Springer Berlin Heidelberg.
Yamins DLK, DiCarlo JJ (2016) Using goal-driven deep learning models to understand
sensory cortex. Nat Neurosci 19:356 -- 365.
Yamins DLK, Hong H, Cadieu CF, Solomon EA, Seibert D, DiCarlo JJ (2014)
Performance-optimized hierarchical models predict neural responses in higher visual
cortex. Proc Natl Acad Sci 111:8619 -- 8624.
14
|
1305.3495 | 1 | 1305 | 2013-05-15T14:37:05 | The overlap of neural selectivity between faces and words: evidences from the N170 adaptation effect | [
"q-bio.NC"
] | Faces and words both evoke an N170, a strong electrophysiological response that is often used as a marker for the early stages of expert pattern perception. We examine the relationship of neural selectivity between faces and words by using a novel application of cross-category adaptation to the N170. We report a strong asymmetry between N170 adaptation induced by faces and by words. This is the first electrophysiological result showing that neural selectivity to faces encompasses neural selectivity to words, and suggests that the N170 response to faces constitutes a neural marker for versatile representations of familiar visual patterns. | q-bio.NC | q-bio | The overlap of neural selectivity between faces and words: evidences
from the N170 adaptation effect
Xiao-hua Cao 1
， Chao Li 1 ， Carl M Gaspar 2 ， Bei Jiang 1
1. Department of Psychology,Zhejiang Normal University,Jinhua 321004,China;
2. Center for Cognition and Brain Disorders, Hangzhou Normal University,Hangzhou
310000,China.
Authors: Xiao-hua Cao, [email protected]
Chao Li, [email protected]
Gaspar M Carl, [email protected]
Bei Jiang ,[email protected]
Abstract: Faces and words both evoke an N170, a strong
electrophysiological response that is often used as a marker for the early
stages of expert pattern perception. We examine the relationship of neural
selectivity between faces and words by using a novel application of
cross-category adaptation to the N170. We report a strong asymmetry
between N170 adaptation induced by faces and by words. This is the first
electrophysiological result showing that neural selectivity to faces
encompasses neural selectivity to words, and suggests that the N170
response to faces constitutes a neural marker for versatile representations
of familiar visual patterns.
Key words: N170; adaptation; visual expertise; neural selectivity
Introduction
The human brain has limited resources for representing the visual
world [2]. It makes sense that different categories of images share their
visual representations to some degree. Many non-face expert stimuli
recruited the face-selective parts of the visual system [7-10,18]. Two of
the most familiar types of visual patterns are faces and written language.
Despite their difference in appearance, current studies suggested that
there may be a strong relationship between the visual representations for
faces and words. For example, reading ability enhanced the left fusiform
activation which induced a small competition with faces at this location
[3]. Recently, it was revealed that that faces and words both evoke an
N170, a strong electrophysiological response that is often used as a
marker for the early stages of expert pattern perception[1,14,17].These
evidences suggested that both faces and words do evoke a representation
that is shared to a certain degree. However, thus far, nothing is known
about the nature of overlap between electrophysiological selectivity to
faces and words. Behavior studies revealed that, although the configural
information critical for identity processing is different between faces and
words, there are many similarities of the two categories , for example,
extensive and long-term exposure, canonical upright orientation, high
level expertise, predominantly individual-identity level processing, and
featural information critical for identity processing, which may contribute
to shape the expert pattern perception[15]. Moreover，another study found
that the expertise increased the functional overlap between face and
car[16]. Thus, these evidences above may be implicated that the neural
selectivity of early face processing is overlap, even encompasses that of
words. For the first time, we used the cross-category adaptation paradigm
[13](face adaptor followed by word test and vice versa) to directly reveal
the relation of the early expert perception between faces and words.
Methods
Participants
Sixteen subjects (7 males) were recruited from local universities and
paid for their participation (age range: 19-29 years, mean 21.8 years). All
subjects were right handed with normal or corrected-to-normal vision.
Informed consent was obtained from all subjects, and the study was
approved by the ethical committee of Zhejiang Normal University.
Stimuli
Grayscale pictures of faces, Chinese characters and houses were used
in this experiment. Faces were images of 72 individuals (36 male and 36
female), displaying neutral facial expression. Using Adobe Photoshop,
these faces were cropped to remove external features (hair, ears, and jaw
line) and replaced with the same oval contour used by Eimer et al.
(2010)[5]. All 72 Chinese characters used in this study had a left-right
configuration, high frequency values (from 0.01372 to 0.62457) (the
modern Chinese frequency dictionary, 1986), with the number of the
strokes varying from 7 to 14, and were presented in Song font. In addition,
72 grayscale images of houses were used which were similar with Eimer
et al (2010) [5]. The face stimuli were 180 × 276 p ixels, subtending an
angle of 4.0° × 6.2° from a viewing distance of 90
cm. The character and
house stimuli were 198 × 198 pixels, subtending an angle of 4.5 ° × 4.5 °
from a viewing distance of 90 cm. All stimuli had the same luminance
contrast.
Procedure
The participants were asked to sit on a chair, with a distance of 90 cm
away from the 17′′ CRT monitor (1024 × 768 pixel resolution), on whic h
all stimuli were presented against a dark grey background. E-Prime
software was used for stimulus presentation and behavioral response
collection (Psychology Software Tools, Pitts-burgh, PA).
Subjects were tested in a dimly lit room. In each trial, an adaptor
stimulus and a test stimulus were presented sequentially for 200 ms each
with a 200 ms inter-stimulus interval and followed by a 1500 ms intertrial
interval, which was consistent with Eimer et al (2010) [5]. Adaptor
stimuli and test stimuli were always different. One of three possible
adaptor stimuli - Faces (F), Chinese characters (C), or houses (H) - was
followed by one of two possible target stimuli (F or C). Therefore, there
were 6 conditions: FF, CF, HF, CC, FC, HC. Equal numbers of each
category were presented in random order in each of the four experimental
blocks. There were 108 trials in each block, 12 of which were target trials
with a red outline shape aligned with the outer contours of the stimulus
shape. These
target
trials were randomly
intermixed with
the
experimental trials and presented with equal probability as adaptor
stimuli or test stimuli. Participants were instructed to press a response
button following the second picture presentation when they detected a
target. Response buttons were counterbalanced across subjects.
EEG Recording and Data Analysis
EEG was recorded using a 128-channel HydroCel Geodesic Sensor Net,
with an electrode placed on the Vertex (Cz) serving as reference for the
online recording. Electrode impedances were kept below 50 kΩ. Signals
were digitized at a 500 Hz sampling rate, and amplified with a 0.1–200
Hz elliptical bandpass filter. EEG data were offline digitally filtered with
a 0.3 –30 Hz band-pass filter and epoched from 200 m s before to 800 ms
after stimuli onset with a 100 ms pre-stimulus baseline. Trials with
artifacts exceeding ±100 µV were rejected. Any subject with more than
30% bad segments would be excluded from the group-average. The
remaining EEG data were re-referenced to the average of channels.
EEG data were analyzed for nontarget. trials only. A group of the
larger response channels over the left hemisphere (58, 64, 65) and right
hemisphere (90, 95, 96) were analyzed. EEG waveforms were averaged
separately for each presentation condition of adaptor or test stimuli.
Based on visual inspection of the individual data, the N170 time-window
was defined as 130-210 ms for adaptor stimuli and 140-220 ms for test
stimuli. Repeated-measures analyses of variance (ANOVA) were
performed on the peak amplitudes and latencies of N170 component with
stimuli category and recording hemisphere serving as within-subject
factors.
Results
Adaptor stimuli data
The results of adaptor stimuli were shown in Figure 1,Table 1 and
Table 2. A two-way repeated-measures ANOVA of N170 peak amplitudes
and latencies was conducted for adaptor categories (faces, characters,
houses) and hemispheres (left hemisphere, right hemisphere). A main
effect of adaptor category on N170 amplitude (F(2, 30) = 14.619, p <
0.001, ηp
2=0.494) and an interaction between adaptor categories and
hemispheres (F(2, 30) = 8.227, p = 0.003, ηp
2=0.354) were found. Post
hoc tests revealed that the N170 amplitudes elicited by faces were much
larger than Chinese characters (p = 0.001) and houses (p < 0.001). There
was a main effect of adaptor category on N170 latency (F (2, 30) =
10.990, p = 0.001, ηp
2=0.423) and an interaction between adaptor and
hemisphere (F (2, 30) = 6.678, p = 0.005, ηp
2=0.308). Paired comparisons
showed that the N170 was earlier for characters relative to faces (p =
0.003) and houses (p = 0.001).
(Insert Fig 2 about here)
(Insert Table 1 about here)
(Insert Table 2 about here)
Test stimuli data
We analyzed the N170 amplitude elicited by the test stimuli in six
combinations of adaptor and test category. One of three possible adaptor
stimuli - Faces (F), Chinese characters (C), or houses (H) - was followed
by one of two possible target stimuli (F or C). Therefore, there were 6
conditions: FF, CF, HF, CC, FC, HC. For example, CF refers to the N170
amplitude to faces in condition CF, after adapting to a character. The
average wave shapes and the topographical map are shown in Figure 2.
The results of the 6 conditions analysis were shown in Table 3 and
Table 4.The repeated-measures ANOVA of N170 peak amplitudes and
latencies were conducted for the 6 conditions (CC, CF, FC, FF, HC, HF)
and hemispheres (left hemisphere, right hemisphere). For the N170 peak
amplitude, a main effect of test category (F(5, 75) = 12.910, p < 0.001,
2=0.463) and an interaction between conditions and hemisphere (F(5, 75)
ηp
= 6.877, p = 0.001, ηp
2= 0.314) were found. Post hoc tests are shown in
Table 5. There was a main effect of 6 conditions on N170 latency (F (5,
75) = 7.295, p< 0.001, ηp
2=0.327).
To test the relationship between face N170 adaptation and word N170
adaptation, we focused the analyses as followed:
The adaptation effect of both face and character-specific N170s are
typically measured relative to a house-induced adaptation effect: HF is
significantly larger than FF in both the left (p< 0.001) and right
hemisphere (p < 0.001), The results replicated the previous studies [4,5]
HC is significantly larger than CC in both the left (p= 0.001) and right
hemisphere (p < 0.001).
The results of the cross-category adaptation effect between faces and
Chinese characters were revealed by the following analysis:
We analyzed the face-specific N170 (FN) adaptation effect in the FF,
CF and HF conditions. The amplitude in the FF condition was
significantly smaller than in the CF (left, p = 0.001;right, p < 0.001) and
HF (left, p < 0.001;right, p < 0.001) conditions. And the amplitude in the
CF was similar to that in the HF condition. It suggested that Chinese
characters and houses have the same effect to the rapid followed face
stimuli, namely, compare with face, both the Chinese characters and
houses almost could not produce the adaptation effect to the faces. The
results of the adaptation effect of the Chinese character N170 (CN) in the
CC, FC and HC conditions are as follows: The amplitude of CN in the
HC condition was significantly larger than those in the FC (left, p =
0.012;right, p < 0.001) and CC (left, p = 0.001;right, p < 0.001)
conditions. Most importantly, the amplitude of CN was similar in the FC
and CC condition. The results suggested that both faces and Chinese
characters produced the similar CN adaptation effect to the following
Chinese characters.
In summary, the dramatic results show that faces produce an N170
adaptation effect on Chinese characters, but Chinese characters cannot
elicit a similar adaptation effect on faces.
(Insert Fig 2 about here)
(Insert Table 3 about here)
(Insert Table 4 about here)
(Insert Table 5 about here)
Discussion
The purpose of this study was to reveal the relationship of the neural
selectivity between faces and words. In this experiment, our results show
an asymmetric N170 adaptation effect between faces and Chinese
characters, that is, faces produced full adaptation to Chinese characters,
not vice versa. This surprising phenomenon provides many insights into
the neural selectivity of the N170 component.
Together with the principle of adaptation [6,11], our results allow us
to examine the nature of the overlap in neural selectivity to faces and
Chinese characters. Just as shown in Figure 3, the first three scenarios
imply the symmetric adaptation, while the last two refer to the
asymmetric one. The fourth scenario (Figure 3-D) shows that the neural
selectivity of CN completely contains the neural selectivity of CN. But
previous studies suggested that N170 elicited by face stimuli was related
to holistic /configural processing [12], which is a critical property of face
perception but not of Chinese character [15]. Moreover, if this scenario is
true, based on the principle of adaptation [6,11], the adaptation effect of
Chinese characters on faces would be stronger than that of faces to
Chinese characters. So this possibility can ’t be su pported. Therefore, the
only possibility is that the neural selectivity of FN completely
encompasses that of CN, as is shown in Figure 3-E, which can reasonably
explain our main results and coincide with our hypothesis. This
possibility also suggested that the complex component with bigger neural
selectivity (e.g. FN) could produce complete adaptation of the smaller
one (e.g. CN).
(Insert Fig 3 about here)
The above analysis demonstrates that the face N170 reflects a broad
range of functions that includes those involved in Chinese character
processing. Future behavioral and brain functional studies may reveal the
nature of the differences in information processing that underlie the
overlap in neural selectivity between faces and Chinese characters.
References
[1] Cao, X. H., Li, S., Zhao, J., Lin, S. E. & Weng, X. C. Left-lateralized early
neurophysiological response for Chinese characters in young primary school
children. Neuroscience Letters 492, 165-169, (2011).
[2] Dehaene, S. & Cohen, L. Cultural recycling of cortical maps. Neuron 56, 384-398,
(2007).
[3] Dehaene, S. et al. How learning to read changes the cortical networks for vision
and language. Science 330, 1359-1364, (2010).
[4] Eimer, M., Gosling, A., Nicholas, S. & Kiss, M. The N170 component and its
links to configural face processing: A rapid neural adaptation study. Brain research
1376, 76-87, (2011).
[5] Eimer, M., Kiss, M. & Nicholas, S. Response profile of the face-sensitive N170
component: a rapid adaptation study. Cerebral Cortex 20, 2442-2452, (2010).
[6] Frazier, W. T., Kandel, E. R., Kupfermann, I., Waziri, R. & Coggeshall, R. E.
Morphological and functional properties of identified neurons in the abdominal
ganglion of Aplysia californica. Journal of Neurophysiology 30, 1288-1351,
(1967).
[7] Gauthier, I., Curran, T., Curby, K. M. & Collins, D. Perceptual interference
supports a non-modular account of face processing. Nature Neuroscience 6,
428-432, (2003).
[8] Gauthier, I., Skudlarski, P., Gore, J. C. & Anderson, A. W. Expertise for cars and
birds recruits brain areas involved in face recognition. Nature Neuroscience 3,
191-197, (2000).
[9] Gauthier, I., Tarr, M. J., Anderson, A. W., Skudlarski, P. & Gore, J. C. Activation
of the middle fusiform 'face area' increases with expertise in recognizing novel
objects. Nature neuroscience 2, 568-573, (1999).
[10] Harel, A., Gilaie-Dotan, S., Malach, R. & Bentin, S. Top-down engagement
modulates the neural expressions of visual expertise. Cerebral Cortex 20,
2304-2318, (2010).
[11] He, S. & MacLeod, D. I. A. Orientation-selective adaptation and tilt after-effect
from invisible patterns. Nature 411, 473-476, (2001).
[12] Itier, R. J. & Taylor, M. J. N170 or N1? Spatiotemporal differences between
object and face processing using ERPs. Cerebral Cortex 14, 132-142, (2004).
[13] Kovács, G. et al. Electrophysiological correlates of visual adaptation to faces and
body parts in humans. Cerebral Cortex 16, 742-753, (2006).
[14] Maurer, U., Zevin, J. D. & McCandliss, B. D. Left-lateralized N170 effects of
visual expertise in reading: Evidence from Japanese syllabic and logographic
scripts. Journal of Cognitive Neuroscience 20, 1878-1891, (2008).
[15] McCleery, J. P. et al. The roles of visual expertise and visual input in the face
inversion effect: Behavioral and neurocomputational evidence. Vision Research 48,
703-715, (2008).
[16] Mckeeff,T., McGugin,R.,Tong,F.,& Gauthier,I. Expertise increases the functional
overlap between face and object perception. Cognition 117,335-360,(2010)
[17] Rossion, B., Joyce, C. A., Cottrell, G. W. & Tarr, M. J. Early lateralization and
orientation tuning for face, word, and object processing in the visual cortex.
NeuroImage 20, 1609-1624, (2003).
[18] Rossion, B., Kung, C.-C. & Tarr, M. J. Visual expertise with nonface objects
leads to competition with the early perceptual processing of faces in the human
occipitotemporal cortex. Proceedings of the National Academy of Sciences of the
United States of America 101, 14521-14526, (2004).
Table 1. Mean peak amplitudes and latencies (standard deviations) of the N170
elicited by the adaptor stimuli at left hemisphere (LH) and right hemisphere (RH).
Amplitudes (µV)
Latencies (ms)
LH RH
LH RH
Character
-6.12 (3.15)
-6.01 (3.71)
156.52 (8.20)
151.94 (9.55)
Face
-7.07 (3.06)
-8.35 (4.53)
161.06 (10.76) 160.77 (9.15)
House
-5.40 (3.11)
-6.27 (4.58)
159.58 (4.29)
160.42 (7.40)
Table 2. Statistical results of a 3 categories × 2 hemispher es repeated-measures
ANOVA on the peak amplitudes and latencies of the N170 component elicited by
adaptor stimuli.
Amplitudes
Latencies
F p
F p
Category
(2, 30) = 14.619 0.000*
(2, 30) = 10.990
0.001*
Hemisphere
(1, 15) = 0.655
0.431
(1, 15) = 0.426
0.524
Category × Hemisphere
(2, 30) = 8.227
0.003*
(2, 30) = 6.678
0.005*
* Indicates p-values lower than 0.05, which was considered as significant.
Table 3. Mean peak amplitudes and latencies (standard deviations) of the N170
elicited by the test stimuli as a function of adaptor stimuli at LH and RH.
Amplitudes (µV)
Latencies (ms)
LH RH
LH RH
-5.43 (2.85)
-5.33 (2.51)
169.46 (11.70)
166.75 (14.17)
-4.66 (2.30)
-5.94 (2.45)
181.67 (17.31)
178.33 (12.21)
-5.82 (3.14)
-5.83 (2.15)
171.17 (12.49)
171.46 (15.43)
-7.20 (2.83)
-9.43 (3.97)
175.58 (10.80)
176.25 (9.88)
-7.24 (3.00)
-7.65 (2.77)
181.08 (16.62)
179.08 (13.15)
-7.06 (2.42)
-8.36 (3.18)
175.42 (10.37)
177.75 (11.68)
CC
FF
FC
CF
HC
HF
Table 4. Statistical results of a 6 categories × 2 hemispher es repeated-measures
ANOVA on the peak amplitudes and latencies of the N170 component elicited by test
stimuli.
Amplitudes
Latencies
F p
F p
Category
(5, 75) = 12.910
0.000*
(5, 75) = 7.295
0.000*
Hemisphere
(1, 15) = 1.470
0.244
(1, 15) = 0.291
0.597
Category × Hemisphere
(5, 75) = 6.877
0.001*
(5, 75 ) = 0.794
0.513
* Indicates p-values lower than 0.05, which was considered as significant.
Table 5. Paired comparisons for the different-adaptation conditions of the N170
amplitudes elicited by the test stimuli.
Left hemisphere
Right hemisphere
Mean difference Sig
Mean difference Sig
CC
HC
CC
FC
FC
HC
FF
FF
HF
CF
1.809
0.385
1.424
2.397
-2.538
CF
HF
-0.141
0.001
0.443
0.012
0.000
0.001
0.747
2.324
0.497
1.827
2.418
-3.486
-1.068
0.000
0.092
0.000
0.000
0.000
0.054
Figure 1. Grand averaged N170 waveforms elicited by adaptor stimuli (Character,
Face and House) at left and right hemispheres separately.
Figure 2. Top row: The averaged N170 waveforms elicited by test stimuli at the left
and right hemispheres separately. Bottom row: The N170 amplitudes elicited by the
test stimuli in different-adaptation conditions.
Figure 3. The five possibilities of the relations between face-specific N170 (FN) and
Chinese character-specific N170 (CN)
|
1204.1558 | 1 | 1204 | 2012-04-06T20:33:47 | Nicotinic {\alpha}7 acetylcholine receptor-mediated currents are not modulated by the tryptophan metabolite kynurenic acid in adult hippocampal interneurons | [
"q-bio.NC",
"q-bio.BM",
"q-bio.CB"
] | The tryptophan metabolite, kynurenic acid (KYNA), is classically known to be an antagonist of ionotropic glutamate receptors. Within the last decade several reports have been published suggesting that KYNA also blocks nicotinic acetylcholine receptors (nAChRs) containing the \alpha7 subunit (\alpha7*). Most of these reports involve either indirect measurements of KYNA effects on \alpha7 nAChR function, or are reports of KYNA effects in complicated in vivo systems. However, a recent report investigating KYNA interactions with \alpha7 nAChRs failed to detect an interaction using direct measurements of \alpha7 nAChRs function. Further, it showed that a KYNA blockade of \alpha7 nAChR stimulated GABA release (an indirect measure of \alpha7 nAChR function) was not due to KYNA blockade of the \alpha7 nAChRs. The current study measured the direct effects of KYNA on \alpha7-containing nAChRs expressed on interneurons in the hilar and CA1 stratum radiatum regions of the mouse hippocampus
and on interneurons in the CA1 region of the rat hippocampus. Here we show that KYNA does not block \alpha7* nACHRs using direct patch-clamprecording of \alpha7 currents in adult brain slices. | q-bio.NC | q-bio | B r i e f R e p o r t
neuroscience
Nicotinic α7 acetylcholine receptor-mediated currents are not
modulated by the tryptophan metabolite kynurenic acid in adult
hippocampal interneurons
Peter Dobelis1*, Andrew L. Varnell1, Kevin J. Staley2, & Donald C. Cooper1
The tryptophan metabolite, kynurenic acid (KYNA), is classically known to be an antagonist of ionotropic glutamate receptors.
Within the last decade several reports have been published suggesting that KYNA also blocks nicotinic acetylcholine
receptors (nAChRs) containing the α7 subunit (α7*). Most of these reports involve either indirect measurements of KYNA
effects on α7 nAChR function, or are reports of KYNA effects in complicated in vivo systems. However, a recent report
investigating KYNA interactions with α7 nAChRs failed to detect an interaction using direct measurements of α7 nAChRs
function. Further, it showed that a KYNA blockade of α7 nAChR stimulated GABA release (an indirect measure of α7 nAChR
function) was not due to KYNA blockade of the α7 nAChRs. The current study measured the direct effects of KYNA on α7-
containing nAChRs expressed on interneurons in the hilar and CA1 stratum radiatum regions of the mouse hippocampus
and on interneurons in the CA1 region of the rat hippocampus. Here we show that KYNA does not block α7* nACHRs using
direct patch-clamp recording of α7 currents in adult brain slices.
Kynurenic acid (KYNA) is produced by the metabolism of
tryptophan via the kynurenine pathway1,3. Classically, KYNA
is known for its antagonist actions at ionotropic glutamate
receptors, showing the greatest a"nity for NMDA-mediated
glutamatergic responses2. Altered levels of KYNA have been
associated with several disease states; increased KYNA levels
are seen with Alzheimer’s disease, Down’s syndrome, and
schizophrenia while decreased KYNA levels are associated
with end stage Parkinson’s and Huntington’s disease3.
Additionally, animal studies indicate that increased brain
levels of KYNA are neuroprotective and anti convulsant,
while decreased KYNA levels are associated with an increased
vulnerability to excitotoxic damage4.
Another action attributed to KYNA is the antagonism
of α7 subunit-containing (α7*) nicotinic acetylcholine
receptors (nAChRs). Several reports from Albuquerque and
colleagues present data demonstrating that KYNA also blocks
the activation of α7* nAChRs4. However, a recent report from
Mok and colleagues examining the e#ects of KYNA on several
di#erent ligand gated ion channels revealed that KYNA had
no e#ect on α7* nAChRs6. Here we present the results of our
investigation of KYNA e#ects on α7* nAChRs expressed on
interneurons in the hilar and CA1 stratum radiatum (SR)
regions of the mouse hippocampus, as well as α7* nAChRs
expressed on interneurons the rat CA1 SR.
RESULTS
Kynurenic acid e!ects on α7* nAChRs expressed on
mouse hilar interneurons
choline-induced
form
$e
results obtained
α7* currents in hilar interneurons are shown in Fig 1.
We initially examined the prevalence of α7* currents in
hilar interneurons. Out of the 23 neurons studied, 20
displayed choline-induced and methyllycaconitine (MLA)
sensitive whole cell currents characteristic of α7* nAChRs.
Furthermore, in experiments using α7* null mutant mice,
no choline-induced currents were detected (Fig 1a and 1b).
Next, we examined the e#ects of KYNA on the choline-
induced α7* currents. In these experiments, stable baseline
b
a
c
d
Fig 1: a. Representative traces for the characterization of α7* currents in mouse
hilar interneurons. $e le% most trace shows the control response to pressure
applied choline. $e next two traces show the response to choline in the pres-
ence of MLA (10 nM) and a%er 30 min washout, respectively. $e last trace
shows the lack of response to choline in α7 null mutant mice. b. Methylly-
caconitine (MLA) completely blocked the choline response (p = 0.0003 con-
trol vs MLA, n = 8), the MLA e#ect was reversed partially a%er a washout
(p = 0.035 washout vs. MLA, p = 0.005 washout vs control, n = 8), and the
choline response for the α7* null mutant mice di#ered signi&cantly from wild
type (p = 0.0003, n = 8). c. Representative traces for the e"ect of the bath ap-
plied KYNA on choline-evoked α7* currents. $e top trace shows the control
response to pressure applied choline; note the overriding glutamatergic spon-
taneous EPSCs. $e middle trace shows the choline response a%er a 30 min
exposure to 1 mM KYNA; note the absence of the spontaneous EPSCs. $e
bottom trace shows the response to choline a%er a 20 min. washout of KYNA;
note the reappearance of the spontaneous EPSCs. d. KYNA failed to produce
any reduction in the choline response (p = 0.97, n = 20).
1Center for Neuroscience, Institute for Behavioral Genetics, Department of Neuroscience, University of Colorado at Boulder, Boulder Colorado 80303, USA
2Department of Neurology, Massachusetts General Hospital, 55 Fruit St, WAC 708-D, Boston, MA 02114, USA
*Correspondence should be sent to [email protected]
Page 1 of 2
cA Attribution 3.0 2011 Nature Precedings
α7* currents were obtained, followed by bath application
of KYNA (1 mM) for 30 min. Choline-evoked α7* currents
were then measured a%er a 20 min washout of KYNA (Fig
1c). $e top trace shows a choline-evoked α7* response with
overriding spontaneous glutamatergic-mediated EPSCs. $e
middle trace shows the choline-evoked α7* response a%er a
30 min bath application of 1 mM KYNA. $e absence of the
EPSCs indicates the presence of KYNA. $e bottom trace
shows the choline-evoked response a%er a 20 min washout
of KYNA. Spontaneous EPSCs returned, indicating the
removal of KYNA from the slice (Fig 1d). Out of the 20 hilar
interneurons, KYNA failed to have any signi&cant e#ect on
the choline-evoked α7*-mediated currents.
KYNA e!ects on α7* nAChRs expressed on mouse and rat
CA1 SR interneurons
To determine whether there are regional and/or species
di#erences in α7* nAChR sensitivity to KYNA, we tested
the ability of KYNA to block α7* nAChR currents in
interneurons in the CA1 SR region in both mice and rats.
$e representative traces for choline-evoked α7* currents
in the mouse CA1 SR appear simular to the KYNA treated
cells, indicating no change (Fig 2a). $e top trace shows the
control response to pressure applied choline (10 mM) and the
bottom trace shows the choline response in the presence of 1
mM KYNA for 30 min. KYNA failed to signi&cantly reduce
choline-evoked α7* currents in the &ve CA1 SR interneurons
we tested (Fig 2b). Representative traces from α7* currents
a%er the pressure applied ACh (1 mM) and KYNA exposure
in rat CA1 SR interneurons also showed no change (Fig 2c).
$e top trace shows the control response, the middle trace
shows the response in the presence of 1 mM KYNA, and the
bottom trace shows the response in the presence of the α7-
selective antagonist MLA (10 mM). Exposure to a 20 min
KYNA bath failed to inhibit α7* currents, while subsequent
exposure to MLA produced a greater than 80% blockade of
α7* currents (Fig 2d).
DISCUSSION
$ese results clearly indicate that KYNA has no e#ect on α7*
nAChRs in the hilar and CA1 regions of adult mouse and rat
hippocampal interneurons. $ese results are consistent with
those reported by Mok and colleagues6. Others have concluded
that KYNA does interact with the α7* nAChRs using indirect
measures of α7* function4. KYNA has recently been shown
to activate the orphan G-protein receptor GPR-351. GPR-35
is coupled to the Gi-o pathway and is expressed throughout the
rodent brain1. $is, in addition to the interactions that KYNA
has with other ligand-gated receptors4 suggests that prior
direct α7* mediated physiological e#ects attributed to KYNA
must be made with caution. Further studies are needed to
resolve the apparent e#ects of KYNA on α7* nAChRs. Until
then the most direct evidence indicates no role for KYNA on
α7* nAChR function.
METHODS
Slice preparation and recordings
Hippocampal slices were prepared as described by Proctor
and colleagues7. Whole-cell recordings were made using a
cA Attribution 3.0 2011 Nature Precedings
B r i e f R e p o r t
standard potassium gluconate based internal solution and
a standard ACSF solution. KYNA and MLA were both bath
applied to the tissue slices. Both choline and acetylcholine
were pressure applied to the slices using established protocols.
http://www.neuro-cloud.net/nature-precedings/dobelis/ .
Expanded methods and recordings can be found at
a
b
c
d
Fig 2. a. Representative traces for choline-evoked α7* currents in mouse
stratum radiatum (SR) interneurons. $e top trace is the control response
to pressure applied choline (10 mM, 50 ms). $e bottom trace shows the
choline response a%er a 30 min. bath exposure to 1 mM KYNA. b. KYNA
failed to reduce the choline response (p = 0.44, n = 5). c. Representative traces
for pressure applied ACh-evoked α7* currents in rat CA1 SR interneurons. $e
top trace shows the control response to 1 mM ACh. $e middle trace shows
the ACh respose a%er the 30 min. bath exposure to 1 mM KYNA. the bottom
trace shows the ACh response a%er a 20 min. bath exposure to 10 nM MLA.
d. $e last &ve traces were averaged for each condition: control, KYNA, and
MLA. KYNA failed to block the ACh-induced α7* currents (p = 0.71) while
MLA signi&cantly blocked the response (p = 0.0001, n = 5).
ACKNOWLEDGEMENTS
$is work was funded and supported by the American Epilepsy
Foundation/Milken Family Foundation Fellowship (P.D.), R21
DA026918 (J. Stitzel), National Institute on Drug Abuse grant R01-
DA24040 (to D.C.C.), NIDA K award K-01DA017750 (to D.C.C.).
PROGRESS AND COLLABORATIONS
To see up to date progress on this project or if you are interested
in contributing to this project visit: http://www.neuro-cloud.net/
nature-precedings/dobelis/
AUTHOR CONTRIBUTIONS
P.D., D.C.C. & K.J.S. designed the experiments. P.D. recorded and
analyzed the data. P.D., A.L.V. & D.C.C. wrote and prepared the
manuscript.
Submitted online at http://www.precedings.nature.com
1.
In Neuro-cloud.net. Retrieved Aug. 7, 2011 from, www.
neuro-clould.net/nature-precedings/dobelis/
2. Stone TW. TIPS, 21:149-154, 2000.
3. Vamos E., et al., J. Neurol. Sci., 283:21-27, 2009
4. Hilmas C., et al., J. Neurosci. 21:7463-7473, 2001.
5. Alkondon M., et al., JPET 337:572–582, 2011.
6. Mok HMS, et al., JNC. Neuropharm. 57:242-249, 2009.
Proctor WR, et al., Br J Pharmacol. 162(6):1351-63,
7.
2011.
Page 2 of 2
|
1601.00987 | 1 | 1601 | 2016-01-05T21:33:24 | Stimulation-based control of dynamic brain networks | [
"q-bio.NC",
"eess.SY"
] | The ability to modulate brain states using targeted stimulation is increasingly being employed to treat neurological disorders and to enhance human performance. Despite the growing interest in brain stimulation as a form of neuromodulation, much remains unknown about the network-level impact of these focal perturbations. To study the system wide impact of regional stimulation, we employ a data-driven computational model of nonlinear brain dynamics to systematically explore the effects of targeted stimulation. Validating predictions from network control theory, we uncover the relationship between regional controllability and the focal versus global impact of stimulation, and we relate these findings to differences in the underlying network architecture. Finally, by mapping brain regions to cognitive systems, we observe that the default mode system imparts large global change despite being highly constrained by structural connectivity. This work forms an important step towards the development of personalized stimulation protocols for medical treatment or performance enhancement. | q-bio.NC | q-bio | Stimulation-based control of dynamic brain networks
Sarah Feldt Muldoon1,2,3, Fabio Pasqualetti4, Shi Gu1,5, Matthew Cieslak6, Scott T. Grafton6, Jean
M. Vettel2,6,1 and Danielle S. Bassett1,7
1Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104
2US Army Research Laboratory, Aberdeen Proving Ground, MD 21005
3Department of Mathematics and Computational and Data-Enabled Science and Engineering Program,
University at Buffalo, SUNY, Buffalo, NY 14260
4Department of Mechanical Engineering, University of California, Riverside, CA 92521
5Applied Mathematics and Computational Science Graduate Program, University of Pennsylvania,
Philadelphia, PA 19104
6Department of Psychological and Brain Sciences, University of California, Santa Barbara, Santa
Barbara, CA 93106
7Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA
19104
arXiv:1601.00987v1 [q-bio.NC] 5 Jan 2016
1
Abstract
The ability to modulate brain states using targeted stimulation is increasingly being employed to
treat neurological disorders and to enhance human performance. Despite the growing interest in
brain stimulation as a form of neuromodulation, much remains unknown about the network-level
impact of these focal perturbations. To study the system wide impact of regional stimulation, we
employ a data-driven computational model of nonlinear brain dynamics to systematically explore
the effects of targeted stimulation. Validating predictions from network control theory, we uncover
the relationship between regional controllability and the focal versus global impact of stimulation,
and we relate these findings to differences in the underlying network architecture. Finally, by
mapping brain regions to cognitive systems, we observe that the default mode system imparts large
global change despite being highly constrained by structural connectivity. This work forms an
important step towards the development of personalized stimulation protocols for medical treatment
or performance enhancement.
2
Brain stimulation is increasingly used to diagnose1, monitor2, and treat neurological3 and psychiatric4
disorders. Non-invasive stimulation, such as transcranial magnetic stimulation (TMS) or transcranial
direct current stimulation (tDCS), is used, for example, in epilepsy5, 6, stroke7, attention deficit
hyperactivity disorder2, tinnitus8, headache9, aphasia10, traumatic brain injury11, schizophrenia12,
Huntington's disease13, and pain14, while invasive deep brain stimulation (DBS) is approved for
essential tremor and Parkinson's disease and is being tested in multiple Phase III clinical trials
in major depressive disorder, Tourette's syndrome, dystonia, epilepsy, and obsessive-compulsive
disorder15. In addition to its clinical utility, emerging evidence suggests that stimulation can also
be used to optimize human performance in healthy individuals16, 17, potentially by altering cortical
plasticity17.
Despite its broad utility, many engineering challenges remain18, from the optimization of stimulation
parameters to the identification of target areas that maximize clinical utility19. Critically, an
understanding of the local effects of stimulation on neurophysiological processes -- and the downstream
effects of stimulation on distributed cortical and subcortical networks -- remains elusive20. This gap
has motivated the combination of stimulation techniques with various recording devices (PET21,
MEG22, fast optical imaging23, EEG17, and fMRI24, 25) to monitor the effects of stimulation on
cortical activity26, 27, 28.
In this context, it has become apparent that there is a critical need for
biologically informed computational models and theory for predicting the impact of focal neurostimulation
on distributed brain networks, thereby enabling the generalization of these effects across clinical
cohorts as well as the optimization of stimulation protocols29, including refinements for individualized
3
treatment and personalized medicine.
To meet this fast-growing need, we utilize network control theory30 to understand and predict the
effects of stimulation on brain networks. The advent of network theory as a ubiquitous paradigm
to study complex engineering systems as well as social and biological models has changed the face
of the field of automatic control and redefined its classic application domain. Control of a network
refers to the possibility of manipulating local interactions of dynamic components to steer the
global system along a chosen trajectory. Network control theory offers a mechanistic explanation
for how specific regions within well-studied cognitive systems may enable task-relevant neural
computations: for example, activity in the primary visual cortex can initiate a trajectory of processing
along the extended visual pathway (V2, V3, etc.), driving visual perception and scene understanding.
More broadly, network control theory also offers a mechanistic framework in which to understand
clinical interventions such as brain stimulation, which elicits a strategic functional effect within the
network for synchronized neural processing. In both cases (regional activity and stimulation), it is
critically important to understand how underlying structural connectivity can constrain or modulate
the functional effect of regional alterations in activity. Yet, the application of control-theoretic
techniques to brain networks is underexplored, even though the questions posed by the neuroscientific
community are uniquely suited to the application of these tools30.
In a novel approach to link network control theory and brain dynamics, we examine the effects of
regional stimulation on brain states using a nonlinear meso-scale computational model, built on
data-driven structural brain networks. We demonstrate that the dynamics of our model is highly
4
variable across subjects, but highly reproducible across multiple scans of the same subject. We
confirm the pragmatic utility of network control theory for nonlinear systems, extending previous
work on linear approaches30, and show that general control diagnostics (average and modal controllability)
are strongly correlated with the density of structural connections linking brain regions. Finally,
we investigate the interplay between functional and structural effects of stimulation by examining
how the global functional activity across brain regions is modulated by region-specific stimulation
(a region's functional effect) and whether the region's structural connectivity accounts for its
influence on the larger brain network (structural effect). Results show that the default mode system
and subcortical regions produce the strongest functional effects; the subcortical structures display
weak structural effects, being diversely connected across many cognitive systems. Collectively,
our results indicate the value of data-driven, biologically motivated brain models to understand
how individual variability in brain networks influences the functional effects of region-specific
stimulation for clinical intervention or cognitive enhancement.
Results
To systematically assess the effects of brain stimulation and its utility in control, we begin by
building a spatially embedded nonlinear model of brain network dynamics. Structural brain networks
are derived from diffusion spectrum imaging (DSI) data acquired in triplicate from 8 healthy adult
subjects. We perform diffusion tractography to estimate the number of streamlines linking N = 83
large-scale cortical and subcortical regions extracted from the Lausanne atlas31 and summarize
these estimates in a weighted adjacency matrix whose entries reflect the density of streamlines
5
Figure 1 Building nonlinear brain networks. (a) Subject-specific structural brain networks
are built based on a parcellation of the brain into 83 anatomically defined brain regions (network
nodes) with connections between regions given by the density of streamlines linking them. (b)
The dynamics of each region are represented by a single Wilson-Cowan oscillator, and these
oscillators are coupled according to the structural connectivity of a single subject. (c) Brain states
are quantified by calculating the pairwise functional connectivity between brain regions.
connecting different regions. Finally, we model regional brain activity using biologically motivated
nonlinear Wilson-Cowan oscillators32, coupled through these data-derived structural brain networks.
We quantify brain states based on functional connectivity obtained from pairwise correlations
between simulated regional brain dynamics (Fig. 1 and Methods).
6
Nodes: 83 brain regions
(Lausanne 2008 atlas)
Edges: white matter tracts
(diffusion spectrum imaging)
b
Network dynamics
c
Brain states
Correlation strength
1
0.5
0
Time
Brain regions
Brain regions
Brain regions
Computational model: coupled
Pairwise functional
Wilson-Cowan oscillators
connectivity between regions
a
Network structure
+
Inter-subject variability of brain dynamics
Our broad goal is to understand the role of regional stimulation to differentially control brain
dynamics. Theoretical predictions of regional controllability can be developed based on the underlying
structural connectivity of the computational model30. Given our subject-specific data-driven approach,
it is therefore important to understand how variability between structural brain networks affects the
dynamics of our model. Wilson-Cowan oscillators are a biologically driven mathematical model
of the mean-field dynamics of a spatially localized population of neurons32, 33, modeled through
equations governing the firing rate of coupled excitatory (E) and inhibitory (I) neuronal populations
(Methods). Here, we measure a single brain region's dynamics by the firing rate of the excitatory
population. An important feature of these Wilson-Cowan oscillators is that an uncoupled oscillator
can exhibit one of three states, depending upon the amount of external current applied to the system
(Fig. 2a-b). When no external current is applied (P = 0), the system relaxes to a low fixed point
(Fig. 2a-b). For moderate amounts of applied current, the oscillator is pushed into an oscillatory
limit cycle, and if sufficiently high amounts of current are applied, the system settles at a high fixed
point.
For this system of coupled oscillators, brain regions (oscillators) can receive current from an
external input (stimulation, i.e., P > 0) or from the activity of other brain regions to which they
are connected. The global coupling parameter, c5, therefore serves to govern the global state of the
system by regulating the overall amplitude of current transmitted between brain regions. For low
values of c5, the system will fluctuate around the low fixed point, whereas for high values of c5, the
7
system will transition into the oscillatory (limit cycle) regime. For a given structural connectivity,
we can systematically increase the global coupling parameter and record the value at which the
system transitions from the low fixed point to the oscillatory regime.
Individual differences in
structural connectivity can cause this transition to occur at different points in the parameter space,
and we therefore use this point of transition to assess the sensitivity of our model to inter subject
differences in the structural connectivity.
By measuring the point of oscillatory transition using structural connectivity matrices obtained
from each of three scans for eight subjects, we see in Fig. 2c that model dynamics are highly
reproducible within scans of a single subject, but show variability across subjects. We quantify the
within versus between subject reproducibility using the intraclass correlation coefficient (ICC).
We see high reproducibility within scans of a single subject (Fig. 2d, ICC = 0.826) and low
reproducibility between different subjects (ICC = −0.006), which also corresponds to a low within
subject variance (V = 0.0019) and high between subject variance (V = 0.0143). Due to the high
within subject reproducibility across the three scans of each subject, the remaining findings are
presented at the subject level by averaging the results over simulations derived from each of the
three single subject scans.
Regional controllability predicts the functional effect of stimulation
In order to elucidate the role of regional stimulation to differentially control brain dynamics, we
first turn to predictions made using linear network control theory30. Linear network control theory
8
Figure 2 Nonlinear brain dynamics and variability. (a) Excitatory-inhibitory phase space plots
depicting behavior for a single Wilson-Cowan oscillator in the presence of no external current input
(left; P = 0; low-fixed point), moderate external current input (middle; P = 1.25; limit cycle), and
high external current input (right; P = 2.5; high fixed point). All simulations are started with initial
conditions E = 0.1, I = 0.1. (b) The corresponding firing rate of the excitatory population plotted
as a function of time for the simulations depicted in (a). (c) Box plots showing the value of global
coupling parameter at which the system transitions from the low fixed-point state to the oscillatory
regime for models derived from three different structural scans from each of eight subjects. (d)
Within and between subject reproducibility (left) and variability (right) for the data shown in (c).
Reproducibility is measured by the intraclass correlation coefficient and is high within subjects, but
low between subjects. This is additionally reflected in the low within subject variability, measured
as the average variance, and high between subject variability.
9
P = 1.25
P = 2.5
0.4
0.3
0.2
0.1
0
0
0.4
Excitatory activity (E)
0.2
0.4
0.3
0.2
0.1
0
0
P = 2.5
200
400
Time (ms)
Inhibitory activity (I)
Excitatory activity (E)
0.4
0.3
0.2
0.1
0
0
0.4
Excitatory activity (E)
0.2
0.4
0.3
0.2
0.1
0
0
P = 1.25
200
400
Time (ms)
Inhibitory activity (I)
Excitatory activity (E)
P = 0
0.4
0.3
0.2
0.1
0
0
0.4
Excitatory activity (E)
0.2
0.4
0.3
0.2
0.1
0
0
P = 0
200
400
Time (ms)
a
Inhibitory activity (I)
b
Excitatory activity (E)
Variance (V)
0.016
0.012
0.008
.004
00
0.8
0.6
0.4
ICC
0.2
0
d
Reproducibility
Variability
B etw e e n
su bject
W ithin
su bject
B etw e e n
su bject
W ithin
su bject
1
2 3 4 5 6 7 8
Subject number
1.5
1.4
1.3
1.2
1.1
1
c
Transition point
assumes a simplified linear model of network dynamics and computes controllability measures
based upon the topological features of the structural network architecture. Here, we assess two
different types of regional controllability derived from our structural brain networks: average
controllability and modal controllability. Regions with high average controllability are capable
of moving the system into many easy to reach states with a low energy input, whereas regions
with high modal controllability can move the system into difficult to reach states but require
a high energy input (see Methods for detailed descriptions of controllability measures and the
Supplementary Information and Fig. S1 for their relationship to the steady state network response
from regional stimulation with a constant current input). As previously described30, we observe a
strong correlation between regional degree and average controllability and a strong inverse correlation
between regional degree and modal controllability that is robust across structural networks derived
from all subjects (Fig. 3).
We predict that brain regions with a high average controllability have the ability to impart large
changes in network dynamics, easily moving the system into many nearby states. However, these
predictions are made under the assumption of linear dynamics, and we know that the brain is in
fact a highly nonlinear system. Our modeling approach allows us to directly test the validity of
these linear controllability predictions in a nonlinear setting by systematically studying the effects
of focal stimulation to brain regions and studying how the system moves.
Using our computational model, we select the value of global coupling that places our brain model
just before the transition to the oscillatory regime such that all brain regions are fluctuating near the
10
Figure 3 Linear regional controllability. (a) Average controllability plotted as a function of
regional degree for each of the 83 brain regions. (b) Average controllability plotted as a function
of regional degree. Controllability predictions were performed for each of the three scans for each
subject and the data points reflect controllability and degree values averaged across scans.
11
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
5
10
15
20
25
Degree
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
5
10
15
20
25
Degree
1.04
1.03
1.02
1.01
1
0
1
0.995
0.99
0.985
0.98
0.975
0.97
0.965
0
a
Average controllability
b
Modal controllability
low fixed point. We then select a single brain region and add an external stimulating current that
brings the selected region into its oscillatory state (Fig. 4a). We can quantify the changes in brain
state due to this stimulation by computing the functional matrix of pairwise correlations between
brain region dynamics during a period before stimulation occurs. We compare this pre-stimulation
matrix to the functional matrix obtained during the stimulation period. By taking the difference
between the functional brain state before and during regional stimulation, we measure the distance
that the system moves. As expected, stimulation of low controllability regions produces smaller
changes in the functional brain state than stimulation of high controllability regions (Fig. 4b-c).
The results of systematically stimulating each brain region are shown in Fig. 5. We quantify
the overall change in brain state configuration by measuring the functional effect of regional
stimulation: the absolute value of the pairwise change in functional connectivity, averaged over
all brain region pairs. We observe that stimulation of regions with a high average controllability
produce a large functional effect, while stimulation of regions with a high modal controllability
result in a low functional effect (Fig. 5a-b).
Structural connectivity differentially constrains the effects of regional stimulation
We next ask how the underlying structural connectivity differentially constrains the functional
effect of the stimulation for different brain regions. We quantify this structural constraint by
calculating the structural effect on network dynamics, which measures the change in spatial correlation
between the structural connectivity matrix and the functional brain state matrix before and during
12
Figure 4 Regional stimulation.
(a) Stimulation of a single region pushes the region from
fluctuations around its low fixed point to the oscillatory state. (b) Example brain regions identified
as having low average controllabllity (pars opercularis, blue), medium average controllability
(post central, green) and high average controllability (isthmus cingulate, orange) (c) Simulation of
example regions in panel (b) differentially move the system into new functional states. Stimulation
applied to regions of high average controllability imparts more change in the functional brain state
than stimulation applied to regions of low average controllability.
13
Δ Functional
Connectivity
0.8
0.6
0.4
.2
00
-0.2
Δ Functional
Connectivity
0.8
0.6
0.4
.2
00
-0.2
Δ Functional
Connectivity
0.8
0.6
0.4
.2
00
-0.2
Brain regions
Brain regions
before
during
b
0.4
0.2
0
a
Excitatory Activity (E)
0
1
2
3
4
Time (s)
ow average controllability: Pars opercularis
c L
Brain regions
Brain regions
Medium average controllability: Post central
Brain regions
Brain regions
High average controllability: Isthmus cingulate
Figure 5 Functional effect of stimulation. (a-b) The functional effect of regional stimulation
plotted as a function of the average (a) and modal (b) controllability for each of the 83 brain regions.
(c-d) The structural effect of regional stimulation plotted as a function of the average (c) and modal
(d) controllability for each of the 83 brain regions. Controllability predictions, simulations, and
calculation of the functional and structural effects were performed for each of the three scans for
each subject and the data points reflect values averaged over scans.
14
0.97
0.975
0.98
0.985
0.99
0.995
1
Modal controllability
0.97
0.975
0.98
0.985
0.99
0.995
1
Modal controllability
b
0.8
0.6
0.4
0.2
Functional effect
0
0.965
0.35d
0.3
0.25
0.2
0.15
0.1
0.05
Structural effect
0
0.965
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
1
1.005 1.01 1.015 1.02 1.025 1.03 1.035 1.04
Average controllability
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
c
Structural effect
1
1.005 1.01 1.015 1.02 1.025 1.03 1.035 1.04
Average controllability
a
0.8
0.6
0.4
0.2
0
Functional effect
stimulation (Methods). Brain regions with a higher structural effect show a greater increase in
similarity between structural and functional matrices when stimulated than regions with a lower
structural effect. Thus, stimulation of these regions is more constrained by the underlying structural
connectivity. As seen in Fig. 5(c-d), the relationship between the structural effect and regional
controllability is opposite that of the functional effect and controllability. Stimulation of regions
with a high average controllability results in a smaller structural effect, while stimulation of regions
with a high modal controllability results in a higher structural effect.
In order to better understand why stimulation of regions with a high average controllability easily
moved the system (high functional effect) and were less constrained by the underlying structure
(low structural effect), we quantified the spread of activation from the regional stimulation. Specifically,
we asked if the stimulation of a given region induced focal or global changes in the brain state by
calculating the fractional activation of the functional connectivity matrix. The fractional activation
is given by the fraction of pairwise regions that experience a change in their functional connectivity
value that is above a given threshold (red pixels in Fig. 6a-b). If the stimulation of a brain region
results in large changes that occur globally throughout the brain, the fractional activation will be
high, but if the stimulation has only a focal effect, the fractional activation will be low.
As one might expect, we observe a strong positive correlation between the functional effect and
the fractional activation (Fig. 6c, Spearman's ρ = .992, p (cid:28) .001). Specifically, a large functional
effect is due to a global effect of regional stimulation that pushes the system into a nearby state,
while a small functional effect is due to a focal effect of regional stimulation that moves the system
15
toward a more distant state. However, the relationship between the structural effect and fractional
activation is more complex as seen in the crescent shaped curve of Fig. 6d. A maximum structural
effect occurs at the curve of the crescent (arrow in Fig. 6d), where the effects of stimulation are
neither focal or global. This result indicates that the underlying structural connections constrain
the effects of stimulation the most in situations when regional stimulation impacts a moderately
sized portion of the brain.
Cognitive systems display varied roles in the structure-function landscape
We next investigated the interplay between the structural and functional effects by examining the
structure-function landscape (Fig. 7). Stimulation of individual brain regions revealed a range
of tradeoffs between structural and functional effect values, with some regions displaying a high
functional effect but low structural effect and others displaying a high structural effect but moderate
or low functional effect. We therefore asked if there was a relationship between the cognitive
function associated with a brain region and its location in the structure-function landscape. Brain
regions were assigned to one of nine cognitive systems (Fig. 7 and Supplementary Material),
eight of which were based on a data-driven clustering of functional brain networks34 that group
regions that perform similar roles across a diverse set of tasks, as well as a group of subcortical
regions. Although in reality no region has a singular function, we use these system assignments
as a pragmatic means to assess whether controllability diagnostics are differentially identified in
distributed brain networks.
16
(a-b) The absolute change in functional connectivity and
Figure 6 Fractional activation.
resulting fractional activation shown for a threshold value of 0.6.
(a) Example of stimulation
applied to a region of low average controllability resulting in a focal effect on the resulting
functional connectivity matrix. (b) Example of stimulation applied to a region of high average
controllability resulting in a global effect on the resulting functional connectivity matrix. (c) The
relationship between functional effect and fractional activation due to regional stimulation. We
observe a high positive correlation between functional effect and fractional activation (Spearman's
ρ = .992, p (cid:28) .001), indicating that a high functional effect corresponds to a global impact
(d)
of stimulation while a low functional effect corresponds to a focal impact of stimulation.
The relationship between structural effect and fractional activation due to regional stimulation.
Calculations of the functional and structural effects and fractional activation were performed for
each of the three scans for each subject and the data points reflect values averaged over scans.
17
Not
Activated
Activated
Brain regions
Brain regions
Absolute ΔFC
0.8
0.6
0.4
.2
00
Brain regions
High average controllability: Global activation
0.05
0.1
0.15
0.2
0.25
0.3
0.35
Structural effect
1
0.8
0.6
0.4
0.2
Fractional activation
0
0
Brain regions
b
d
Not
Activated
Activated
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
Functional effect
Low average controllability: Focal activation
Brain regions
Brain regions
Absolute ΔFC
0.8
0.6
0.4
.2
00
Brain regions
1
0.8
0.6
0.4
0.2
0
0
a
Brain regions
c
Fractional activation
In Fig. 7a, we can see that some cognitive systems cover a wide range of the structure-function
landscape, whereas other systems tend to be more localized. Although the subcortical regions are
well connected structurally, stimulation to any single subcortical region produces a functional effect
that is less constrained by the underlying structural connections (low structural effect). In contrast,
regions in other systems remain similarly constrained by the underlying structural connections, but
vary in their ability to impart a functional effect on the system. Clusters of these nine cognitive
systems emerge, suggesting that the parcellation scheme may be too fine-grained to understand the
general organizing principles across the structure-function landscape; for example, there may be
general network principles for sensory processing that are independent of modality (sight, sound,
touch). Consequently, we created a more coarse-grained grouping of four cognitive systems (three
functional and one structural) to examine broad-stroke differences among well-studied cognitive
systems: the sensorimotor cortex, higher order cognitive, medial default mode network, and subcortical
regions. As seen in Fig. 7b, regions in the sensorimotor cortex and higher order cognitive regions
show a wide variation in their ability to impart a large functional effect when stimulated. Interestingly,
although regions in the default mode network are similarly constrained by structure when compared
to the sensorimotor and higher order cognitive regions, regions within the default mode network
consistently impart a large functional effect on the system. Stimulation of subcortical regions also
results in a large functional effect, but these regions are less constrained by structure than those in
the default mode and therefore these two systems occupy different spaces in the structure-function
landscape (Fig. 7b; two sample, two-dimensional Kolmogorov-Smirnov test, p = 0.0003). This
separation in the structure-function landscape could reflect the fact that the default mode network
18
represents a functionally defined system, whereas subcortical regions are functionally diverse despite
being well connected structurally.
In our final analysis, we examined why the default mode and subcortical regions had a stronger
functional effect, and interestingly, why the subcortical structures showed a lower structural effect.
We hypothesized that differential properties of these four systems can be better understood by
investigating the density of connections between brain regions within a system versus the density
of connections between systems. Both the sensorimotor cortex and the higher order regions are
defined functionally and are composed of multiple distributed structural systems. They therefore,
as a whole, have a low density of structural connections both within the system and between
the system and the rest of the brain (Fig. 7c).
In contrast, subcortical regions form a highly
connected structural subnetwork while remaining well connected to the rest of the brain. Thus,
stimulation to these regions is globally distributed, resulting in a high functional effect. The
medial default mode network also forms a well-connected subnetwork, but is less connected to
the sensorimotor cortex and higher order cognitive regions than the subcortical regions (see Table
1). While stimulation to regions in the default mode network results in a large functional effect,
regions in this subnetwork remain more constrained by the underlying structure. This latter result
is consistent with previous work showing that brain regions in the default mode network display
the highest correlation between structural and functional connectivity35. These results capture a
mechanism for the subcortical regions to strongly influence regions across cerebral cortex where
the default mode has a strong but targeted functional effect within its network, which enables these
19
Sensorimotor Higher Order Cognitive Default Mode
Subcortical
Sensorimotor
Higher Order Cognitive
Default Mode
Subcortical
0.054
0.025
0.059
0.094
0.025
0.067
0.046
0.068
0.059
0.046
0.257
0.137
0.094
0.068
0.137
0.412
Table 1 Average density of connections between and within subnetworks of four cognitive
systems.
regions to quickly adapt from rest to a wide variety of task states30.
Discussion
As neuromodulation is increasingly used to treat neurological disorders, it is essential to develop
an understanding of the network-wide effects of focal stimulation. Such knowledge would directly
inform the development of targeted protocols that effectively and efficiently maximize therapeutic
benefits while minimizing the potential for adverse effects on brain dynamics and cognition. An
initial step towards achieving this goal lies in the examination of neuroimaging data through the
lens of linear network control theory, a mathematical framework that predicts highly controllable
brain regions from the pattern of underlying structural connectivity. However, these techniques
rely on the assumption that brain dynamics are linear, when in reality they are highly nonlinear.
Here, we developed a computational modeling approach to investigate whether the predictions
of a region's controllability drawn from a linear model could be validated in a nonlinear model.
Specifically, we built data-derived structural connectivity matrices to computationally explore the
effects of regional stimulation on network dynamics and functional states. Using this model,
20
Figure 7 Structure-function landscape.
(a) Structural and functional effect values for
stimulation of individual brain regions sorted into 9 cognitive systems. Colored ellipses are
centered on the mean structural and functional effects for a given cognitive system and the major
and minor axis of the ellipse represent the standard error of the mean for the associated system.
(b) Same as (a) but data is further course grained into 4 cognitive system types as in36. The
colored regions indicate the convex hull surrounding the data points associated with the given
system. Simulations, and calculations were performed for each of the three scans for each subject
and the data points reflect values averaged over both scans and individuals. (c) Average density
of connections within and between the four cognitive system types. Colors correspond to system
assignments in (b) and dark shades represent the average density of connections between regions
within a single cognitive system while light shades represent the average density of connections
between regions within that system and regions outside of the system.
21
auditory
visual
motor and
somatosensory
ventral temporal
association
attention
fronto-parietal
cingulo-opercular
medial default mode
subcortical
0.6
0.7
0.8
sensorimotor
cortex
higher order
cognitive
medial default
mode
subcortical
0.6
0.7
0.8
0.4
0.3
Functional effect
0.5
0.3
0.4
0.5
Functional effect
Within system
Between system and brain
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0
0.1
0.2
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
a
Structural effect
b
Structural effect
c
Average density of connections
Sensorimotor
Higher order
Medial default
Subcortical
cortex
cognitive
mode
we confirmed predictions from linear control theory, showing that stimulation of high average
controllability regions resulted in global activation that produced little change in the topology of the
functional connectivity, underscoring their role in moving the brain to nearby states. Furthermore,
we investigated the interplay between functional and structural effects of stimulation by examining
how the global functional activity across brain regions was modulated by region-specific stimulation
(a region's functional effect) and whether the region's structural connectivity accounted for its
influence on the larger brain network (structural effect). We observed that the underlying network
connectivity differentially constrained the effects of stimulation: regions of high average controllability
(strongly connected hubs30) displayed a high functional effect -- meaning that they greatly increased
the magnitude of functional connectivity -- while regions of low average controllability (weakly
connected areas30) did not. Yet, stimulation that led to larger changes in functional connectivity
magnitude (functional effect) induced global changes in functional connectivity topology (fractional
activation), moving the system towards easily reachable states as opposed to the more distant states
accessed through focal activation due to stimulation of low controllability regions. Interestingly,
when we parsed brain regions into cognitive systems, we found that stimulation of the medial
default mode network showed both high structural and functional effects, differentiating it from
other subnetworks by its ability to move the system while remaining influenced by the underlying
network connectivity.
Perhaps one of the most striking observations from these data lies in the tradeoff between two
competing consequences of stimulation:
the magnitude of changes in functional connectivity,
22
and the spatial specificity of changes in functional connectivity. We observed that stimulation
to network hubs, predominantly located in default mode and subcortical structures31, 37, 30, induces
widespread increases in the magnitude of functional connectivity between brain regions. Yet, this
broad impact is affected only at the expense of spatial specificity. In contrast, stimulation targeted
to weakly connected areas (low average controllability) predominantly located in fronto-parietal
regions30, induces focal changes in functional connectivity. The differential impact of stimulation
to these two strongly versus weakly connected regions suggests the possibility of two different
classes of therapeautic interventions: (i) a broad reset, in which brain dynamics are globally altered,
and (ii) a focal change, in which brain dynamics of a few regions are altered. These differential
outcomes may offer mechanistic insights into the role of stimulation in distinguishing fine-scale
differences between the concepts38 versus broadly altering general cognitive processes39, 40 or brain
states41.
The differences in the impact of stimulation to hubs versus non-hubs further supports the predictions
of linear network control theory42 and their recent applications to neuroimaging data30. In this
prior work, theoretical insights from structural controllability43 were used to make the prediction
that network hubs -- particularly in the default mode system -- facilitate the movement of the
brain to many easily reachable states.
In contrast, weakly connected nodes of the network --
particularly in cognitive control systems -- were predicted to facilitate the movement of the brain
to difficult-to-reach states. Our results confirm these predictions and offer further insights into
the mechanisms of these control strategies. Specifically, easily reachable states are those that
23
display patterns of functional connectivity that are very similar to those observed in the initial
state, while difficult-to-reach states are those that display patterns of functional connectivity that
are very different from those observed in the initial state.
In addition to these large-scale observations of network state change, we also probed the degree
to which the pattern of functional connectivity (whether focal or global) was correlated with the
pattern of structural connectivity. Prior work has focused largely on the relationship between
resting state functional connectivity and structural connectivity44, 45, 46, under the assumption that
the brain's resting baseline might be highly constrained by anatomy. However, structural connections
likely constrain functional connectivity present in all brain states, irrespective of the cognitive
process at play30.
Indeed, a few recent studies have demonstrated the non-trivial relationships
between individual differences in the pattern of structural connections and the observed functional
connectivity across multiple cognitive states47, 48. Here we observe that the similarity between
structural connectivity and observed functional connectivity depends significantly on the brain
region that was stimulated. Critically, this relationship was not driven by the average controllability
of the region. Along with prior links between average controllability and degree30, these results
suggest that structural constraints on stimulation-elicited functional connectivity can not easily be
predicted by whether a region is a hub or a non-hub.
The relative independence of the functional and structural effects is particularly evident across
large-scale cognitive systems. Indeed, we observe an inverted U-shaped curve between these two
variables: systems that display a middling change in the magnitude of functional connectivity
24
with stimulation tend to display functional connectivity patterns that are most reminiscent of the
underlying structural connectivity.
In contrast, systems that display a very large or very small
change in the magnitude of functional connectivity with stimulation tend to display functional
connectivity patterns that are very different from the structural connectivity. Intuitively, stimulation
to hubs or non-hubs produces brain states that are far from those simply predicted by structural
connectivity, and potentially thus far from normative48. It will be interesting in future to determine
which brain states elicited by stimulation are consistent versus inconsistent with states observed
in normative brain dynamics. Such a question is reminiscent of similar work in control theory
identifying so-called allowable transitions49, 50. More broadly, an understanding of potential brain
states elicited by stimulation is key to deploying stimulation in such a way as to maximize clinical
benefit while minimizing pathological configurations of the network.
There are several important methodological considerations pertinent to this work. While this work
represents an important step in characterizing the effects of stimulation and control in nonlinear
brain networks, it should be noted that the nonlinear model of brain dynamics employed here,
while biologically inspired, is a simplified mean-field model of neuronal dynamics. Additionally,
we have performed only a rough partitioning of the brain into regions and finer scales of regional
partitioning could lead to greater distinction between regional roles as more subtle patterns of brain
connectivity are revealed. Future work is necessary to confirm the effects of spatial resolution on
models of targeted stimulation.
Finally, an important finding of this study was that while we observed a high level of reproducibility
25
in simulations run using connectivities derived from separate scans within a single subject, we
observed variance across subjects in the range of coupling corresponding to the fixed point and
oscillatory regimes of the model. While measures of controllability, functional, and structural
effects collapsed to the same curves across subjects (Fig. 5), the shifting of the oscillatory regime in
the coupling parameter space (Fig. 2c) indicates that the model is sensitive to variances in structural
connectivities between individuals. Although the set of 8 subjects studied here is underpowered to
study individual variation in connectivity and its impact on model performance, this encouraging
result suggests the utility of such approaches to understand individual variability in structural
connections and how it may change the functional effect of stimulation. This modeling approach
provides a means to individualize stimulation protocols for personalized medical treatments or
performance enhancements.
Methods
Human DSI data acquisition and preprocessing
Diffusion spectrum images (DSI) were acquired from a total of 8 subjects in triplicate (mean age
27± 5 years, 2 female, 2 left handed) along with a T 1 weighted anatomical scan at each scanning
session51. DSI scans sampled 257 directions using a Q5 half shell acquisition scheme with a
maximum b value of 5000 and an isotropic voxel size of 2.4mm. We utilized an axial acquisition
with the following parameters: T R = 11.4s, T E = 138ms, 51 slices, FoV (231,231,123 mm).
All participants volunteered with informed consent in accordance with the Institutional Review
26
Board/Human Subjects Committee, University of California, Santa Barbara.
DSI data were reconstructed in DSI Studio (www.dsi-studio.labsolver.org) using q-space diffeomorphic
reconstruction (QSDR)52. QSDR first reconstructs diffusion weighted images in native space and
computes the quantitative anisotropy (QA) in each voxel. These QA values are used to warp the
brain to a template QA volume in MNI space using the SPM nonlinear registration algorithm. Once
in MNI space, spin density functions were again reconstructed with a mean diffusion distance of
1.25 mm using three fiber orientations per voxel. Fiber tracking was performed in DSI Studio with
an angular cutoff of 55◦, step size of 1.0 mm, minimum length of 10 mm, spin density function
smoothing of 0.0, maximum length of 400 mm and a QA threshold determined by DWI signal
in the CSF. Deterministic fiber tracking using a modified FACT algorithm was performed until
100,000 streamlines were reconstructed for each individual.
Anatomical scans were segmented using FreeSurfer53 and parcellated according to the Lausanne
2008 atlas included in the connectome mapping toolkit31. A parcellation scheme including 83
regions was registered to the B0 volume from each subject's DSI data. The B0 to MNI voxel
mapping produced via QSDR was used to map region labels from native space to MNI coordinates.
To extend region labels through the gray/white matter interface, the atlas was dilated by 4mm.
Dilation was accomplished by filling non-labeled voxels with the statistical mode of their neighbors'
labels. In the event of a tie, one of the modes was arbitrarily selected. Each streamline was labeled
according to its terminal region pair.
27
Construction of structural brain networks
We define structural brain networks by subdividing the entire brain into 83 anatomically distinct
brain areas (network nodes)54. Consistent with prior work37, 47, 48, we connect nodes by the number
of white matter streamlines identified by a commonly used deterministic tractography algorithm
described above51 and normalized by the sum of the volumes of the nodes. This procedure results in
sparse, weighted, undirected structural brain networks for each subject (N = 8) and each scanning
session (n = 3), where network connections represent the density of white matter tracts between
brain regions. The definition of structural brain networks based on tractography data in humans
follows from our primary hypothesis that control features of neural dynamics are in part determined
by the structural organization of the brain's white matter tracts.
Mathematical model of brain dynamics
We employ a data-driven, nonlinear model of brain dynamics that allows for the systematic study
of the effects of stimulation to different nodes in the network. Brain regions (modeled by a
single network node) are governed by Wilson-Cowan equations32 representing the population-level
activity of the jth region as follows:
dE j
dt
dIj
dt
= −E j(t) + (Se_max − E j(t))Se c1E j(t)− c2Ij(t) + c5∑
= −Ij(t) + (Si_max − Ij(t))Si (c3E j(t)− c4Ij(t)) + σv j(t)
k
τ
τ
A jkEk(t − τk
d ) + Pj(t)! + σw j(t)
28
where E(t)/I(t) represent the firing rate of the excitatory/inhibitory population respectively, τ = 8
ms is a time constant, P(t) is an external stimulus parameter, and w j(t) and v j(t) are drawn from
a standard normal distribution and act as additive noise to the system with σ = 0.00001. Regions
are coupled through the excitatory population with a connectivity matrix A derived from a single
scan of an individual subject's tractography data, parcellated to give a total of 83 brain regions, and
normalized by region size. Delays, τd, between regions are calculated as a function of physical
distance between identified brain regions, assuming a transmission velocity of 10 m/s (range τd =
0.8− 14.8 ms). The transfer function is given by the sigmoidal function
.
Se/i(x) =
1
1
1 + e(−ae/i(x−θe/i)) −
1 + eae/iθe/i
Model constants are set as in32 to be c1 = 16, c2 = 12, c3 = 15, c4 = 3, ae = 1.3, ai = 2, θe = 4,
θi = 3.7. These parameter choices imply that an uncoupled oscillator has three states: a low fixed
point, limit cycle, and high fixed point. For a single oscillator, increasing the current input into
the oscillator, P(t), will allow it to transition between the three states. In our model of coupled
oscillators, the global coupling parameter, c5, serves as an additional mechanism for allowing the
system to transition between these states for fixed values of P(t). To simulate brain activity, we set
P(t) = 0 for all regions, and we model stimulation to single brain region, j, by setting Pj(t) = 1.25.
This value of P implies that an uncoupled oscillator will enter into limit cycle activity with a
frequency near 20 Hz which is inline with the biological range for oscillatory brain activity. All
simulations are allowed to initially stabilize for 1s before analysis, and the temporal dynamics of
the jth brain region are given by the firing rate of the excitatory population, E j(t).
29
Evaluating oscillatory transition parameters
In order to asses the point in the parameter space at which the system switched from the low
fixed point to the oscillatory regime, we ran 1s simulations in which no stimulating current was
applied (P = 0) for increasing values of the global coupling parameter (c5 = 1.0 to c5 = 1.5 in step
sizes of 0.05). The average firing rate of the excitatory population for each region was recorded
as a function of the global coupling parameter and is shown for stimulations using a single scan
from two separate subjects in Fig. S2. At a certain value of the global coupling parameter that
varied between scans and subjects, we see a sudden increase in the firing rate across most regions,
indicating the transition of the system from the fixed point to the oscillatory regime. We use the
value of c5 at which this transition occurs to assess the reproducibility and variability within and
between subjects. For all simulations assessing the impact of regional stimulation, the value of c5
immediately before this transition was used.
Inter- and intra-subject reproducibility and variability
Inter- and intra-subject reproducibility were calculated using the intraclass correlation coefficient
(ICC), and assuming a random effect model as in55. The between subject ICC is estimated using
the mean squares (MS) obtained by applying ANOVA with the subject as the factor to the value of
global coupling at which the system switches to the oscillator state obtained as described above:
ICCB =
J(SMS− EMS)
J ∗ SMS + I ∗ RMS + (IJ − I − J)EMS
.
(1)
30
Similarly, the within subject ICC [jmv6]is given by
ICCW =
I(RMS− EMS)
J ∗ SMS + I ∗ RMS + (IJ − I − J)EMS
,
(2)
where I is the number of subjects, J is the number of scans, SMS is the MS between scans, RMS is
the MS between subjects, and EMS is the MS due to error.
The between subject variability was defined to be the variance of the mean value of the transition
point obtained for each subject, and the within subject variability was defined to be the within
subject variance, averaged across all subjects.
Linear network control theory
To study the theoretical ability of a certain brain region to influence other regions in arbitrary ways
we adopt the control theoretic notion of controllability. Controllability of a dynamical system refers
to the possibility of driving the state of a dynamical system to a specific target state by means of
an external control input56. As in30, we employ a simplified noise-free linear discrete-time and
time-invariant network model:
x(t + 1) = Ax(t) + BK uK (t),
(3)
where x describes the state of brain regions over time, A is a normalized structural connectivity
matrix derived from tractography data as described above. The Supplementary Material contains
a discussion of the method used to normalize matrices across subject data sets. The input matrix
BK identifies the control points, and uK denotes the control strategy.
31
(4)
AτBK BT
K Aτ .
∞∑τ
=0
WK =
Classic results in control theory ensure that controllability of the network (3) from the set of
network nodes K is equivalent to the controllability Gramian WK being invertible, where
We utilize this framework to choose control nodes one at a time, and thus the input matrix B in fact
reduces to a one-dimensional vector.
We examine 2 diagnostics of controllability utilized in the network control literature: average
controllability and modal controllability30. Average controllability of a network equals the average
input energy from a set of control nodes and over all possible target states57, 58. As in30, we adopt
Trace(WK) as a measure of average controllability. Regions with high average controllability are,
on average, most influential in the control of network dynamics over all nearby target states with
least energy. Modal controllability refers to the ability of a node to control each evolutionary mode
of a dynamical network59, and can be used to identify states that are difficult to control from a
set of control nodes. Modal controllability is computed from the eigenvector matrix V = [vi j] of
the network adjacency matrix A. Regions with high modal controllability are able to control all
the dynamic modes of the network, and hence they can drive the dynamics towards hard-to-reach
configurations.
The stimulation paradigm that we study is a constant input over a short period of time, whereas the
controllability metrics discussed above assume a more general time-varying input. We treat this
constant current stimulation as a a simple approximation of the more general stimulation paradigm
32
traditionally studied in the control literature, and therefore use average and modal controllability
to assess control architecture. Because our stimulation paradigm is given by a constant input, one
can also examine the steady state response of the network. In the Supplementary Information and
Fig. S1, we study the more constrained steady state response of the network, and demonstrate that
our approximation presented here is an accurate one. However, it should be noted that the statistics
we use here (average/modal controllability) may not always be an adequate approximation of the
steady state response of a network, and therefore care should be taken in using these statistics in
other studies without first demonstrating their consistency with the steady state response statistics.
Quantifying functional brain states
We quantify brain states by calculating the pairwise maximum normalized cross-correlation60, 61
between the firing rate of the excitatory populations, Ei(t) and E j(t), for brain regions i and j. All
calculations are performed using a 1 s window and a maximum lag of 250 ms.
Functional effect of stimulation
Simulations of neural dynamics are first allowed to stabilize for 1s to reach the stable activity from
the global coupling parameter and then the remaining time is divided into two parts: a stimulation
free period of 1s, followed by a 1s period of stimulation. During the stimulation period, a single
region, s, is selected and a stimulus is applied to this region by setting Ps(t) = 1.25, while P(t) =
0 for all other regions. Functional brain states are computed separately for the stimulation-free
('before' in Fig. 4) and stimulation ('during' in Fig. 4) periods. We assess the pairwise change in
33
functional brain states by subtracting the correlation values obtained in the stimulation-free time
window before the stimulation is applied from the correlation values obtained in the time window
during the stimulation. We then measure the average change in functional brain states, termed the
functional effect, as the absolute value of this difference averaged over all region pairs. The greater
the deviation of the functional effect from zero, the greater the effect of stimulation on brain state
reconfiguration.
Structural effect on network dynamics
In order to assess how the structural brain connectivity constrains network dynamics, we calculate
the similarity, defined by the 2-dimensional correlation coefficient, between the structural connectivity
matrix used as a basis of the simulation and the functional matrix describing the resultant brain
state. This calculation is done first using the initial functional brain state before stimulation, and
then using the functional brain state during the stimulation period. Stimulation of a single region
increases the similarity between the structural connectivity and functional brain state, and we
quantify this increase, termed the structural effect, by subtracting the correlation obtained before
stimulation from that obtained during stimulation. Thus, a high structural effect indicates that the
underlying connectivity structure constrains the functional effects of stimulation to this region.
Fractional activation
The fractional activation due to the stimulation of a single brain region is defined to be the fraction
of pairwise regions that experience a change in their functional connectivity value that is above a
34
given threshold. The change in functional connectivity is calculated as described above to calculate
the functional effect by assessing the absolute change in pairwise correlation values before and
during regional stimulation. If the stimulation of a brain region results in large changes that occur
globally throughout the brain, the fractional activation will be high, but if the stimulation has only
a focal effect, the fractional activation will be low. Here, we report findings using a threshold value
of 0.6, however, results were similar across a range of thresholds (Fig. S3.)
35
References
1. Volz, L. J. et al. Motor cortex excitability and connectivity in chronic stroke: a multimodal
model of functional reorganization. Brain structure & function 220, 1093 -- 1107 (2015).
2. Helfrich, C. et al. Monitoring cortical excitability during repetitive transcranial magnetic
stimulation in children with ADHD: a single-blind, sham-controlled TMS-EEG study. PloS
one 7, e50073 (2012).
3. Tierney, T. S., Vasudeva, V. S., Weir, S. & Hayes, M. T.
Neuromodulation for
neurodegenerative conditions. Frontiers in bioscience (Elite edition) 5, 490 -- 499 (2013).
4. Temel, Y. et al. Neuromodulation in psychiatric disorders.
International review of
neurobiology 107, 283 -- 314 (2012).
5. Terra, V. C. et al. Vagus nerve stimulator in patients with epilepsy:
indications and
recommendations for use. Associação Arquivos de Neuro-Psiquiatria (2013).
6. DeGiorgio, C. M. & Krahl, S. E. Neurostimulation for Drug-Resistant Epilepsy. Continuum :
Lifelong Learning in Neurology 19, 743 -- 755 (2013).
7. Di Lazzaro, V. et al. Inhibitory theta burst stimulation of affected hemisphere in chronic stroke:
A proof of principle, sham-controlled study. Neuroscience Letters 553, 148 -- 152 (2013).
8. Vanneste, S., Fregni, F. & De Ridder, D. Head-to-Head Comparison of Transcranial
Random Noise Stimulation, Transcranial AC Stimulation, and Transcranial DC Stimulation
for Tinnitus. Frontiers in psychiatry 4, 158 (2013).
36
9. Jürgens, T. P. & Leone, M. Pearls and pitfalls: neurostimulation in headache. Cephalalgia :
an international journal of headache 33, 512 -- 525 (2013).
10. Shah, P. P., Szaflarski, J. P., Allendorfer, J. & Hamilton, R. H. Induction of neuroplasticity
and recovery in post-stroke aphasia by non-invasive brain stimulation. Frontiers in human
neuroscience 7, 888 (2013).
11. Bonnì, S., Mastropasqua, C., Bozzali, M., Caltagirone, C. & Koch, G. Theta burst stimulation
improves visuo-spatial attention in a patient with traumatic brain injury. Neurological sciences
: official journal of the Italian Neurological Society and of the Italian Society of Clinical
Neurophysiology 34, 2053 -- 2056 (2013).
12. Hasan, A., Falkai, P. & Wobrock, T. Transcranial brain stimulation in schizophrenia: targeting
cortical excitability, connectivity and plasticity. Current medicinal chemistry 20, 405 -- 413
(2013).
13. Berardelli, A. & Suppa, A. Noninvasive brain stimulation in Huntington's disease. Handbook
of clinical neurology 116, 555 -- 560 (2013).
14. Andrade, D. C., Borges, I., Bravo, G. L., Bolognini, N. & Fregni, F. Therapeutic time window
of noninvasive brain stimulation for pain treatment: inhibition of maladaptive plasticity with
early intervention. Expert review of medical devices 10, 339 -- 352 (2013).
15. Lozano, A. M. & Lipsman, N. Probing and regulating dysfunctional circuits using deep brain
stimulation. Neuron 77, 406 -- 424 (2013).
37
16. Meinzer, M. et al. Transcranial direct current stimulation over multiple days improves learning
and maintenance of a novel vocabulary. Cortex; a journal devoted to the study of the nervous
system and behavior 50, 137 -- 147 (2014).
17. Ferreri, F. & Rossini, P. M. TMS and TMS-EEG techniques in the study of the excitability,
connectivity, and plasticity of the human motor cortex. Reviews in the neurosciences 24,
431 -- 442 (2013).
18. Johnson, M. D. et al. Neuromodulation for brain disorders: challenges and opportunities.
IEEE Trans Biomed Eng 60, 610 -- 624 (2013).
19. Chaieb, L. et al.
Short-duration transcranial random noise stimulation induces blood
oxygenation level dependent response attenuation in the human motor cortex. Experimental
brain research 198, 439 -- 444 (2009).
20. Hess, C. W. Modulation of cortical-subcortical networks in Parkinson's disease by applied
field effects. Frontiers in human neuroscience 7, 565 (2013).
21. Grafton, S. T. et al. Normalizing motor-related brain activity: subthalamic nucleus stimulation
in Parkinson disease. Neurology 66, 1192 -- 1199 (2006).
22. Soekadar, S. R. et al. In vivo assessment of human brain oscillations during application of
transcranial electric currents. Nature Communications 4, 2032 (2013).
23. Parks, N. A. et al.
Examining cortical dynamics and connectivity with simultaneous
38
single-pulse transcranial magnetic stimulation and fast optical imaging. NeuroImage 59,
2504 -- 2510 (2012).
24. Pascual-Leone, A. et al. Characterizing Brain Cortical Plasticity and Network Dynamics
Across the Age-Span in Health and Disease with TMS-EEG and TMS-fMRI. Brain Topogr
24, 302 -- 315 (2011).
25. Leitão, J., Thielscher, A., Werner, S., Pohmann, R. & Noppeney, U. Effects of parietal TMS on
visual and auditory processing at the primary cortical level -- a concurrent TMS-fMRI study.
Cerebral Cortex 23, 873 -- 884 (2013).
26. Levit-Binnun, N., Handzy, N. Z., Moses, E., Modai, I. & Peled, A. Transcranial Magnetic
Stimulation at M1 disrupts cognitive networks in schizophrenia. Schizophrenia Research 93,
334 -- 344 (2007).
27. Grefkes, C. & Fink, G. R. Reorganization of cerebral networks after stroke: new insights from
neuroimaging with connectivity approaches. Brain : a journal of neurology 134, 1264 -- 1276
(2011).
28. Arzouan, Y., Moses, E., Peled, A. & Levit-Binnun, N.
Impaired network stability in
schizophrenia revealed by TMS perturbations. Schizophrenia Research 152, 322 -- 324 (2014).
29. Bestmann, S. & Feredoes, E. Combined neurostimulation and neuroimaging in cognitive
neuroscience: past, present, and future. Annals of the New York Academy of Sciences 1296,
11 -- 30 (2013).
39
30. Gu, S. et al. Controllability of structural brain networks. Nature Communications 6, 8414
(2015).
31. Hagmann, P. et al. Mapping the structural core of human cerebral cortex. PLoS Biology 6,
e159 (2008).
32. Wilson, H. R. & Cowan, J. D. Excitatory and inhibitory interactions in localized populations
of model neurons. Biophysj 12, 1 -- 24 (1972).
33. Wilson, H. R. & Cowan, J. D. A mathematical theory of the functional dynamics of cortical
and thalamic nervous tissue. Kybernetik 13, 55 -- 80 (1973).
34. Power, J. D. et al. Functional Network Organization of the Human Brain. Neuron 72, 665 -- 678
(2011).
35. Horn, A., Ostwald, D., Reisert, M. & Blankenburg, F. The structural -- functional connectome
and the default mode network of the human brain. NeuroImage 102, 142 -- 151 (2014).
36. Gu, S. et al. Emergence of system roles in normative neurodevelopment. Proceedings of the
National Academy of Sciences of the United States of America 112, 13681 -- 13686 (2015).
37. Bassett, D. S., Brown, J. A., Deshpande, V., Carlson, J. M. & Grafton, S. T. Conserved and
variable architecture of human white matter connectivity. NeuroImage 54, 1262 -- 1279 (2011).
38. Lecce, F., Walsh, V., Didino, D. & Cappelletti, M. 'how many' and 'how much' dissociate in
the parietal lobe. Cortex 73, 73 -- 79 (2015).
40
39. Kraft, A. et al. TMS over the right precuneus reduces the bilateral field advantage in visual
short term memory capacity. Brain Stimul 8, 216 -- 223 (2015).
40. Bonni, S. et al. TMS evidence for a selective role of the precuneus in source memory retrieval.
Behav Brain Res 282, 70 -- 75 (2015).
41. Crescentini, C., Di Bucchianico, M., Fabbro, F. & Urgesi, C. Excitatory stimulation of the
right inferior parietal cortex lessens implicit religiousness/spirituality. Neuropsychologia 70,
71 -- 79 (2015).
42. Pasqualetti, F., Zampieri, S. & Bullo, F. Controllability metrics, limitations and algorithms for
complex networks. IEEE Trans. Control Netw. Syst. 1, 40 -- 52 (2014).
43. Kailath, T. Linear Systems (Prentice-Hall, 1980).
44. Honey, C. J. et al. Predicting human resting-state functional connectivity from structural
connectivity. Proc Natl Acad Sci USA 106, 2035 -- 40 (2009).
45. Honey, C. J., Thivierge, J. P. & Sporns, O. Can structure predict function in the human brain?
Neuroimage 52, 766 -- 776 (2010).
46. Goñi, J. et al. Resting-brain functional connectivity predicted by analytic measures of network
communication. Proc Natl Acad Sci U S A 111, 833 -- 838 (2014).
47. Hermundstad, A. M. et al. Structural foundations of resting-state and task-based functional
41
connectivity in the human brain. Proceedings of the National Academy of Sciences of the
United States of America 110, 6169 -- 6174 (2013).
48. Hermundstad, A. M. et al. Structurally-Constrained Relationships between Cognitive States
in the Human Brain. PLoS Computational Biology 10, e1003591 (2014).
49. Cornelius, S. P., Kath, W. L. & Motter, A. E. Realistic control of network dynamics. Nat
Commun 4, 1942 (2013).
50. Sun, J. & Motter, A. E. Controllability transition and nonlocality in network control. Phys Rev
Lett 110, 208701 (2013).
51. Cieslak, M. & Grafton, S. T. Local termination pattern analysis: a tool for comparing white
matter morphology. Brain imaging and behavior 8, 292 -- 299 (2014).
52. Yeh, F.-C. & Tseng, W.-Y. I. NTU-90: a high angular resolution brain atlas constructed by
q-space diffeomorphic reconstruction. NeuroImage 58, 91 -- 99 (2011).
53. Dale, A. M., Fischl, B. & Sereno, M. I. Cortical surface-based analysis. I. Segmentation and
surface reconstruction. NeuroImage 9, 179 -- 194 (1999).
54. Cammoun, L. et al. Mapping the human connectome at multiple scales with diffusion spectrum
MRI. Journal of Neuroscience Methods 203, 386 -- 397 (2012).
55. Wei, X. et al. Functional MRI of auditory verbal working memory: long-term reproducibility
analysis. NeuroImage 21, 1000 -- 1008 (2004).
42
56. Kalman, R. E., Ho, Y. C. & Narendra, K. S. Controllability of linear dynamical systems .
Contributions to Differential Equations 1, 189 -- 213 (1963).
57. Marx, B., Koenig, D. & Georges, D. Optimal sensor and actuator location for descriptor
systems using generalized Gramians and balanced realizations 3, 2729 -- 2734 vol.3 (2004).
58. Shaker, H. R. & Tahavori, M. Optimal sensor and actuator location for unstable systems.
Journal of Vibration and Control 19, 1077546312451302 -- 1920 (2012).
59. Hamdan, A. M. A. & Nayfeh, A. H. Measures of modal controllability and observability
for first- and second-order linear systems. Journal of Guidance, Control, and Dynamics 12,
421 -- 428 (1989).
60. Kramer, M., Eden, U., Cash, S. & Kolaczyk, E. Network inference with confidence from
multivariate time series. Physical Review E 79, 061916 (2009).
61. Feldt, S., Osterhage, H., Mormann, F., Lehnertz, K. & Zochowski, M.
Internetwork and
intranetwork communications during bursting dynamics: Applications to seizure prediction.
Physical Review E 76, 021920 (2007).
Acknowledgements
Research was sponsored by the Army Research Laboratory and was accomplished under Cooperative
Agreement Number W911NF-10-2-0022. DSB acknowledges support from the John D. and Catherine
T. MacArthur Foundation, the Alfred P. Sloan Foundation, the Army Research Office (W911NF-14-1-0679),
43
the National Institute of Mental Health (2-R01-DC-009209-11), the National Institute of Child
Health and Human Development (1R01HD086888-01), the Office of Naval Research, and the
National Science Foundation (BCS-1441502 and BCS-1430087). FP acknowledges support from
the National Science Foundation (BCS-1430279) and STG acknowledges support from the Institute
for Collaborative Biotechnologies through the US Army Research Office (W911NF-09-0001). The
content is solely the responsibility of the authors and does not necessarily represent the official
views of any of the funding agencies.
Author Contributions
SFM, FP, JV, and DSB conceptualized the project; SFM, SG, MC, and STG contributed code/data;
SFM analyzed the data; and SFM, FP, JV, and DSB wrote the manuscript.
44
Supplementary Information for "Stimulation-based control of
dynamic brain networks"
Sarah Feldt Muldoon1,2,3, Fabio Pasqualetti4, Shi Gu1,5, Matthew Cieslak6, Scott T. Grafton6, Jean
M. Vettel2,6,1 and Danielle S. Bassett1,7
1Department of Bioengineering, University of Pennsylvania, Philadelphia, PA 19104
2US Army Research Laboratory, Aberdeen Proving Ground, MD 21005
3Department of Mathematics and Computational and Data-Enabled Science and Engineering Program,
University at Buffalo, SUNY, Buffalo, NY 14260
4Department of Mechanical Engineering, University of California, Riverside, CA 92521
5Applied Mathematics and Computational Science Graduate Program, University of Pennsylvania,
Philadelphia, PA 19104
6Department of Psychological and Brain Sciences, University of California, Santa Barbara, Santa
Barbara, CA 93106
7Department of Electrical and Systems Engineering, University of Pennsylvania, Philadelphia, PA
19104
arXiv:1601.00987v1 [q-bio.NC] 5 Jan 2016
1
Normalization of structural matrices across subjects
When doing controllability calculations, we must ensure that the structural adjacency matrix describing
the tractography connectivity is Schur stable. In Gu et al.1, this was obtained by normalizing each
structural matrix by one plus its largest singular value. Because each matrix effectively received
a slightly different normalization factor, regional controllability results were ranked in order to
perform averaging across scans and subjects. To avoid the loss of information that results from
ranking the data, we instead employed a different form of normalization. We first calculated
the maximum eigenvalue for each structural matrix in the data set. From this pool of maximum
eigenvalues, we then selected the maximum value, and divided all structural matrixes by two times
this quantity. This ensures Schur stability and allows us to compare directly compare regional
controllability values obtained from different structural matrices. The controllability results presented
in this paper therefore represent the resultant numerical values of regional controllability calculations,
as opposed to the ranked values presented in Gu et al1.
Average controllability and the steady state response
In linear network control theory, the controllability of a network refers to the possibility of altering
the configuration of the network nodes via external stimuli and in a predictable way. To quantify
the degree of controllability of a network, we first model the dynamical interaction among network
nodes by means of a discrete-time, linear, time-invariant system:
x(t + 1) = Ax(t) + BKu(t).
2
In the equation above, x is a vector containing the states of the network nodes, A = AT is a (stable)
weighted adjacency matrix of the network, u is the external control signal, and BK identifies the
control nodes; see also1, 2.
In our simulations, we use a constant input current to stimulate brain regions, and therefore are
interested in the steady state response of the network. With a constant control input, the network
steady state is
xsteady = (I − A)−1BKuconstant,
where uconstant is the value of the constant input. Thus, the steady state effect of a constant input
to the i-th region is characterized by the i-th column of the matrix (I − A)−1. The largest value of
the i-th column gives steady state response of the region maximally effected by the input, while the
average of the i-th column gives the average steady state effect over all regions.
The use of a constant input as regional stimulation is a special case of the more general paradigm
of a time-varying input normally used to define network control statstics such as those used in the
main manuscript (average/modal controllability). We therefore would like to relate this steady state
response to the average controllability which describes the more general paradigm. The degree of
controllability of a network can be quantified in different ways2, but in this paper, we use the
classical definition of the Controllability Gramian, that is,
AτBKBT
KAτ ,
∞∑τ
=0
WK =
3
and measure the average degree of controllability as Trace(WK), which has a specific system
AτBKBT
KAτ!
Trace(cid:16)A2τBKBT
K(cid:17)
= Trace ∞
K!
= Trace(cid:16)(I − A2)−1BKBT
K(cid:17)
= ∑
i∈K
(I − A2)−1
A2τBKBT
∑τ
=0
∞∑τ
=0
=
.
ii
=0
∑τ
Trace(WK) = Trace ∞
theoretic interpretation1, 2, 3. Moreover, notice that
In other words, the controllability degree with control nodes K equals the sum of the diagonal
entries of (I − A2)−1 indexed by K. Because of the normalization of the adjacency matrix adopted
in this work, it can be verified that
(I − A)−1 ≈ (I − A2)−1,
so that the average controllability information can be reconstructed from the steady state response
matrix (I−A)−1. Specifically, for stimulation of a single region, the largest entry of the i-th column
of the steady state matrix will be approximately equal to the average controllability.
As seen in Fig. S1a, when we plot the functional effect of stimulation as a function of the largest
steady state value, we do indeed reproduce the results of Fig. 5a. Additionally, we see a similar
result when plotting the functional effect of stimulation for the average steady state value (Fig. S1b).
4
Therefore, in the main manuscript, we present our findings in terms of the more general regional
controllability values instead of the steady state response, which also allows for comparison with
previous work using these measures to study the properties of structural brain networks1. However,
it should be noted that the average/modal controllability may not always be an adequate approximation
of the steady state response of a network, and therefore care should be taken in using these
statistics in other studies without first demonstrating their consistency with the steady state response
statistics.
Mapping regions to cognitive systems
Similar to the assignment of brain regions in Gu et al.1 and inspired by Power et al.4, we initially
assigned each of the 83 brain regions to one of 9 cognitive systems. Our only divergence from
Gu et al. was the creation of a "ventral temporal association" category to bin perceptual regions
associated with invariant object representations and multisensory activation. For further analysis,
this assignment was coarse grained into 4 cognitive systems5. The placement of each region into
each cognitive system is summarized in Table S1.
5
Figure 1 Steady state response and functional effect. The functional effect resulting from
regional stimulation plotted as a function of (a) the largest steady state value (b) the average steady
state value.
6
1.02
1.04
Largest steady state response
1.03
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
1.05
1.06
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
0
1
1.01
0.8
0.6
0.4
0.2
b
Functional effect
a
0.8
0.6
0.4
0.2
Functional effect
0
0.01
0.015
0.02
0.025
0.03
0.035
Average steady state response
Region Name
Lateral Orbitofrontal
Pars Orbitalis
Frontal Pole
Medial Orbitofrontal
Pars Triangularis
Pars Opercularis
Rostral Middle Frontal
Superior Frontal
Caudal Middle Frontal
Precentral
Paracentral
Rostral Anterior Cingulate
Caudal Anterior Cingulate
Posterior Cingulate
Isthmus Cingulate
Post Central
Supramarginal
Superior Parietal
Inferior Parietal
Precuneus
Cuneus
Pericalcarine
Lateral Occipital
Lingual
Fusiform
Parahippocampal
Entorhinal Cortex
Temporal Pole
Inferior Temporal
Middle Temporal
Bank of the Superior Temporal Sulcus
Superior Temporal
Transverse Temporal
Insula
Thalamus
Caudate
Putamen
Pallidum
Nucleus Accumbens
Hippocampus
Amygdala
Brainstem
9 system assignment
attention
cingulo-opercular
fronto-parietal
fronto-parietal
fronto-parietal
cingulo-opercular
cingulo-opercular
medial default mode
fronto-parietal
motor and somatosensory
motor and somatosensory
cingulo-opercular
cingulo-opercular
medial default mode
medial default mode
motor and somatosensory
cingulo-opercular
attention
fronto-parietal
medial default mode
visual
visual
visual
visual
ventral temporal association
ventral temporal association
ventral temporal association
ventral temporal association
ventral temporal association
ventral temporal association
ventral temporal association
auditory
auditory
fronto-parietal
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
4 system assignment
higher order cognitive
higher order cognitive
higher order cognitive
higher order cognitive
higher order cognitive
higher order cognitive
higher order cognitive
medial default mode
higher order cognitive
sensorimotor cortex
sensorimotor cortex
higher order cognitive
higher order cognitive
medial default mode
medial default mode
sensorimotor cortex
higher order cognitive
higher order cognitive
higher order cognitive
medial default mode
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
sensorimotor cortex
higher order cognitive
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
subcortical
Table 1 Assignment of brain regions to cognitive systems.
7
Figure 2 Transition to oscillatory regime. (a-b) Examples of the transition to the oscillatory
regime in simulations from a single scan obtained from two different subjects. In (a) the transition
occurs for a global coupling value of c5 = 1.4 whereas in (b) the transition occurs for c5 = 1.25.
8
Average firing rate, E
0.4
0.3
0.2
0.1
0
Average firing rate, E
0.4
0.3
0.2
0.1
0
1.25
1.15
Global coupling parameter, c5
1.35
1.45
1.25
1.15
Global coupling parameter, c5
1.35
1.45
1.05
1.05
10
20
30
40
50
60
70
80
10
20
30
40
50
60
70
80
Region number
Region number
a
b
Figure 3 Fractional activation for varied threshold values.
(a-b) The fractional activation
calculated using a threshold value of 0.2 shown as a function of Functional Effect (a) and Structural
Effect (b). (c-d) The same as (a-b) but using a threshold value of 0.4. These results are comparable
to that shown in Fig. 6c-d of the main manuscript.
9
0.05
0.1
0.15
0.2
Structural effect
0.25
0.3
0.35
0.05
0.1
0.15
0.2
Structural effect
0.25
0.3
0.35
1
0.8
0.6
0.4
0.2
Fractional activation
0
0
1
0.8
0.6
0.4
0.2
Fractional activation
0
0
b
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
d
Subject 1
Subject 2
Subject 3
Subject 4
Subject 5
Subject 6
Subject 7
Subject 8
0.6
0.7
0.8
0.6
0.7
0.8
0.4
0.3
Functional effect
0.5
0.4
0.3
Functional effect
0.5
0
0
0.1
0.2
0
0
0.1
0.2
c
1
0.8
0.6
0.4
0.2
Fractional activation
1
0.8
0.6
0.4
0.2
Fractional activation
a
References
1. Gu, S. et al. Controllability of structural brain networks. Nature Communications 6, 8414
(2015).
2. Pasqualetti, F., Zampieri, S. & Bullo, F. Controllability metrics, limitations and algorithms for
complex networks. IEEE Trans. Control Netw. Syst. 1, 40 -- 52 (2014).
3. Kailath, T. Linear Systems (Prentice Hall, 1980).
4. Power, J. D. et al. Functional Network Organization of the Human Brain. Neuron 72, 665 -- 678
(2011).
5. Gu, S. et al. Emergence of system roles in normative neurodevelopment. Proceedings of the
National Academy of Sciences of the United States of America 112, 13681 -- 13686 (2015).
10
|
1601.02948 | 1 | 1601 | 2016-01-12T16:38:14 | An Efficient Coding Theory for a Dynamic Trajectory Predicts non-Uniform Allocation of Grid Cells to Modules in the Entorhinal Cortex | [
"q-bio.NC"
] | Grid cells in the entorhinal cortex encode the position of an animal in its environment using spatially periodic tuning curves of varying periodicity. Recent experiments established that these cells are functionally organized in discrete modules with uniform grid spacing. Here we develop a theory for efficient coding of position, which takes into account the temporal statistics of the animal's motion. The theory predicts a sharp decrease of module population sizes with grid spacing, in agreement with the trends seen in the experimental data. We identify a simple scheme for readout of the grid cell code by neural circuitry, that can match in accuracy the optimal Bayesian decoder of the spikes. This readout scheme requires persistence over varying timescales, ranging from ~1ms to ~1s, depending on the grid cell module. Our results suggest that the brain employs an efficient representation of position which takes advantage of the spatiotemporal statistics of the encoded variable, in similarity to the principles that govern early sensory coding. | q-bio.NC | q-bio | Authors
Correspondence
A
Theory
for
Efficient
Coding
of
a
Dynamic
Trajectory
Predicts
non-‐Uniform
Allocation
of
Grid
Cells
to
Modules
in
the
Entorhinal
Cortex
Noga
Weiss
Mosheiff1,
Haggai
Agmon2,
Avraham
Moriel1,
and
Yoram
Burak1,2
1
Racah
Institute
of
Physics,
Hebrew
University,
Edmond
J.
Safra
Campus,
Jerusalem,
Israel
2
Edmond
and
Lily
Safra
Center
for
Brain
Sciences,
Hebrew
University,
Edmond
J.
Safra
Campus,
Jerusalem,
Israel
[email protected]
September
20,
2015
Grid
cells
in
the
entorhinal
cortex
encode
the
position
of
an
animal
in
its
environment
using
spatially
periodic
tuning
curves
of
varying
periodicity.
Recent
experiments
established
that
these
cells
are
functionally
organized
in
discrete
modules
with
uniform
grid
spacing.
Here
we
develop
a
theory
for
efficient
coding
of
position,
which
takes
into
account
the
temporal
statistics
of
the
animal’s
motion.
The
theory
predicts
a
sharp
decrease
of
module
population
sizes
with
grid
spacing,
in
agreement
with
the
trends
seen
in
the
experimental
data.
We
identify
a
simple
scheme
for
readout
of
the
grid
cell
code
by
neural
circuitry,
that
can
match
in
accuracy
the
optimal
Bayesian
decoder
of
the
spikes.
This
readout
scheme
requires
persistence
over
varying
timescales,
ranging
from
~1ms
to
~1s,
depending
on
the
grid
cell
module.
Our
results
suggest
that
the
brain
employs
an
efficient
representation
of
position
which
takes
advantage
of
the
spatiotemporal
statistics
of
the
encoded
variable,
in
similarity
to
the
principles
that
govern
early
sensory
coding.
SUMMARY
INTRODUCTION
A
central
goal
of
systems
neuroscience
is
to
unravel
the
principles
of
encoding
in
the
brain.
In
primary
sensory
areas,
it
has
been
conjectured
that
the
neural
circuitry
implements
coding
schemes
that
maximize
information
about
sensory
inputs,
while
constraining
neural
resources
such
as
the
number
of
cells
or
the
rate
of
spikes.
This
hypothesis
(Barlow,
1961)
has
been
particularly
successful
in
explaining
neural
responses
in
early
visual
and
auditory
areas
(Atick
and
Redlich,
1992;
Bell
and
Sejnowski,
1997;
Fairhall
et
al.,
2001;
Laughlin,
1981;
Olshausen
and
Field,
1997;
Smith
and
Lewicki,
2006).
More
recently,
it
was
proposed
that
grid
cells
in
the
entorhinal
cortex
(Hafting
et
al.,
2005)
implement
an
efficient
code
for
an
internally
computed
quantity,
the
position
of
an
animal
in
its
environment
(Fiete
et
al.,
2008;
Mathis
et
al.,
2012a;
2012b;
Sreenivasan
and
Fiete,
2011;
Wei
et
al.,
2015).
According
to
this
proposal,
the
neural
code
for
position
possesses
a
dynamic
range
(defined
as
the
ratio
between
the
representable
range
and
the
resolution)
that
depends
exponentially
on
the
number
of
encoding
neurons
(Burak,
2014).
Thus,
the
dynamic
range
of
the
grid
cell
code
vastly
exceeds
that
of
unimodal
coding
schemes,
such
as
the
encoding
of
position
in
the
hippocampus
(Hartley
et
al.,
2014),
or
the
encoding
in
head
direction
cells
(Taube,
2007).
Most
theoretical
treatments
of
grid
cell
coding
assumed
a
uniform
distribution
of
neurons
across
grid
cell
modules
(Fiete
et
al.,
2008;
Mathis
et
al.,
2012b),
or
deduced
that
such
a
distribution
is
expected
based
on
an
optimization
principle
(Mathis
et
al.,
2012a;
Wei
et
al.,
2015)
(here,
a
module
is
defined
as
a
group
of
grid
cells
that
share
the
same
grid
spacing
and
orientation).
However,
in
a
systematic
characterization
of
grid
cell
parameters
(Stensola
et
al.,
2012),
many
more
cells
were
found
in
modules
with
small
spacing,
compared
to
modules
with
larger
spacing
(Fig.
1A).
The
study
by
Stensola
et
al
(Stensola
et
al.,
2012)
did
not
attempt
to
quantify
precisely
the
population
sizes
in
different
modules,
and
the
reported
numbers
were
likely
influenced
by
experimental
biases,
but
the
trend
observed
in
this
study
is
so
pronounced,
that
it
is
highly
suggestive
of
a
non-‐uniform
distribution
of
grid
cells
across
modules.
This
observation
raises
two
questions:
First,
is
a
non-‐uniform
distribution
of
grid
cells
across
modules
compatible
with
the
efficient
coding
hypothesis?
Second,
if
the
number
of
cells
in
modules
with
large
spacing
is
indeed
very
small
as
hinted
by
the
existing
experimental
evidence,
can
these
cells
support
at
all
a
precise
neural
representation
of
position?
Here
we
propose
that
the
entorhinal
cortex
is
adapted
to
represent
a
dynamic
quantity,
the
animal’s
trajectory
in
its
environment,
while
taking
into
account
the
temporal
statistics
of
this
variable
in
addition
to
the
requirements
on
spatial
range
and
resolution.
This
hypothesis
leads
to
the
prediction
that
grid
cell
population
sizes
should
decrease
sharply
as
a
function
of
grid
spacing,
whereas
grid
spacings
should
follow
approximately
a
geometric
series
as
predicted
previously
(Mathis
et
al.,
2012b;
2012a;
Wei
et
al.,
2015)
–
in
agreement
with
experimental
evidence
(Stensola
et
al.,
2012).
We
show,
in
addition,
that
fairly
simple
neural
circuitry
can
reliably
read
out
a
neural
code
with
these
properties,
while
taking
into
account
the
temporal
statistics
of
the
animal’s
location.
This
reasoning
leads
to
an
interesting
prediction
on
the
processing
of
spikes
downstream
of
the
entorhinal
cortex:
the
characteristic
integration
time
of
spikes
in
neurons
that
implement
a
readout
of
position
is
expected
to
increase
sharply
as
a
function
of
the
grid
spacing
of
the
presynaptic
grid
cell.
RESULTS
Encoding
of
a
dynamic
location
First,
we
briefly
review
the
theoretical
considerations
relevant
to
the
representation
of
a
static
variable.
Imagine
an
ideal
observer
that
attempts
to
read
out
position
from
the
spikes
generated
by
all
the
neurons
in
one
module
with
grid
spacing
𝜆,
over
a
time
interval
𝛥𝑇.
If
the
rate
of
spikes
is
sufficiently
large,
the
posterior
distribution
over
position
is
approximately
given
by
a
periodic
array
of
Gaussians
(Fig.
1B).
The
spatial
periodicity
of
this
distribution
is
a
consequence
of
the
single
neuron
tuning
curves,
which
all
share
the
same
periodic
structure.
If
the
individual
receptive
fields
are
isotropic
and
compact,
the
Gaussians
are
isotropic
as
well
(Mathis
et
al.,
2015).
Their
characteristic
width
reflects
the
precision
at
which
position
can
be
read
out
locally
within
a
unit
cell
of
the
periodic
lattice.
For
independent
Poisson
spikes,
(1)
where
𝛥!
is
the
mean
variance
of
the
periodic
Gaussians,
the
factor
2
comes
from
the
two
dimensions,
and
𝐽,
the
Fisher
information
rate
(in
each
direction
in
space,
see
Supplemental
Information)
is
independent
of
Δ𝑇
for
Poisson
neurons,
and
can
be
written
as:
𝛥!= 2𝐽𝛥𝑇
𝐽=𝛼𝑛𝜆! .
(2)
Here
𝑛
is
the
number
of
neurons
in
the
module,
and
the
proportionality
constant
𝛼
depends
on
the
detailed
shape
of
the
firing
fields
(see
Supplemental
Information
for
a
derivation
of
𝛼
for
Gaussian
receptive
fields).
We
assumed
that
neurons
within
the
module
cover
densely
and
uniformly
all
possible
phases
of
the
periodic
tuning
curve,
and
that
the
Cramér-‐Rao
bound
is
saturated.
The
dependence
of
𝐽
on
𝜆
can
be
deduced
based
on
dimensional
analysis,
relying
on
the
observations
that
the
maximal
firing
rate,
𝑟max,
is
approximately
constant
in
different
modules,
and
that
firing
fields
of
grid
cells
scale
in
proportion
to
the
grid
spacing
(Hafting
et
al.,
2005)
(therefore,
𝜆
is
the
only
spatial
length
scale
characterising
the
response
in
each
module).
Note
that
the
precision
of
readout
depends
on
the
choice
of
the
observation
interval,
and
that
the
MSE
is
inversely
proportional
to
the
number
of
neurons
(Fig.
1D).
Based
on
Eqs.
(1)
and
(2),
a
uniform
allocation
of
neurons
to
modules
implies
that
the
ratio
𝛥/λ,
the
precision
of
readout
relative
to
the
grid
spacing,
is
the
same
for
all
modules.
Intuitively,
this
is
a
plausible
requirement,
and
indeed
this
relation
was
postulated
(or
derived)
in
previous
works:
for
example,
consider
a
nested
coding
scheme
(Mathis
et
al.,
2012b;
2012a;
Wei
et
al.,
2015),
in
which
the
grid
spacings
follow
a
geometric
series.
Let
us
denote
by
𝜆!
the
grid
spacings,
ordered
sequentially
(𝜆! >
𝜆! >⋯),
and
by
𝛥!
the
corresponding
precision
of
readout
from
each
module.
Uniformity
of
𝛥!/𝜆!
across
modules
implies
also
that
the
ratio
𝛥!/𝜆!!!
is
uniform
across
modules.
A
sufficiently
small
value
of
this
ratio
ensures
that
readout
from
each
module
is
accurate
enough
to
avoid
ambiguities
arising
from
the
periodicity
of
response
in
the
successive
module
with
smaller
spacing.
Thus,
by
choosing
a
fixed
(and
sufficiently
small)
ratio
𝛥!/𝜆!!!,
it
is
possible
to
ensure
that
ambiguities
do
not
arise
in
the
readout
of
the
code
at
any
scale.
To
see
why
the
dynamic
aspect
of
the
trajectory
is
consequential,
let
us
suppose
that
the
animal’s
trajectory
follows
the
statistics
of
a
simple
random
walk.
We
imagine
that
each
neuron
fires
as
an
inhomogeneous
Poisson
process
with
a
rate
determined
by
the
tuning
curve
of
the
neuron,
evaluated
at
the
instantaneous
position
of
the
animal.
Consider
an
ideal
observer,
attempting
to
estimate
the
animal’s
position
at
time
𝑡,
based
on
the
spikes
from
all
neurons
in
a
single
module,
emitted
up
to
that
time
(Fig.
1C).
In
the
𝛥!=2 2𝐷𝐽 ,
Supplemental
Information
we
show
that
the
local
mean
square
error
(MSE)
of
such
an
optimal
estimator
is
given
by
(3)
instead
of
Eq.
(1)
(see
also
Fig.
1E).
Thus
the
MSE
is
proportional
to
𝑛!!/!,
instead
of
the
𝑛!!
dependence
of
the
static
case
(compare
Figs.
1
D
and
E).
This
difference
in
scaling
with
n
may
seem
minor,
but
using
Eqs.
(2)
and
(3)
we
find
that
in
order
to
achieve
a
fixed
relative
precision
(𝛥!/𝜆!)
for
all
modules,
it
is
now
necessary
to
have
(4)
Thus,
far
fewer
neurons
are
required
in
modules
with
large
spacing,
compared
to
modules
with
small
spacing.
This
result
can
be
easily
explained
in
qualitative
terms:
the
relative
position
of
the
animal,
in
relation
to
the
period
of
the
grid
response,
varies
more
slowly
in
the
modules
with
large
𝜆
compared
to
modules
with
small
𝜆.
Thus,
an
ideal
decoder
can
rely
on
spikes
emitted
within
a
longer
period
of
time
in
order
to
estimate
the
relative
position
within
a
unit
cell
of
the
periodic
grid.
The
validity
of
this
interpretation
is
further
demonstrated
below
(Biological
implications
for
dynamic
readout).
Optimal
module
population
sizes
To
see
how
these
principles
impact
a
more
detailed
theory
for
the
allocation
of
grid
cells
to
modules,
we
consider
a
nested
code
(Mathis
et
al.,
2012a;
Wei
et
al.,
2015;
Weiss
et
al.,
2012),
in
which
position
can
be
read
out
sequentially
starting
from
the
module
with
the
largest
spacing,
progressing
sequentially
to
modules
with
smaller
grid
spacings. We
follow
a
similar
line
of
argumentation
as
in
(Wei
et
al.,
2015).
Our
goal
is
to
minimize
𝛥!,
the
resolution
of
readout
from
the
smallest
module,
while
constraining
the
largest
grid
spacing
𝜆!
and
the
number
of
neurons
𝑁
(equivalently,
it
is
possible
to
minimize
the
number
of
neurons
while
constraining
the
readout
resolution).
Additionally,
we
require
that
ambiguities
about
position
do
not
arise
at
any
one
of
the
refinement
steps.
Therefore,
we
impose
a
relation
between
the
readout
resolution
and
grid
spacing,
(5)
Here,
𝛽
should
be
sufficiently
small
such
that
the
range
of
likely
positions,
inferred
from
module
𝑖,
does
not
contain
multiple
periods
of
the
response
from
module
𝑖+1.
Below,
𝑛∼1𝜆! .
𝛥!=𝛽𝜆!!!.
𝑛!∼2!
,
the
value
of
𝛽
is
set
as
described
in
the
Experimental
Procedures
and Fig. S3.
Crucially,
we
use
Eq.
(3)
for
the
resolution
of
readout
at
each
step,
since
we
hypothesize
that
grid
cells
encode
a
dynamic
position
with
random
walk
statistics.
Additional
details
of
the
optimization
are
described
in
the
Experimental
Procedures
and
the
Supplemental
Information.
The
requirement
of
unambiguous
reconstruction,
combined
with
Eq.
(3),
leads
to
several
salient
results.
First,
we
find
that
in
the
optimized
code,
the
module
population
sizes
precisely
follow
a
geometric
progression
(6)
where
𝑛!
is
the
number
of
neurons
in
module
𝑖,
Fig.
2A.
Second,
we
find
that
the
ratios
between
subsequent
grid
spacings
are
approximately
constant
in
the
modules
with
small
spacing.
The
optimal
ratio
approaches
a
limit
given
by
2≃1.41
for
the
smallest
modules
(Supplemental
Information
and
Fig.
2C).
This
prediction
is
in
close
agreement
with
the
ratio
of
grid
spacings
in
subsequent
modules,
measured
in
(Stensola
et
al.,
2012)
and
averaged
across
animals,
approximately
1.42.
Note
that
the
ratios
were
measured
only
for
the
first
few
modules
with
lowest
grid
spacings,
hence
the
theory
is
in
very
good
agreement
with
the
existing
measurements.
The
properties
listed
above
are
independent
of
the
total
number
of
neurons,
the
shape
of
the
tuning
curve,
and
the
parameter
𝛽.
Moreover,
these
properties
remain
intact
even
if
we
relax
the
assumption
of
an
optimal
estimator,
but
assume
the
relation
𝛥!∼𝐽!!/!
,
as
in
Eq.
(3).
Other,
more
detailed
aspects
of
the
results
do
depend
on
parameters.
In
Fig.
2
we
assumed
that
the
total
number
of
grid
cells
is
either
10!
(blue)
or
10!
(red),
leading
to
differences
in
the
ratios
between
subsequent
modules
–
but
not
in
the
ratios
obtained
for
the
smallest
modules.
Most
importantly
for
our
discussion
on
the
allocation
of
grid
cells
to
modules,
the
module
population
sizes
are
given
precisely
by
Eqs.
(6)
and
(8),
irrespective
of
parameters.
In
particular,
about
half
of
the
neurons
are
allocated
to
the
module
with
the
smallest
grid
spacing
(Fig.
2A).
It
may
seem
surprising
that
accurate
readout
is
possible
at
all
with
only
a
handful
of
neurons
in
the
modules
with
the
largest
spacing.
Figure
3A
shows
the
estimate
of
position
obtained
by
an
optimal
Bayesian
decoder
(see
Supplemental
Information),
in
response
to
simulated
spikes
from
ten
modules
with
the
above
parameters
(altogether
10!
grid
cells,
and
ten
neurons
in
the
module
with
the
largest
spacing).
The
root
mean
square
error
(RMSE)
of
this
estimator
is
1.276±0.004
cm.
It
is
instructive
to
compare
this
result
with
the
performance
under
two
other
allocations
of
grid
cells
to
modules:
if
the
neurons
are
allocated
with
equal
proportion
to
all
modules,
the
RMSE
is
increased
by
a
factor
of
about
1.5
(Fig.
3B).
If
the
allocation
of
neurons
to
modules
is
reversed,
such
that
most
neurons
participate
in
the
modules
with
larger
grid
spacing,
the
RMSE
becomes
larger
by
a
factor
of
about
3.4
(Fig.
3C).
In
summary,
the
hypothesis
that
grid
cells
are
adapted
to
efficiently
encode
a
dynamic
position
predicts
a
sharp
decrease
in
the
number
of
grid
cells
allocated
to
modules
with
high
grid
spacing,
compared
to
modules
with
smaller
spacing,
while
remaining
compatible
with
previous
theories,
which
predicted
a
geometric
progression
in
the
grid
spacings.
Biological
implications
for
dynamic
readout
The
analysis
of
grid
cell
activity
from
the
perspective
of
an
ideal
observer
is
relevant
for
coding
in
the
brain
only
if
neural
circuitry
can
implement
an
efficient
decoding
scheme
of
the
grid
cell
code,
while
taking
into
account
the
statistics
of
the
animal’s
motion.
The
direct
computations
involved
in
a
precisely
optimal
decoder
[Eq.
(10)]
are
elaborate,
but
we
show
next
that
fairly
undemanding
processing
of
the
spikes
can
produce
near
optimal
readout
of
position
from
each
module.
We
analyze
a
simple
readout
scheme
in
which
spikes
emitted
by
grid
cells
are
interpreted
as
if
the
position
of
the
animal
is
static.
For
a
truly
static
position,
all
the
spikes
emitted
in
the
past
are
informative
about
the
current
position.
Here,
however,
we
consider
an
estimate
of
position
which
is
constructed
based
only
on
spikes
from
the
recent
history,
weighted
by
an
exponential
kernel
with
time
constant
𝜏
(Fig.
4A).
An
estimator
that
treats
the
position
of
the
animal
as
if
it
is
static,
has
a
simple
structure:
the
log
likelihood
to
be
at
position
𝑥
can
be
expressed
as
a
linear
function
of
the
spike
counts
(Supplemental
Information).
A
single
non-‐linearity
is
then
sufficient
to
select
the
position
𝑥
at
which
the
log
likelihood
is
maximal.
Since
the
trajectory
of
the
animal
is
in
fact
dynamic,
the
above
estimator
is,
in
general,
suboptimal.
Its
best
performance
is
obtained
by
choosing
𝜏
as
follows
(Supplemental
Information),
𝜆2𝐷𝛼𝑛 .
𝜏= 12𝐷𝐽=
(7)
This
choice
balances
two
contributions
to
the
error
of
the
estimator,
with
opposing
dependencies
on
𝜏:
first,
the
ambiguity
in
the
decoding
of
position
due
to
the
stochasticity
of
spikes,
which
becomes
large
when
𝜏
is
small
(and
few
spikes
contribute
to
the
estimate).
The
second
contribution
to
the
error
is
due
to
the
animal’s
motion.
This
contribution
increases
with
𝜏,
since
the
simple
decoder
ignores
the
animal’s
motion
altogether.
In
the
Supplemental
Information
we
show
that
despite
its
simplicity,
the
above
estimator
achieves
the
same
performance
as
the
optimal
Bayesian
decoder,
Eq.
(3),
when
the
readout
time
𝜏
is
chosen
according
to
Eq.
(7).
According
to
Eqs.
(4)
and
(7),
the
time
scale
𝜏
should
decrease
in
sequential
modules
by
a
factor
of
2
for
the
modules
with
smaller
grid
spacings,
where
Eq.
(4)
is
approximately
valid.
Characteristic
values
of
𝜏
are
shown
in
Fig.
4B,
where
the
parameters
are
the
same
as
in
Fig.
2.
In
this
example,
𝜏
varies
from
~1
ms
to
~600
ms,
depending
on
the
grid
spacing.
Based
on
these
results,
we
consider
a
simple
model
for
readout
of
the
grid
cell
code,
in
which
place
cells
in
the
hippocampus
approximate
the
log
likelihood
of
position
based
on
incoming
spikes
from
the
entorhinal
cortex.
We
model
the
activity
of
each
place
cell
as
linearly
determined
from
incoming
spikes,
with
an
exponential
temporal
kernel
whose
characteristic
time
constant
𝜏
depends
on
the
grid
spacing
of
the
presynaptic
input
cell,
Fig.
4C.
A
single
exponential
nonlinearity
is
then
sufficient
to
obtain
an
approximation
of
the
likelihood.
In
addition,
lateral
inhibitory
connectivity
in
the
place
cell
network,
not
modeled
explicitly
here,
could
implement
winner-‐take-‐all
dynamics
(Dayan
and
Abbott,
2001)
which
would
serve
to
select
a
unique
estimate
for
the
maximum-‐likelihood
position.
With
the
readout
time
constants
set
by
Eq.
(7),
and
with
appropriately
chosen
synaptic
weights,
selecting
the
cell
with
the
maximal
activation
yields
an
estimate
for
position
with
a
MSE
which
is
close
to
that
of
an
optimal
Bayesian
decoder
(compare
Fig.
4D
and
Fig.
3).
An
interesting
prediction
follows
for
the
readout
of
position
in
the
hippocampus
(or
in
other
brain
areas),
based
on
inputs
from
grid
cells:
Spikes
in
grid
cells
are
expected
to
influence
the
activity
of
a
postsynaptic
readout
cell
over
an
integration
time
that
depends
on
the
grid
spacing.
The
integration
time
should
increase
monotonically
with
grid
spacing,
as
predicted
by
Eq.
(7).
Optimization
for
other
trajectory
statistics
Within
the
above
readout
scheme,
it
is
possible
to
adjust
the
module
properties
in
order
to
optimize
the
resolution
of
readout
for
trajectories
that
deviate
from
simple
random
walk
statistics.
Let
us
suppose,
for
example,
that
the
variance
of
motion
increases
quadratically
with
time,
instead
of
the
linear
dependence
that
characterizes
a
simple
random
walk.
This
scenario
corresponds
to
motion
at
a
constant
velocity,
and
at
a
random
direction
which
is
unknown
to
the
decoder
of
the
spikes.
It
straightforward
to
evaluate
the
readout
error
of
the
simple
estimator
in
each
module,
under
this
type
of
motion
(see
SI),
and
to
find
the
value
of
𝜏
that
minimises
its
MSE,
Eq.
(17).
We
thus
repeat
our
optimization
scheme
for
the
number
of
cells
in
each
module
and
the
grid
spacings,
while
using
Eq.
(14)
instead
of
the
expression
for
random
walk
statistics,
Eq.
(3).
The
predicted
ratio
in
the
number
of
cells
between
successive
modules
is
1.5
instead
of
2
(Fig.
5A),
the
asymptotic
ratio
between
the
spacings
in
successive
modules
is
1.5
instead
of
2≃1.4
(Fig
5C),
and
the
range
of
optimal
readout
times
is
somewhat
narrower
than
obtained
under
the
assumption
of
random
walk
statistics
(Fig.
5D).
Nevertheless,
the
qualitative
conclusions
are
very
similar
under
the
two
scenarios:
module
population
sizes
decrease
sharply
with
grid
spacing,
and
the
ratios
of
successive
grid
spacings
are
approximately
constant
(and
similar
in
the
two
scenarios)
for
the
modules
with
small
spacing.
Thus,
the
qualitative
conclusions
are
valid
for
a
wide
range
of
statistics
that
may
characterize
the
motion.
DISCUSSION
In
summary,
we
propose
that
the
representation
of
position
in
the
entorhinal
cortex
takes
advantage
of
the
continuous
temporal
statistics
of
motion,
in
order
to
efficiently
encode
the
animal’s
position.
This
is
possible
due
to
the
multiscale
structure
of
the
code:
in
modules
with
larger
grid
spacing,
the
encoded
variable
varies
less
rapidly
than
in
modules
with
smaller
spacing.
Spikes
from
these
modules
remain
informative
about
the
current
position
over
a
longer
time
scale,
allowing
for
an
efficient
encoder
to
allocate
a
smaller
number
of
cells
to
these
modules.
The
relevance
of
this
argumentation
[and
related
reasoning
(Mathis
et
al.,
2012a;
Wei
et
al.,
2015)]
to
the
encoding
of
position
in
the
entorhinal
cortex,
is
suggestive
of
an
intriguing
possibility,
that
the
principle
of
efficient
coding
(Barlow,
1961)
extends
in
the
brain
well
beyond
the
realm
of
early
sensory
processing.
From
the
theoretical
perspective,
it
is
interesting
to
consider
the
dynamical
range
of
the
representation:
the
ratio
between
the
represented
range
and
a
measure
of
the
resolution,
such
as
the
MSE
(Burak,
2014).
Optimizing
this
quantity
is
a
difficult
theoretical
problem,
even
when
assuming
that
the
encoded
variable
is
static
(Mathis
et
al.,
2012a;
Wei
et
al.,
2015).
Our
goal
here
was
not
to
fully
solve
this
problem,
but
to
explore
the
salient
consequences,
arising
from
a
hypothesis
that
the
code
is
adapted
to
the
dynamics
of
the
animal’s
trajectory.
We
focused
our
analysis
on
nested
codes,
and
assumed
that
the
range
of
representation
of
the
grid
cell
code
matches
the
largest
grid
spacing.
However,
we
expect
that
the
principles
revealed
here
for
the
neural
representation
of
a
dynamic
trajectory
apply
also
if
the
range
of
positions
encoded
by
grid
cells
is
much
larger
(Fiete
et
al.,
2008).
The
most
important
prediction
arising
from
our
hypothesis,
is
the
highly
non-‐uniform
distribution
of
grid
cells
across
modules.
Previous
experiments
(Stensola
et
al.,
2012)
strongly
hint
that
this
is
indeed
a
prominent
feature
in
the
organization
of
the
entorhinal
cortex,
but
additional
experiments
are
necessary
in
order
to
establish
this
conclusion
more
firmly,
and
to
obtain
quantitative
estimates
of
the
distribution.
Another
intriguing
prediction
arises
from
the
identification
of
a
relatively
simple
decoding
scheme
that
takes
into
account
the
dynamic
aspect
of
motion:
action
potentials
of
grid
cells
are
expected
to
affect
the
activity
of
postsynaptic
readout
cells
over
varying
time
scales,
which
increase
with
the
grid
spacing.
Predicted
time
scales
span
about
three
orders
of
magnitude,
from
~1ms
to
~1s,
assuming
that
the
largest
grid
spacing
spans
a
few
meters.
Integration
time
scales
up
to
~100ms
can
be
implemented
in
neural
circuitry
by
the
dynamics
of
synaptic
integration.
Longer
time
scales
of
integration
in
the
order
of
1s
require
other
mechanisms
for
persistence
(Barak
and
Tsodyks,
2014):
these
can
potentially
rely
on
recurrent
connectivity,
on
short
term
synaptic
plasticity
(Mongillo
et
al.,
2008)
or
perhaps
on
intrinsic
cellular
persistence.
It
is
noteworthy
that
intrinsic
persistence,
with
characteristic
time
scales
of
seconds
has
been
widely
observed
in
the
hippocampal
formation,
and
specifically
in
the
entorhinal
cortex
(Egorov
et
al.,
2002)
and
the
hippocampus
(Knauer
et
al.,
2013).
Our
model
for
readout
of
grid
cell
activity
was
deliberately
simplified,
in
order
to
emphasize
the
main
principles
governing
the
readout
of
a
dynamic
variable.
Thus,
we
described
the
readout
as
occurring
in
a
single
feedforward
layer.
We
speculate
that
the
functional
organization
along
the
dorso-‐ventral
axis
of
the
hippocampus
may
be
helpful
in
implementing
different
time
scales
for
integration
at
different
spatial
scales.
Furthermore,
several
lines
of
experimental
evidence
suggest
that
place
cells
are
driven
by
environmental
sensory
inputs
that
are
independent
of
grid
cell
activity
(Bush
et
al.,
2014;
Hales
et
al.,
2014).
Nevertheless,
there
is
also
compelling
evidence
that
grid
cells
contribute
to
the
activity
of
place
cells,
perhaps
most
prominently
when
direct
sensory
cues
are
absent,
and
that
the
MEC
and
hippocampus
form
together
a
processing
loop
responsible
jointly
for
spatial
representation,
computation,
and
memory
(Bush
et
al.,
2014;
Hales
et
al.,
2014)
[see,
also,
(Sreenivasan
and
Fiete,
2011)
for
a
discussion
of
possible
implications
from
a
theoretical
perspective].
In
short
term
memory
networks,
the
fidelity
of
the
neural
code
is
consequential
for
self-‐
maintenance
of
the
persistent
state,
in
addition
to
its
significance
for
downstream
readout
(Burak
and
Fiete,
2012).
If
the
entorhinal
cortex
autonomously
maintains
short-‐
term
memory
of
position,
as
required
for
idiothetic
path
integration
(Burak
and
Fiete,
2009;
Hafting
et
al.,
2005;
McNaughton
et
al.,
2006),
we
hypothesize
that
recurrent
connectivity
within
the
entorhinal
cortex
may
follow
principles
similar
to
those
proposed
here
for
readout
in
the
hippocampus.
EXPERIMENAL
PROCUDURES
Details
of
the
optimization
are
provided
in
the
Supplemental
Information.
The
optimal
number
of
neurons
in
a
module
𝑛!
(Fig.2A),
and
the
ratio
𝜆! !!!!
(Fig.2C)
are
given
by:
(8)
𝑛!= 12 !!!!!
1− 12 ! 𝑁 ,
where
𝑁
is
the
total
number
of
grid
cells,
and
Optimized
code
𝜆! !!!!=𝜆1 12 𝑖
8𝐷𝛼𝑁𝛽4 1− 12 𝑚 − 12 𝑖+12−2−𝑖+1 𝑚+2+12 .
(9)
For
simplicity,
we
assume
Gaussian
receptive
fields,
for
which
the
factor
𝛼= 43𝜋𝑟max
in
Eq.
(2)
(see
SI).
The
parameter
𝛽 [Eq.
(5)]
was
chosen
as
follows.
This
parameter
should
be
set
small
enough
to
ensure
that
there
are
no
global
ambiguities,
since
the
minimization
of
𝛥!
affects
only
the
local
inference
error.
We
applied
the
minimization
procedure
with
various
values
of
𝛽
to
select
𝑛!
and
𝜆!,
and
then
evaluated
the
MSE
of
the
exponential
kernel
estimator.
As
expected,
very
small
values
of
𝛽
led
to
large
MSE,
since
𝛽
sets
a
limit
on
the
degree
of
reduction
in
𝛥
from
module
to
module.
Large
values
of
𝛽
also
led
to
large
MSE
due
to
errors
arising
from
global
ambiguities
(Fig.
S3).
We
chose
in
all
simulations
𝛽=0.1,
as
this
value
provides
an
MSE
close
to
minimal
(we
did
not
attempt
to
find
the
precise
optimal
value
of
𝛽, but
note
that
the
choice
of
this
parameter
does
not
affect
any
of
the
qualitative
results).
The
posterior
probability
distribution
used
by
the
optimal
Bayesian
decoder
(Fig.
3),
is
obtained
using
the
dynamic
update
rule:
(10)
𝑝(𝑥,𝑡+𝑑𝑡)=!!
where
𝑝!(𝑥𝑥′)
is
the
probability
for
the
random
walk
to
reach
𝑥
at
time
𝑡+𝑑𝑡
from
position
𝑥′
at
time
𝑡,
and
𝑝!"#$%!(𝑥,𝑡)
represents
the
likelihood
of
the
spikes
observed
within
the
short
time
interval,
given
the
position
𝑥
(see
SI
for
more
details).
The
optimal
Bayesian
decoder
estimates
the
location
of
the
animal
by:
(11)
𝑥!"=argmax!𝑝(𝑥,𝑡)
.
The
temporal
exponential
kernel
decoder
estimates
the
location
of
the
animal
as
follows:
Optimal
Bayesian
decoder
𝑑𝑥′𝑝!(𝑥𝑥′)𝑝(𝑥′,𝑡)𝑝!"#$%!(𝑥,𝑡)
,
Exponential
kernel
decoder
.
.
!
𝑥=argmax!
!
!
!
!∈!
𝑓!(𝑥−𝑥!) ,
𝑝𝑥,𝑡 =1𝑍exp
𝑑𝑡′ℎ!(𝑡!)𝜉!(𝑡−𝑡′)log
Optimized
code
for
variance
of
motion
which
increases
quadratically
with
time
(12)
where
the
second
sum
is
over
neurons
𝜇
that
belong
to
module
𝑖,
𝑓!(𝑥)
is
the
shape
of
the
tuning
curve,
characterising
the
receptive
field
of
the
neurons
in
the
𝑖th
module,
𝑥𝜇
is
the
center
of
the
receptive
field
of
neuron
𝜇,
𝜉!(𝑡)
is
a
series
of
delta
functions
that
represents
the
spikes
of
neuron
𝜇,
and
ℎ!(𝑡)
is
the
temporal
kernel
of
module
𝑖.
In
our
case:
ℎ!𝑡 =exp −𝑡𝜏!
The
posterior
probability
distribution
illustrated
in
Fig.
4D
is
given
by
(13)
𝑑𝑡′ℎ!(𝑡!)𝜉!(𝑡−𝑡′)log𝑓!(𝑥−𝑥!)
!
!
!∈!
In
this
case
Eq.
(3)
is
replaced
by
(see
SI)
(14)
𝛥!=3⋅ 𝑣2𝐽 !! .
Consequently,
the
optimal
number
of
neurons
in
module
𝑛!
(Fig
.5A)
are
given
by
(see
SI)
(15)
𝑛!=12 23 !!!!!
1− 23 ! 𝑁 ,
the
ratio
λ!/𝜆!!!
(Fig
.5C)
is
given
by
(16)
1−12𝑚+3 23 𝑖
,
12 23 𝑖 32
𝜆! !!!!=
!1𝛽3𝛼𝑁
332𝑣1− 23 𝑚
and
the
optimal
time
constant
for
readout
𝜏
is
given
by
(17)
12𝐽𝑣! !!.
𝜏=
SUPPLEMENTAL
INFORMATION
AUTHOR
CONTRIBUTIONS
ACKNOWLEDGMENTS
The
supplemental
Information
is
available
from
the
authors
on
request.
Y.B.
designed
the
research,
N.W.M
and
Y.B.
developed
the
theory,
designed
the
numerical
simulations,
and
wrote
the
paper.
H.A.
designed
the
numerical
simulations,
and
edited
the
manuscript.
N.W.M,
H.A.,
and
A.M.
performed
the
numerical
simulations.
We
thank
Tor
Stensola
and
Edvard
Moser
for
permission
to
reproduce
a
figure
from
their
paper,
and
Michael
Hasselmo
and
Motoharu
Yoshida
for
helpful
correspondence.
This
research
was
supported
by
the
Israel
Science
Foundation
grant
No.
1733/13
and
(in
part)
by
grant
No.
1978/13. We
acknowledge
support
from
the
Gatsby
Charitable
Foundation.
An
earlier
version
of
this
work
appeared
in
abstract
form
(Weiss
et
al.,
2015).
Atick,
J.J.,
Redlich,
A.N.,
1992.
What
Does
the
Retina
Know
About
Natural
Scenes.
Neural
computation
4,
196–210.
Barak,
O.,
Tsodyks,
M.,
2014.
Working
models
of
working
memory.
Current
Opinion
in
Neurobiology
25,
20–24.
Barlow,
H.B.,
1961.
Possible
principles
underlying
the
transformation
of
sensory
messages,
in:
Rosenbluth,
W.A.
(Ed.),
Sensory
Communications.
Bell,
A.J.,
Sejnowski,
T.J.,
1997.
The
“independent
components”
of
natural
scenes
are
edge
filters.
Vision
Res.
37,
3327–3338.
Burak,
Y.,
2014.
Spatial
coding
and
attractor
dynamics
of
grid
cells
in
the
entorhinal
cortex.
Current
Opinion
in
Neurobiology
25C,
169–175.
Burak,
Y.,
Fiete,
I.R.,
2012.
Fundamental
limits
on
persistent
activity
in
networks
of
noisy
neurons.
Proc
Natl
Acad
Sci
USA
109,
17645–17650.
REFERENCES
Burak,
Y.,
Fiete,
I.R.,
2009.
Accurate
path
integration
in
continuous
attractor
network
models
of
grid
cells.
PLoS
Comput
Biol
5,
e1000291.
Bush,
D.,
Barry,
C.,
Burgess,
N.,
2014.
What
do
grid
cells
contribute
to
place
cell
firing?
Trends
in
Neurosciences
37,
136–145.
Dayan,
P.,
Abbott,
L.F.,
2001.
Theoretical
Neuroscience.
MIT
Press.
Egorov,
A.V.,
Hamam,
B.N.,
Fransén,
E.,
Hasselmo,
M.E.,
Alonso,
A.A.,
2002.
Graded
persistent
activity
in
entorhinal
cortex
neurons.
Nature
420,
173–178.
Fairhall,
A.L.,
Lewen,
G.D.,
Bialek,
W.,
van
Steveninck,
R.,
2001.
Efficiency
and
ambiguity
in
an
adaptive
neural
code.
Nature
412,
787–792.
Fiete,
I.R.,
Burak,
Y.,
Brookings,
T.,
2008.
What
grid
cells
convey
about
rat
location.
Journal
of
Neuroscience
28,
6858–6871.
Hafting,
T.,
Fyhn,
M.,
Molden,
S.,
Moser,
M.-‐B.,
Moser,
E.I.,
2005.
Microstructure
of
a
spatial
map
in
the
entorhinal
cortex.
Nature
436,
801–806.
Hales,
J.B.,
Schlesiger,
M.I.,
Leutgeb,
J.K.,
Squire,
L.R.,
Leutgeb,
S.,
Clark,
R.E.,
2014.
Medial
entorhinal
cortex
lesions
only
partially
disrupt
hippocampal
place
cells
and
hippocampus-‐dependent
place
memory.
Cell
Rep
9,
893–901.
Hartley,
T.,
Lever,
C.,
Burgess,
N.,
O'Keefe,
J.,
2014.
Space
in
the
brain:
how
the
hippocampal
formation
supports
spatial
cognition.
Philos
Trans
R
Soc
Lond,
B,
Biol
Sci
369,
20120510.
Knauer,
B.,
Jochems,
A.,
Valero-‐Aracama,
M.J.,
Yoshida,
M.,
2013.
Long-‐lasting
intrinsic
persistent
firing
in
rat
CA1
pyramidal
cells:
A
possible
mechanism
for
active
maintenance
of
memory.
Hippocampus
23,
820–831.
Laughlin,
S.,
1981.
A
simple
coding
procedure
enhances
a
neuron's
information
capacity.
Z.
Naturforsch.,
C,
Biosci.
36,
910–912.
Mathis,
A.,
Herz,
A.V.M.,
Stemmler,
M.B.,
2012a.
Optimal
population
codes
for
space:
grid
cells
outperform
place
cells.
Neural
computation
24,
2280–2317.
Mathis,
A.,
Herz,
A.V.M.,
Stemmler,
M.B.,
2012b.
Resolution
of
Nested
Neuronal
Representations
Can
Be
Exponential
in
the
Number
of
Neurons.
Phys
Rev
Lett
109,
018103.
Mathis,
A.,
Stemmler,
M.B.,
Herz,
A.V.M.,
2015.
Probable
nature
of
higher-‐dimensional
symmetries
underlying
mammalian
grid-‐cell
activity
patterns.
Elife
4.
McNaughton,
B.L.,
Battaglia,
F.P.,
Jensen,
O.,
Moser,
E.I.,
Moser,
M.-‐B.,
2006.
Path
integration
and
the
neural
basis
of
the
“cognitive
map.”
Nat
Rev
Neurosci
7,
663–
678.
Mongillo,
G.,
Barak,
O.,
Tsodyks,
M.,
2008.
Synaptic
theory
of
working
memory.
Science
319,
1543–1546.
Olshausen,
B.A.,
Field,
D.J.,
1997.
Sparse
coding
with
an
overcomplete
basis
set:
a
strategy
employed
by
V1?
Vision
Res.
37,
3311–3325.
Smith,
E.C.,
Lewicki,
M.S.,
2006.
Efficient
auditory
coding.
Nature
439,
978–982.
Sreenivasan,
S.,
Fiete,
I.R.,
2011.
Grid
cells
generate
an
analog
error-‐correcting
code
for
singularly
precise
neural
computation.
Nat
Neurosci
14,
1330–1337.
Stensola,
H.,
Stensola,
T.,
Solstad,
T.,
Frøland,
K.,
Moser,
M.-‐B.,
Moser,
E.I.,
2012.
The
entorhinal
grid
map
is
discretized.
Nature
492,
72–78.
Taube,
J.S.,
2007.
The
head
direction
signal:
Origins
and
sensory-‐motor
integration.
Annu
Rev
Neurosci
30,
181–207.
Wei,
X.-‐X.,
Prentice,
J.,
Balasubramanian,
V.,
2015.
A
principle
of
economy
predicts
the
functional
architecture
of
grid
cells.
Elife
4,
e08362.
Weiss,
N.,
Agmon,
H.,
Moriel,
A.,
Burak.
Y,
2015.
An
efficient
coding
theory
for
a
dynamic
trajectory
predicts
non-‐uniform
allocation
of
grid
cells
to
modules
in
the
entorhinal
cortex.
Program
No.
727.02.
2015
Neuroscience
Meeting
Planner.
Chicago,
IL:
Society
for
Neuroscience,
2015.
Online.
FIGURE
LEGENDS
Figure
1.
Modular
organization
and
dynamic
decoding.
(A)
Experimental
evidence
for
the
modular
organisation
of
grid
cells
[reproduced
with
permission
of
the
authors
from
(Stensola
et
al.,
2012)]:
grid
spacing
in
a
single
rat,
where
each
dot
corresponds
to
a
single
cell.
Right,
kernel
smoothed
density
(KSD)
estimate
of
the
distribution.
Red
text,
spacing
in
cm
for
the
estimated
peaks.
Note
the
dramatic
decline
in
the
number
of
cells
with
larger
grid
spacings.
See
also
additional
Figs.
in
(Stensola
et
al.,
2012).
(B)
Schematic
illustration
of
the
posterior
distribution
over
position,
inferred
from
spikes
generated
by
all
cells
in
a
single
module.
The
posterior
has
the
same
periodicity
𝜆
as
the
single
neuron
tuning
curves,
the
local
variance
of
the
peaks
is
denoted
by
𝛥!.
(C)
Schematic
illustration
of
a
decoder
for
a
dynamic
variable,
which
follows
the
statistics
of
a
simple
random
walk
(shown
for
simplicity
in
one
dimension).
Black:
a
random
walk
trajectory
𝑥(𝑡).
Red
lines
represent
spikes
emitted
by
a
population
of
neurons
with
different
tuning
curves,
where
the
red
y-‐axis
represents
the
neuron
index.
The
decoder
estimates
the
animal’s
position
at
time
𝑡!,
based
on
all
the
spikes
that
occurred
up
to
that
time.
(D-‐E)
MSE’s
of
an
optimal
decoder,
estimating
position
based
on
spikes
from
a
single
module,
as
a
function
of
the
number
of
neurons,
for
a
static
two-‐dimensional
variable
(D),
and
a
dynamic
random
variable,
following
the
statistics
of
a
simple
random
walk
in
two
dimensions
(E).
Blue
dots:
measurements
of
the
MSE
from
simulations
of
an
optimal
decoder,
responding
to
spikes
generated
by
neural
populations
of
varying
size.
Each
dot
represents
an
average
over
300
realizations,
where
in
(D)
the
averaging
is
over
a
single
time
interval
from
each
simulation
lasting
100
ms,
and
in
(E)
we
average
the
MSE
in
each
simulation
also
over
time
(realizations
lasting
at
least
~200
ms)
.
Error
bars:
1.96
standard
deviations
of
the
MSEs
obtained
from
each
simulation,
divided
by
square
root
of
the
number
of
simulations.
The
receptive
fields
of
the
cells
are
Gaussians
with
𝑟!"#=10Hz,
𝜆=2.82 m,
𝛥𝑇=100ms
for
the
static
case
(D),
and
𝐷=0.0125!!!,
for
the
dynamic
case
(E).
Red
lines:
theoretical
predictions
from
Eq.
(1)-‐(2)
(D)
and
Eq.
(2)-‐(3)
(E).
Figure
2.
Optimized
code:
analytical
results.
(A)
Number
of
neurons
in
a
module
as
a
function
of
the
module
index
(10
modules,
ordered
by
grid
spacing
starting
from
the
largest
spacing).
The
total
number
of
neurons
is
𝑁=10!
(blue)
and
𝑁=10!
(red).
(B)
Grid
spacings
in
the
optimized
code.
(C)
Ratios
between
grid
spacings
in
successive
modules.
The
ratio
approaches
2
in
the
smallest
modules.
In
all
three
panels,
𝜆!=5m,
𝐷=0.05!!!,
𝛽=0.1,
and
the
receptive
fields
of
the
cells
are
Gaussians
with
𝑟!"#=
10Hz
.
Figure
3.
Optimal
Bayesian
decoder.
The
posterior
probability
distribution
obtained
using
Eq.
(10)-‐
(Experimental
Procedures)
in
simulations,
shifted
by
the
true
position
of
the
animal,
and
averaged
over
1350
time
points.
Three
different
allocations
of
neurons
to
modules
are
shown:
(A)
optimal
allocation
as
in
Fig.
2A,
(B)
equal
number
of
neurons
in
each
module,
and
(C)
reversed
allocation.
The
MSE
and
margins
of
error
noted
on
the
left
bottom
of
each
panel
were
computed
as
in
Fig.
1D-‐E,
based
on
100
simulations
each
lasting
∼1.4s.
All
the
parameters
are
the
same
as
in
Fig.
2,
with
𝑁=10!.
Figure
4.
Simplified
estimator.
(A)
An
illustration
of
the
temporal
exponential
kernel,
and
the
characteristic
time
𝜏.
A
spike
at
time
𝑡′
contributes
to
the
estimate
of
position
at
time
𝑡
with
weight
𝑒!!!!!!
.
(B)
Integration
time
constant
𝜏
as
a
function
of
the
module
index,
as
obtained
from
Eq.
(7),
substituting
the
values
of
𝜆!
and
𝑛!
from
Fig.
2
(𝑁=
10!).
(C)
Schematic
illustration
of
a
model
for
readout
(e.g.
by
place
cell
in
the
hippocampus).
Each
place
cell
approximate
the
log
likelihood
to
be
at
a
particular
position
given
the
spikes
of
multiple
grid
cells,
as
a
linear
summation
of
the
spikes
with
integration
times
that
vary
depending
on
the
grid
spacing.
(D)
Performance
of
the
simplified
estimator,
measured
in
the
same
way
as
in
Fig.
3.
All
the
parameters
are
the
same
as
in
Figs.
2
and
3,
with
𝑁=10!.
Figure
5.
Optimization for other statistics of motion. Optimized
parameters
for
encoding
and
decoding
by
grid
cells,
as
in
Fig.
2
&
Fig.
4B,
but
assuming
that
the
variance
of
motion
increases
quadratically
with
time
(see
SI
for
details),
and
that
readout
is
performed
by
an
the
simplified
estimator.
(A)
Number
of
neurons
in
a
module
as
a
function
of
the
module
index.
(B)
The
grid
spacing.
(C)
Ratios
between
grid
spacings
in
subsequent
modules.
This
ratio
approaches
1.5
in
the
smallest
modules.
(D)
Integration
time
𝜏
as
a
function
of
the
module
index,
Eq.
(17).
All
the
parameters
are
the
same
as
in
Fig.2,
with
𝑁=10!
and
velocity
𝑣=1m/s.
Figure
1
(Weiss-‐Mosheiff
et
al.)
A
B
C
N
x
e
d
n
i
n
o
r
u
e
n
1
)
t
(
X
X(t0spikes) ?
t0
t
Δ
λ
D
)
2
m
(
E
S
M
4 10-2
2 10-2
10-2
5 10-3
3 10-3
Static Location
Simulation
Theory
E
Dynamic Location
Simulation
Theory
4 10-2
2 10-2
10-2
5 10-3
3 10-3
102
n
103
102
n
103
Figure
2
(Weiss-‐Mosheiff
et
al.)
A
105
104
i
n
103
102
101
0
N=104
N=105
~2i
5
module
10
B
)
r
e
t
e
m
(
i
λ
C
i
1
+
λ
/
λ
i
5
4
3
2
1
0
0
5
module
10
2.6
2.4
2.2
2
1.8
1.6
1.4
0
N=104
N=105
√2
5
module
10
Figure
3
(Weiss-‐Mosheiff
et
al.)
Optimal Allocation
B
Equal Allocation
C
Reversed Allocation
A
0.1
0
s
r
e
t
e
m
√MSE =1.276±0.004 cm
√MSE =1.851±0.007 cm
√MSE =4.28±0.03 cm
-0.1
-0.1
0
meters
0.1 -0.1
0.1 -0.1
0
meters
0
meters
0.08
0.06
0.04
0.02
0
0.1
Figure
4
(Weiss-‐Mosheiff
et
al.)
Figure
5
(Weiss-‐Mosheiff
et
al.)
B
)
r
e
t
e
m
(
i
λ
5
4
3
2
1
0
0
5
module
10
~(1.5)i
~(1.5)i
~(1.5)i
5
module
10
D
102
)
s
m
(
τ
101
5
module
10
100
0
5
module
10
A
105
104
i
n
103
C
i
1
+
λ
/
λ
i
102
101
0
2.4
2.2
2
1.8
1.6
1.4
0
1.5
|
1705.08756 | 1 | 1705 | 2017-05-23T09:08:14 | Short and random: Modelling the effects of (proto-)neural elongations | [
"q-bio.NC",
"q-bio.PE",
"q-bio.TO"
] | To understand how neurons and nervous systems first evolved, we need an account of the origins of neural elongations: Why did neural elongations (axons and dendrites) first originate, such that they could become the central component of both neurons and nervous systems? Two contrasting conceptual accounts provide different answers to this question. Braitenberg's vehicles provide the iconic illustration of the dominant input-output (IO) view. Here the basic role of neural elongations is to connect sensors to effectors, both situated at different positions within the body. For this function, neural elongations are thought of as comparatively long and specific connections, which require an articulated body involving substantial developmental processes to build. Internal coordination (IC) models stress a different function for early nervous systems. Here the coordination of activity across extended parts of a multicellular body is held central, in particular for the contractions of (muscle) tissue. An IC perspective allows the hypothesis that the earliest proto-neural elongations could have been functional even when they were initially simple short and random connections, as long as they enhanced the patterning of contractile activity across a multicellular surface. The present computational study provides a proof of concept that such short and random neural elongations can play this role. While an excitable epithelium can generate basic forms of patterning for small body-configurations, adding elongations allows such patterning to scale up to larger bodies. This result supports a new, more gradual evolutionary route towards the origins of the very first full neurons and nervous systems. | q-bio.NC | q-bio |
Short and random: Modelling the effects of (proto-)neural
elongations
Oltman O. de Wiljesa,b, R.A.J. van Elburgc, and Fred A. Keijzera,b
aTheoretical Philosophy, University of Groningen
bResearch School of Behavioural and Cognitive Neurosciences, University of Groningen
cInstitute of Artificial Intelligence, University of Groningen
Abstract
To understand how neurons and nervous systems first evolved, we need an account of the origins
of neural elongations: Why did neural elongations (axons and dendrites) first originate, such that they
could become the central component of both neurons and nervous systems? Two contrasting conceptual
accounts provide different answers to this question. Braitenberg's vehicles provide the iconic illustration
of the dominant input-output (IO) view. Here the basic role of neural elongations is to connect sensors
to effectors, both situated at different positions within the body. For this function, neural elongations
are thought of as comparatively long and specific connections, which require an articulated body
involving substantial developmental processes to build.
Internal coordination (IC) models stress a
different function for early nervous systems. Here the coordination of activity across extended parts
of a multicellular body is held central, in particular for the contractions of (muscle) tissue. An IC
perspective allows the hypothesis that the earliest proto-neural elongations could have been functional
even when they were initially simple short and random connections, as long as they enhanced the
patterning of contractile activity across a multicellular surface. The present computational study
provides a proof of concept that such short and random neural elongations can play this role. While
an excitable epithelium can generate basic forms of patterning for small body-configurations, adding
elongations allows such patterning to scale up to larger bodies. This result supports a new, more
gradual evolutionary route towards the origins of the very first full neurons and nervous systems.
Keywords: Early nervous systems neural elongations nervous system evolution computational
modelling internal coordination
Introduction
To understand how the very first neurons and ner-
vous systems evolved, we need an account how
each of neurons' most central characteristics came
about:
(a) their electrical signalling, (b) their
synaptic connections and (c) their elongations (ax-
ons and dendrites). All three are central to ner-
vous system functioning and each has evolved into
a wide variety of forms and modes of operation
within the huge group of animals now known as
the neuralia (Nielsen, 2008). From these three
characteristics, graded and action potentials go
back to unicellular organisms (Naitoh and Eckert,
1969; Greenspan, 2007; Liebeskind et al., 2011),
and the same applies to macromolecular compo-
nents of both the pre- and postsynaptic organiza-
tion (Burkhardt, 2015; Ryan and Grant, 2009).
More difficulties remain with explaining how sep-
arate pre- and postsynaptic components came to-
gether to form synapses between separate cells, as
well as how neural elongations first evolved. Both
are tied to the multicellular morphology and func-
tionality of proto- and early neuralia. To under-
stand such early multicellular organizations, ge-
nomic and molecular evidence gives insufficient
1
guidance (Smith et al., 2014; Nielsen, 2013; Hejnol,
2015), while other evidence remains inconclusive.
Fossils of neuralia go back to the beginning of the
Cambrian, 542 Ma (Million years ago) (Valentine,
2004) therefore synapses and neural elongations
must have originated earlier. Body and trace fossils
from the preceding Ediacaran (635 to 542 Ma) tend
to be very different from modern animals and are
difficult to interpret (Budd, 2015; Fedonkin et al.,
2007; Brasier, 2009). Molecular clock studies fur-
ther suggest that the first neuralia evolved long be-
fore this period (Cunningham et al., 2017; Erwin
et al., 2011). Leaving no fossils, such proto- and
early neuralia could have lived as small meiofauna
with sizes up to 1 mm (Wray, 2015; Erwin, 2015).
It is also suggested that neurons and nervous sys-
tems evolved several times independently (Moroz,
2009; Moroz et al., 2014). At present, very little
can be said with certainty about the first neuralia,
either their form, size or how and when they lived.
A systematic investigation and explication of po-
tential evolutionary transitions from basic proto-
neural configurations to neural ones will be bene-
ficial here: this involves articulating possible tra-
jectories that specify sequences of organismal orga-
nizations that span the transition from non-neural
multicellular organizations to neural ones. Each
step should consist of a functioning organism, while
the consecutive steps from one organization to the
next should be gradual, each one providing some
improvement on the existing functionality (Calcott,
2009).
Investigating specific hypothetical transi-
tion trajectories will aid interpreting the limited
empirical evidence and formulating more specific
questions concerning this evidence.
A well-known example of such an idealized tra-
jectory for nervous systems is provided by Brait-
enberg (Braitenberg, 1984). He formulated a se-
quence of configurations starting with a single neu-
ral connection between a sensor and an effector
to which more connections could be added, even-
tually leading to increasingly complex neural cir-
cuits. Figure 1, for example, sketches a configura-
tion with two connections. Braitenberg's proposed
trajectory is based on an Input-Output (IO) view
on (early) neural evolution (J´ekely et al., 2015).
IO views stress the functioning of neurons -- and
whole nervous systems -- as connections between
sensors and effectors.
Initially these connections
may have been simple and direct, but over evolu-
tionary time they have become increasingly com-
plex neural circuits governing behaviour. Neu-
ral elongations function here as specific and of-
ten long-distance connections between specific loci
within an organism (sensors, effectors, or other neu-
rons). While IO views seem well-suited for mod-
ern nervous systems, they start with organization-
ally and developmentally complex bodily organiza-
tions, which raises doubts about their suitability as
a primitive condition (but see J´ekely (2011) for a
possible approach). In addition, an IO view does
not readily fit the most primitive extant nervous
systems consisting of surface-distributed nerve nets
such as those found in cnidarians and acoelomor-
pha (Hejnol, 2015).
In contrast, we focus here on an alternative In-
ternal Coordination (IC) view (J´ekely et al., 2015),
which stresses the need to acquire multicellular
(bodily) coordination as an initial key task for
early -- and modern -- nervous systems. Coordinat-
ing contractile (muscle) tissue for motility and re-
versible changes in body-shape is a central exam-
ple here (Pantin, 1956; Keijzer et al., 2013; Keijzer
and Arnellos, 2017) that imposes different func-
tional demands on early nervous systems and neu-
ral elongations. Rather than focusing on neural
elongations as a way to provide specifically targeted
connections, an IC view opens up the possibility of
acquiring neural elongations in a more gradual way.
We performed a modelling study to test the IC
idea that simple -- short and randomly connected --
neural elongations can have played a significant be-
havioural role for proto-neuralia that we take here
to be of limited (meiofaunal) size consisting of a few
hundreds to thousands of cells. The starting point
for this study consists of an excitable and contrac-
tile epithelium (a myoepithelium) that can provide
primitive contractile motility (Mackie, 1970). My-
oepithelia with electrical connections between the
constituting cells exist in extant animals (Mackie
and Passano, 1968). We hypothesize that proto-
neuralia could have possessed myoepithelia with
similar electrical connections or with chemical sig-
naling to conduct electrical activity from one cell
to the next, either by juxtacrine signaling or basic
chemical synapses (Keijzer et al., 2013; De Wiljes
et al., 2015). Importantly, this configuration could
have provided a scaffold to bring separate pre- and
post-synaptic elements in adjacent cells together as
a full synapse. However, here we only focus on the
2
citable epithelium provided a way to initiate
ring-shaped spatiotemporal patterns of excitation
running along the long axis of the body-tube
(De Wiljes et al., 2015). However, the presence of
such patterns at the whole body scale depended on
the form of the body-tube -- a large length to width
ratio was required -- as well as the size of the body:
ring-shaped activity only occurred for smaller tubes
and was lost in larger ones. These patterns are
an emergent feature resulting from the topology of
the system, its limited size in terms of cells, ran-
dom activations, and of the refractionary nature of
the cells. For larger body-tubes, signalling to adja-
cent cells provides only travelling wave-fronts that
remain thin and insignificant on the scale of the en-
tire animal, while they do not travel fast enough to
entrain excitation at the level of the whole body.
With this setting, we investigated whether neural
elongations could provide a mechanism to overcome
the limitations of nearest neighbour signalling. We
hypothesized that the patterns of activation across
a body-tube would change when neural elongations
were added. To keep the change as generic and ba-
sic as possible, we focused on the effects of elonga-
tions that are relatively short -- that is, connecting
cells that have only a few intermediate cells between
them -- while the connections are made randomly.
Such short and random connectivity is undemand-
ing in terms of body morphology, cell differentia-
tion, and developmental patterning as it does not
require preset destinations for these connections.
Short and random elongations can therefore pro-
vide a small and plausible step in an evolutionary
trajectory towards very primitive nervous systems.
To test the validity of this idea we investigated a
number of variations of randomly connected config-
urations: we varied the fraction of cells with elonga-
tions, the length of the elongations, and the size of
the body-tube. The presence of ring-shaped pat-
terns of excitation travelling along the length of
the tube was used as a biologically plausible form
of coordinated activity that also could be measured
both qualitatively and quantitatively under various
modelling conditions.
Results
Central examples of our main results are presented
in figure 2. The panes on the left feature small
Figure 1: A comparison between an input-output
(IO) view as exemplified by a Braitenberg vehicle,
top, and an internal coordination (IC) view, repre-
sented by an excitable (myo)epithelium with short
and random connections, bottom. Behaviour of the
respective systems is shown on the left; the wiring
on the right.
impact of neural elongations on an excitable epithe-
lium where electrical activity is transmitted from
cell to cell (see figure 1).
The model is kept as abstract as possible: a
worm-like body-tube consisting of a single tubu-
lar sheet of cells, representing an excitable ep-
ithelium; electrical excitability generic enough to
represent either a calcium-based mechanism or a
sodium/potassium based one, and abstract exci-
tatory synaptic transmission to represent either
electrical or chemical transmission. The model
does not include external senses, proprioception, or
pacemakers. A Poisson process randomly initiates
activity in single cells. We consider traveling rings
of activity across the body-tube that could enable
a basic form of peristalsis to be relevant activity
(Figure 1).
Earlier modelling work showed that an ex-
3
t = 0t = 1t = 2t = 0t = 1t = 2 Figure 2: Panes A1-3 and B1-3 each represent individual cases of sequences of activity across the body-
tube for six different experiments (above), together with condensed representations spanning longer
sequences (below).
Individual time frames are identified by frame numbers, both above and below.
Panes A1-3 show a small (32 by 8) body-tube, panes B1-3 a large one (128 by 32). Panes A1 and B1
represent nearest-neighbour connections only; A2 and B2 have in addition elongations of length 2; A3
and B3, have elongations of length 4. In all experiments shown, 50% of the cells have elongations. The
graphs in panes A4 and B4 represent the average patternedness for the conditions shown above them
(marked in the graph) together with the results for five more elongation lengths. The points scattered
around the graph represent the 10 individual model iterations which were performed for each condition.
4
150160170180190200210220Black = 0 spikes; white = 16 spikesA1: 32x8; NN onlyB1: 128x32; NN onlyBlack = 0 spikes; white = 81 spikes150160170180190200210220176177178179180181176177178179180181A2: 32x8; length 2B2: 128x32; length 2Black = 0 spikes; white = 190 spikes150160170180190200210220Black = 0 spikes; white = 16 spikes150160170180190200210220176177178179180181176177178179180181Black = 0 spikes; white = 16 spikesBlack = 0 spikes; white = 249 spikesA3: 32x8; length 4B3: 128x32; length 4150160170180190200210220150160170180190200210220176177178179180181176177178179180181A4B4Elongation length01234PatternednessNN1.52346102.53.50123PatternednessNN1.52346102.53.5B1B2B3Elongation lengthA1A2A3 systems measuring 32 cells in length and 8 in cir-
cumference. Those on the right represent large sys-
tems of length 128 and circumference 32. All sys-
tems have nearest neighbour (NN) signalling. Ad-
ditionally, A2 and B2 have elongations of length
2, and A3 and B3 of length 4. These lengths re-
fer to each cell's elongation length, implying that
cells up to maximally 4 and 8 cells apart can be-
come connected. Panes A1-3 and B1-3 each pro-
vide a detailed picture of the randomly initiated
electrical activity across the cut open body-tube
during six time steps of three milliseconds. Below
is a condensed representation of this same activity
across 100 time steps, providing a temporally ex-
tended overview of this activity by compressing all
activity at a time-step into a line of 16 segments,
scaled to the size of the body-tube (see figure 4).
As white reflects high levels of activity, travelling
waves are shown as diagonal lines, unpatterned ac-
tivity as smudges, and synchronous activation of
all cells as a vertical line. Together, these panes
give an indication of both the details and the more
abstracted differences between the patterning re-
sulting from the various conditions.
In addition to these qualitative results, panes A4
and B4 show graphs for the same small and large
body-tubes, representing more extensive parameter
scans, involving the NN condition and eight differ-
ent elongation lengths (on the x-axis). The y-axis
shows the measure of patternedness calculated as
outlined in the section. The line represents the av-
erage patternedness for the given condition and the
points the individual model runs. The conditions
represented in detail in A1-3 and B1-3 are marked
as such within the two graphs.
As previously found (De Wiljes et al., 2015),
small systems without added elongations exhibit
ring-shaped patterns (A1), while these patterns are
lost when the system is larger (B1). As hypothe-
sized, when neural elongations are added to such
large systems, ring-shaped patterns return (B2-
3). However, such elongations are detrimental for
smaller systems (A2-3), where they have a nega-
tive effect on patternedness (A4). This can be seen
both by inspecting the examples and the quantita-
tive results presented in A4 and B4.
The comparison between A4 and B4 shows the
main effect of elongations on different topologies:
For the small systems, elongations lower the pat-
ternedness (A4). For the large systems, elongations
increase the patternedness, though returns do di-
minish for the longer elongations (B4).
The results discussed so far are based on sys-
tems where the probability of each cell having an
elongation is .5 (i.e. the chance that any given cell
has an elongation on top of nearest-neighbour con-
nectivity is 50%) while we focused on two sizes of
the body-tube. To investigate the effect of various
elongation fractions (hereafter referred to as f ) on
patternedness, we also compared elongation frac-
tions f across various body sizes and elongation
lengths as described above. We also did these ex-
periments for two extra body sizes: 64 by 16 and
256 by 64. The results are presented in figure 3,
which is similar to panes A4 and B4 of figure 2 (in-
cluded, respectively, as the lines marked with cir-
cles and diamonds in the pane marked 'f = 0.5')
but with additional conditions regarding f as well
as the additional body sizes.
Figure 3 shows the effect of f on patternedness.
From top to bottom, f increases. for larger systems
(darker red lines, marked with diamonds and tri-
angles), that having a higher fraction of cells with
elongations improves patternedness. For the small-
est system (yellow circles) elongations appear to
be detrimental to patternedness. For the second-
smallest system (orange squares), higher fractions
(lower graphs) have a beneficial effect on patterned-
ness for all but the longest two. However,
low
fractions (top graphs) show no such positive effect
on patternedness -- yet on the other hand, the low
fractions also harm less for the longer elongation
lengths.
Interestingly, for the larger body sizes, there ap-
pear to be diminishing returns to higher fractions.
For the largest system (darkest line), patternedness
for f = 0.5 is only marginally better than f = 0.2
for all elongation lengths.
Discussion
Our model represents an IC approach to early ner-
vous system evolution (see figure 1). Our aim was
to investigate internal coherence scenarios that in-
volve tissue configurations that are intermediate
between non-neural and neural ones.
In particu-
lar, we asked whether the presence of neural elon-
gations providing random connections over very
short distances could have functioned to enhance
5
internal coordination for comparatively small meio-
faunal proto-neuralia. Our results show how such
elongations can indeed enhance coordinated activ-
ity across very basic multicellular configurations.
We discuss four specific implications of the mod-
elling results.
First, short and randomly directed neural elon-
gations allow patterned activity within larger mul-
ticellular organisms. Direct connections between
adjacent cells within an excitable epithelium can
provide a way to initiate and maintain coordinated
patterns of excitability across a body surface. Such
patterns could have enabled organized contraction.
However, while this mechanism works for small
body-sizes, such patterning deteriorates for larger
bodies (De Wiljes et al., 2015). The model pre-
sented here shows that short and random neural
elongations can support similar forms of pattern-
ing for increasingly larger multicellular organisms.
Thus such very primitive forms of neural elonga-
tions provide a mechanism for adapting patterned
activity to changes in body-size.
Second, neural elongations provide a way to scale
the activity patterns themselves with respect to the
size of the organism. An epithelial configuration
only allows the spread of activation to adjacent
cells, which limits the width of the patterns in the
travelling direction to one or two cells (see Pane
A1 and B1 of figure 2). With elongations, the trav-
elling patterns of activity can extend across more
cells at every time-step. This allows the patterns
themselves to become wider and to scale up with
larger body sizes (see Pane B2 and B3).
Third, for larger systems (See Pane B4 of fig-
ure 2), patternedness tends to correlate positively
with the length of the elongations, although even
short elongations already provide an improvement
of patterning. Thus, while even short elongations
can be beneficial, lengthening them over evolution-
ary time provides a gradual path to further improve
such patterning capacities.
Fourth, while the influence of elongations on pat-
terning tends to be more prominent when a larger
fraction of cells have them, even a small fraction
of cells with elongations can have significant effects
depending on the size of the organism (see the top
graph of figure 3). Again, this provides an evo-
lutionary path for a gradual improvement of the
system.
Together these results provide a proof of concept
6
Figure 3: Comparison of the effect on patterned-
ness of different elongation lengths and body sizes
for six elongation fractions, f = .02 through f =
1.0. Line color and point shape represents differ-
ent body sizes. Patternedness is shown on the each
individual y-axis, the different elongation lengths
are on the x-axis and individual plots represent the
various values of f . The points scattered around
the graph line indicate the 10 individual model it-
erations which were performed for each condition
and the line itself represents the average.
Body size32 x 864 x 16128 x 32245 x 6401234Patternedness01234Patternedness01234Patternedness01234Patternedness01234Patternedness01234Patternedness012340123401234012340123401234NN1.52346102.53.5f = 0.02f = 0.05f = 0.1f = 0.2f = 0.5f = 1Elongation length that an organismal configuration relying on pat-
terned activity across an excitable epithelium can
use very basic neural elongations to maintain and
improve patterning capacity for larger body-sizes.
These changes can occur in small incremental steps
allowing for a gradual evolutionary route towards
increasingly complex neural elongations.
These findings have a broad conceptual relevance
for understanding the very early evolution of both
neurons and nervous systems. Rather than assum-
ing that neurons came first and through assembly
constituted the first nervous systems, here the se-
quence is reversed. With the IC view developed
here, 'nervous system functioning' can be produced
without full neurons by epithelia acting as a 'proto-
nervous system'. The latter can provide a scaffold
for gradually evolving full neurons and nervous sys-
tems. In this way, tissue configurations spanning
the gap between non-neural and neural tissues be-
come conceivable. The model presented here shows
that this speculative idea is indeed consistent and
opens up new avenues for looking at the early evo-
lution of nervous systems.
In comparison to the standard IO view that as-
sumes the beneficial presence of long and specif-
ically targeted neural connections,
the results
from this IC-based model provides new evolu-
tionary scenarios that bring the neuron's three
main features -- electrical signalling, synapses, and
elongations -- together in a gradual and plausible
way.
• Each cell has superthreshold connections to its
direct neighbors;
• Each cell has a superthreshold per-cell Poisson
process driver. In accord with De Wiljes et al.
(2015) the rate of Poisson process is set at 0.1
Hz;
• We used a worm-like topology. This is biologi-
cally plausible and functionally sufficient. The
length to circumference ratio of the folded tube
was fixed at 4:1. While we varied overall size
leading to length by circumference combina-
tions, respectively, of 32 by 8, 64 by 16, 128 by
32, and 256 by 64;
• A fraction of the cells were given straight elon-
gations in random directions which provide
them with connectivity to each cell visited by
the elongation; details can be found in supple-
ment . We performed a parameter scan over
various elongation lengths: 1.5, 2, 2.5, 3, 3.5,
4, 6, and 10 grid spacings. Additionally, we
included a condition without any elongations,
but as in all cases above keeping nearest neigh-
bour connectivity;
• Experiments have varying fractions of elon-
gated cells. We performed a parameter scan
over various elongated cell fractions. For frac-
tions < 1, cells to be elongated were picked
randomly. The fractions used are .02, .05, .1,
.2, .5, and 1.
Materials and Methods
Pattern Quantification
Computational Model
We reimplemented the model of De Wiljes et al.
(2015) using the brian2 package in Python (Good-
man and Brette, 2008). All experiments use the
following model:
• A 2D sheet of cells placed equidistantly in a
triangular grid;
• The sheet is rolled into a cylinder to create a
tube;
• Each cell is modelled by an integrate-and-fire
The system as described above produces spatiotem-
poral patterns. Previous work shows that given
suitable noise rates, nearest-neighbour connectiv-
ity, and activation with a refractory period, travel-
ling ring-shaped activity patterns appear. To quan-
tify the degree and relevance of ring-shaped activity
on the surface of the animal we developed a pat-
tern detection method. We detect patternedness
by comparing the activity in ring-shaped segments
with average activity over all segments. The ac-
tivity in a segment is simply the number of cells
active in a given ring-shaped segment in a given
time frame.
We summarize this activity for a given run as
model with a refractory period;
follows:
7
shaped segment travelling along the length of the
animal and showing up in the next segment in a
later time interval is a diagonal move.
To quantify the deviation in activity we per-
formed the following steps. First, we define a per
run normalized activity (cid:101)Cs,j:
(cid:80)
(cid:101)Cs,j =
Cs,j
Cs,j
(1)
∆t
nTend
s,j
Using the normalized activities we can define
a new patternedness measure P , which detects
how activity deviates from homogeneity during
each time steps and then calculates the root mean
square:
(cid:80)
((cid:101)Cs,j − ¯Cj)2
P =
s
(n · (Tend/∆t) − 1)
,
(2)
(cid:118)(cid:117)(cid:117)(cid:117)(cid:116) (cid:80)
(cid:80)
(cid:101)Cs,j
j
s
where ¯Cj = 1
n
A high P implies low homogeneity over time and
ring-shaped segments and thus strong patterned-
ness over time and segments, allowing us to easily
compare these values for various experimental con-
ditions.
This measure compares the activity of a given
segment not to the overall mean activity level but
to the activity level in a particular time step. This
way we disqualify situations where the whole body
shows high and homogeneous activity in one time
interval and low homogeneous activity in other time
intervals ('flickering' behaviour).
References
Braitenberg, V. (1984). Vehicles: Experiments in
synthetic psychology. MIT Press, Cambridge,
MA.
Brasier, M. (2009). Darwin's lost world: The hid-
life. Oxford University
den history of animal
Press, Oxford.
Budd, G. E. (2015). Early animal evolution and
the origins of nervous systems. Phil. Trans. R.
Soc. B, 370(1684):20150037.
Burkhardt, P. (2015). The origin and evolution of
synaptic proteins -- choanoflagellates lead the way.
8
Figure 4: This graphic shows how a condensed ac-
tivity representation is calculated. Time steps t
and t+1 show the process in detail. For each time-
step, the body-tube -- cut open in grey on the left --
is divided into 16 ring-shaped segments (in this
case, all 1 cell wide) for each of which the number
of active cells -- in red -- is tallied. This numerical
score is translated to greyscale values ranging from
white (maximum) to black (minimum) providing a
single vector for each time step.
• Divide the tube in n segments. A cell i is in
segment s if (s − 1)/n · ltube ≤ xi < s/n.
• A spike for a cell i in segment s takes place at
a time t. The simulation is divided into time
bins [j∆t, (j + 1)∆t).
• For each segment and time bin we can now
count how many spikes take place: the matrix
Cs,j indicates the number of spikes in segment
s during time bin j.
We set ∆t to 3ms, larger than the average trans-
mission delay between cells (2ms) but shorter than
the refractory period (20ms). Within a time bin
transmission between connected cells can occur but
a single cell cannot fire twice. Empty and incom-
plete intervals at the start and the end of each run
are discarded. The number of segments (n) is set
to 16 for all systems in order to be able to com-
pare measurements between the various body di-
mensions.
The ratio Tend/∆t determines the number of in-
tervals included in the analysis.
Visualizing Cs,j as shown in figure 4 results in
the condensed activity sequence plots found on
the lower sides of panes A1-3 and B1-3 in fig-
ure 2. Ring-shaped activity travelling down the
tube shows up in this format as diagonal stripes,
since a large amount of activation in a single ring-
tt+1t+2t+3t+4t+...004000000002220040000000032123 The Journal of experimental biology, 218(4):506 --
514.
Calcott, B. (2009).
Lineage explanations:
ex-
plaining how biological mechanisms change. The
British Journal for the Philosophy of Science,
60(1):51 -- 78.
Cunningham, J. A., Liu, A. G., Bengtson, S., and
Donoghue, P. C. (2017). The origin of animals:
Can molecular clocks and the fossil record be rec-
onciled? Bioessays, 39(1):1 -- 12.
De Wiljes, O. O., Van Elburg, R. A., Biehl, M., and
Keijzer, F. A. (2015). Modeling spontaneous ac-
tivity across an excitable epithelium: Support for
a coordination scenario of early neural evolution.
Frontiers in computational neuroscience, 9.
Keijzer, F. and Arnellos, A. (2017). The animal
sensorimotor organization: a challenge for the
environmental complexity thesis. Biology & Phi-
losophy, pages 1 -- 21.
Keijzer, F., Van Duijn, M., and Lyon, P. (2013).
What nervous systems do: Early evolution,
input-output, and the skin brain thesis. Adap-
tive Behavior, 21(2):67 -- 84.
Liebeskind, B. J., Hillis, D. M., and Zakon, H. H.
(2011). Evolution of sodium channels predates
the origin of nervous systems in animals. Proc
Natl Acad Sci U S A, 108(22):9154 -- 9159.
Mackie, G. O. (1970). Neuroid conduction and the
evolution of conducting tissues. Quarterly Re-
view of Biology, pages 319 -- 332.
Erwin, D. H. (2015). Early metazoan life: diver-
gence, environment and ecology. Phil. Trans. R.
Soc. B, 370(1684):20150036.
Mackie, G. O. and Passano, L. M. (1968). Ep-
ithelial conduction in hydromedusae. Journal of
General Physiology, 52(4):600.
Erwin, D. H., Laflamme, M., Tweedt, S. M., Sper-
ling, E. A., Pisani, D., and Peterson, K. J.
(2011). The cambrian conundrum: early diver-
gence and later ecological success in the early his-
tory of animals. science, 334(6059):1091 -- 1097.
Fedonkin, M. A., Gehling, J., Grey, K., Narbonne,
G., and Vickers-Rich, P. (2007). The rise of an-
imals: evolution and diversification of the king-
dom Animalia. JHU Press.
Goodman, D. and Brette, R. (2008). Brian: a simu-
lator for spiking neural networks in python. Fron-
tiers in neuroinformatics, 2.
Greenspan, R. J. (2007). An introduction to ner-
vous systems. Cold Spring Harbor Laboratory
Press, Cold Spring Harbor, NY.
Hejnol, A. (2015). Acoelomorpha. Structure and
evolution of invertebrate nervous systems, pages
56 -- 61.
J´ekely, G. (2011). Origin and early evolution of
neural circuits for the control of ciliary loco-
motion. Proceedings of the Royal Society B,
278(1707):914 -- 922.
Moroz, L. L. (2009). On the independent origins
of complex brains and neurons. Brain, Behavior
and Evolution, 74(3):177 -- 190.
Moroz, L. L., Kocot, K. M., Citarella, M. R., Do-
sung, S., Norekian, T. P., Povolotskaya, I. S.,
Grigorenko, A. P., Dailey, C., Berezikov, E.,
Buckley, K. M., et al. (2014). The ctenophore
genome and the evolutionary origins of neural
systems. Nature, 510(7503):109 -- 114.
Naitoh, Y. and Eckert, R.
Ionic
mechanisms controlling behavioral responses of
paramecium to mechanical stimulation. Science,
164(3882):963 -- 965.
(1969).
Nielsen, C. (2008). Six major steps in animal evo-
lution: are we derived sponge larvae? Evolution
& Development, 10(2):241 -- 257.
Nielsen, C. (2013). Life cycle evolution: was the eu-
metazoan ancestor a holopelagic, planktotrophic
gastraea? BMC evolutionary biology, 13(1):171.
Pantin, C. F. A. (1956). The origin of the nervous
system. Pubblicazioni della Stazione Zoologica di
Napoli, 28:171 -- 181.
J´ekely, G., Keijzer, F., and Godfrey-Smith, P.
(2015). An option space for early neural evolu-
tion. Phil. Trans. R. Soc. B, 370(1684):20150181.
Ryan, T. J. and Grant, S. G. N. (2009). The ori-
gin and evolution of synapses. Nature Reviews
Neuroscience, 10(10):701 -- 712.
9
Smith, C. L., Varoqueaux, F., Kittelmann, M., Az-
zam, R. N., Cooper, B., Winters, C. A., Eitel, M.,
Fasshauer, D., and Reese, T. S. (2014). Novel cell
types, neurosecretory cells, and body plan of the
early-diverging metazoan trichoplax adhaerens.
Current Biology, 24(14):1565 -- 1572.
Valentine, J. W. (2004). On the origin of phyla.
University of Chicago Press.
Wray, G. A. (2015). Molecular clocks and the early
evolution of metazoan nervous systems. Phil.
Trans. R. Soc. B, 370(1684):20150046.
10
Supplementary Material 1: Mikado
implementation details
Regarding implementation details, consider the
case of cells arranged in a plane. Every elon-
gated cell extends an axodendritic elongation p
with length lp in an arbitrary direction φp, rela-
tive to the x-axis. Where elongations cross each
other they create bidirectional connections between
the originating cells. Given any two cells i and j
with coordinates (xi, yi) and (xj, yj) and elongation
directions relative to the x-axis φi and φj respec-
tively, a triangle emerges, with at one point cell i,
another point cell j and the third point the crossing
of their elongations. Two sides of this triangle con-
sist of the elongations themselves. The third side,
d, is the line between the cells, with an angle rela-
tive to the x-axis: φd. Assuming φi (cid:54)= φj (cid:54)= φd
1, the
sine-rule allows us to locate the crossing location,
given the angles and the locations of the cells.
d
sin(φj − φi)
=
ρi
sin(π − φj + φd)
=
ρj
sin(φi − φd)
(3)
2-dimensional plane, its contiguousness results in
some particular cases: elongations which may wrap
around the cylinder.
If we assume the cylinder
is formed by making the x-direction periodic, we
should consider not only the crossing of elonga-
tions starting at (xi, yi) and (xj, yj) but also at
(xi + ki · (cid:12), yi) and (xj + kj · (cid:12), yj), with k repre-
senting the number of windings and (cid:12) the circum-
ference of the tube. We can limit the number of
windings we need to consider, since the maximum
elongation distance is limited: k < (2 · lp/(cid:12)) + 1.
The +1 is needed because the nearest copy might
be on the other side of the tube's period. Another
particular of the tube as opposed to a torus is that
the crossing point may lie off the tube on the finite
end, so we need to check whether the y-coordinate
of the crossing, ycrossing = ρi · sin φi + yi is on the
tube: −0.5 ≤ ycrossing ≤ (ltube − 1) · 1
3 + 0.5,
√
where ltube is the number of rings on the tube and
1
3 is the height of an equilateral triangle with
2
sides of length 1.
√
2
Furthermore,
sin(π − φj + φd) = − sin(−φj + φd) = sin(φj − φd)
(4)
Cell coordinates allow us to calculate d and φd:
(cid:113)
φd = arctan
d =
(xi − xj)2 + (yi − yj)2
(cid:17)
(cid:16) yj − yi
xj − xi
Then we can solve for ρi and ρj:
ρi = d · sin(φj − φd)
sin(φj − φi)
ρj = d · sin(φi − φd)
sin(φj − φi)
(5)
(6)
(7)
(8)
If, for a given pair of cells, ρi and ρj are shorter than
lp, so 0 < ρi < lp and 0 < ρj < lp , the cells are
connected. We loosened this constraint to − 1
2 <
ρi < lp and − 1
2 < ρj < lp to account for crossings
occurring on the cell body of the originating cell.
All the above, however, is predicated on an in-
finite plane, while our system is finite and tube-
shaped. Although a tube is essentially a curved
1In practice, given the granularity of the random numbers
used, this assumption always holds.
11
Figure 5: Illustration of variation in segment num-
ber affecting the patternedness measure. Individ-
ual graphs represent segmentation options: the top
graph represents 8 segments whereas the bottom
graph represents 8 segments. Different body sizes
are differentiated by color and point shape. Again,
patternedness is shown on the y-axis, the different
elongation lengths are on the x-axis. The points
scattered around the graph indicate the 10 individ-
ual model iterations which were performed for each
condition and the line represents the average.
Since the aim of our study is to establish qualita-
tive changes in whole-body coordination, having a
measure that is constant across multiple body sizes
is important, and dividing the body-tube in a fixed
number of segments accomplishes this. Whether
that is 8 or 16 is of secondary importance for the
purposes of this study.
Supplementary Material 2: Addi-
tional notes on pattern quantification
The constant n, denoting the number of segments
into which the modelled body is divided, requires
more scrutiny. The analysis yields different out-
comes for different values of n. This supplement
aims to provide an explanation of why we chose to
set the number of segments to 16 instead of iterat-
ing over it or choosing another value.
First of all, we use a power of two when setting
the length of our model. We force our segments
to all have the same length and thus our possible
choices for the number of segments are limited to
powers of two as well. The shortest length used is
32, that is 25 and thus we have 2, 4, 8, 16 and 32
to pick from. The values 2 and 4 are theoretically
allowable but do not fit our intuition on wave prop-
agation. The value 32 is also allowed on theoretical
grounds, but already nearest neighbor connections
alone will spread out of the bin within a single time
bin. This leaves us with two prima facie suitable
candidates: n = 8 and n = 16, which we evaluated
here.
Succinctly put, the qualitative results are similar
when using different number of segments. Figure 5
shows that there is a difference in actual values but
hardly a difference in overall shape. The effect of a
lower number of segments is twofold: first, it causes
patternedness to be higher for faster ring-shaped
patterns. The optimum values will thus be biased
towards longer elongation lengths, as those provide
faster pattern propagation. Second, since more seg-
ments allows finer measurement the higher number
of segments is able to show higher values overall.
Even neat patterns within a large segment result
in some lost signal. This also results in the effect
visible in figure 5, where 8 segments (top graph) for
small systems (yellow circles) produces an improve-
ment in patternedness between not having elonga-
tions at all and having elongations of length 2, since
the broader patterns generated by elongations al-
low the segments to fill out whereas smaller seg-
ments get filled out even with patterns 1 cell wide.
12
01234Patternedness01234Patternedness0123401234Body size32 x 864 x 16128 x 32245 x 64n = 8Elongation lengthNN1.52346102.53.5n = 16 |
1210.7495 | 3 | 1210 | 2016-02-27T14:45:52 | Illustrating a neural model of logic computations: The case of Sherlock Holmes' old maxim | [
"q-bio.NC",
"cs.AI"
] | Natural languages can express some logical propositions that humans are able to understand. We illustrate this fact with a famous text that Conan Doyle attributed to Holmes: 'It is an old maxim of mine that when you have excluded the impossible, whatever remains, however improbable, must be the truth'. This is a subtle logical statement usually felt as an evident truth. The problem we are trying to solve is the cognitive reason for such a feeling. We postulate here that we accept Holmes' maxim as true because our adult brains are equipped with neural modules that naturally perform modal logical computations. | q-bio.NC | q-bio | Published in THEORIA 31/1 (2016): 7-25
Illustrating a neural model of logic computations:
The case of Sherlock Holmes' old maxim*
Eduardo Mizraji
Group of Cognitive Systems Modeling, Biophysics Section, Facultad de Ciencias,
Universidad de la República, Iguá 4225, Montevideo 11400, Uruguay
E-mails: [email protected], [email protected]
ABSTRACT: Natural languages can express some logical propositions that humans
are able to understand. We illustrate this fact with a famous text that Conan Doyle
attributed to Holmes: "It is an old maxim of mine that when you have excluded the
impossible, whatever remains, however improbable, must be the truth". This is a subtle
logical statement usually felt as an evident truth. The problem we are trying to solve
is the cognitive reason for such a feeling. We postulate here that we accept Holmes'
maxim as true because our adult brains are equipped with neural modules that
naturally perform modal logical computations.
Keywords: Neural computations; Natural language; Models of reasoning; Modal
logics
* Previous version in arXiv:
"The neural computation of Sherlock Holmes' old maxim"
1
1. Introduction
Human language is used to express logical and mathematical computations whose
cognitive bases still remain unexplained. We ask ourselves if language acquisition
involves a kind of implicit logical and mathematical programming that could explain
such performances. Examples of these performances are some logical propositions,
transferred by natural language, valid for different languages and for large
populations of humans sharing similar cultural traditions. In the present work we
choose –as a largely accepted logical statement– one of the most cited expressions
that Arthur Conan Doyle attributed to Sherlock Holmes, the "old maxim" mentioned
by the character in the story "The Adventure of the Beryl Coronet". For an allusion to
this maxim in a scientific context see Cairns-Smith (1990). We enunciate different
versions of this "old maxim", and the reader can explore by himself the origin of the
conviction that these expressions usually provoke (see Doyle 1988). This maxim is a
subtle logical statement and the interesting (and even astonishing) point is that
many readers feel that this maxim is an evident truth. One problem we are trying to
explain is the following: which is the cognitive reason for such a feeling? The
method we follow here implies to keep the theories as simple as possible. We avoid
in this first approach the temptation to deeply expand Bayesian theories, logical
formalisms or the mathematical technicalities of neural models. Our aim is to
explore the possibility of establishing an acceptable link between very different
disciplines that are, however, all connected with our problem.
Perhaps the most important question about the "old maxim" concerns the
consistency of the statement. At the same time, the most enigmatic aspect of this
maxim is the natural acceptability triggered by it in our mind. We are going to
analyze two approaches to consistency: the
logical consistency and the
"neurocomputational consistency". We propose to accept the logical consistency
whenever the statement can be immersed in normal well-formed formulas of basic
propositional and modal
logic.
In addition, we propose
to accept
the
2
neurocomputational acceptability if the logical computation of the statement can be
executed by a neural network model. In this work we intend to show that both
consistency conditions are satisfied. We begin by describing a Bayesian approach to
the maxim. Then, we provide a logical framework to Holmes' statement. After this,
we show how these logical approaches can be computed in model neural modules
capable to integrate modular networks (ie. networks of networks) that execute a
variety of logical operations. With these purposes, we begin describing some basic
concepts of propositional calculus and modal logic that confer the logical support of
Holmes' maxim. Then, we present a kind of modular neural model based on matrix
algebra, and describe how logic operations can be very naturally incorporated into
these matrix modules. Finally, we put together this material to provide a neural
model that is capable to represent Holmes' maxim. This approach allows us to
propose an explanation of the reason why the Holmes' maxim seems so naturally
acceptable by us. Our explanation is based on modeling some particular neural
networks capable of representing the cognitive computation of modalities.
2. Holmes' Old Maxim
Perhaps the most cited version is the following: "It is an old maxim of mine that when
you have excluded the impossible, whatever remains, however improbable, must be the
truth" (in Appendix 1 we present three versions of this maxim with the
corresponding sources). In the articles compiled by Eco and Sebeok (1988) there are
many allusions to these texts, as well as some formalizations of the methods
attributed by Conan Doyle to Holmes using techniques coming from mathematical
logic.
The involvement of probability in this text invites us to approach the meaning of
these statements using a naïve Bayesian description. In this context, the probability
)BA(P
i
of
iA being the cause of event B is given by
3
)BA(P
i
=
i
)A(P)AB(P
i
n
∑
=
1j
)A(P)AB(P
j
j
(1)
where it is assumed that the set {
being interpreted as causes of the event B (eg. B can be a symptom of a disease and
{
includes disjoint events susceptible of
the set of potential pathological agents). Each
measures the
,A …
}n
,A …
A,
1
)AB(P
j
}n
A,
1
conditional probability of event B given
jA , and each
)A(P
j
measures (or
estimates) the a priori probability of event
jA .
In this Bayesian framework, Holmes' maxim can be interpreted as follows: after a
research process the investigator can establish that for all
j „
i
)AB(P
j
0
: As a
consequence, only
0)AB(P
i
. Hence, at this stage, Bayes' formula adopts the
following aspect:
P(A B)
i
+
0
+
⋯
P(B A ) P(A )
i
P(B A ) P(A )
i
i
i
+
⋯
+
0
(2)
Consequently,
)BA(P
i
1
independently of
the value of
the a priori
probability
)A(P
i
. This deliberately simplified Bayesian approach gives us a good
insight into the meaning of Holmes' maxim.
It is important to remark that, together with allusions to probability, Holmes' maxim
includes the words "impossible" and "truth". These two words point to a logical
framework, and particularly "impossible" leads us to the domain of modal logic.
Consequently, we are going to assume that understanding our cognitive acceptation
of the maxim requires to start from logic, and to include a posteriori a probabilistic
argument inside the logical frame. Moreover, we need to include this logical frame
4
fi
„
»
fi
inside a neural model in order to explore how a biological device such as the brain
could become prone to accept the validity of Holmes' maxim. We introduce the
neural modeling framework in Section 4.1 and the probabilistic approach returns in
Section 4.3.
3. Logical Consistency
Some of the fundamental ideas about modal logic were exposed by Aristotle in his
short text "On Interpretation" (Aristotle, around 350 BC, edition E.M. Edghill). The
two basic modal operators "possibility" and "necessity" are formally related by a
postulate clearly stated by Aristotle. We provide here with a version of this postulate
(Aristotle, 350 BC, Chapter 13):
"It remains, therefore, that the proposition 'it is not necessary that it should not be'
follows from the proposition 'it may be' ".
A modern formalization of modal logic can be found in Hugues and Cresswell 1972.
The symbolic representation for the previous postulate is the following,
◊
(Q)
(
Q
)
□
(3)
where ◊ represents the modal proposition "It is possible", □ means "It is necessary",
is the negation, and "
represents the logical equivalence. Q represents any
proposition.
An equivalent representation is
(Q)
□
◊ (cid:216)
(
Q
)
(4)
5
"
(cid:216)
(cid:216)
(cid:216)
"
(cid:216)
A corollary, based on the fact that the double negation corresponds to an
affirmation, is
[
]
◊
(Q)
(cid:216)
□
(
Q
)
(5)
Note that if Q is an arbitrary proposition, its negation (
)Q(cid:216)
can be interpreted as
"whatever remains" once excluded Q. Hence, this corollary (5) shows the proximity
of Aristotle's postulate with Holmes' maxim: If it is true that Q is impossible, it is
true that what remains,
(
)Q(cid:216)
, is necessary.
Remark that the equivalences (3), (4) and (5) are true in all cases, notwithstanding
the truth-value of the modal evaluations of Q (i.e., both members of the equivalence
can be false). The relation with Holmes' maxim restricts to the case in which both
members of the equivalence are true.
The works of De Morgan, Gregory, Boole and Peirce, among others, were
fundamental to transform logic into a discipline susceptible of being supported by
the techniques of mathematics. Their work produced a variety of mathematical
representations of logical calculus. We note here that, deeply influenced by the
symbolic methods developed in the field of differential equations (his main area of
expertise), Boole established the basic conditions that allow to map the logic truth-
values on mathematical variables and to transform the logical statements in
mathematical functions (Boole 1847, 1854). In addition, in his book "The Laws of
Thought" (1854) Boole attempted to link logical procedures with probability theory.
From this Boolean approach, it was possible to define functional relations for all the
fundamental logical operations (e.g., negation, disjunction, conjunction, implication,
exclusive-or, equivalence).
6
(cid:216)
"
When logic is immersed into this "Boole's Universe" (BU), the truth-values define a
set
t =
2
{
t , f
}
where t and f are abstract (even arbitrary) mathematical objects corresponding to
the truth-values "true" and "false" respectively. In this BU, the monadic functions
Mon are applications
Mon : t
2
2
(the negation Not is an example of these monadic functions, being Not (t) = f and
Not (f) = t). The dyadic functions Dyad are applications
Dyad : t
2
2
;
2
where the symbol ·
indicates the Cartesian product. The truth tables used to
represent the logical operations (e.g. disjunction or implication) are examples of the
use of these dyadic mappings.
Following the methods created by Boole, other researchers tried to represent modal
operations as mathematical functions. Nevertheless, in a famous work Łukasiewicz
(1930) demonstrated the impossibility to represent "possibility" and "necessity" as
mathematical functions inside the two-valued logic defined in BU (for details, see
Łukasiewicz 1930, Mizraji 2008). Consequently, the search for truth-functional
representations for these logical modalities leads Łukasiewicz to extend the truth-
value space by adding a third value "u" corresponding to uncertain or undecidable
propositions. In this new Łukasiewicz's Universe (LU), the logical monadic and
dyadic functions are built up over the set
7
fi
t
·
t
fi
t
{
t, f , u
}
t =
3
where
Mon : t
and
Dyad : t
3
3
3
3
.
3
Inside this LU, the classical modalities become monadic logical functions, and
"possibility" and "necessity" can be respectively expressed by the functions
possibility
and necessity
(
)x◊
)x□
(
x t
,
3
, defined as follows:
" ◊
(
)
u
t ;
□
t ;
)
u
(
( )
t
( )
t
◊
□
=
f
◊
( )
f
( )
f
□
f
,
The negation is defined in the LU as follows:
( )
t
f ;
(
u
)
u ;
( )
f
t
.
With this formal repertoire, we can represent Aristotle's postulate with a truth-
functional equivalence:
◊
[
]
x(Q)
(cid:216)
□
(
x(Q)
)
(6)
8
fi
t
·
t
fi
t
"
"
"
"
(cid:216)
"
(cid:216)
"
(cid:216)
"
"
(cid:216)
where
)Q(x
3
is the truth-value corresponding to an abstract proposition Q.
Equation (6) transforms the Aristotelian postulate into a theorem. Being the
negation (cid:216)
, we can deduce the following
an idempotent operator
)
x
]
(x)
(
[
equivalence from (6) :
[
◊
x(Q)
]
□
(
x(Q)
)
(7)
We are going to assume the following Axiom:
AXIOM :
(
)
x(Q)
(
x Negation Q
)
;
Observe that, in this truth-functional format, our logical operators apply only to
truth-values; instead "Negation Q" designates the negation of a proposition. For
instance, the truth-value of Q = "3 is an even number" is f and its negation is t; the
truth-value of Negation Q = "3 is not an even number" is t; "Negation Q" is not a
mathematical variable but a proposition and can be considered as a propositional
definition of negation. This Axiom can be proved as a Lemma (Mizraji and Lin 2011)
if we consider that Q refers to a category W (a set) of propositions, and that the
proposition is true if Q belongs to this category and false if it belongs to the negation
of this category (the complement of the set W).
Using this Axiom, the logical equivalence (7) can be transformed as follows:
[
◊
x(Q)
]
(
x Negation Q
)
□
(8)
If it is verified that
9
t
(cid:216)
(cid:216)
"
(cid:216)
"
(cid:216)
(cid:216)
"
(cid:216)
"
[
◊
x(Q)
]
t
(9)
then the equivalences (7) or (8) can be considered as partial mathematical models of
Holmes' statement. In fact, this mathematical result seems too simple because
equivalence (9), due to Łukasiewicz functional definitions of negation (cid:216)
possibility ◊ , implies that
and, hence, that x(Q) = f. Does this excessive
and
[
◊
]
x(Q)
f
simplicity disqualify the usefulness of the logical model? If one considers that the
problem for Holmes is to develop a research process able to abolish uncertainty, the
answer is negative because the statement x(Q) = f represents the success of such
process. In fact, the introduction of the third truth-value in Łukasiewicz logic was
essential to guide the neural modeling of modal logic. As we are going to see in the
next Section, modalities can be computed in the matrix neural models using two
different ways to represent the uncertain truth-value in the network: on the one
hand, defining a specific "truth-vector" to characterize uncertain truth-values
(Section 4.2), and on the other, as a probabilistic weighting of the "true" and "false"
vectors (Section 4.3).
The modal formats previously established are extremely important for the
interpretation of Holmes' maxim, but they do not solve our problem because they do
not constitute cognitive models and probabilities are absent. However, the modal
relation (5) and the corresponding truth-valued representations (7) to (9) give us a
powerful formal construction that shares with Holmes' postulate the enigmatic
"feeling" of correctness that it produces. In fact, this formalism is a fundamental link
to connect these modal equivalences with mathematical models of logical operations
derived from neural models. Consequently, this approach covers some logical
aspects of the statement, but a more comprehensive modeling should include the
following points: a) a connection with the neural structures that produce and decode
Holmes' maxim, and b) a link with probability concepts. As we are going to show in
10
(cid:216)
"
"
the next Section, matrix algebra is a fundamental framework for the construction of
these models.
4. A neural approach to logic computations
The human brain is a spontaneous, pre-theoretical, computing device capable of
performing sophisticated information processing, including mathematical and
geometrical calculations. It is "pre-theoretical" in the sense that the human brains
display many computational performances with no need of any explicitly
programmed procedures or techniques. Clearly, language is used to express logical
and mathematical computations whose cognitive bases still remain unexplained. We
can ask ourselves if language acquisition involves a kind of implicit logical and
mathematical programming that could explain such performances.
It is a common experience that the brain can solve problems concerning visual
patterns without using any pre-existing mathematical knowledge. These problems
are usually stated verbally. To illustrate this point, we can consider the image of
Figure 1 representing a road.
Figure 1. The points A and B over the design of a curved road have very different
curvatures.
11
If you ask adults in which zone, A or B, the road shows larger curvature, the majority
of the answers indicate point B. Of course, this conclusion is not based on the use of
the classical mathematical formula
)x(K
=
2
yd
2
dx
1
+
dy
dx
2
3
2
that gives the curvature of point (x, y) in a plane trajectory with Cartesian equation
y(x). On the contrary, almost surely this equation was strongly inspired by the pre-
existing cognitive notion of curvature.
Logical judgments usually integrate the repertoire of human cognition, and even if
many human actions are not submitted to the logic, the logical procedures are used
for some crucial tasks that involve rational decisions. Consequently, it is particularly
relevant to
investigate the relation between cognitive models and
logical
performances (for an analysis of this point, see Binazzi 2012).
It is important to note that the three-valued logic, defined in Section 3, assumes the
existence of well pre-classified data according to three cognitive categories: true,
false, uncertain. The possibility of this classification is the consequence of clear
diagnoses about the nature of the facts (e.g.: due to the lack of documents some
historical facts can be transitorily classified in these three categories). But an open
and evolving inquiry is full of transitory conjectures. In this situation, a sharp
classification is not possible and the investigators must explore actual facts and build
up their conclusions trying to decide if the conjectures are true or false in an
environment full of uncertainties. In fact, this is the normal situation for many of the
decisions adopted by humans in their natural environments.
4.1 Matrix Associative Memories
12
-
Many different approaches have been proposed to describe different aspects of
neural function (Arbib 1995). In particular, the relations between brain dynamics
and linguistic processes have been the subject of important investigations (see, for
instance, Blutner 2004, beim Graben 2008). Let us mention the importance of
matrices in the development of the theory of neural associative memories, a theory
mainly developed around 1970 (Anderson 1972, Kohonen 1972, Cooper 1974). The
matrix associative memories are able to model important facts about biological
memories established
in many experimental and clinical
investigations.
Comprehensive descriptions of this theory are included in Kohonen (1977) and
Anderson (1995). These matrix memories are considered "distributed memories"
because the residence of the information is a large set of synaptic contacts between
neurons. This information is scattered and partitioned prior to be stored. This fact
produces a desirable robustness of the model, in the sense that stored data can
persist even in the presence of damages that produce a loss of neurons or synapses;
this robustness of the model is a desirable fact because biology had revealed the
existence of a relative tolerance to damage in real memories (Anderson 1995).
This theory assumes that the cognitive patterns correspond to neural activities that
map on large dimensional vectors. A memory is a matrix that associates pairs of
column vectors
)f,g(
i
i
, i = 1,2, …, K, where
if corresponds to the pattern i that
enters the memory (e.g. the image of a person), and
ig is the associated output (e.g.,
a name associated with the input image). As the theory shows, these vectors are
composed by the electrochemical signals used by neurons to code information; these
signals are generated in parallel by thousands of firing axons (Anderson 1995). The
simplest form of the matrix that stores those pairs of vectors is as follows:
=
A
K
∑
=
1i
i fg
T
i
. (10)
13
(the superindex T indicates transposition). Usually it is assumed that the set of
stored input vectors is orthonormal (i.e. the
if are orthogonal between them and
with lengths equal one). This assumption implies that the similarity between
patterns is measured by the angle (equal patterns are parallel and completely
different patterns are orthogonal). When a pattern kf
enters the memory A, it is
processed and generates an output. The following equation illustrates the
mechanism:
fA
k
=
K
∑
=
1i
f,f
i
k
g
i
(11)
where
f,f
i
k
=
f
T
i
f
k
is the scalar product (an operation that directly produces the
cosines of the angle between this multidimensional unitary vectors; this cosines
measure the angle and, consequently, the similarity between the patterns). If the
{ }i
f
, we have
input pattern belongs to the set stored into the memory, i.e.
f
k
fA =
k g
k
, (12)
a perfect association.
To include semantic contexts in the framework of this theory, different approaches
involving a complex integration between inputs and contexts have been proposed
(Arbib 1995). One of these approaches (Mizraji and Lin 2011) uses the Kronecker
product to integrate inputs and contexts. In this framework, the matrix memory can
be expressed as
=
M
∑
i, j
g ( f
ij
i
T
p )
j
(13)
14
jp is the context associated with the input
where
associated with the contextualized input. The symbol
represents the Kronecker
(or tensor) product; in Appendix 2 we describe the basic properties of this operation
if , and
ijg is the output
(we use this tensor product inside a neural model, but for a foundation of this
operation based on cognitive science see Smolensky 1990).
According to the algebraic rules involved in matrix algebra and Kronecker products
(Graham 1981), the response of matrix memory M in the presence of an input and its
context is
f(M
k
)p
h
∑=
j,i
f,f
i
k
p,p
j
h
g
ij
(14)
with exact associations if the sets { }if
and {
}jp are orthonormal, and if
f
k
{ }i
f
and
p
h
{
}j
p
.
If matrix memories (10) and (13) represent biological associative memory modules,
they are usually rectangular matrices of large dimensionality.
4.2 A Matrix-Vector Logic
The neural models previously described, provide us with a simple and powerful way
to represent a large variety of logical operations1. Based on these memory modules,
some years ago a matrix formalism named 'vector logic', that connects elementary
propositional and modal logics with matrix neural models, was developed (see, for
instance, Mizraji 2008). Inside this neural theory, the logical gates map on matrices
1 For an historical account of the links between logical theory and neural models see
Eduardo Mizraji (2013) En Busca de las Leyes del Pensamiento (Second Edition),
Montevideo: Trilce-Dirac.
15
and the truth values on vectors. The procedure to create the maps begins by
mapping the truth values on orthogonal unitary vectors. Inside this neural
formalism, the number of vector truth-values is a priori only limited by the
dimensionality of the neural vectors. A large variety of many-valued logics can in
principle be developed and sustained by matrix memory modules, because normally
it is biologically plausible to assume that a neural vector has hundreds or thousands
of components (Anderson 1995). In what follows, we describe two- and three-valued
matrix-vector logics. In Section 4.3 we are going to show how a translation of the
classical neural representation of modalities described by McCulloch-Pitts (1943)
neuronal circuits can be adapted to produce an infinite-valued logic from two truth
values and probabilistic weights.
Thus, a two-valued logic requires mapping
orthonormal q-dimensional vectors; hence
s
t ֏ and
{
2 =
n
f ֏ , with s and n being
}n,s
this vector
. Using
representation for the truth-values, the monadic and the dyadic two-valued gates
respectively become the functions
2:)2(Mon
Dyad
2:)2(
2
2
,
.
2
A three-valued logic is defined over the vector set
{
}h,n,s
3 =
, where h is a q-
dimensional normal vector (orthogonal to vectors s and n) corresponding to the
uncertain truth-value u (
u ֏ ). This vector set allows building up matrix versions
h
for monadic and dyadic three-valued logical operators:
3:)3(Mon
Dyad
3:)3(
3
3
,
.
3
16
t
t
fi
t
t
fi
t
·
t
t
t
fi
t
t
fi
t
·
t
As simple examples of Mon(2) we have the following matrices, 2I
and
2N , that
correspond in this matrix framework to the logical two-valued Identity and Negation
:
I
2
====
T
ss
++++
T
nn
, (15)
====
T
ns
++++
T
sn
N
2
. (16)
Important examples of Dyad(2) operators are the matrix conjunction
2C and
disjunction
2D , given by the expressions:
====
====
s(s
s(s
C
2
D
2
T
)s
++++
s(n
T
)n
++++
n(n
T
)s
++++
n(n
T
)n
, (17)
T
)s
++++
s(s
T
)n
++++
n(s
T
)s
++++
n(n
T
)n
. (18)
Using these equations it can be easily proved that these operators execute vector
versions of the classical operations. For instance,
s(C2
====
)s
s
;
s(C
2
====
n(C)n
2
====
n(C)s
2
====
)n
n
.
It is important to see that the monadic operators (15) and (16) are particular cases of
memory modules (10), and that the dyadic operators (17) and (18) correspond to
memory modules (13). The vector logic provides explicit expressions for some of the
Mon(3) and Dyad(3) matrix operators. We mention the fact that, under this
formalism, the modalities possibility and necessity become very simple monadic
matrices (a consequence of the Łukasiewicz functional definitions described in
Section 3). Thus, for the three-valued vector system of logic, the identity 3I , the
3N , the matrix Pos (that represents in this formalism the possibility ◊ )
negation
17
(cid:216)
and the matrix Nec (that represents the necessity □ ) can be expressed by the
following simple formulas:
=
I
2
I
3
+
T
hh
, (19)
N
3
=
N
2
+
T
hh
, (20)
Pos
=
I
2
+
T
sh
, (21)
Nec
=
I
2
+
T
nh
, (22)
where 2I
and
2N are the operators given by equations (15) and (16)
Inside this matrix formalism, some basic theorems on logical modalities can be
expressed as vector-matrix equalities. For instance, the postulate expressed by the
equivalence (3) and its algebraic version given by (6), can be expressed by the matrix
equation
Pos
3=
N)Q(Val
N[Nec
3
)]Q(Val
, (23)
with Val(Q)
}h,n,s
)Q(Val
{
representing
the
truth-value assigned
to proposition Q,
. In addition, in this case the rules of matrix calculus allow us to
express Aristotle's postulate as a product between the matrix operators involved:
, (24)
Pos =
N
3
NNec
3
an identity not dependent on any particular value of the logical variable. These
modal operators can be used as a way to represent logical modalities in terms of
memory modules; for a discussion of the biological situations where these
representations can be operative see Mizraji (2008).
18
The point of contact with Holmes' maxim is given by the matrix versions of
equations (7) and (9):
N
3
Pos
)Q(Val
=
N[Nec
3
)]Q(Val
(25)
and
N3
Pos
)Q(Val
= . (26)
s
Consequently,
N[Nec
3
)]Q(Val
= . (27)
s
We can adapt the Axiom of Section 3 to this context by writing
3N Val(Q) Val (Ne gation Q)
=
. (28)
Note that Q is not a vector-matrix variable and we cannot apply the matrix negation
on it. Instead, Negation Q is a proposition with a vector truth-valuation
Val(Negation Q). Under this condition, equation (18) can be restated as
3N Pos Val(Q) Nec[Val (Ne gation Q)]
=
= . (29)
s
This is a way to express that the impossibility of a proposition is equivalent to the
necessity of the complement of that proposition. This result establishes a close
contact with Holmes' maxim.
Obviously, equation (27) implies an extremely simple fact: Val (Negation Q) = s. As
we mentioned previously, this seems a trivial conclusion, but it can be the result of a
19
non-trivial process that was capable of eliminating uncertainty and of converging to
this assertion. In fact, the valuation represented by Val(Q) (or Val(Negation Q))
involves a neural substrate apt to sustain the very complex cognitive process
required to diagnose the truth-value of a proposition.
Equation (27) helps us to put the modal logical apparatus in terms of neural models,
and to provide a new point of view to approach the spontaneous understanding that
we feel when we are confronted with Holmes' old maxim. But this formalism does
not consider any probabilistic evaluation of propositions, except the global
assignment of uncertainties via the inclusion of a third truth-value. In the next
Section, we are going to connect the previous matrix-logical approach with matrix
models for associative memories capable of sustaining probabilistic operations.
It is easy to see that the matrix logical operations of this Section are particular cases
of the Anderson-Kohonen distributed memories. In equation (25) Holmes' maxim is
represented using modal logical operators and truth-evaluations expressed with a
matrix-vector formalism. These logical operators are interpretable as memory
modules that integrate a modular network (in fact a network of networks, because
each one of the logical modules is by itself a neural network).
4.3 Guessing probabilities from "Neuro-Logic"
Many decisions are taken in the presence of uncertainty. These decisions usually rely
heavily on modal operations. The decision adopted out of a group of choices must
take into account a set of evaluations about possibilities. For instance, a gambler
may say, 'I play roulette if I believe it is possible to win', and by gambling the risk-
taker intentionally ignores the mathematical odds against winning. Considering this
propensity, it is natural that Holmes' maxim was embedded in its modal frame.
20
However, Conan Doyle enlarges and refines this frame including a subjective
probabilistic evaluation: "whatever remains, however improbable". Probability is a
well defined mathematical construct (see Feller 1968) and an interesting open
problem is the accuracy of subjective estimations of a probability (for an important
analysis of this point, see Pearl 2000). Nevertheless, we are continuously guessing
probabilities, using a lot of accessible databases that help us to roughly estimate
frequencies. Probability as a mathematical construct is one of the basis of the
Bayesian explanation for Holmes' maxim sketched in Section 2, where the a priori
probabilities involved are subjective guesses. How to deal with such probabilistic
guesses in a logical theory? Many theoretical approaches connecting logic and
probability have been published (e.g. Boole 1854, Keynes 1921) and these approaches
have been connected with the problems of plausible reasoning (in particular by
Polya 1990).
The neural models of logical operators have the potentiality to permit two different
ways for the computation of the logical modalities. One way has been described in
Section 4.2, with uncertainty "conceptualized" by a specific vector h. The other way
is based on a recursive approach to logical modalities. We are going to adopt the
vector-logic
formalism to establish a connection with cognitive estimated
probabilities and the recursive modal logic formalism. In the framework of this
"neuro-logic", a way of introducing a probability guess in the logical formalism is to
assign a numerical weight to the truth-value of an uncertain proposition. For
instance, somebody can enunciate the proposition "giving the clouds I am seeing, and
the direction of the wind, I can forecast rain for the next two hours" and based on my
own experience about weather (i.e.: screening the databases installed in my memory
modules), I can establish that such a proposition has an 80 % of probability of being
true. This assignment (obviously not a probabilistic measure but a conjecture) in
some way touches probability theory because it implies that the complementary
21
situation (no rain) has a conjectural probability of 20 %. In the framework of the
vector formalism that procedure can be modeled assuming that
Val(Q) =
a + b
s
n
,
,
a +b =
[0,1] ,
1
, (30)
s and n being the vector truth-values defined in Section 4.2.
Let us recall how inside the algebraic logic, recursive processes based on disjunction
and conjunction (see Blanché 1968) can define the classical modalities. In this
abstract approach, it is assumed the existence of an infinite set Q of propositions
iQ .
This set can be mapped on a set of binary evaluations
(
)Q◊
. The proposition "Q is possible",
}f,t{)Q(Val
=
q
i
i
q,q{
1
,...,
q
n
2
,...}
, with
, can be symbolically
represented by
◊
(Q)
=
q
1
q
2
.... q
n
....
,
that is an informal representation of the recursive process
◊
+
n 1
=
(Q) q
+
n 1
(cid:218) ◊
(Q)
n
n = 1, 2, ...
(31)
with
fin
=
◊
(Q) q
1
1
. The symbol (cid:218)
. In this process, the possibility
◊
(Q)
is the limit of
n◊
(Q)
for
represents the dyadic disjunction.
In this formalism, the necessity is defined as follows. The proposition "Q is
necessary",
□
(Q)
, can be represented using a concatenated conjunction (cid:217)
,
(Q)
□
=
q
1
q
2
.... q
n
....
,
22
a
b
(cid:218)
(cid:218)
(cid:218)
(cid:218)
¥
(cid:217)
(cid:217)
(cid:217)
(cid:217)
or by the limit for
fin
of the recursive process
, n = 1, 2, ... (32)
□
+
n 1
=
(Q) q
+
n 1
(Q)
□
n
with
□
1=
(Q) q
1
.
It is important to mention that these recursive processes reported in Mizraji (2008)
using a matrix-vector formalism, were originally implemented by McCulloch and
Pitts (1943) with formal neurons capable of executing OR and AND. Obviously, in
the context of any neural model that pretends to describe a physical reality,
recursions become finite. The formalism described in the previous section allows to
represent these recursions using the conjunction and disjunction matrices (17) and
(18) . Let Nec (Q) describe a neural system that recursively evaluates possibilities
exploring the information stored in a finite set {
}iQ of propositions evaluated by
vector truth-values
u
i =
)Q(Val
i
. The matrix version of this process is as follows:
Pos
+
1n
=
u(D]u[
2
+
1n
Pos
n
])u[
(33)
with
]u[
Pos
1
=
u
1
, in general being
a=
u
i
-+
1(
s
i
,n)
i
i
]1,0[
. If we project this
recursive process on vector s (the projection of a vector u on s is given by the scalar
usT ) we obtain some interesting results. The scalar projection of
lim
n
Pos
]u[
n
is given by the product
product
=
]u[Pos
T
s
1(1]u[Pos
-=
1)(
1
1)(
2
…)
3
. (34)
For a large number of data stored in a memory, this product can be approximated by
a quasi-infinite recursion, and interpreted as a geometrical mean expression:
23
¥
(cid:217)
a
a
¥
fi
a
-
a
-
a
-
. (35)
T
s Pos [u]
lim [1 (1
n
n
) ]
Consequently
T
s Pos[u]
a =
0 iff
»
1 iff
0
0
.
For the necessity operation, the matrix version is:
Nec
+
1n
u(C]u[
2
=
+
1n
Nec
n
])u[
(36)
with
]u[Nec
1
=
u
1
(
u
i
a=
-+
1(
s
i
,n)
i
i
). Using the previous quasi-infinite
]1,0[
=
lim
n
approximation, the scalar projection of
]u[Nec
Nec
]u[
n
gives the product
T
s
]u[Nec
a=
…321
(37)
Note that, due to the fact that u is the result of a recursive process, we have the
following result:
T
s Nec[N u]
2
=
(1
)(1
1
= b
)(1
2
)...
3
1 2 3
…
(38)
This recursion can be averaged using the limit geometrical mean
T
s Nec[u]
n
lim
n
. (39)
24
fi
¥
»
-
-
a
a
„
a
a
¥
fi
a
a
-
a
-
a
-
a
b
b
fi
¥
»
a
that gives the classical scalar expression for the necessity
T
s Nec[u]
0 iff
»
1 iff
a =
1
1
.
Note that these modal operators calculated from two valued operators, but with
"probabilistic" truth-values given by equation (30), satisfy the theorem:
2=
N]u[Pos
]uN[Nec
2
. (40)
A version of equation (8) is given by
N
2
]u[Pos
=
]uN[Nec
2
=
s
, (41)
that gives us another formal "neuro-logical" version of Holmes' maxim, but now with
the possibility of establishing contact with subjective probabilities (obviously only
estimated probabilities from the mathematical point of view).
Remark that, in the scheme of this section, a research process implies the existence
of a fact F that must be explained from a potential set of causes
. We can
}iQQ =
{
assume that a priori highly improbable causes do not belong to Q; consequently,
they belong to the complement or negation of Q, that we represent symbolically by
Negation Q. Remark that the elements of the set Q are not necessarily unlinked nor
exhaustive: (a) they can be linked and (b) they are not exhaustive. Concerning (a),
they can be linked because if, for instance,
3Q represents the name of a possible
guilty of a crime (say Jean) then
7Q (say Jacques) can be the name of the same
criminal (Jean-Jacques) or the name of his associate, both corresponding to the
25
a
„
searched cause. Concerning (b), we mention that in general we may expect that
∑
i
. The assumption here is that
1)Q(obPr
i
)Q(Val
i
a=
b+
s
i
,n
i
,
i
i
,]1,0[
b+a
i
=
1
i
. (42)
Hence,
)Q(obPr
i
a=
1
i
. This is the only assumption concerning probabilities. It
implies a kind of conservation inside the judgment (conservation mapped in the
complementarity of assigned probabilities for the two canonical truth-values s and n,
and in turn supported by the complementarity of the associated conceptual sets
(Mizraji and Lin 2011)). In this work we will leave as an open problem the link
between these "cognitive probabilities" and the formal probabilities involved in the
Bayesian treatment illustrated in Section 2.
Using equation (41) we can now establish a modal-probabilistic version of Holmes'
maxim. Let us rewrite this equation as follows:
N
2
]e[Pos
=
]eN[Nec
2
, (43)
with e being a composed event described by a vector. This vector e emerges from the
recursive process triggered by the search of evidences (evidences obtained from
previous knowledge of the researcher or from fresh data coming from the external
reality). We define the link between modalities and the corresponding binary
probabilities as follows:
Pr ob(u)
=
T
s Nec[u]
For the modal evaluation describing Holmes' maxim, we have
26
„
b
a
£
2
N
]e[Pos
]eN[Nec
2
⇒=
s
⇒=
s
0
)e(obPr
.1)eN(obPr
=
=
2
This is the final point of our neural version of Holmes' old maxim. According to
equation (38), we have
s Nec[N e] =
T
2
(1
)(1
1
=
)(1
2
)...
3
1 2 3
… ,
then
Pr ob(N e) 1= implies
2
=
,1
i
i
. This very simple conclusion indicates that,
b were very small, the a posteriori result, after a
even if the pre-judged probabilities *
i
ie to be impossible, induces the necessity of the
research process showing the events
complementary events and rise their probability to
i =
1
.
5. Discussion
From the very beginning of the mathematical theory of neuronal networks, the
relation between logical reasoning and its neuronal bases has been a subject of
primordial interest. The pioneering work by McCulloch and Pitts (1943) shows how
some basic logical gates as NOT, AND and OR can be represented on the basis of
binary neuronal elements by adjusting the thresholds and the synaptic weights of
these formal neurons; other gates (e.g.: XOR) required a small network to be
computed. In addition, McCulloch and Pitts showed that modalities ◊ and □ could
be computed with recurrent networks based on the neuronal gates OR and AND
respectively. These McCulloch-Pitts "logical neurons" had strong influence in the
very important works published by Kleene (1951) and von Neumann (1956). Recently,
the investigation of the links between reasoning and neural models acquired a new
impulse and new perspectives promoted in part from the advances in cognitive
sciences and the interest in neural computations. Between these new approaches, we
want to mention the panoramic contribution of Stenning and van Lambalgen (2008,
in particular Chapter 8) linking logic, cognition and biological evolution, and the
27
-
a
-
a
-
a
b
b
b
"
b
b
investigations of d'Avila Garcez, Lamb and Gabbay concerning the representation by
artificial neural networks of a variety of symbolic logical processes, including
modalities and temporal logic (d'Avila Garcez 2007, d'Avila Garcez, Lamb and
Gabbay 2009).
In the present work, we opted to use models based on Anderson-Kohonen matrix
memories, where the
logical computations are performed by networks of
interconnected matrix modules (Mizraji and Lin 2011). In turn, each modular unit is
composed by a large set of interconnected neurons represented by high dimensional
matrices and can be programmed or instructed via a learning algorithm (eg:
Widrow-Hoff algorithm, see Anderson 1995, chapter 9). This kind of matrix models
present many aspects that enhance their biological plausibility (a point analyzed
with detail in Anderson 1995), including their reliability in the presence of failures,
their ability to create statistical averages from their learned inputs (interpretable as a
source of conceptualizations, see Cooper 1974), and also their capacity to sustain
logical gates and to display many-valued logics in the presence of uncertain data. It
is now important to challenge these models with interesting problems linked with
natural neural computations ("natural" in the sense that an adult human brain with
a normal education and linguistic performances can do these computations). The
Holmes' old maxim is one of these challenges and we think that these models
provide interesting answers.
Using the neural-inspired logical formalism (Mizraji 2008 and Mizraji and Lin 2011),
we showed in equation (29) how Holmes' maxim is represented using modal logical
operators and how truth-evaluations are expressed with a matrix-vector formalism.
As was previously mentioned, these logical operators are interpretable as memory
modules that integrate a modular network (in fact a network of networks, because
each one of the logical modules is by itself a neural network). The final point of our
argument, equation (43), involves the matrices
,N2
,Pos
Nec
and the vector e. The
28
matrix operations implement the abstract algebraic relationships constructed in
order to formalize a basic modal logic, using the Boolean theoretical framework.
Clearly, these operators are simplified models of biological neural devices. In this
formalism, vectors are the 'stuff cognitive decisions are made on' inside the neural
realm. Obviously, if mathematical approach helps us to explain our cognitive
acceptability of Holmes' maxim, but on the other hand mathematics itself (being a
cognitive construction) remains unexplained, then our eventual explanation can be
considered incomplete and provisional, a flaw that we accept as a transitory step in
our understanding of this kind of problem. We want to mention that many of the
modern works concerning the neural representation of reasoning, emphasizes the
need of nonmonotonic logics capable to deal with adaptive planning (see Stenning
and van Lambalgen 2008). In Section 4.3 we saw how in our neural representation of
Holmes' maxim, all the process –even if it is guided by the framework of modal logic
truth-functionality – is dependent on the adaptive evaluations of probabilities. We
can ask if, formally, this process represents a class of nonmonotonic reasoning (as
described for instance in d'Avila Garcez, Lamb and Gabbay 2009, Chapter 2) but we
do not have still a clear answer. Surely the possibility to formalize this process as a
nonmonotonic logic is an issue that deserves further exploration.
The argument developed in this work suggests that we accept Holmes' maxim as
true because our brains are capable to activate neural modules able to perform
modal logical computations. Our formal approach is neutral in the debate about the
"Nature versus Nurture" origin of our logical abilities: these neural logical modules
can be the result of a genetically coded ontogenetic process, or, on the contrary, they
can be the result of a learning process occurring in a particular cultural environment
(for a discussion of this point, see Mizraji and Lin 2011). We argue that the
spontaneous computation involved in the understanding of Holmes' maxim is only
one example among many others; all of them emerging from a natural biological
design that obviously includes our cognitive brains, together with our sensory and
29
motor systems. In fact, language uses some computational codes that trigger
complex cognitive procedures. For instance, a preposition like "in" can induce the
brain to represent a complex spatial relationship between an object and a container.
Prepositions and logical words are crucial linguistic constructions that can act as
passwords that give access to sophisticated neurocomputational operations.
30
APPENDIX 1.
Sherlock Holmes' old maxim.
We transcribe three well-known versions of Holmes' maxim (Doyle, Penguin Edition
1981). In "The Adventure of the Beryl Coronet", Holmes says: "It is an old maxim of
mine that when you have excluded the impossible, whatever remains, however
improbable, must be the truth".
Another version is in the novel "The Sign of Four" where Holmes says to Dr. Watson:
"How often I said to you that when you have eliminated the impossible, whatever
remains, however improbable, must be the truth?".
The last version we want to reproduce here is the one included in "The Adventure of
the Bruce-Partington Plans", where Holmes comments: "We must fall back upon the
old maxim that when all other contingencies fail, whatever remains, however
improbable, must be the truth".
APPENDIX 2
Kronecker Product (Graham 1981).
For the matrices
=
A a
ij
m n
R
and
B
=
b
ij
p q
R
, the Kronecker product is a
matrix
A B R
mp nq
defined by
=
A B
a B
ij
.
Some important properties of the Kronecker product are the following:
A1. (A A ')
+
A2.
(A B)
=
T
+
(B B') A B A B' A ' B A ' B'
+
+
+
=
A3. (A B)(A ' B')
T
A
=
T
B
(AA ')
(BB')
31
·
·
·
Acknowledgments: I thank Julio A. Hernández, Juan C. Valle-Lisboa and Andrés
Pomi for discussions and comments. This work was partially supported by
PEDECIBA (Uruguay), CSIC, (UdelaR).
References
Anderson, James A. 1972. A simple neural network generating an interactive
memory. Mathematical Biosciences 14: 197-220.
Anderson, James A. 1995. An Introduction to Neural Networks. Cambridge, USA: The
MIT Press.
Arbib, Michael A. (Ed.) 1995. The Handbook of Brain Theory and Neural Networks.
Cambridge, USA: The MIT Press.
Aristotle. 350 BC. On Interpretation, Translated by E. M. Edghill, Provided by The
Internet Classics Archive, Available:
http://classics.mit.edu//Aristotle/interpretation.html. Accessed: 28 October 2012
beim Graben, Peter, Dimitry Pinotsis, Douglas Saddy, Roland Potthast. 2008.
Language processing with dynamic fields. Cognitive Neurodynamics 2: 79-88.
Binazzi, Alberto (2012) Cognizione logica e modelli mentali. Studi sulla formazione 1:
69-84, ISSN 2036-6981 (online).Firenze University Press.
Available:
Accessed: 4 November 2013
Blanché Robert. 1968. Introduction à la Logique Contemporaine. Paris: Armand Colin.
Blutner, Reinhard. 2004. Nonmonotonic inferences and neural networks. Synthese
142: 143-174.
Boole, George. 1847. The Mathematical Analysis of Logic. Cambridge: Macmillan.
Boole, George. 1854. An Investigation of the Laws of Thought, on which are Founded
the Theories of Logic and Probabilities. London; New York: Macmillan (Dover
Reedition, 1958).
Cairns-Smith, Alexander G. 1990. Seven Clues to the Origin of Life: A Scientific
Detective Story. Cambridge, UK: Cambridge University Press.
http://fupress.net/index.php/sf/article/viewFile/11671/11092#page=69.
32
Cooper, Leon N. 1974. A possible organization of animal memory and learning.
Proceedings of the Nobel Symposium on Collective Properties of Physical Systems.
Sweden: Aspensagarden. pp 252-264
d'Avila Garcez, Artur S. 2007. Advances in neural-symbolic learning systems: Modal
and temporal reasoning. Studies in Computational Intelligence 77: 265-282.
d'Avila Garcez, Artur S, Luís C. Lamb, Dov M. Gabbay. 2009. Neural-Symbolic
Cognitive Reasoning. Berlin: Springer.
Doyle, Arthur Conan. 1981. The Penguin Complete Sherlock Holmes. London:
Penguin.
Eco, Umberto, Thomas A. Sebeok Eds. 1988. The Sign of Three: Dupin, Holmes,
Peirce. Bloomington: Indiana University Press.
Feller, William. 1968. An Introduction to Probability Theory and Its Applications, Vol.
1. New York: Wiley.
Graham, Alexander. 1981. Kronecker Products and Matrix Calculus with Applications.
Chichester: Ellis Horwood.
Hugues, George E., Max J. Cresswell. 1972. An Introduction to Modal Logic. London:
Methuen,.
Keynes, John Maynard. 1921. A Treatise on Probability. London: Macmillan.
Kleene, Stephen C. 1951. Representation of events in nerve nets and finite automata.
The Rand Corporation, RM-704.
Kohonen, Teuvo. 1972. Correlation matrix memories. IEEE Transactions on
Computers C-21: 353-359.
Kohonen, Teuvo. 1977. Associative Memory: A System-Theoretical Approach. New
York: Springer.
Łukasiewicz, Jan. 1930. Philosophical remarks on many-valued logic. Reprinted in Jan
Łukasiewicz, Selected Works, edited by L. Borkowski, pp.153-178, Amsterdam: North-
Holland, 1980.
McCulloch, Warren S., Walter Pitts. 1943. A logical calculus of the ideas immanent in
nervous activity. Bulletin of Mathematical Biophysics 5: 115-133.
33
Mizraji, Eduardo. 2008. Vector Logic: A Natural Algebraic Representation of the
Fundamental Logical Gates. Journal of Logic and Computation 18: 97-121
Mizraji, Eduardo, Juan Lin. 2011. Logic in a Dynamic Brain. Bulletin of Mathematical
Biology 73: 373-397
Pearl, Judah. 2000. Causality. Cambridge, UK: Cambridge University Press.
Polya, George. 1990. Mathematics and Plausible Reasoning, Volume 2: Patterns of
Plausible Inference. Princeton: Princeton University Press. .
Smolensky, Paul. 1990. Tensor product variable binding and the representation of
symbolic structures in connectionist systems. Artificial Intelligence 46: 159-216.
Stenning, Keith, Michiel van Lambalgen. 2008. Human Reasoning and Cognitive
Science. Cambridge, USA: The MIT Press,.
von Neumann, John. 1956. Probabilistic logics and the synthesis of reliable organisms
from unreliable components. Automata Studies, (C. E. Shannon and J. McCarthy
Eds). Princeton: Princeton University Press.
34
|
1812.03122 | 2 | 1812 | 2019-01-25T14:58:38 | Trade-off between sensitivity and selectivity in olfactory receptor neuron | [
"q-bio.NC",
"physics.bio-ph"
] | It was observed before that due to convergence in the olfactory system a possible amplification can be as large as the degree of convergence. This is in the case when a single impulse from the converging inputs is enough to trigger the secondary neuron. On the other hand, if a number of impulses are required for triggering, a gain in discriminating ability may be obtained along with decrease in sensitivity gained due to the convergence. We discuss this trade-off in terms of concrete estimates using olfactory sensory neuron and the set of its receptor proteins as an example of system with convergence. | q-bio.NC | q-bio |
Trade-off between sensitivity and selectivity in olfactory
receptor neuron
Alexander Vidybida1,a)
1Bogolyubov Institute for Theoretical Physics, Metrologichne str. 14-B, 03680 Kyiv, Ukraine
a)Corresponding author: [email protected]
URL: http://vidybida.kiev.ua
Abstract. It was observed before that due to convergence in the olfactory system a possible amplification can be as large as the
degree of convergence. This is in the case when a single impulse from the converging inputs is enough to trigger the secondary
neuron. On the other hand, if a number of impulses are required for triggering, a gain in discriminating ability may be obtained
along with decrease in sensitivity gained due to the convergence. We discuss this trade-off in terms of concrete estimates using
olfactory sensory neuron and the set of its receptor proteins as an example of system with convergence.
INTRODUCTION
In a neural system, a convergent organization of inter-neuronal connections is a typical pattern. E.g., the degree of
convergence observed for rat hippocampal pyramidal cells is about 12 000, [1]. In the olfactory system of mouse, see
Fig. 1, about 5 000 olfactory receptor neurons (ORN) send their output spikes to a single mitral cell, [2]. Our purpose
is to figure out what role the convergence may have in forming sensitivity and selectivity in a sensory system. For this
purpose we use olfactory system as an example.
FIGURE 1. Schematic example of convergent structure in olfactory system. Here as much as several thousands [3] (the concrete
number is species-dependent) of olfactory sensor neurons may converge through a single glomerulus onto a single mitral cell.
PROBABILISTIC DESCRIPTION
Real neurons are noisy [4]. This has a consequence for a sensory system under weak stimulation. Namely, if a stimulus
has low level, the response to it becomes probabilistic. The dose-response dependence becomes probabilistic as well:
the higher is the dose, the higher is the response probability, [5, 6]. The sensitivity is considered to be higher when the
same stimulus evokes a response (output spike) with higher probability.
mitralcellglomerulusORNcilium SENSITIVITY GAIN DUE TO CONVERGENCE
It was observed, [7], that threshold odor concentration, which is able to evoke a response at the level of whole animal
can be 10-100 times lower than that at the level of ORN. This might be due to convergence in the olfactory sensory
pathways. At the stage of ORN-mitral cell communication, the explanation has been proposed by W. van Drongelen
et. al., [5, 6]. Namely, if under a certain stimulation (odor concentration) an ORN output activity can be modeled as
Poisson stochastic process with intensity λ, then compound input of N ORN outputs into a single mitral cell can be
modeled as Poisson process of intensity Nλ. The probability that a single ORN emits a spike during time interval T
is PORN = 1 − e−λT ≈ λT. The latter approximation is valid for a short enough T. Expect that a single input spike
is able to trigger the secondary neuron. Then the probability that the secondary neuron emits a spike during T is
Pm = 1 − e−NλT ≈ NλT. In this case we have sensitivity gain equal to NλT
= N. Actually, a real mitral cell requires
more than a single input impulse for triggering. Denote the required for triggering number of input impulses as N0. It
can be proven that the mitral cell output rate Rm in this case satisfies the following estimate:
λT
Rm ≤ N
N0
λ.
The equality holds here if the mitral cell is modeled as perfect integrator. The corresponding sensitivity gain Gn is as
follows:
Gn =
.
(1)
Rm
λ
≤ N
N0
SELECTIVITY GAIN INSIDE THE OLFACTORY RECEPTOR NEURON
The olfactory system has the remarkable capacity to discriminate among a wide range of odor molecules. This begins
with the ORN, which performs the task of converting information contained in the odor molecules into information
contained in membrane signals and neural space, [8]. Discriminating ability of ORN starts to build up already at the
level of individual receptor proteins. The primary act of odor perception happens when odor molecule physically
contacts with a receptor protein integrated in the ORN's membrane. During this contact, the molecule can be bound
to the receptor with some probability p. The probability p depends on the chemical nature of the molecule, which
determines its affinity to a given receptor. It is namely due to this dependence that an ORN is able to discriminate
between different odors.
Selectivity of Single Receptor Protein
The discriminating ability of a receptor protein can be defined as follows. Let during two separate experiments two
different odors O and O' are presented to a given receptor or a set of them at the same concentration. Denote as p
and p(cid:48) the probability to find a receptor bound with odor O and O'. If p (cid:44) p(cid:48) then we say that this receptor is able
to discriminate between O and O'. Otherwise, it cannot. The same is with the ORN expressing this receptor. Suppose
that p > p(cid:48). Then the discriminating ability at the level of single receptor protein can be characterized numerically as
µ = log
p
p(cid:48) .
Selectivity of ORN
(2)
In the ORN expressing concrete receptor protein, one may observe convergence of signals from individual receptor
proteins onto the ORN's interior, see Fig. 2, and farther onto the axonal hillock where it is decided whether to fire or
not. The degree of convergence is characterized by the total number N of receptors per neuron. The firing threshold can
be expressed as the minimal number N0 of receptors which must be bound with odor in order to ensure depolarization
necessary for triggering. Since odor binding-releasing is driven by the Brownian motion, the firing threshold will be
achieved irregularly and this will result in irregular/random spiking of ORN.
In a simplified model of ORN, see Fig. 2, the ORN's discriminating ability between the O and O' can be defined
as follows. Let the ORN during two separate experiments is exposed to O and O' at equal concentrations. Denote F
FIGURE 2. Simplified model of olfactory receptor neuron as single cilium with firing threshold. N -- is the total number of
receptors, N0 -- is the threshold number. The neuron starts firing if the number of bound receptors exceeds N0.
and F(cid:48) the ORN's mean firing rate if O and O' are presented. Then the ORN's ability δ to discriminate between O and
O' can be defined as follows
It was shown, [9], that in situation when the dose-response dependence is of threshold type and probabilistic,
a considerable selectivity gain is possible. For the model ORN, the following mathematically rigorous estimate has
been obtained, [10, 11, 12]:
δ = log
F
F(cid:48) .
p0 − p
1 − p
δ >
N µ,
p0 =
N0
N .
(3)
(4)
The numbers N and N0 can be high. E.g., N = 2 500 000, N0 = 250 for moth, (J.-P. Rospars, private communication).
For frog, N = 25 000, N0 = 35, [13]. Estimate (4) suggests that ORN's selectivity can be much higher than that
of its receptor proteins, provided p0 − p > 0. The latter is achieved if the odors O and O' are presented at sub-
threshold concentration. Namely, the mean number of bound receptors is below the threshold one. In this case, the
firing threshold is achieved due to random fluctuations, and this process appeared to be more selective than random
binding-releasing at a single receptor. Conclusion (4) has been checked by means of direct numerical simulation of
odor binding-releasing in a set of receptors, [14].
TRADE-OFF BETWEEN SENSITIVITY AND SELECTIVITY
Define the selectivity gain, Gl, as follows
Gl = δ/µ.
Then from (1) and (4) one obtains the following estimate
1
Gn
− p
1 − p
N.
Gl >
(5)
(6)
The right-hand side of inequality (6) is the decreasing function of Gn. Therefore, the better is the sensitivity gain the
less optimistic are estimates (4), (6) for selectivity. Consider situation when all N receptors should be bound to ensure
triggering (N0 = N). In this case Gn = 1 (no sensitivity gain) and we have from (6)
Gl > N,
which is quite optimistic for selectivity.
The estimate (6) depends on the probability p that a receptor is bound with odor O. This probability can be calcu-
lated based on association-dissociation rate constants k+, k− between the odor and receptor, and odor's concentration,
c:
p =
c k+
c k+ + k− .
If concentration c is very low, then p is very low as well and (6) turns into the following:
Gl Gn > N,
which again demonstrates the trade-off between selectivity and sensitivity. Finally, if concentration c increases drawing
p to the 1
Gn
, then the estimate (6) turns into Gl > 0, which promises nothing as regards selectivity gain.
✟✟✟❜❜❜✟✟✟❜❜❜✟✟✟❜❜❜✟✟✟❜❜❜✟✟✟✟✟✟❜❜❜❜❜❜✟✟✟✟✟✟❜❜❜❜❜❜✟✟✟❜❜❜✟✟✟❜❜❜✟✟✟✟✟✟✟✟✟❜❜❜❜❜❜❜❜❜✟✟✟✟✟✟✟✟✟❜❜❜❜❜❜❜❜❜❆❆❆✔✔✔❆❆❆✔✔✔receptorproteins✏✏✏✏✶✁✁✕❍❍❍❨❳❳❳❳❳❳❳②N0<N CONCLUSIONS AND DISCUSSION
We have considered here a set of identical receptor proteins belonging to a single ORN as converging on its interior and
made some conclusions about sensitivity and selectivity gain in the ORN itself, see Equations (1), (4), (6). The main
conclusion is that due to convergence one may have either high selectivity or high sensitivity. Increasing one of them
results in decreasing the other one. While the sensitivity gain does not depend on the stimulus intensity, the selectivity
decreases with increasing concentration, see Equations (4), (6). This conforms with experimental observations, [15,
16]. Concentration at which a selectivity gain might be expected, must be sub-threshold: the firing threshold can be
achieved due to fluctuations, but the time-averaged concentration is less than the threshold one. This strong limitation
on the concentration range may cast doubt on a possibility that the mechanism discussed here could operate in a
real biological system. In this connection I would like to say that the suitable concentration depends on the threshold
N0, which itself could be adjustable. The evident mechanisms for this are adaptation and inhibition of the ORN.
Interesting, that in the olfactory system, ORN's output is subjected to feedback presynaptic inhibition, [17]. For the
secondary neurons, this inhibition acts similarly as elevation of ORN's firing threshold.
Conclusions made here about the trade-off between sensitivity and selectivity and the selectivity gain itself are
obtained by means of mathematical analysis of the binding-releasing stochastic process, [12].
As to my knowledge, no experimental attempt was made to compare the ORN selectivity with that of its receptor
proteins. This is not surprising because measuring selectivity of a receptor protein belongs to chemistry, while dis-
criminating ability of ORN -- to sensory biology. The selectivity gain predicted in [12] and here is made possible
due to the threshold manner of the ORN response to converging inputs from its receptors. This offers an experimen-
tal procedure for checking the prediction. Namely, the selectivity of individual receptor protein (µ) can be estimated
while measuring the ORN output in the form of receptor potential, with spike-triggering mechanism blocked. The
ORN selectivity (δ) can be estimated through mean firing rate as defined in (3).
ACKNOWLEDGMENTS
This research was supported by theme grant of department of physics and astronomy of NAS Ukraine "Dynamic
formation of spatial uniform structures in many-body system" PK0118U003535.
Also, I would like to thank to the 30th Annual Biophysics Congress (International), organized by Turkish Bio-
physical society, and to Prof. Mehmet Can Akyolcu personally for organizing a nice Congress and inviting me.
REFERENCES
[1]
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
P. Andersen, M. Raastad, and J. F. Storm, in Cold Spring Harbor Symposia on Quantitative Biology (Cold
Spring Harbor Laboratory Press, 1990), pp. 81 -- 86.
K. J. Ressler, S. L. Sullivan, and L. B. Buck, Cell 79, 1245 -- 1255 (1994).
L. B. Buck, Cell 100, 611 -- 618 (2000).
E. T. Rolls and G. Deco, The Noisy Brain: Stochastic Dynamics as a Principle of Brain Function (OXFORD
UNIVERSITY PRESS, 2010).
v. W. Drongelen, The Journal of Physiology 277, 423 -- 435 (1978).
v. W. Drongelen, A. Holley, and K. B. Døving, Journal of theoretical Biology 71, 39 -- 48 (1978).
K.-E. Kaissling and E. Priesner, Naturwissenschaften 57, 23 -- 28 (1970).
G. M. Shepherd, Neuron 13, 771 -- 790 (1994).
A. K. Vidybida, Journal of theoretical Biology 152, 159 -- 164 (1991).
A. K. Vidybida, Biological Cybernetics 81, 469 -- 473 (1999).
A. K. Vidybida, BioSystems 58, 125 -- 132 (2000).
A. K. Vidybida, Sensors.and.Actuators.A:.Physical. 107, 233 -- 237 (2003).
V. Bhandawat, J. Reisert, and K.-W. Yau, Proceedings of the National Academy of Sciences 107, p. 18682
(2010).
A. K. Vidybida, A. S. Usenko, and J.-P. Rospars, Journal of Biological Systems 16, 531 -- 545 (2008).
A. Duchamp and G. Sicard, Chem.Senses 8, 355 -- 366 (1984).
P. Duchamp-Viret, A. Duchamp, and G. Sicard, Brain Research 517, 256 -- 262 (1990).
J. P. McGann, Chemical.Senses 38, 459 -- 474 (2013).
|
1609.04245 | 2 | 1609 | 2016-12-13T17:59:01 | Non-random network connectivity comes in pairs | [
"q-bio.NC"
] | Overrepresentation of bidirectional connections in local cortical networks has been repeatedly reported and is in the focus of the ongoing discussion of non-random connectivity. Here we show in a brief mathematical analysis that in a network in which connection probabilities are symmetric in pairs, $P_{ij} = P_{ji}$, the occurrence of bidirectional connections and non-random structures are inherently linked; an overabundance of reciprocally connected pairs emerges necessarily when the network structure deviates from a random network in any form. | q-bio.NC | q-bio | Non-random network connectivity comes in pairs
Felix Z. Hoffmann1,2,* and Jochen Triesch1
1Frankfurt Institute for Advanced Studies (FIAS), Johann Wolfgang Goethe University,
2International Max Planck Research School for Neural Circuits, Max Planck Institute for
Brain Research, Frankfurt am Main, Germany
Frankfurt am Main, Germany
*Email: [email protected]
Overrepresentation of bidirectional connections in local cortical networks
has been repeatedly reported and is in the focus of the ongoing discus-
sion of non-random connectivity. Here we show in a brief mathematical
analysis that in a network in which connection probabilities are symmet-
ric in pairs, Pij = Pji, the occurrence of bidirectional connections and
non-random structures are inherently linked; an overabundance of recip-
rocally connected pairs emerges necessarily when the network structure
deviates from a random network in any form.
Introduction
Increasing evidence shows that cortical microcircuitry is highly structured [1,
2]. Not every connection is equally likely to be established, rather some pairs
of neurons are more likely connected than others. In this context, the relative
occurrence of bidirectionally connected pairs has been of particular interest.
Using data obtained from paired whole-cell recordings in cortical slices, the
amount of bidirectionally connected pairs was compared to the number of
reciprocal pairs as one would expect in a random network with the same
overall connection probability. Connectivity of layer 5 pyramidal neurons in
the rat visual cortex [1] and somatosensory cortex [3, 2] was shown to have a
significantly stronger reciprocity than expected.
The prevalence of bidirectional connectivity has since been established as an
important indicator for the non-randomness of a network [4, 5]. However,
the exact relationship between non-randomness and relative reciprocity has
not been explained. Here, we model cortical circuitry as random networks
in which each possible connection has a separate probability to exist. Using
this model we're able to show that any non-random connectivity, expressed as
higher connection probabilities in some edges and lower probabilities in others,
necessarily induces a relative overrepresentation of bidirectional connections
1
6
1
0
2
c
e
D
3
1
]
.
C
N
o
i
b
-
q
[
2
v
5
4
2
4
0
.
9
0
6
1
:
v
i
X
r
a
as long as connection probabilities remain symmetric within pairs. Quanti-
tatively, we analyze reciprocity in networks with a discrete and a continuous
distribution in connection probabilities to demonstrate that a relative occur-
rence of bidirectional connections as reported from experimental studies can
be easily obtained in these models.
Results
The emergence of non-random connectivity patterns can be modeled by as-
signing each possible connection in a random graph a separate probability to
exist. In such a model some connections are more likely to be realized than
others, allowing for the encoding of patterns within the specific probabilities
of each connection.
In the limiting case each connection either exists or is
absent with certainty, representing a blueprint for the network architecture.
To analyze the effect of non-random structures within a network, specifically
on the statistics of bidirectionally connected pairs found in the network, we
consider a random graph model of N neurons in which the probability of node
i to connect to node j is modeled by a random variable Pij. For this we
assume Pij for i, j = 1, . . . , N with i (cid:54)= j to be identically distributed random
variables in [0, 1], yielding a probability of connection for each ordered pair of
nodes in the graph. Outside of pairs the random variables Pij are assumed
to be independent, that is non-equal Pij and Pkl are independent as long as
i (cid:54)= l or j (cid:54)= k. Finally, we explicitly exclude self-connections in this model
and assume at all times that i (cid:54)= j.
Given the distributions of connection probabilities, what is then the prob-
ability in this model for a randomly selected node to have a projection to
another randomly selected node? As the random variables Pij are identically
distributed, we compute this overall connection probability µ easily as the
expected value of Pij,
µ = E(Pij).
(1)
For example, if the Pij have a probability density function f with essential
support in [0, 1], we can compute the connection fraction as
(cid:90) 1
µ =
xf (x) dx.
(2)
0
In this work we are interested in the probability Pbidir of a bidirectional con-
2
nection to exist in a random pair of neurons. We determine Pbidir as the
expected value of the product of Pij and Pji,
Pbidir = E(PijPji).
(3)
The relative occurrence of such reciprocally connected pairs compares Pbidir
with the occurrence of bidirectional pairs in an Erdos-R´enyi graph, in which
each unidirectional connection is equally likely to occur with probability µ [6,
7]. The probability of a particular bidirectional connection to exist in such
a random graph is simply µ2 and we obtain the relative occurrence as the
quotient
=
Pbidir
µ2 =
E(PijPji)
E (Pij)2 .
(4)
Experimental studies in local cortical circuits of rodents have repeatedly re-
ported a relative occurrence of bidirectional connections > 1 [3, 1, 2]. To
understand in which cases such an overrepresentation occurs, we consider two
cases. In the first case, assume that connection probabilities are independently
determined in pairs as well, meaning that the random variables Pij and Pji
are independent. Then, as Pij and Pji are identically distributed,
E(PijPji) = E(Pij) E(Pji) = E(Pij)2,
(5)
and we expect to observe no overrepresentation of reciprocal connections, =
1. In the second case, assume that connection probabilities are symmetric in
pairs, Pij = Pji. In this case,
Pbidir = E(P 2
ij),
and the expected relative occurrence of reciprocal connections becomes
=
E(P 2
ij)
E (Pij)2 .
(6)
(7)
We note that now any distribution of Pij with a nonvanishing variance will
lead to a relative occurrence that deviates from the Erdos-R´enyi graph, as
Var(Pij) = E(P 2
(8)
Moreover, since x (cid:55)→ x2 is a strictly convex function, Jensen's inequality [8, 9]
yields
ij) − E (Pij)2 .
E(P 2
ij) ≥ E(Pij)2,
(9)
3
and we find that ≥ 1 in networks with symmetric connection probabilities.
Jensen's inequality further states that equality in (9), and thus = 1, holds
if and only if Pij follows a degenerate distribution, that is if all Pij take the
identical value µ. In the other case, where the Pij take on more than one value
with non-zero probability, we speak of a non-degenerate distribution.
As a central result of this study we thus find that any non-degenerate distribu-
tion of symmetric connection probabilities (Pij = Pji) necessarily induces an
overrepresentation of bidirectional connections in the network, > 1. In other
words, in a network in which both directions of connection are equally likely
within any given pair, but where some pairs are more likely to be connected
than others, the count of expected reciprocally connected pairs is strictly un-
derestimated by the statistics of an Erdos-R´enyi graph with same the overall
connection probability E(Pij) = µ.
Upper bound for
The overrepresentation of bidirectional connections in a network is maximal
when every connected pair is already a reciprocally connected pair. In terms
of the model defined above, this is the case when
E(PijPji) = E(Pij).
The relative occurrence of reciprocal connections from (4) then becomes
=
1
E(Pij)
=
1
µ
(10)
(11)
Thus, for local cortical circuits of L5 pyramidal neurons with a typical con-
nection probability of µ = 0.1 [10, 1], the network model yields a maximal
overrepresentation of = 10. While this theoretical maximum is unlikely to
exist in actual cortical networks, the precise degree of overrepresentation will
depend on the specific distribution of connection probabilities in the network.
In the following, we study two generic examples.
Two-point distribution
The simplest non-degenerate distribution of connection probabilities is a dis-
tribution that takes two values x, y with probability p and 1− p, respectively,
as illustrated in Figure 1A. This distribution may be seen as a crude approxi-
mation to the connection probabilities recently observed in visual cortex as a
4
function of the neurons' absolute difference in orientation preference, where a
"high" connection probability was reported for a difference between 0° -- 45° and
a "low" probability was seen for cells with a difference of 45° -- 90° in orientation
tuning [11]. Formally, let x, y ∈ [0, 1] with x > y and 0 < p < 1. A random
variable X follows the two-point distribution T (p, x, y) if P (X = x) = p and
P (X = y) = 1 − p.
In our network model let then the Pij be T (p, x, y) distributed. The overall
connection probability µ is
µ = E(Pij) = px + (1 − p)y.
(12)
Assume again that Pij = Pji. The relative occurrence of bidirectional connec-
tions is given by
Solving (12) for p as
=
E(P 2
ij)
µ2 =
px2 + (1 − p)y2
µ2
.
µ − y
x − y
p =
(13)
(14)
and inserting into equation (13) yields an expression for the relative overrep-
resentation depending on x, y and µ (see Supplementary Information SI1),
=
x + y
µ
− xy
µ2 .
(15)
Here we fix µ = 0.1 in accordance with the overall connection probability
found in local circuits of pyramidal cells in the rat visual cortex [1] and obtain
the relative occurrence dependent on the two connection probability values x
and y. Given x ≥ µ it follows that y ≤ µ (see Supplementary Information SI2)
and the possible values for x and y are 0.1 ≤ x ≤ 1 and 0 ≤ y ≤ 0.1. Figure 1B
shows contours of for the (x, y) pairings illustrating how different values for
the relative overrepresentation of reciprocal connections can be induced by
two-point distributed connection probabilities. We find that in such networks
higher values of are easily obtained with reasonable network configurations.
For example, a relative overrepresentation of = 4 could be achieved by a
two-point distribution of connection probabilities where one group of neuron
pairs is highly connected with probability x = 0.7, while the other group of
neuron pairs is sparsely connected with probability y = 0.05. Collectively, the
highly connected pairs then make up less than 8% of all neuron pairs, showing
that it is sufficient to have a small subgroup of highly connected neuron pairs
5
Figure 1: Relative overrepresentation of bidirectional connections in net-
works with a fraction of pairs connected with a high probability
x and the rest of the pairs connected with a low probability y.
A Diagram illustrating the targets connecting with a high chance
x (thick arrows) and targets connecting with a low probability y
(thin arrows) for a single source node (hatched). B Different
pairings of x and y can induce a high relative overrepresentation
in a network with two-point distributed connection probabili-
ties, Pij ∼ T ( µ−y
x−y , x, y), and a fixed overall connection probability
µ = 0.1. The dashed line marks an overrepresentation of bidirec-
tional connections of = 4 as observed for layer 5 pyramidal neurons
in the rat visual cortex [1].
to induce a high overrepresentation of bidirectionally connected pairs in the
network. For more densely connected networks, µ > 0.1, the effect that two
distinct connection probabilities have on the overrepresentation of reciprocal
connections is reduced (c.f. Figure S1), as one would intuitively expect from
the dependence of the maximal overrepresentation on µ in (11).
Gamma distribution
Next, we analyze the relative overrepresentation of bidirectional connections in
a network with continuously distributed connection probabilities. The gamma
distribution Γ(α, β) with probability density function
fα,β(x) =
(16)
(cid:40) 1
βαΓ(α) xα−1 e−x/β x ≥ 0
0
otherwise,
6
Pij=xPij=yA0.10.20.40.60.81.0x0.000.020.040.060.080.10y=1.1=1.3=1.6=2.0=3.0=4.0=6.0=8.0µ=0.1B allows the variation of the variance Var(X) = αβ2 of a gamma distributed
random variable X ∼ Γ(α, β), while keeping its mean E(X) = αβ constant
[12]. The exponential distribution emerges as a special case of the gamma
distribution (α = 1).
To ensure that the randomly drawn connection probabilities lie within the
interval [0, 1], we here consider a modification to the traditional gamma dis-
tribution in the form of a truncated version. Let α, β > 0. A random variable
X follows the truncated gamma distribution ΓT (α, β) if it has the probability
density function
f T
α,β(x) =
Kα,β
0
βαΓ(α) xα−1 e−x/β
1
0 ≤ x ≤ 1
otherwise.
(17)
(cid:40)
The factor Kα,β is the inverse of the cumulative probability that x ≤ 1 of the
untruncated gamma distribution,
(cid:18)(cid:90) 1
(cid:19)−1
Kα,β =
and is needed to ensure that (cid:90)
fα,β(x) dx
,
0
f T
α,β(x) dx = 1.
(18)
(19)
Consider then the above network model in which the connection probabili-
ties P T
ji . We compute the relative
overrepresentation numerically from
ij are ΓT (α, β) distributed and P T
ij = P T
µ = E(cid:0)P T
(cid:1) =
2(cid:17)
(cid:16)
ij
E
P T
ij
=
(cid:90) 1
(cid:90) 1
0
0
xf T
α,β(x) dx,
x2f T
α,β(x) dx.
(20)
(21)
Pairings of the shape parameter α and the scale parameter β were chosen such
that the overall connection probability reflects connectivity statistics in local
cortical networks, µ = 0.1 [1, 10]. Probability density functions and resulting
relative overrepresentation of reciprocal connections for four representative
α, β pairs are shown in Figure 2A. Here, β was determined to yield µ = 0.1
for the given α, following the relationship shown in Figure 2B (solid curve).
In the sparse networks we modeled, the tail of the gamma distribution is near
ij = 1 (see Figure 2A). Thus Kα,β ≈ 1 and the truncated gamma
zero at P T
distribution can be well approximated by the untruncated version. Assuming
7
Figure 2: Relative occurrence of bidirectional connections in networks with
gamma distributed connection probabilities. A Probability density
functions of the truncated gamma distribution ΓT (α, β) for different
shape parameters α and the induced relative overrepresentation
in a network with such distributed connection probabilities Pij. For
a given α, the scale parameter β was chosen such that µ = 0.1. Plot
to the right continues the density functions on a different scale. B
Contour of α, β pairings that yield an overall connection probability
of µ = 0.1. The dashed line shows the approximation β = µ
α , where
µ = 0.1. C Relative occurrence as a function of α for fixed µ = 0.1.
For α ≥ 1 this relationship is well approximated by ≈ 1 + 1
α .
connection probabilities to be standard gamma distributed, Pij ∼ Γ(α, β), we
have
E(P 2
ij) = Var(Pij) + E(Pij)2 = αβ2 + α2β2,
(22)
8
0.000.050.100.150.200.250.30connectionprobabilityPTij0246810121416fT(PTij)α=0.248,=4α=1,=2α=2,=1.5α=15,=1.070.30.50.81.00.00.10.20.30.40.5µ=0.1A0.2110100shapeparameterα0.00.20.40.60.8scaleparameterβµ=0.1β=µαB0.2110100shapeparameterα1.02.03.04.05.06.0relativeoccurrenceµ=0.11+1αC and thus
(cid:16)
(cid:16)
E
E
P T
ij
P T
ij
2(cid:17)
(cid:17)2 ≈ E(cid:0)Pij
2(cid:1)
E (Pij)2 =
=
α2β2
α2β2 +
αβ2
α2β2 = 1 +
1
α
=: .
(23)
The approximation ≈ = 1 + 1
α works well for α ≥ 1 as shown in Figure 2C.
To induce a high overrepresentation of reciprocal pairs in the network, the
gamma distribution of connection probabilities takes a highly skewed shape.
In order to obtain = 4, only 57% of pairs are expected to have a higher
connection probability than 0.01 (α = 0.248, β = 0.487). Such a situation
in which a large part of all neuron pairs have a small connection probability
while some few pairs have a high chance to be connected is likely if, e.g.,
the connection probability strongly depends on the spatial separation of the
neuron, as it was found in layer 5 excitatory circuits of the rat somatosensory
cortex [2]. Then only nearby neurons are likely to be connected, while the
larger part of more distant neurons has a low probability of connection.
Symmetry of connection probabilities in neural circuits
In neural circuits, connection probabilities that are equal within pairs but dif-
fer across the network are plausible from both an anatomical and a functional
perspective. From the anatomical point of view, the distance dependency of
connection probabilities mentioned above is a characteristic of cortical circuits
that necessarily leads to symmetric probabilities: the distance from the first
neuron's soma to the second neuron's soma is the same as the distance from
the second to the first, resulting in equal probabilities within a pair of neurons
when inter-neuron distance determines connection probabilities. Regarding
the functional perspective, connection probabilities may also depend on func-
tional properties of the cells in the network. For example, the probability
of connection of orientation tuned cells in the mouse primary visual cortex
depends on their absolute difference in orientation tuning [11, 13]. Since the
absolute difference in orientation tuning will be the same in both directions,
connection probabilities can be expected to be equal within a pair of orienta-
tion tuned cells.
However, even when connection probabilities within pairs do not match ex-
actly, an overrepresentation of reciprocal connections is still likely to be ob-
served when connection probabilities follow a non-degenerate distribution. To
see this, consider that connection probabilities Pij are distributed according
to some probability density function fPij (x). As before we assume that the Pij
9
are independent outside of pairs. In the following we also assume that i > j
without loss of generality. The expected probability of a reciprocal connection
within a pair can then be expressed as
(cid:90) 1
(cid:90) 1
0
0
(cid:90) 1
0
E(PijPji) =
xy fPij ,Pji(x, y) dx dy,
(24)
where fPij ,Pji(x, y) is the joint probability density function of Pij and Pji,
fPij ,Pji(x, y) = fPjiPij (y x)fPij (x).
(25)
In the case that Pij and Pji are independent we have fPjiPij (y x) = fPji(y)
and in the case of Pij = Pji it is fPjiPij (y x) = δ(y − x). Here we propose
a model for the conditional density function that transitions between the two
extreme cases by multiplying fPji(y) with the density function of a normal
distribution centered around x,
fPjiPij (y x) =
1
Nσ(x)
fPji(y)
√
1
2π
σ
(y−x)2
2σ2
e
,
(26)
where the additional factor Nσ(x)−1 makes sure that fPjiPij (y x) integrates
to one,
Nσ(x) =
fPji(z)
√
1
2π
σ
e
(z−x)2
2σ2 dz.
(27)
Indeed, as the standard deviation σ of the modulating normal distribution
increases fPjiPij (y x) approaches fPji(y) and in the limit σ → 0 we have
fPjiPij (y x) = δ(y − x).
lim
σ→0
(28)
In Figure 3A, conditional density functions for various σ are shown for the
truncated gamma distribution. For low values of σ the conditional density
function resembles a narrow Gaussian around x, reflecting approximately sym-
metric connection probabilities. For σ > 1 on the other hand fPjiPij (y x)
becomes virtually indistinguishable from f T
α,β(y), reflecting independence of
Pji from Pij.
Finally we employ the model to examine how the relative overrepresentation
of bidirectional connections changes with the degree of symmetry in the
connection probabilities within a pair of neurons. For this is computed as a
10
Figure 3: Relative overrepresentation of bidirectional connections is sus-
tained when connection probabilities are only approximately sym-
metric in pairs. A Illustration how the conditional density func-
tion fPjiPij (y x) of (26) transitions from equality of the random
variables Pij and Pji to independence with increasing σ. We use
fPij (y) = f T
α,β(y), with α = 0.248 and β such that E(Pij) = 0.1. For
the illustration Pij was fixed as x = 0.15. Already for σ = 1 the
conditional density function becomes visually indistinguishable from
f T
α,β(y). B Relative occurrence of reciprocally connected pairs as
a function of σ. The curves for α = 1 and α = 2 show numerical so-
lutions of (29) with fPij (y) = f T
α,β(y), where β was chosen such that
E(Pij) = 0.1. Relative reciprocal pair counts from generated net-
works following the model matched these theoretical curves (data
not shown). For α = 0.248 random variables with the respective
probability density functions were sampled and the average was
computed via (29) using the sample means. Error bars show SEM,
the curve for α = 0.248 (solid line) was fitted to the data points and
is purely for illustrative purposes.
function of σ as for a given distribution of Pij as
=
E(PijPji)
µ2
,
(29)
where the numerator is given by (24) with (26) and the overall connection
probability µ is calculated as
µ =
1
2
xfPij (x) dx +
1
2
fPij (x)
y fPjiPij (y x) dy dx.
(30)
(cid:90) 1
0
(cid:90) 1
0
(cid:90) 1
0
Figure 3B shows the change of with σ for connection probabilities Pij follow-
ing a truncated gamma distribution ΓT (α, β). For the three parameter sets
11
0.0x=0.150.30.4connectionprobabilityy0246810121416fPjiPij(yx)fPij(y)=fTα,β(y),α=0.248σ→0σ=0.025σ=0.065σ=1A0.00.10.20.30.40.50.6widthσ1.01.52.02.53.03.54.0relativeoccurrencefPij(y)=fTα,β(y)α=0.248α=1α=2B chosen we see that a strong overrepresentation of bidirectional connections is
sustained when connection probabilities are only approximately symmetric in
pairs. Furthermore, as long as Pji is at least somewhat biased to take similar
values to Pij an overrepresentation of > 1 can be observed, implying that
effects such as distance-dependency or the dependence on the absolute differ-
ence in orientation tuning of connection probabilities will tend to increase the
relative occurrence of bidirectional connections, even when other effects are
also influencing the neurons' connection probabilities.
Discussion
Experimental evidence suggests that any pair of excitatory cells within a cor-
tical column has contact points between axon and dendrite close enough to
support a synaptic connection between the cells [14, 15]. Despite this potential
"all-to-all" connectivity, only a small fraction of the contacts are realized as
functional synapses. Uncovering the underlying principles of which contact
points get utilized for synaptic transmission is crucial for our understanding
of the structure and function of the local cortical circuits in the mammalian
brain.
The emerging local networks in the rat visual and somatosensory cortex have
been shown to feature non-random structure [1, 2] and much attention was
given to bidirectionally connected neuron pairs that are occurring more often
than expected from random connectivity [5, 16, 17]. In this study we have
shown a condition under which non-random network structure and the occur-
rence of reciprocally connected pairs are inherently linked; a relative overrep-
resentation of bidirectional connections arises necessarily in networks with a
non-degenerate distribution of symmetric connection probabilities. Absence
of an overabundance of reciprocal pairs on the other hand, as for example
found in the intra-layer connectivity of the mouse C2 barrel column [4], points
towards either a truly random network or an asymmetry in the connection
probabilities.
Quantitatively, a network in which connection probabilities take on one of two
values is easily able to account for even the highest values of overrepresen-
tation reported. A network with such a two-point distribution of connection
probabilities might occur naturally, where the probability of connection de-
pends on whether a given pair of neurons shares a certain feature, for example
has a similar orientation preference or not [11].
A continuous distribution in connection probabilities on the other hand might
12
occur when pair connectivity depends on a continuous parameter such as the
inter-neuron distance or the neurons' age. We showed that networks in which
connection probabilities follow a gamma distribution can as well have a high
relative occurrence of reciprocally connected pairs, however in this case a larger
fraction of pairs remain unconnected with a very high probability.
It is likely that a combination of such effects determines the connection prob-
abilities in local cortical networks. Importantly, we showed that as long as
this probability is symmetric for pairs, any such effect that creates a non-
degenerate distribution of probabilities will cause an increase of the reciprocity
in the network.
Our results confirm the intuitive notion that reciprocity is favored in sym-
metric networks, whereas asymmetric probabilities of connection inhibit the
occurrence of bidirectionally connected pairs. Network models with symmet-
ric connectivity such as Hopfield nets generally excel at memory storage and
retrieval through fixed point attractor dynamics [18], while asymmetric net-
work models such as synfire chains are suitable for reliable signal transmission
[19, 20]. This suggests the intriguing possibility that one may be able to infer
the nature of the computations in a neural circuit based on certain statistics
of its connectivity such as the abundance of bidirectionally connected pairs.
In conclusion, the present study puts the overrepresentation of bidirectional
connections found in local cortical circuits in a new light. If connection proba-
bilities are symmetric in pairs, the overrepresentation emerges as a symptom of
any form of non-random connectivity. It is thus crucial for both future exper-
imental and modeling studies to develop a more refined view of non-random
network connectivity that goes beyond simple pair statistics. Focusing on
higher order connectivity patterns and taking into account the actual synap-
tic efficacies seem promising avenues for future research into the non-random
wiring of brain circuits.
13
Supplementary Material
The supplementary information document for references SI1 and SI2 and Fig-
ure S1 is available online at doi: 10.6084/m9.figshare.3501860. Python
code for the numerical computations is available as a GitHub repository and
was archived including the generated data at doi: https://doi.org/10.
5281/zenodo.200368. A website documenting the code is found at https:
//non-random-connectivity-comes-in-pairs.github.io/.
Acknowledgements
The authors would like to thank the anonymous reviewers for their helpful
and constructive comments on earlier versions of this article. JT is supported
by the Quandt foundation.
References
[1] S. Song, P. J. Sjostrom, M. Reigl, S. Nelson, and D. B. Chklovskii.
Highly Nonrandom Features of Synaptic Connectivity in Local Cor-
tical Circuits. In: PLoS Biol 3.3 (2005), e68. doi: 10.1371/journal.
pbio.0030068.
[2] R. Perin, T. K. Berger, and H. Markram. A Synaptic Organizing
Principle for Cortical Neuronal Groups. In: Proceedings of the Na-
tional Academy of Sciences 108.13 (2011), pp. 5419 -- 5424. doi: 10.
1073/pnas.1016051108.
[3] H. Markram, J. Lubke, M. Frotscher, A. Roth, and B. Sakmann.
Physiology and Anatomy of Synaptic Connections between Thick
Tufted Pyramidal Neurones in the Developing Rat Neocortex. In:
The Journal of Physiology 500.Pt 2 (1997), pp. 409 -- 440.
[4] S. Lefort, C. Tomm, J. C. Floyd Sarria, and C. C. H. Petersen. The
Excitatory Neuronal Network of the C2 Barrel Column in Mouse
Primary Somatosensory Cortex. In: Neuron 61.2 (2009), pp. 301 --
316. doi: 10.1016/j.neuron.2008.12.020.
14
[5] M. A. Bourjaily and P. Miller. Excitatory, Inhibitory, and Struc-
tural Plasticity Produce Correlated Connectivity in Random Net-
works Trained to Solve Paired-Stimulus Tasks. In: Frontiers in Com-
putational Neuroscience 5 (2011), p. 37. doi: 10.3389/fncom.2011.
00037.
[6] E. N. Gilbert. Random Graphs. In: The Annals of Mathematical
Statistics 30.4 (1959), pp. 1141 -- 1144. doi: 10.1214/aoms/1177706098.
[7] P. Erdos and A. R´enyi. On Random Graphs, I. In: Publicationes
Mathematicae (Debrecen) 6 (1959), pp. 290 -- 297.
[8] J. L. W. V. Jensen. Sur les fonctions convexes et les in´egalit´es entre
les valeurs moyennes. In: Acta Mathematica 30.1 (1906), pp. 175 -- 193.
doi: 10.1007/BF02418571.
[9] T. M. Cover and J. A. Thomas. Elements of Information Theory 2nd
Edition. 2 edition. Hoboken, N.J: Wiley-Interscience, 2006.
[10] A. M. Thomson, D. C. West, Y. Wang, and A. P. Bannister. Synaptic
Connections and Small Circuits Involving Excitatory and Inhibitory
Neurons in Layers 2 -- 5 of Adult Rat and Cat Neocortex: Triple In-
tracellular Recordings and Biocytin Labelling In Vitro. In: Cerebral
Cortex 12.9 (2002), pp. 936 -- 953. doi: 10.1093/cercor/12.9.936.
[11] W.-C. A. Lee, V. Bonin, M. Reed, B. J. Graham, G. Hood, et al.
Anatomy and Function of an Excitatory Network in the Visual Cor-
tex. In: Nature (2016). doi: 10.1038/nature17192.
[12] R. V. Hogg and A. T. Craig. Introduction to Mathematical Statistics.
4th ed. New York: Macmillan, 1978.
[13] H. Ko, S. B. Hofer, B. Pichler, K. A. Buchanan, P. J. Sjostrom, et al.
Functional Specificity of Local Synaptic Connections in Neocortical
Networks. In: Nature 473.7345 (2011), pp. 87 -- 91. doi: 10 . 1038 /
nature09880.
[14] A. Stepanyants, G. Tam´as, and D. B. Chklovskii. Class-Specific Fea-
tures of Neuronal Wiring. In: Neuron 43.2 (2004), pp. 251 -- 259. doi:
10.1016/j.neuron.2004.06.013.
[15] N. Kalisman, G. Silberberg, and H. Markram. The Neocortical Mi-
crocircuit as a Tabula Rasa. In: Proceedings of the National Academy
of Sciences of the United States of America 102.3 (2005), pp. 880 --
885. doi: 10.1073/pnas.0407088102.
[16] C. Clopath, L. Busing, E. Vasilaki, and W. Gerstner. Connectivity
Reflects Coding: A Model of Voltage-Based STDP with Homeostasis.
In: Nature Neuroscience 13.3 (2010), pp. 344 -- 352. doi: 10.1038/nn.
2479.
15
[17] D. Miner and J. Triesch. Plasticity-Driven Self-Organization under
Topological Constraints Accounts for Non-Random Features of Cor-
tical Synaptic Wiring. In: PLOS Computational Biology 12.2 (2016).
Ed. by O. Sporns, e1004759. doi: 10.1371/journal.pcbi.1004759.
[18] J. J. Hopfield. Neural Networks and Physical Systems with Emergent
Collective Computational Abilities. In: Proceedings of the National
Academy of Sciences of the United States of America 79.8 (1982),
pp. 2554 -- 2558.
[19] M. Abeles. Local Cortical Circuits an Electrophysiological Study. Berlin,
Heidelberg: Springer Berlin Heidelberg, 1982.
[20] M. Diesmann, M.-O. Gewaltig, and A. Aertsen. Stable Propagation of
Synchronous Spiking in Cortical Neural Networks. In: Nature 402.6761
(1999), pp. 529 -- 533. doi: 10.1038/990101.
16
|
1506.04899 | 1 | 1506 | 2015-06-16T10:17:24 | Inference based method for realignment of single trial neuronal responses | [
"q-bio.NC"
] | Neuronal responses to sensory stimuli or neuronal responses related to behaviour are often extracted by averaging neuronal activity over large number of experimental trials. Such trial-averaging is carried out to reduce noise and to reduce the influence of other signals unrelated to the corresponding stimulus or behaviour. However, if the recorded neuronal responses are jittered in time with respect to the corresponding stimulus or behaviour, averaging over trials may distort the estimation of the underlying neuronal response. Here, we present an algorithm, named dTAV algorithm, for realigning the recorded neuronal activity to an arbitrary internal trigger. Using simulated data, we show that the dTAV algorithm can reduce the jitter of neuronal responses for signal to noise ratios of 0.2 or higher, i.e. in cases where the standard deviation of the noise is up to five times larger than the neuronal response amplitude. By removing the jitter and, therefore, enabling more accurate estimation of neuronal responses, the dTAV algorithm can improve analysis and interpretation of the responses and improve the accuracy of systems relaying on asynchronous detection of events from neuronal recordings. | q-bio.NC | q-bio | Inference based method for realignment of
single trial neuronal responses
Tomislav Milekovic1, 2, 3,*, Carsten Mehring1, 2, 3
1 Bernstein Center Freiburg, University of Freiburg, Hansastr. 9A, 79104 Freiburg, Germany
2 Faculty of Biology, University of Freiburg, 79104 Freiburg, Germany
3 Department of Bioengineering and Department of Electrical and Electronic Engineering, Imperial
College London, South Kensington Campus, SW7 2AZ London, United Kingdom
* current address: Center for Neuroprosthetics and Brain Mind Institute, School of Life Sciences,
Swiss Federal Institute of Technology (EPFL), 1015 Lausanne, Switzerland
Corresponding author: Tomislav Milekovic; email: [email protected]
Abstract. Neuronal responses to sensory stimuli or neuronal responses related to behaviour are
often extracted by averaging neuronal activity over large number of experimental trials. Such trial-
averaging is carried out to reduce noise and to reduce the influence of other signals unrelated to the
corresponding stimulus or behaviour. However, if the recorded neuronal responses are jittered in
time with respect to the corresponding stimulus or behaviour, averaging over trials may distort the
estimation of the underlying neuronal response. Here, we present an algorithm, named dTAV
algorithm, for realigning the recorded neuronal activity to an arbitrary internal trigger. Using
simulated data, we show that the dTAV algorithm can reduce the jitter of neuronal responses for
signal to noise ratios of 0.2 or higher, i.e. in cases where the standard deviation of the noise is up to
five times larger than the neuronal response amplitude. By removing the jitter and, therefore,
enabling more accurate estimation of neuronal responses, the dTAV algorithm can improve analysis
and interpretation of the responses and improve the accuracy of systems relaying on asynchronous
detection of events from neuronal recordings.
1. Introduction
Many neurophysiological studies are investigating neuronal responses to external events. These
studies range from simple stimulus evoked neuronal responses in the corresponding primary sensory
areas, e.g. neuronal responses to light flashes in the primary visual cortex [1], to neuronal activity
correlated to complex behaviours, e.g. neuronal correlates of abstract problem solving [2,3]. In such
studies, neuronal responses are usually extracted by averaging the neuronal signal in order to
reduce the “noise”, i.e. parts of the neuronal signal that are not correlated to the stimulus or
behaviour that is being investigated. This procedure relies on the assumption that neuronal
responses are time locked to the corresponding stimulus or behaviour. This assumption can be
challenged, however, as neuronal responses show temporal variability in relation to the
corresponding stimulus or behaviour [4-8]. Depending on the amount of the temporal jitter, the
underlying neuronal response estimated by averaging may be distorted (Figure 1), possibly leading
to mistakes in subsequent analyses and incorrect conclusions.
Here, we propose an inference based algorithm for realignment of neuronal responses (dTAV
algorithm). First, we demonstrate that the reduction of variability across trials can be used as a
measure of jitter reduction and, therefore, as a measure of how well the neuronal responses have
been aligned. This property is crucial for the operation of the dTAV algorithm as it identifies the
underlying neuronal response by reducing the variability. Then, we built a simple model of neuronal
responses and demonstrated the usability of the dTAV algorithm for various noise levels using
simulated data. Furthermore we compare the dTAV algorithm to another non-parametric
realignment algorithm [9], referred to as MaxCorr in the rest of the text.
Figure 1. Effect of single trial jitter on the estimation of the underlying neuronal response. A: Neuronal responses are
related to the event (E), but are triggered (R) by an internal process, which is not precisely time-locked to the onset of
the external event. B: A certain amount of noise is recorded together with the relevant neuronal responses. C: During
the experiment, the external event occurs multiple times, while the neuronal activity is recorded. D: When neuronal
responses are aligned on the response start, the trial average response (F: blue line) is a good approximation of the real
neuronal response. However, the response onset is unknown. The trial-averaged response aligned on the event onset
triggers (F: red line) does not correctly reproduce the real neuronal response. In addition, the standard deviation across
trials calculated using the event onset triggers (F: blue and red shaded tubes) is an incorrect estimate of the variability of
neuronal responses.
2. Methods
The Methods are presented in the following order. First, we describe our simple model of neuronal
responses to an external event (stimulus or behaviour). Second, we present the analytic tools used
to predict the jitter reduction. Since the exact triggers of neuronal responses are not known in a real-
world application of the algorithm, it is necessary to design such a measure of jitter reduction in
order to optimize the parameters of the dTAV algorithm. Third, we describe the dTAV realignment
algorithm in detail. Finally, we describe the details of simulated data used to assess the performance
of the dTAV realignment algorithm and to compare it to a previously published MaxCorr realignment
algorithm [9].
2.1. A measure of jitter reduction
We assume that the neuronal signal is a superposition of neuronal responses r(t) evoked at response
onset times ti plus the Gaussian white noise signal η:
+
( )
signal t
(0,
η
η
N
∈
−
=
(
r t
( )
t
( )
t
)
s
η
)
t
i
∑
i
where ση is the standard deviation of the Gaussian white noise process and N(µ,σ) is a normal
distribution with a mean of µ and a standard deviation of σ. After recording the neuronal signal and
if neuronal response onset times are known, one can estimate the neuronal response by calculating
the response-triggered average ( )r t :
( )
r t
=
1
M
M
∑
i
1
=
(
signal t
+
)
t
i
=
( )
r t
+
1
M
M
∑
i
1
=
η
i
( )
t
(
η η
( )
t
=
i
t
+
)
t
i
s
η
( )
t
=
1
M
M
∑
η
i
i
1
=
( )
t
N
∈
0,
s
η
M
where M is the number of responses used to calculate the average; ηi(t) is the noise in i-th trial; and
( )
s
tη
is the random variable drawn from a Gaussian distribution that follows from the presence of
for variance,
noise. In the following derivations, we will use operator
V
response ( )r t in the presence of the noise is given by:
for sample covariance. The sample variance of the neural
for sample variance and
for expectation,
( )V
( )E
∧
Cov
( )
( )
( )
V r t
(
)
=
1
−
1
M
M
∑
i
1
=
(
(
signal t
+
)
t
i
−
( )
r t
2
)
=
1
−
1
M
M
∑
i
1
=
(
η
i
( )
t
−
s
η
( )
t
)2
and is distributed as a χ2 distribution [10]:
(
M
−
)
1
)
(
( )
V r t
2
s
η
−
2
χ
M
1
−
For a large number of degrees of freedom, i.e. M being large, the χ2 distribution can be
approximated by a normal distribution using the following transformation [11]:
a
−
2
χ
k
⇒
a k
−
2
k
−
N
(
)
0,1
(1)
(2)
(3)
(4)
(5)
(6)
Using this transformation, equation (5) becomes:
( )
V r t
(
)
=
( )
S t
V
( )
S t
V
2
N
∈
η ηs s
,
2
2
−
1
M
Experiments are usually designed in such a way that the studied neuronal responses are elicited by
events, e.g. sensory stimuli or certain types of behaviour. However, the onset times of the neuronal
responses are not known, since these are triggered internally by the brain. The time shift between
an event tEi and the onset of the neuronal response ti may not be constant and can be regarded as a
stochastic process. In our model, the difference between these two time points is modelled by a
Gaussian distribution:
− ∈
,
N µs
J
(
)
J
t
Ei
t
i
(7)
(8)
where µJ and σJ are the mean and the standard deviation of the distribution. If one has access to
event times only, as it is the case in a real experiment where the time points ti when the brain
triggers a response are unknown, one can estimate the neuronal response by calculating the event-
triggered sample mean
:
( )
t
( )
r t
J
=
1
M
(
signal t
+
t
Ei
)
=
1
M
M
∑
i
1
=
(
r t
+
t
Ei
−
t
i
)
+
1
M
M
∑
i
1
=
η
i
( )
t
=
( )
r t
+
s
η
( )
t
(9)
Jr
M
∑
i
1
=
is the average signal in the presence of the jitter but no noise (ση=0). The sample
Jr
is given by:
( )
t
( )
r t
where
variance of
( )
V r t
(
J
)
=
=
=
M
M
M
(
(
M
1
1
1
i
=
M
∑
∑
1
−
1
−
1
−
1
=
∧
(
Cov r t
∑
1
(
1
i
=
M
i
+
2
(
signal t
+
t
Ei
)
−
( )
r t
J
2
)
(
r t
+
t
Ei
−
t
i
)
−
( )
r t
+
η
i
( )
t
−
s
η
( )
t
2
)
(
η
i
( )
t
−
s
η
( )
t
2
)
+
1
−
1
M
M
∑
i
1
=
(
(
r t
+
t
Ei
−
t
i
)
−
( )
r t
2
)
(10)
+
t
Ei
−
t
i
)
,
η
i
( )
t
)
( )
V r t
(
)
in the absence of jitter and depends
The first term of equation (10) is the sample variance
only on the noise ηi(t) in a way given by equation (7). The second term is the variance contribution
arising from the jitter in the absence of noise and depends only on the jitter. The third term is the
sample covariance between the signal and the noise. We can rewrite equation (10) as:
2
s
J
( )
t
=
M
+
t
Ei
−
t
i
)
−
( )
r t
2
)
1
−
M
i
(
1
=
(
r t
∑
1
∧
(
(
Cov r t
( )
2
S t
s
+
V
J
( )
S
t
Cov
( )
V r t
(
J
=
)
2
=
t
+
Ei
( )
t
)
(11)
−
+
)
t
i
S
Cov
( )
,
t
η
i
( )
t
For normally distributed and small jitters ti - tEi, we can use a Taylor series expansion to express
2Js
( )
t
:
( )
r t
=
2
s
J
( )
t
=
≈
≈
≈
∑
1
=
i
M
M
M
1
=
1
1
M
M
∑
i
1
−
1
1
M
−
( )
dr t
dt M
( )
dr t
dt
i
∑
1
=
1
−
(
r t
+
t
Ei
−
t
i
)
≈
M
∑
i
1
=
( )
r t
+
(
t
Ei
−
t
i
)
( )
dr t
dt
=
( )
r t
+
( )
dr t
1
dt M
M
∑
i
1
=
(
t
Ei
−
t
i
)
1
M
)
(
(
r t
+
t
Ei
−
t
i
−
( )
r t
2
)
( )
r t
+
(
t
−
t
i
Ei
( )
r t
−
M
∑
j
1
=
(
t
E j
−
t
j
2
)
(12)
M
∑
i
1
=
1
)
(
t
Ei
−
t
i
)
(
⇒
(
V t
−
t
i
Ei
( )
dr t
1
dt M
2
−
t
i
)
−
2
χ
M
1
−
M
)
−
−
( )
dr t
dt
1
∑
M
1
i
=
( )
)
1
2
M
s
t
−
J
( )
dr t
2
s
J
dt
(
Ei
t
For large M, we can use transformation (6) and obtain:
2
s
J
( )
t
N
∈
( )
dr t
dt
2
s
J
,
( )
dr t
dt
2
s
J
M
2
−
1
CovS
( )
t
Since r(t + tEi - ti) and the noise η are not correlated, the expectation of
variance depends on the shape of the neuronal response and, thus, cannot be precisely estimated.
is zero while its
Using equations (7) and (12), the expectation of
expressed as:
( )
JV r t
(
)
for large M and small jitters can be
(
E
( )
V r t
)
We can now see from equation (14) how comparing
one of the two σJ values, e.g. σJ’ and σJ’’, is larger than the other. If σJ’ > σJ’’, the corresponding
expectations of the
′′ , will follow this relationship:
)
values may be used to infer which
( )
dr t
dt
)
( )
JV r t
′ and
)
( )
JV r t
( )
JV r t
2
ηs
2
s
J
=
+
,
(
)
(
(
(
)
J
′′>
s s
J
′
J
⇒
E
(
( )
JV r t
( )
V r t
(
J
′
)
>
E
( )
V r t
(
J
′′
)
(13)
(14)
(15)
)
(
CovS
( )
JV r t
VS and
To optimize the alignment, it is necessary to reduce the amount of jitter without the knowledge of
the real neuronal response triggers ti. A possible approach minimizes the variance of the stimulus
triggered response
(equation 10) at one particular time point since a reduction of jitter
may result in a decrease of Js and, hence, lead to a smaller variance (equation 16). However, since
the
terms depend on the noise in the signal and, thus, come from a stochastic process,
their values might go up by chance and, therefore, mask the reduction of Js . Neighbouring time
points of the neuronal response are correlated in time and so is the variance term Js arising from
the jitter. On the other hand, the noise may be correlated on a smaller time scale. Therefore, the
stochastic values of
may average out across time if the variance is averaged across a
sufficiently large time window. A more reliable indicator of jitter reduction may, therefore, be the
decrease of the time-averaged variance, TAV:
VS and
CovS
TAV
=
( )
V r t
(
J
)
=
1
−
T
S
)
(
T
E
T
E
∫
T
S
V r t dt
( )
(
)
J
(16)
As discussed before, averaging over time can reduce the variance of the TAV, thereby increasing the
reliability of the measure. On the other hand, integrating over periods of time where the neuronal
response is small compared to the noise or completely absent may not increase the reliability of TAV
but instead increase the variance of TAV. This trade-off means that the integration window across
which the variance is averaged should neither be too long nor too short.
In this study, we used the difference of TAV, dTAV, as a measure of jitter reduction:
dTAV
TAV
TAV
−
=
(
′′
s s
J
′
,J
(
′′
s
J
(
′
s
J
)
)
)
(17)
We used dTAV to optimize parameters of our realignment algorithm by assuming the following is
true:
dTAV
dTAV
p
(
′′
s s
J
(
′′
s s
J
)
(
′′′
s s
J
(
′′
s s
J
⇒
′′′
J
′′′
J
>
>
<
>
p
)
)
)
′
J
,
′
J
,
(18)
If equation (19) is correct, we can choose those parameters of our realignment algorithm which lead
to the strongest decrease of the time-averaged variability, i.e. we select the parameters for which
dTAV is largest. While equations (16) and (17) indicate that equation (19) is correct, analytical
derivation of such relationship will depend on the neural responses and may not always hold. In the
next section, we use numerical simulations to show that such relationship holds for two simulated
neural responses that have a mono-phasic and a bi-phasic shape.
2.2. Numerical analysis of the reliability of dTAV as a measure of jitter
reduction
To demonstrate the conditions for which the dTAV is a reliable measure of the reduction in jitter, we
performed a set of simulations, each composed of 2000 repetitions of an experiment composed of
100 trials. The neuronal responses were simulated as mono-phasic and bi-phasic functions
composed from Gaussian functions (Figure 2):
r
mono
( )
t
=
2
t
−
22
e o
R
1
2
πs
R
r
bi
( )
t
=
−
2
t
2
2
o
R
e
−
1.5
⋅
1
2
πs
R
1
⋅
2
π s
R
3
Figure 2. Simulated mono-phasic (left) and bi-phasic (right) neuronal responses.
−
(
5
t
−
s
R
(
2 3
⋅
s
R
2
)
2
)
e
(19)
(20)
where σR was taken to be 100ms. The shifts of the neural responses, ti - tEi, were drawn from a
Gaussian distribution with zero mean and standard deviation σJ:
− ∈
N s
J
0,
(
)
t
Ei
t
i
(21)
Noise was modelled as white Gaussian noise with zero mean and standard deviation ση (equation 1).
Our simulations used discrete time with a time step of 1ms. The TAV was calculated for each
combination of σJ, ση and the integration time TI; and for each simulation run k using the following
equation:
(
TAV T
I
k
,
s s
J
,
η
)
=
T
I
1
+ ∑
2
1
T
I
=−
T
I
t
(
( ;
V r t
k
J
s s
J
,
η
)
)
(22)
Each simulation was performed by selecting a combination of σJ, ση and TI values. The used σJ values
ranged from 0ms to 120ms in steps of 1ms and the simulation for σJ = 60ms was performed twice
because a dataset with 60ms jitter was used as the starting point for the simulated realignment. TI
values ranged from 30ms to 990ms in steps of 30ms. The ση values were selected to model different
signal to noise ratios (SNRs), defined as the ratio of the maximum absolute value of the neuronal
response and the standard deviation of the noise ση. We used ση values that yielded SNR values of
0.03, 0.05, 0.08, 0.13, 0.20, 0.32, 0.50, 0.79, 1.26 and 2.00 for both mono and bi-phasic responses.
We used these simulations to emulate an experiment where the jitter standard deviation of the
dataset was σJ’ = 60ms before the realignment. This initial dataset was compared to datasets with
jitter standard deviations σJ” ranging from 0ms (no jitter) to 120ms (doubled jitter) which
represented the dataset after the realignment. dTAV was then calculated for each combination of k,
σJ”, ση and TI.
Ranges of dTAV values varied across different orders of magnitude for different ση and TI values. We
therefore normalized dTAV values by dividing them by the maximum of the absolute value of the 1st
and 99th percentile.
(
T
I
,
s
η
)
)
(22)
99%
(
1%
dTAV
T
I
(
(
dTAV T
I
dTAV
99%
N
T
I
(
ndTAV T
I
k
k
)
=
,
s
η
=
)
=
(
1%
(
P dTAV T
I
′′
,
s
J
(
99%
(
P dTAV T
I
′′
,
k
s
J
(
(
max
T
I
(
dTAV T
)
I
′′ =
,
s s s
J
,
s
η
)
dTAV
,
η
,
s
η
′
J
,
k
,
1%
,
s
η
)
′′
′
,
,
,
s s s
J
J
η
(
)
,
dTAV T
s
I
η
N
k
′′
,
s s s
J
,
η
′
J
,
′′
,
s s s
J
,
η
′
J
,
)
)
)
)
dTAV
)
X %
P is the X-th percentile operator acting over variables a and b; and max is the maximum
where
,a b
value operator. The normalized dTAV (ndTAV) was binned in 50 equally wide bins spanning the
space from -1 to 1. Binned values were used to calculate the probability of jitter reduction,
(
p s s′
′′>
, for different ndTAV values, while keeping ση and TI constant. To show how the
J
reliability of dTAV as a measure of jitter reduction depends on SNR and TI, we calculated the
(
s s′
′′>
p
contours in the space spanned by ndTAV and TI for each value of
J
SNR separately. We also calculated the joint probabilities for each combination of ση and TI values,
p
, in order to verify that the relationship in equation (19) holds.
(
′
s s s
J
ndTAV
ndTAV
0.9
)0
>
=
−
)
(
′′
J
)
)
′
J
/
,
J
J
2.3. dTAV realignment algorithm
The dTAV realignment algorithm (Figure 3) relies on the assumption that the distribution of shifts in
the recorded neuronal signal is unimodal and that the neuronal responses can be represented by a
small number of features (f1,...,fn). In our case, these features were neuronal signals recorded at
different equidistant time points around the time of the event (τ1,...,τn):
signal t t
k
+
=
(
)
,k i
Ei
f
(23)
In the first step, we parameterized the (τ1,...,τn) set by the following three parameters: (i) time of the
first feature t1, (ii) number of features n and (iii) temporal distance between the last and the first
feature tn- t1.
The second step of our dTAV realignment algorithm is to select a subset of trials S that are already
fairly well aligned (Figure 3b). We selected a subgroup of trials containing half of the total number of
trials that has a small variance in the Euclidean space spanned by the features when compared to
other such subsets. The selection was performed using an iterative selection algorithm that with
initited with one trial, α1, and then consecutively adds a trial to the set of selected trials until this set
contains half of the trials. The selection process operates as follows: Let Ψ be the set of all trials and
Θ be the set of selected trials. In one iteration the trial that is closest to the mean of the set of
selected trials is added to the set:
F
l
=
l
f
1,
f
,
n l
,
µ
k
=
1
k
k
∑
i
1
=
F
α
i
α
k
+ =
1
minarg
l
F
l
−
µ
k
2
,
=
F
l
i
k
f
1
k
1,
α
i
∑
1
=
∑
∈ Ψ Θ
1
k
,
α
i
1
=
f
/
n
k
i
This iterations proceeds until half of the trials have been selected. We then calculated the variance
V(α1) of this set. The algorithm is then started again, this time using a different initial trial. This
procedure is repeated until all trials have been used once to initiate the selection algorithm. The
subset S with the smallest variance was selected:
(
V α
i
min arg
∈ Ψ
α
i
=
S
)
(
)
i
(24)
(25)
(26)
In the third step (Figure 3c), the selected trials were used to build a sliding window detection model.
Features of the selected trials constituted the “response” class (Rclass). The “baseline” class (Bclass)
was constituted from features taken from the same trial subset at times that differed from (τ1,...,τn)
by integer multiples of 1ms.
)
,
+
f
1,
,
i
n i
(
signal t
}
t
+ ∆
∈ ∆ = ⋅
(
signal t
,
k ms k
R
class
B
class
t
+ ∆
∈
,...,
,...,
t
n
,
i
t
1
}
/{0}
{
(
{
(
∈
,
1
+
=
=
S
S
)
)
)
Ei
Ei
f
i
t
(27)
These two classes were then used to train a quadratic discriminant analysis (QDA) [12] by fitting
Gaussian distributions to each of the classes. The QDA was then applied to neural feature vectors to
calculate the probabilities that the feature vectors belong to one of the classes.
In the next step (Figure 3d), the QDA model is used to give a probability pR that a feature set from a
given trial belongs to the response class. For each of the trials, we calculated the time of the
maximum probability in a certain time range:
Figure 3. Processing steps of the dTAV realignment algorithm. First step (A): The response is represented by a small
number of features. Left: The selected features are determined by the chosen array of equidistant time points
parameterized by the time of the first feature t1, the number of features n and the temporal distance between the last
and the first feature tn- t1. Right: Five example trials and the values of the chosen response features (blue and magenta
circles). Second step (B): The subset of half of the trials with the smallest within-subset variance is selected. Left: The
values of the chosen response features for every trial shown in the feature space. Green stars show the chosen subset.
Right: Five example trials with trials belonging to the chosen subset shown in green. Red circles in the left panel show
the feature values of the five example trials. Third step (C): The QDA model is built from the population of values
extracted by the time sliding template. Left: Feature values used to calibrate the model are taken only from the selected
subset of trials. Right: The response class is made by extracting features using the previously chosen array (left: blue and
magenta dots; right: red stars), while the “baseline” class is made by sliding the same array in time (left: first feature -
black dots; second feature - grey dots at the temporal distance of tn- t1 from the black dots; light blue brackets connect
first and second features of a pair; right: feature pairs represented by black stars). Blue ellipses show the estimated
standard deviation for the “response” and “baseline” class. Fourth step (D): The posterior probability of belonging to the
“response” class was calculated by sliding the chosen array in every trial. Left: Extracted features in one trial and in
feature space. Red circle and red star represent the feature with the maximum probability for the “response” class.
Right: Features with maximum probability for “response” class were found in each trial. Fifth step (E): Left: Trials were
realigned to the points of maximum probability. Right: Five example trials realigned to the point of maximum
probability.
t
,
i MAX
=
max arg
t
p
R
(
signal t
+
t
1
+
)
t
Ei
(
signal t
+
t
n
Ei
+
)
t
∈
R
class
,
t
S
< <
t
t
E
and used it to re-align the stimulus-triggered neuronal responses and calculate dTAV (Fig. 2e):
dTAV
=
( )
r t
1
M
∑
M =
1
i
(
)
TAV
t t =
1
,...,
n
(
signal t
−
t
Ei
−
t
,
i MAX
)
−
(
TAV t
1,
,...,
t
MAX
M MAX
,
)
START
(28)
(29)
(30)
This procedure was repeated for a different selection of features representing the neuronal
response.
In addition to parameters that determine feature extraction, it is possible to modify other
parameters in the realignment algorithm, such as the integration time used to calculate the TAV or
the size of the trial subgroup used to build the QDA model.
To train the QDA, a larger or smaller trial subset could be used. However, if the number of selected
trials is small, it will be difficult to reliably estimate the Gaussian distributions. On the other hand, if
the number of selected trials is very large and close to the total number of trials, the QDA will be
trained on many not well aligned trials and therefore, not be able to reduce the jitter. We assumed
that using half of the trials may be a good compromise but this value may be adjusted when the
method is applied to other datasets.
After all the parameter values have been exhausted, a set of time shifts was chosen by maximizing
dTAV:
(
t
1,
,...,
t
MAX
2.4. Simulated data
)
,
M MAX Chosen
=
max arg
parameters
(
dTAV
)
(31)
We used dTAV realignment algorithm to realign the neuronal responses in a range of simulated
experiments. In addition to dTAV algorithm, we also used the MaxCorr algorithm [9] and compared
the results of the two algorithms.
Simulations were made for two neuronal responses, mono-phasic (rM) and bi-phasic responses (rB),
whose shape resembled reported neurophysiological responses. For both types of responses, we
performed 100 simulated experiments, each consisting of 200 trials. In each trial, the single channel
neuronal response to an arbitrary stimulus was recorded at 1KHz. Neuronal responses were
modelled as follows (Figure 2):
r
M
( )
t
=
(
2
)
2
250
t
−
(
)
2 83
⋅
for 0
e
≤ <
0 otherwise
t
500
(32)
( )
r t
B
=
2
)
(
125
t
−
2
(
)
2 25
⋅
2
(
)
250
t
−
2
(
)
2 83
⋅
e
-1.5
e
≤ <
0 otherwise
for 0
t
500
Neural responses were simulated in the following way:
− +
( )
data t
=
t
i
η
( )
t
∈
N
(0,
s
,
η
i
);
t
j
1
+
(
∑
r t
j
t
− ∈
j
j
j
t
t
−
)
(10 ,10 );
t
(
s
s
Ej
)
j
+
η
( )
t
− ∈
t
Ej
N
(0,0.1 )
s
N
(33)
(34)
Noise in the recordings was simulated as additive Gaussian noise with zero mean and different
standard deviations ση: 31.62, 19.95, 12.59, 7.94, 5.01, 3.16, 2.00, 1.26, 0.79 and 0.50, which
corresponded to SNRs of 0.03, 0.05, 0.08, 0.13, 0.20, 0.32, 0.50, 0.79, 1.26 and 2.00. Temporal
distances between the stimulus times were drawn from a Gaussian distribution with a mean of 10s
and a standard deviation of 10s. To keep neuronal responses from overlapping, temporal distances
below 3s were redrawn. The neuronal response offset (i.e. the temporal jitters) were drawn from a
Gaussian distribution with zero mean and a standard deviation of 0.1s. To avoid occasional very
large jitters, all jitters with an absolute value above 0.3s were redrawn.
To correctly simulate the outcome of the experiment, we assumed that the person analysing the
data would filter the data using a low-pass filter, given that low-frequencies dominate the simulated
neuronal responses. We filtered the simulated recordings using 2nd order symmetric Savitzky-Golay
filters [13,14] with different time windows of 100ms, 250ms, 500ms or 1000ms.
The time points used to extract the neuronal features were varied using three parameters: time of
the first feature relative to the time of the event τ1 (values ranged from -125ms to 1324ms in steps
of 63ms), temporal distance between the first and the last feature τ1- τn (100ms, 250ms, 500ms or
1000ms) and the number of features n (2, 4, 8 or 12). We used an integration time TI of 350ms and
the maximum probability was identified in the time window ranging from 300ms before the stimulus
till 300ms after the stimulus.
2.5. MaxCorr realignment algorithm
Realignment results obtained using dTAV algorithm were compared to results obtained using the
MaxCorr algorithm [9]. The MaxCorr algorithm works by approximately maximizing cross-
correlations between each pair of trials in three steps. First, for N trials, N(N-1)/2 cross-correlations
CXij for all possible trial pairs i and j and time lags (λi- λj) up to half of the trial length are calculated.
Second, a parabolic function Fij(λi- λj) is fitted to the crosscorrelation between the i-th and the j-th
trial around the time lag (λi- λj)MAX for which the crosscorrelation is at maximum.
CX
ij
(
λ λ
i
j
−
)
≈
F
ij
(
λ λ
i
j
−
)
=
b
0
+
b
1
(
λ λ
i
j
−
)
+
b
2
(
λ λ
i
j
−
)2
(35)
To fit the parabolic functions Fij(λi- λj), we used a neighbourhood of ±10ms around (λi- λj)MAX. We
tested other neighbourhoods of up to 60ms and found that the size of the neighbourhood did not
influence the results (data not shown). In the last step, all parabolic functions are summed up to
derive a new function F(λ2,…, λN), which is quadratic in all of its variables and, therefore, has a
unique global maximum (λ2,…, λN) MAX. The values of (λ2,…, λN) MAX are calculated by solving the
system of linear equations obtained by applying partial derivatives to the parabolic function F.
(λ2,…, λN) are used to realign the trials relative to the first trial. MaxCorr algorithm was applied to
the same signal used to evaluate the realignment of the dTAV algorithm and we evaluated jitter
reduction metrics (σJ’ - σJ’’) / σJ’ for both algorithms for comparison.
3. Results
3.1. dTAV as a measure of jitter reduction
Using the model of neuronal responses described in section 2.1, we calculated the dependence of
dTAV on the reduction of the jitter for different SNR levels and integration times TI. Results are
summarized in Figure 4 for the mono-phasic signal and in Figure 5 for the bi-phasic signal.
For an integration time of TI=300ms, selected for presentation in Figures 4a-c and 5a-c, the
expectation of the dTAV increased monotonously with the amount of reduction of the jitter and did
hardly depend on the SNR (Figure 4a, Figure 5a). On the other hand, the standard deviation of dTAV
(Figure 4b, Figure 5b) depended strongly on the SNR and was comparable or larger than the
expected value of the dTAV. For high SNRs, the expectation of dTAV surpassed the standard
deviation of dTAV even for small reduction in jitter (see Figure 4a-b and 5a-b). In such cases, dTAV is
a reliable measure of the amount of jitter reduction, even when the amount of jitter reduction is
small.
Figure 4. Reliability of dTAV as a measure of jitter reduction for mono-phasic neuronal responses. A: Expectation of
dTAV as a function of the reduction of jitter standard deviation. Lines drawn only for values of SNR of 0.2 and higher.
For lower SNR, 2000 repetitions were insufficient to provide a reliable estimate of the expected value of dTAV due to
the high noise level. For the shown SNR range, the expected value of dTAV is independent of the SNR. B: The standard
deviation (std) of dTAV as a function of the amount of jitter reduction for different SNRs. Standard deviations of dTAV
for SNR of 0.2 and lower are above 10-6 and are, therefore, not shown. C: Probability of jitter reduction as a function of
ndTAV for different SNRs. Panels A, B and C are shown for integration time TI of 300ms. D: Values of jitter reduction
and integration times for which the probability of correct dTAV prediction reaches 90%. For jitter reductions and
(
integration times above the line, the probability for correct dTAV prediction,
s s′
′′>
J
For SNRs of 0.13 and lower, the probability of correct dTAV prediction never reached 90%.
, is above 90%.
)0
ndTAV
>
p
J
To measure how well we can rely on the dTAV as a measure of jitter reduction, we calculated the
probability of correctly predicting that the jitter was reduced based on the normalized dTAV values
(Figure 4c, Figure 5c). As the SNR was increased, the probability increased up to 1, even for the
smallest jitter reductions. For low SNR values the probability never reached 1, even when the jitter
was completely removed. To reach a substantial increase of the probability above 0.5, an SNR of
about 0.20 or higher was needed.
To provide an insight into the dependence of the probability of correct dTAV prediction on the
integration time, we calculated the values of integration times and reductions of jitter standard
deviation for which the probability of correct dTAV prediction reached 90% (Figure 4d, Figure 5d).
For high SNRs the performance of the dTAV prediction was nearly independent on the integration
time whereas for low SNRs the integration time had a stronger influence on the performance of the
dTAV prediction. The performance of the dTAV prediction decreased faster for integration times
below the optimal integration time, while it decreased more slowly for integration times bigger than
the optimal integration time. Therefore, choosing a short integration time could be more
disadvantageous than choosing a longer integration time.
Figure 5. Reliability of dTAV as a measure of jitter reduction for bi-phasic neuronal response. See caption of Figure 4 for
details.
So far we have computed the probability of jitter reduction given a reduction in dTAV. Next, we
investigated whether the difference of dTAV is also predictor of the expected amount of jitter
reduction, i.e. does a larger reduction in dTAV also indicate a larger amount of jitter reduction. If so,
we can use dTAV to optimize parameters of our re-alignment algorithm by selecting parameter
values that gave the highest dTAV values. To assess whether this is the case, we calculated the joint
probability distribution of jitter reduction and dTAV for different SNR values for the mono-phasic
neural response (Figure 6). For SNRs of 0.13 and lower, dTAV provided no or only very little
information of the amount of jitter reduction but as the SNR increased (to values of about 0.2 and
higher), the relation between dTAV and jitter reduction became less variable and dTAV became an
increasingly good predictor of the amount of jitter reduction. For high SNRs (1.26 and higher), even
small differences in dTAV indicated increased jitter reduction with high certainty. This suggests that
dTAV is a good predictor of the amount of jitter reduction for sufficiently high SNRs (of about 0.2
and higher) and can be used to optimize the parameters of our realignment algorithm in such cases.
The parameter selection based on dTAV will improve with increasing SNR.
Figure 6. Joint probability distribution of jitter reduction and ndTAV for different SNR values for the mono-phasic
neuronal response. An integration time window of Ti=300ms was used. For low SNR, dTAV is uninformative as a
measure of jitter reduction. As the SNR increases, dTAV becomes more informative of the jitter reduction, i.e the ability
to differentiate different levels of jitter reduction based on dTAV improves substantially.
3.2. Re-alignment of simulated data
We used the dTAV and the MaxCorr realignment algorithms to realign neuronal responses in 100
simulated experiments for either mono-phasic or bi-phasic neuronal responses and for different
levels of noise (Figure 7). To calculate dTAVwe used an integration window that captured the
majority of the signal, starting at TS = 0s and ending at TE = 1s, both in respect to the stimulus times
tE. We intentionally did not want to use the results of the dTAV reliability analysis (Figure 4d and 5d)
to estimate the optimal integration window since this would bias our comparison with the MaxCorr
algorithm. However, we did use the general finding of the dTAV reliability analysis that using an
integration window wider than the response diminished the dTAV reliability less than using a
window that is narrower than the response. In other words, we used a window that would certainly
be wider than the optimal window, knowing that this does only weakly influence the reliability of
dTAV as a measure of jitter reduction. We used a filter window length of 250ms. For low SNRs
(mono-phasic signal: SNR<0.2; bi-phasic signal: SNR<0.32) both algorithms increased the amount of
jitter, rather than decreasing it. For intermediate SNRs (mono-phasic signal: 0.2<SNR<0.75; bi-phasic
signal: 0.32<SNR<1.26; filter window 250ms) the dTAV algorithm outperformed the MaxCorr
algorithm (p<10-9, Mann–Whitney–Wilcoxon signed test). For high SNRs (mono-phasic signal:
0.8<SNR; bi-phasic signal: 1.5<SNR), both algorithms removed almost all jitter from the recorded
signal. In some of the high SNR cases, the MaxCorr algorithm achieved a higher jitter reduction, but
the difference was very small (<0.08) and significant only in the case of the mono-phasic response
for SNR=0.79 (P<0.05, Mann–Whitney–Wilcoxon signed test).
For both algorithms, the re-alignment worked best when a filter length of 250ms was used. For the
dTAV algorithm, the differences in the jitter reduction between using a filter window length of
250ms and using filter window lengths of 100ms and 500ms were very small. Only for a very long
filter window (1000ms), the re-alignment performance decreased substantially. For the MaxCorr
algorithm, the filter window length had a stronger influence on the amount of jitter reduction.
When the filter window length is included as an additional parameter whose value was chosen by
maximizing dTAV, the final jitter reduction was not significantly different from the jitter reduction
that was obtained by the optimal filter length (250ms). Thus, the dTAV algorithm can automatically
select the proper filter length if filter length is included as one of the parameters.
Figure 7. Reduction of jitter standard deviation for the dTAV algorithm (full lines) and the MaxCorr algorithm (dashed
lines; [9]) for mono-phasic (left) and bi-phasic (right) neural responses. Reduction is shown for different low-pass filters
and for the dTAV algorithm when the filter length is optimized as one of the algorithm parameters (dotted black line).
All results are averaged over 100 simulation repetitions; error bars depict the standard errors of the mean.
4. Discussion
In this paper, we presented and evaluated a novel method for the realignment of neuronal
responses. The algorithm uses the difference of time-averaged variance (dTAV) as a measure of
jitter reduction, and the quadratic linear discriminant analysis to infer the temporal shifts of
neuronal responses in individual trials. We showed that, by using the dTAV algorithm, it is possible
to realign simulated mono and bi-phasic single trial neuronal responses for noise levels several times
larger than the neuronal response amplitudes.
We used mono and bi-phasic potentials composed out of one and two Gaussian functions as
examples of neuronal responses. The success of the method in reducing the jitter for intermediate
and low noise levels (SNR>0.32) and for both mono and bi-phasic neuronal responses, suggests that
the algorithm can be successfully applied in a large number of cases. At the same time our results
show differences between mono and bi-phasic response shapes, suggesting that the signal shape
can affect the performance of the dTAV algorithm.
Our simulations assumed that the recorded neuronal signal is a continuously modulated signal,
which is valid for example for local field potentials, electro-corticographic signals (ECoG), electro-
/magnetoencephalographic recordings (EEG/MEG) as well as for near-infrared spectroscopy (NIRS)
and functional magnetic resonance imaging (fMRI) signals. Furthermore, our algorithm can also be
applied to spike trains by estimating instantaneous neuronal firing from the spike times [15,16] and
using the instantaneous neuronal firing rates as the neuronal signal in our algorithm. In addition, the
algorithm can be applied to continuously modulated signals which were derived from the
aforementioned neuronal signals, such as time-resolved spectral amplitudes (e.g. extracted using
short-time Fourier transform) or crosscorelation measures. Recordings that depend on other
variables than time, such as space or frequency, can also be realigned by exchanging the time
variable by the corresponding variable (e.g. spatial coordinate or frequency).
We compared the performance of our dTAV algorithm to the performance of the MaxCorr algorithm
[9] which was also designed to reduce the temporal jitter in the neuronal responses. The dTAV
algorithm outperformed the MaxCorr algorithm substantially for intermediate SNRs and yielded
similar performance for low and high SNRs.
The dTAV algorithm works with a large number of parameters whose values have to be determined
through testing different parameter sets. This can be achieved by defining admissible values for each
parameter (e.g. within an interval) and scan all possible combinations. If the number of admissible
values for each parameter is very high, this process becomes computationally demanding and,
therefore, may be time consuming. In our simulations we obtained good realignments within a
reasonable timeframe (less than an hour), by optimizing the parameters across a limited number of
values. Indeed, our results show (e.g. Figures 4 and 5) that the fine optimization of parameters may
not be necessary and the required computational time for parameter optimization may therefore
not be problematic.
In general, realignment algorithms can be used to improve the analysis of neurophysiological
experiments by improving the estimation of neuronal responses. The obvious case is the estimation
of the neuronal response by calculating trial averages. As shown in our example (Figure 1), even if
neuronal responses are constant across trials and the noise is uncorrelated to the signal, averaging
the jittered single-trial responses can lead to a distorted estimation of the response and the
incorrect estimation of the noise. Removing the jitter by the proposed algorithm can improve the
estimation of the neuronal response and improves the estimation of the noise. A large number of
neuroscience studies investigate neuronal responses related to sensory stimuli. When neuronal
responses are well locked to the stimulus [17] realignment methods might be of limited use. On the
other hand, neuronal responses may not be locked to the stimulus but, in addition to the stimulus,
may also be affected by the internal neuronal state [18,19] and, therefore be temporally jittered
relative to the stimulus onset. Our algorithm can be used to find out whether the responses are
locked to such internal events and compute the approximate timing of these events. Furthermore,
neuronal responses related to behaviour may also be jittered. For example, neuronal responses
related to movement planning [20,21] may be jittered with respect to the times when the
movements were initiated if the movements were either self-paced or triggered by another
stimulus. Realignment algorithms could be used to align noisy individual trials in order to improve
the accuracy of determining the underlying neuronal response (Figure 1). These more accurately
determined neuronal responses would then also facilitate a comparison of responses between
studies, allowing for identification of neuronal response parts shared between classes of responses.
Computing the realignment times may give us also insights into the timing of internal events and
into the variability of timings in internal cognitive processes.
Additionally, brain-machine interface systems that detect events based on neuronal recordings [22-
29] may benefit from re-alignment algorithms. Such systems require a certain number of trials
containing the neuronal responses to build the model used to detect the events from continuous
neuronal recordings. If the jitter or the neuronal responses used to train the model is reduced, the
detection may perform better.
In summary, we showed that the dTAV realignment algorithm can reduce the jitter of simulated
neuronal responses for response waveforms commonly observed in neurophysiological recordings
and noise levels many times higher than the neuronal response itself. Hence, the application of the
dTAV algorithm can improve analysis and interpretation of neuronal responses and improve the
performance of asynchronous detection of events from neuronal recordings.
Acknowledgements
This work was supported by the German Federal Ministry of Education and Research (BMBF) grant
01GQ0830 to BFNT Freiburg and Tübingen.
Bibliography
1. Dagnelie G, Spekreijse H, van Dijk B (1989) Topography and homogeneity of monkey V1 studied
through subdurally recorded pattern-evoked potentials. Vis Neurosci 3: 509-525.
2. Gaona CM, Sharma M, Freudenburg ZV, Breshears JD, Bundy DT, et al. (2011) Nonuniform high-
gamma (60-500 Hz) power changes dissociate cognitive task and anatomy in human cortex. J
Neurosci 31: 2091-2100.
3. Gunduz A, Brunner P, Daitch A, Leuthardt EC, Ritaccio AL, et al. (2011) Neural correlates of visual-
spatial attention in electrocorticographic signals in humans. Front Hum Neurosci 5: 89.
4. Radons G, Becker JD, Dulfer B, Kruger J (1994) Analysis, classification, and coding of multielectrode
spike trains with hidden Markov models. Biol Cybern 71: 359-373.
5. Seidemann E, Meilijson I, Abeles M, Bergman H, Vaadia E (1996) Simultaneously recorded single
units in the frontal cortex go through sequences of discrete and stable states in monkeys
performing a delayed localization task. J Neurosci 16: 752-768.
6. Seal J, Commenges D, Salamon R, Bioulac B (1983) A statistical method for the estimation of
neuronal response latency and its functional interpretation. Brain Res 278: 382-386.
7. Requin J, Riehle A, Seal J (1988) Neuronal activity and information processing in motor control:
from stages to continuous flow. Biol Psychol 26: 179-198.
8. Vaadia E, Kurata K, Wise SP (1988) Neuronal activity preceding directional and nondirectional cues
in the premotor cortex of rhesus monkeys. Somatosens Mot Res 6: 207-230.
9. Nawrot MP, Aertsen A, Rotter S (2003) Elimination of response latency variability in neuronal
spike trains. Biol Cybern 88: 321-334.
10. Knight K (2000) Mathematical statistics. Boca Raton: Chapman & Hall/CRC Press. 481 p. p.
11. Canal L (2005) A normal approximation for the chi-square distribution. Computational Statistics
& Data Analysis 48: 803-808.
12. Hastie T, Tibshirani R, Friedman JH (2009) The elements of statistical learning : data mining,
inference, and prediction. New York: Springer.
13. Savitzky A, Golay MJE (1964) Smoothing and Differentiation of Data by Simplified Least Squares
Procedures. Analytical Chemistry 36: 1627-1639.
14. Steinier J, Termonia Y, Deltour J (1972) Smoothing and Differentiation of Data by Simplified Least
Square Procedure. Analytical Chemistry 44: 1906-1909.
15. Cunningham JP, Yu BM, Sahani M, Shenoy KV (2007) Inferring neural firing rates from spike trains
using {G}aussian processes. Neural Coding, Computation and Dynamics (NCCD).
16. Nawrot M, Aertsen A, Rotter S (1999) Single-trial estimation of neuronal firing rates: from single-
neuron spike trains to population activity. J Neurosci Methods 94: 81-92.
17. Richmond BJ, Optican LM (1990) Temporal encoding of two-dimensional patterns by single units
in primate primary visual cortex. II. Information transmission. Journal of Neurophysiology 64:
370-380.
18. Churchland MM, Yu BM, Cunningham JP, Sugrue LP, Cohen MR, et al. (2010) Stimulus onset
quenches neural variability: a widespread cortical phenomenon. Nat Neurosci 13: 369-378.
19. Churchland MM, Yu BM, Ryu SI, Santhanam G, Shenoy KV (2006) Neural variability in premotor
cortex provides a signature of motor preparation. J Neurosci 26: 3697-3712.
20. Milekovic T, Truccolo W, Grun S, Riehle A, Brochier T (2015) Local field potentials in primate
motor cortex encode grasp kinetic parameters. Neuroimage.
21. Confais J, Kilavik BE, Ponce-Alvarez A, Riehle A (2012) On the anticipatory precue activity in
motor cortex. J Neurosci 32: 15359-15368.
22. Levine SP, Huggins JE, BeMent SL, Kushwaha RK, Schuh LA, et al. (2000) A direct brain interface
based on event-related potentials. IEEE Trans Rehabil Eng 8: 180-185.
23. Bashashati A, Ward RK, Birch GE (2007) Towards development of a 3-state self-paced brain-
computer interface. Comput Intell Neurosci: 84386.
24. Hwang EJ, Andersen RA (2009) Brain control of movement execution onset using local field
potentials in posterior parietal cortex. J Neurosci 29: 14363-14370.
25. Awwad Shiekh Hasan B, Gan JQ (2010) Unsupervised movement onset detection from EEG
recorded during self-paced real hand movement. Med Biol Eng Comput 48: 245-253.
26. Achtman N, Afshar A, Santhanam G, Yu BM, Ryu SI, et al. (2007) Free-paced high-performance
brain-computer interfaces. J Neural Eng 4: 336-347.
27. Solis-Escalante T, Muller-Putz G, Pfurtscheller G (2008) Overt foot movement detection in one
single Laplacian EEG derivation. J Neurosci Methods 175: 148-153.
28. Pistohl T, Schmidt TS, Ball T, Schulze-Bonhage A, Aertsen A, et al. (2013) Grasp Detection from
Human ECoG during Natural Reach-to-Grasp Movements. PLoS One 8: e54658.
29. Milekovic T, Ball T, Schulze-Bonhage A, Aertsen A, Mehring C (2013) Detection of Error Related
Neuronal Responses Recorded by Electrocorticography in Humans during Continuous
Movements. PLoS One 8: e55235.
|
1703.00698 | 1 | 1703 | 2017-03-02T10:19:17 | A mean-field model for conductance-based networks of adaptive exponential integrate-and-fire neurons | [
"q-bio.NC"
] | Voltage-sensitive dye imaging (VSDi) has revealed fundamental properties of neocortical processing at mesoscopic scales. Since VSDi signals report the average membrane potential, it seems natural to use a mean-field formalism to model such signals. Here, we investigate a mean-field model of networks of Adaptive Exponential (AdEx) integrate-and-fire neurons, with conductance-based synaptic interactions. The AdEx model can capture the spiking response of different cell types, such as regular-spiking (RS) excitatory neurons and fast-spiking (FS) inhibitory neurons. We use a Master Equation formalism, together with a semi-analytic approach to the transfer function of AdEx neurons. We compare the predictions of this mean-field model to simulated networks of RS-FS cells, first at the level of the spontaneous activity of the network, which is well predicted by the mean-field model. Second, we investigate the response of the network to time-varying external input, and show that the mean-field model accurately predicts the response time course of the population. One notable exception was that the "tail" of the response at long times was not well predicted, because the mean-field does not include adaptation mechanisms. We conclude that the Master Equation formalism can yield mean-field models that predict well the behavior of nonlinear networks with conductance-based interactions and various electrophysiolgical properties, and should be a good candidate to model VSDi signals where both excitatory and inhibitory neurons contribute. | q-bio.NC | q-bio |
A mean-field model for conductance-based networks of adaptive
exponential integrate-and-fire models
Yann Zerlaut and Alain Destexhe
September 12, 2018
1 Abstract
Voltage-sensitive dye imaging (VSDi) has revealed fundamental properties of neocortical pro-
cessing at mesoscopic scales. Since VSDi signals report the average membrane potential, it
seems natural to use a mean-field formalism to model such signals. Here, we investigate
a mean-field model of networks of Adaptive Exponential (AdEx) integrate-and-fire neurons,
with conductance-based synaptic interactions. The AdEx model can capture the spiking re-
sponse of different cell types, such as regular-spiking (RS) excitatory neurons and fast-spiking
(FS) inhibitory neurons. We use a Master Equation formalism, together with a semi-analytic
approach to the transfer function of AdEx neurons. We compare the predictions of this mean-
field model to simulated networks of RS-FS cells, first at the level of the spontaneous activity
of the network, which is well predicted by the mean-field model. Second, we investigate the
response of the network to time-varying external input, and show that the mean-field model
accurately predicts the response time course of the population. One notable exception was that
the "tail" of the response at long times was not well predicted, because the mean-field does
not include adaptation mechanisms. We conclude that the Master Equation formalism can
yield mean-field models that predict well the behavior of nonlinear networks with conductance-
based interactions and various electrophysiolgical properties, and should be a good candidate
to model VSDi signals where both excitatory and inhibitory neurons contribute.
2
Introduction
Recent advances in imaging technique, in particular voltage-sensitive dye imaging (VSDi), have revealed
fundamental properties of neocortical processing (Arieli et al., 1996; Contreras and Llinas, 2001; Petersen
and Sakmann, 2001; Ferezou et al., 2006; Civillico and Contreras, 2012): subthreshold responses to sensory
inputs are locally homogeneous in primary sensory areas, depolarizations tend to spread across spatially
neighboring regions and responses to sensory stimuli are strongly affected by the level of ongoing activity. It
1
also appears as a great tool to unveil how the spatio-temporal dynamics in the neocortex shape canonical
cortical operations such as normalization (Reynaud et al., 2012).
On the other hand, the literature lacks, to the best of our knowledge, theoretical models that provides a
detailed account of those phenomena with a clear relation between the biophysical source of the VSDi signal
and network dynamics at that spatial scale (i.e. at the millimeters or centimeters scale). Detailed model of a
neocortical column (i.e. ∼0.5mm2 scale) have been recently proposed, see Chemla and Chavane (2010) for
the link with the VSDi signal or more generally Markram et al. (2015), but their computational cost impedes
the generalization to higher spatial scale. The aim of the present communication is therefore to design a
theoretical model of neocortical dynamics with the following properties: 1) it should describe the temporal
scale of optical imaging as well as easily extend to its spatial scale and 2) it should have a correlate in terms
of single-cell dynamics (in particular membrane potential dynamics), so that the model can directly generate
predictions for the signal imaged by the VSDi technique (Berger et al., 2007).
More specifically, our study focuses on network dynamics in activated cortical states, thus the desired
model should describe neocortical computation in the asynchronous regime, where cortical activity is
characterized by irregular firing and strong subthreshold fluctuations at the neuronal level (Steriade et al., 2001;
Destexhe et al., 2003). The strategy behind the present model is to take advantage of the mean-field
descriptions of network dynamics in this regime. Via self-consistent approaches, those descriptions allow
to capture the dynamical properties of population activity in recurrent networks (Amit and Brunel, 1997;
Brunel and Hakim, 1999; Brunel, 2000; Latham et al., 2000; El Boustani and Destexhe, 2009). The present
model will thus consider randomly connected network of 10000 neurons as a unit to describe a cortical column
and we will compare its behavior to network simulations.
3 Material and Methods
We describe the equations and parameters used for the neuronal, synaptic and network modeling. We present our
heuristic treatment of the neuronal transfer functions: the quantity that accounts for the cellular computation in
mean-field models of population activity. Then, we present the specific markovian model of population activity used
in this study. Finally, we compare this analytical model to numerical simulations of network dynamics.
3.1 Single neuron models
The neuronal model used in this study is the adaptative exponential and fire (AdExp) model (Brette and Gerstner,
2005). The equation for the membrane potential and the adaptation current therefore reads:
Cm
τw
dV
dt
dIw
dt
= gL (EL − V ) + Isyn(V, t) + kae
V −Vthre
ka
− Iw
b δ(t − ts)
(1)
= −Iw + X
ts∈{tspike}
where Isyn(V, t) is the current emulating synaptic activity that will create the fluctuations, Iw reproduces the
Im current (McCormick et al., 1985). The spiking mechanism is the following: when V (t) reaches Vthre + 5 ka, this
2
Table 1: Model parameters.
Parameters
Parameter Name
Symbol Value Unit
cellular properties
excitatory cell
inhibitory cell
synaptic properties
numerical network
ring model
leak conductance
leak reversal potential
membrane capacitance
leak reversal potential
AP threshold
refractory period
adaptation time constant
gL
EL
Cm
EL
Vthre
τref rec
τw
sodium sharpness
adaptation current increment
adaptation conductance
sodium sharpness
adaptation current increment
adaptation conductance
ka
b
a
ka
b
a
excitatory reversal potential
Ee
inhibitory reversal potential
Ei
excitatory quantal conductance Qe
inhibitory quantal conductance Qi
excitatory decay
τe
inhibitory decay
τi
nS
10
-65 mV
150
pF
-65 mV
-50 mV
5 ms
500 ms
2 mV
20
pA
nS
4
0.5 mV
pA
nS
0
0
0 mV
-80 mV
nS
1
5
nS
5 ms
5 ms
cell number
connectivity probability
fraction of inhibitory cells
external drive
total extent
excitatory connectivity extent
inhibitory connectivity extent
propagation delay
Ntot
g
νdrive
e
Ltot
lexc
linh
vc
10000
5%
20%
4 Hz
40 mm
5 mm
1 mm
300 mm/s
triggers a spike ts ∈ {tspike}, this increases the adaptation variable Iw by b, the membrane potential is then clamped
at EL for a duration τ refrac=5ms. We consider two versions of this model: a regular spiking neuron for the excitatory
cells and a fast spiking neuron for the inhibitory cells (see Figure 2). The parameters of those two models can be
found on Table 1.
3.2 Synaptic model
The time- and voltage-dependent current that stimulate the neuron is made of the sum of an excitatory and inhibitory
currents (indexed by s ∈ {e, i} and having a reversal potential Es):
Isyn(V, t) = X
X
τs (Es − V )H(t − ts)
− t
Qs e
(2)
where H is the Heaviside function.
s∈{e,i}
ts∈{ts}
This synaptic model is referred to as the conductance-based exponential synapse. The set of events {te} and {ti}
are the set of excitatory and inhibitory events arriving to the neuron. In numerical simulations of single neurons
3
(performed to determine the transfer function F of either excitatory or inhibitory neurons), it will be generated by
stationary Poisson processes. On the other hand, in numerical simulations of network dynamics it will correspond to
the set of spike times of the neurons connecting to the target neurons, both via recurrent and feedforward connectivity.
3.3 Numerical network model
[Figure 1 about here.]
All simulations of numerical network were performed with the brian2 simulator (Goodman and Brette, 2009), see
http://brian2.readthedocs.org. For all simulations, the network was composed of Ntot=10000 neurons, separated in
two populations, one excitatory and one inhibitory with a ratio of g=20% inhibitory cells. Those two populations we
recurrently connected (internally and mutually) with a connectivity probability =5%.
Because this network did not display self-sustained activity (see Figure 3, in contrast to Vogels and Abbott (2005)),
an excitatory population exerted an external drive to bring the network out of the quiescent state. This population
targeted both the excitatory and inhibitory neurons. Note that the firing rate of this population was linearly increased
to avoid a too strong initial synchronization (see Figure 4). Finally, when studying responses to external inputs, an
excitatory population of time varying firing rate was added to evoke activity transients in the population dynamics.
This last stimulation targeted only the excitatory population. The number of neurons in those two excitatory
populations was taken as identical to the number of excitatory neurons (i.e. (1 − g) Ntot) and created synapses onto
the recurrent network with the same probability . After temporal discretization, the firing rates of those afferent
populations were converted into spikes by using the properties of a Poisson process (i.e. eliciting a spike at t with a
probability ν(t) dt). All simulations were performed with a time-step dt=0.1ms.
3.4 Estimating the transfer functions of single neurons
The transfer function F of a single neuron is defined here as the function that maps the value of the stationary excitatory
and inhibitory presynaptic release frequencies to the output stationary firing rate response, i.e. νout = F(νe, νi). Note
the stationary hypothesis in the definition of the transfer function (see discussion in main text).
Because an analytical solution of this function for the single neuron models considered in our study is a very
challenging mathematical problem, we adopted a semi-analytical approach. We performed numerical simulations of
single cell dynamics at various excitatory and inhibitory presynaptic frequencies (νe and νi respectively) (see the
output in Figure 2) on which we fitted the coefficients of an analytical template to capture the single cell model's
response.
The procedure relied on fitting a phenomenological threshold V ef f
thre that accounts for the single neuron non-linearities
(spiking and reset mechanism, adaptation mechanisms) on top of the subthreshold integration effects (Zerlaut et
al., 2016). This phenomenological threshold is then plugged-in into the following formula (analogous to Amit and
Brunel (1997)) to become our firing response estimate:
νout = 1
2 τV
thre − µV
· Erf c( V ef f
√
2 σV
)
(3)
Where (µV , σV , τV ) are the mean, standard deviation and autocorrelation time constant of the membrane potential
fluctuations. How to calculate those quantities as a response to a stationary stimulation is the focus of the next
section.
4
The phenomenological threshold was taken as a second order polynomial in the three dimensional space (µV , σV , τV ):
thre(µV , σV , τ N
V ef f
V ) = P0 + X
Pxy ·(cid:16) x − x0
X
x∈{µV ,σV ,τ N
V
Px ·(cid:16) x − x0
(cid:17)
(cid:17)(cid:16) y − y0
δx0
}
δx0
δy0
(cid:17)
+
(4)
x,y∈{µV ,σV ,τ N
V
}2
V =10mV, σ0
V =4mV, δσ0
V = 6mV, τ N0
V =0.5 and δτ N0
V =-60mV, δµ0
Where the normalization factors µ0
V = 1 arbitrarily
delimits the fluctuation-driven regime (a mean value x and an extent δx, ∀x ∈ {µV , σV , τ N
V }). They render the fitting
of the phenomenological threshold easier, as they insure that the coefficients take similar values. It is kept constant
all along the study. The phenomenological threshold was taken as a second order polynomial and not as a linear
threshold, for two reasons: 1) unlike in an experimental study (Zerlaut et al., 2016), we are not limited by the number
of sampling points, the number of fitted coefficients can thus be higher as the probability of overfitting becomes
negligible 2) it gives more flexibility to the template, indeed the linear threshold was found a good approximation in
the fluctuation-driven regime, i.e. when the diffusion approximation holds, however, for low values of the presynaptic
frequencies, we can be far from this approximation, the additional coefficients are used to capture the firing response
in those domains.
The fitting procedure was identical to Zerlaut et al. (2016), it consisted first in a linear regression in the
phenomenological threshold space of Equation 4, followed by a non-linear optimization of Equation 3 on the firing
rate response. Both fitting were performed with the leastsq method in the optimize package of SciPy.
3.5 Calculus of the subthreshold membrane potential fluctuations
Here, we detail the analytical calculus that translate the input to the neuron into the properties of the membrane
potential fluctuations. The input is made of two Poisson shotnoise: one excitatory and one inhibitory that are both
convoluted with an exponential waveform to produce the synaptic conductances time courses.
3.5.1 Conductances fluctuations
From Campbell's theorem (Papoulis, 1991), we first get the mean (µGe, µGi) and standard deviation (σGe, σGi) of the
excitatory and inhibitory conductance fluctuations:
r
r
µGe(νe, νi) = νe Ke τe Qe
σGe(νe, νi) =
νe Ke τe
2
Qe
µGi(νe, νi) = νi Ki τi Qi
σGi(νe, νi) =
νi Ki τi
2
Qi
(5)
The mean conductances will control the input conductance of the neuron µG and therefore its effective membrane
time constant τm:
5
µG(νe, νi) = µGe + µGi + gL
τm(νe, νi) = Cm
µG
(6)
3.5.2 Mean membrane potential
Following Kuhn et al. (2004), the mean membrane potential is obtained by taking the stationary solution to static
conductances given by the mean synaptic bombardment (for the passive version of Equation 1, i.e. removing the
adaptation and spiking mechanisms). We obtain:
µV (νe, νi) = µGe Ee + µGi Ei + gL EL
µG
(7)
We will now approximate the driving force Es − V (t) of synaptic events by the level resulting from the mean
conductance bombardment: Es − µV . This will enable an analytical solution for the standard deviation σV and the
autocorrelation time σV of the fluctuations.
3.5.3 Power spectrum of the membrane potential fluctuations
Obtaining σV and τV is achieved by computing the power spectrum density of the fluctuations. In the case of Poisson
processes, the power spectrum density of the fluctuations resulting from the sum of events P SPs(t) at frequency Ks νs
can be obtained from shotnoise theory (Daley and Vere-Jones, 2007):
PV (f) = X
s∈{e,i}
Ks νs k PSPs(f)k2
(8)
where
this paper rely on the following convention for the Fourier transform: F (f) =R
PSPs(f) is the Fourier transform of the time-varying function PSP(t). Note that the relations presented in
R F (t) e−2iπf t dt.
After fixing the driving force to Es − µV , the equation for a post-synaptic membrane potential event s around µV
is
where Us = Qs
µG
Its solution is:
τm
d PSPs
dt
+ PSPs = Us H(t) e
−t
τs
(Es − µV ) and H(t) is the Heaviside function.
(cid:0)e
−t
τm − e
τs(cid:1)H(t)
−t
PSPs(t) = Us
τs
τm − τs
We take the Fourier transform:
PSPs(f) = Us
τs
τm − τs
(cid:0)
τm
2 i π f τm + 1 −
τs
2 i π f τs + 1
(cid:1)
We will need the value of the square modulus at f = 0:
As well as the integral of the square modulus:
k PSP(0)k2 = (Us · τs)2
6
(9)
(10)
(11)
(12)
Z
R
df k PSP(f)k2 = (Us · τs)2
2 (τ effm + τs)
3.5.4 Standard deviation of the fluctuations
The standard deviation follows:
R
Using Equation 13, we find the final expression for σV :
(σV )2 =
df PV (f)
Z
sX
σV (νe, νi) =
Ks νs
(Us · τs)2
2 (τ effm + τs)
3.5.5 Autocorrelation-time of the fluctuations
s
We defined the global autocorrelation time as (Zerlaut et al., 2016):
Using Equations 13 and 12, we find the final expression for τV :
R PV (f) df
PV (0)
(cid:0)R
(cid:1)−1
(cid:0)Ks νs (Us · τs)2(cid:1)
P
(cid:0)Ks νs (Us · τs)2/(τ effm + τs)(cid:1)(cid:17)
s
τV = 1
2
(cid:16)
P
s
τV (νe, νi) =
(13)
(14)
(15)
(16)
(17)
Therefore the set of Equations 7, 15 and 17 translate the presynaptic frequencies into membrane fluctuations
properties µV , σV , τV .
The previous methodological section allowed to translate the fluctuations properties µV , σV , τV into a spiking
probability thanks to a minimization procedure. The combination of the present analytical calculus and the previous
fitting procedure (on numerical simulations data) constitute our semi-analytical approach to determine the transfer
function of a single cell model: νout = F(νe, νi).
3.6 Master equation for local population dynamics
An analytical description of the cellular transfer function is the core of theoretical descriptions of asynchronous
dynamics in sparsely connected random networks (Amit and Brunel, 1997; Brunel, 2000; Renart et al., 2004).
Because we will investigate relatively slow dynamics (τ>25-50ms) (and because of the stationary formulation
of our transfer function), we will use the Markovian description developed in El Boustani and Destexhe (2009), it
describes network activity at a time scale T , for which the network dynamics should be Markovian. The choice of
the time-scale T is quite crucial in this formalism, it should be large enough so that activity can be considered as
memoryless (e.g. it can not be much smaller than the refractory period, that would introduce memory effects) and
small enough so that each neuron can fire statistically only once per time interval T . Following El Boustani and
Destexhe (2009), we will arbitrarily take T =5ms all along the study as it offers a good compromise between those two
constraints.
The formalism describes the first and second moments of the population activity for each populations. We consider
here two populations: one excitatory and one inhibitory, the formalism thus describes the evolution of five quantities:
7
the two means νe(t) and νi(t) of the excitatory and inhibitory population activity respectively (the instantaneous
population firing rate, i.e. after binning in bins of T =5ms, see discussion in El Boustani and Destexhe (2009)), the
two variances cee(t) and cii(t) of the the excitatory and inhibitory population activity respectively and the covariance
cei(t) between the excitatory and inhibitory population activities. The set of differential equations followed by those
quantities reads (El Boustani and Destexhe, 2009):
T
T
∂νµ
∂t
∂cλη
∂t
=(Fµ − νµ) + 1
=Aλη + (Fλ − νλ) (Fη − νη)+
∂2Fµ
∂νλ∂νη
2 cλη
∂Fµ
∂νλ
cλµ
+ cµη
∂Fµ
∂νη
− 2cλη
Fλ (1/T − Fλ)
Nλ
if λ = η
0
otherwise
(18)
(19)
with:
Aλη =
Note that, for the concision of the expressions, we used Einstein's index summation convention: if an index is
repeated in a product, a summation over the whole range of value is implied (e.g. we sum over λ ∈ {e, i} in the first
equation, note that, consequently, λ does not appear in the left side of the equation). Also the dependency of the
firing rate response to the excitatory and inhibitory activities has been omitted: yielding Fµ instead of Fµ(νe, νi),
∀µ ∈ {e, i}.
We will also use the reduction to first order of this system (for the phase-space analysis, see Results). This yields:
T
∂νµ
∂t
= Fµ − νµ
(20)
3.7 Afferent stimulation
In some simulations, an afferent input was present and was represented by the following piecewise double Gaussian
waveform:
(cid:16)
(cid:17)
νaf f
e
(t) = A
−( t−t0√
2τ1
e
)2H(t0 − t) + e
−( t−t0√
2τ2
)2H(t − t0)
(21)
In this afferent input, we can independently control: 1) the maximum amplitude A of the stimulation, its rising
time constant τ1 and its decay time constant τ2.
4 Results
The results are organized as follows. We construct the analytical model that describes the dynamics of a
single cortical column. We start by describing the semi-analytical workflow that enables the derivation of the
cellular transfer function: the core of this population model. Next, we investigate whether the analytical
description accurately describe population dynamics by comparing its prediction to numerical simulations.
Finally, we investigate the response of the network model subject to an external input.
8
4.1 Modeling a single cortical column
[Figure 2 about here.]
Because optical imaging presumably sample most of its signals from superficial layers, we model here the
layer II/III network: it is characterized by a strong recurrent connectivity and an important cellular diversity,
in particular one finds many types of interneurons (Markram et al., 2004; Ascoli et al., 2008). We adopt here
a very simplistic description of this network, it is made of two neuronal population: one excitatory and one
inhibitory comprising 8000 and 2000 neurons respectively. All neurons within the two population synaptically
interconnect randomly to each other with a connectivity probability of 5%. The excitatory and inhibitory
cells have the same passive properties. We nonetheless include an asymmetry between the excitatory and
inhibitory populations: because the inhibitory population includes Fast-Spiking cells that can exhibit very
high firing frequencies (Markram et al., 2004), we set its spiking mechanism sharper (more precisely its
sodium activation activation curve is steeper, see Methods) than that of excitatory cells, additionally we add
a strong spike-frequency adaptation current in excitatory cells that is absent in inhibitory cells. Those two
effects render the inhibitory neurons more excitable (see the different responses to the same current step in
Figure 2). All parameters of the cortical column can be found in Table 1.
4.2 A Markovian model to describe population dynamics
We now want to have an analytical description of the collective dynamics of this local network. We adopted
the formalism presented in El Boustani and Destexhe (2009). Two reasons motivated this choice: 1) because
10000 neurons is still far from the large network limit, finite-size effects could have a significant impact on the
dynamics and 2) because of the relative complexity of the cellular models, an analytic treatment of the type
Amit and Brunel (1997) is, to our knowledge, not accessible and would be extremely challenging to derive.
The Markovian framework proposed in El Boustani and Destexhe (2009) positively respond to those two
constraints: it is a second-order description of population activity that describes fluctuations emerging from
finite-size effects and it is applicable to any neuron model as long as its transfer function can be characterized.
In a companion study (Zerlaut et al., 2016), we developed a semi-analytical approach to characterize those
transfer functions (see next section), we will therefore incorporate this description into the formalism.
Nonetheless, the study of El Boustani and Destexhe (2009) only investigated the ability of the formalism to
describe 1) the stationary point of the network activity and 2) in a situation where the neuronal models models
had an analytic estimate for the transfer function (current-based integrate-and-fire model). Investigating
whether this description generalizes to transient dynamics and transfer functions estimated with a semi-
analytical approach is investigated in the next sections.
4.3 Transfer functions of excitatory and inhibitory cells
We briefly describe here the semi-analytical approach used to characterize the transfer function (see details in
the Methods).
9
The transfer function F of a single neuron is defined here as the function that maps the value of the
stationary excitatory and inhibitory presynaptic release frequencies to the output stationary firing rate
response, i.e. νout = F(νe, νi). This kind of input-output functions lie at the core of mean-field models
of population dynamics, reviewed in Renart et al. (2004) and is consequently the main ingredient of the
formalism adopted here (El Boustani and Destexhe, 2009). Note here that the formulation of the transfer
function imply a stationary hypothesis: both for the input (stationary Poisson processes) and the output
firing (a stationary firing rate). We will study in the following what are the limitations introduced by this
stationary hypothesis in the description of the temporal dynamics of network activity.
In a previous communication (Zerlaut et al., 2016), we found that the firing rate response of several
models (including the adaptative exponential integrate and fire considered in this study) would be captured
by a fluctuations-dependent threshold in a simple approximation of the firing probability (see Methods).
The semi-analytical approach thus consisted in making numerical simulations of single-cell dynamics for
various presynaptic activity levels (i.e. scanning various νe, νi configurations) and measuring the output firing
rate νout. All those configurations corresponded to analytical estimates of (µV , σV , τV ), we then fitted the
fluctuations-dependent threshold that bring the analytical estimate to the measured firing response. This
procedure resulted in the analytical estimates shown in Figure 2 and compared with the results of numerical
simulations.
4.4 Spontaneous activity in the cortical column
[Figure 3 about here.]
[Figure 4 about here.]
[Figure 5 about here.]
The combination of the transfer function and the markovian formalism (Equation 18 in the Methods)
yields our analytical description of the layer II-III population dynamics in a single cortical column.
We first use this analytical description to look for a physiological configuration of spontaneous activity.
There exists two qualitatively different types of spontaneous asynchronous activity (Vogels and Abbott, 2005;
Kumar et al., 2008): either the network is dominated by inhibition and the network needs an asynchronous
external excitatory drive to exhibit spontaneous activity (Amit and Brunel, 1997; Brunel, 2000) or the
network exhibits an asynchronous self-sustained activity state and just needs an initial "kick" to exit from the
quiescent state (Vogels and Abbott, 2005; Kumar et al., 2008; El Boustani and Destexhe, 2009). In the latter
case, the network is globally dominated by excitation and a strong shunting conductance effects prevents the
network from an excitatory runaway (Kuhn et al., 2004; Kumar et al., 2008). Those two behaviors are thus
determined by the membrane, synaptic and connectivity parameters. We therefore investigate how the chosen
network parameters in this study would determine the qualitative nature of the spontaneous activity state.
In the case of a single electrophysiological type (e.g. excitatory and inhibitory neurons taken as the same
integrate-and-fire model), it was shown that a simple mean-field analysis allow to predict in which situation
10
the network parameters corresponds (Brunel, 2000; Kumar et al., 2008), here we generalized this approach to
the two populations considered in this study and we investigate the behavior of our network model given the
parameters of Table 1. To this purpose, we simplified the dynamical system describing population activity
(Equation 18) to its first order so that we get a two dimensional system describing the population spiking
activity νe(t) and νi(t). We then plotted the vector field of the time evolution operator in the phase space of
the dynamical system direction and launched some trajectories with different initial conditions (see Figure
3). The result of this analysis is that, in absence of external input (νdrive
=0Hz), the only fixed point of the
system is the quiescent state (see Figure 3A). This prediction of the mean-field analysis was indeed confirmed
by numerical simulations, whatever the initial external "kick", the activity rapidly decayed (T<50ms) to the
quiescent state.
e
We conclude that, given the parameters of Table 1, our network model does not have the ability to
self-sustain activity and will need an external excitatory drive to exhibit spontaneous activity (note that
this is also consistently with recent in vivo observations in mice visual cortex, see Reinhold et al. (2015)).
Indeed, when raising the external drive, a non-quiescent fixed point appears (see Figure 3B for νdrive
=4Hz).
Numerical simulations confirmed the existence of such a fixed point at those levels of activity (see Figure 4).
The particularity of this stationary fixed-point is its asymmetry in terms of population activity, it
corresponds to νe=1.6Hz and νi=8.9Hz (i.e. corresponding to a factor 5-6 between the their respective firing
rates). The origin of this asymmetry is very naturally the asymmetry in electrophysiological properties as the
excitatory and inhibitory neurons sample statistically the same recurrent and external input. This phenomena
has been observed in extracellular recordings in human cortex (Peyrache et al., 2012), cells categorized
as Fast-Spiking (such as our inhibitory cells) were shown to fire 6-7 times more than cells categorized as
Regular-Spiking (such as our excitatory cells), an asymmetry in excitabilities thus naturally provides a
putative explanation for this phenomena (rather than specific circuitry).
e
4.5 Accuracy of the description of the spontaneous activity state
We compare more closely the numerical simulation (Figure 4) to the prediction of the Markovian description.
First, we see that there is a transient period of ∼ 400ms resulting from the onset of the external drive
(see Figure 4B-D), we will therefore evaluate stationary properties after discarding the first 500ms of the
simulation.
After this initial transient, the population activities (νe and νi) fluctuates around the stationary levels (see
Figure 4). The Markovian description predicts this phenomena as it contains the impact of finite size effects
(the network comprises 10000 neurons). In Figure 5A, we can see that the distributions of the excitatory
and inhibitory population activities are rather well predicted by the formalism (it slightly overestimates the
means of the population activities).
We also investigated whether the average neuronal and synaptic quantities were well predicted by the
Markovian formalism. Indeed, we found a very good match for all quantities (see Figure 5B,C, mean and
11
variance of membrane potential and synaptic conductances). Only the standard deviation of the membrane
potential fluctuations was underestimated (Figure 5C), presumably because of residual synchrony in the
dynamics whereas the Markovian formalism assumes a purely asynchronous regime.
4.6 Description of the response to time-varying input
[Figure 6 about here.]
We now examine whether the formalism captures the response to time-varying input. Here again, we set
the input and examine the response after 500ms of initial simulation to discard transient effects.
We first choose an afferent input of relatively low frequency content (∼ [5-20]Hz, τ1=60ms and τ2=100ms
in Equation 21). The afferent input waveform, formulated in terms of firing rate, was translated into individual
afferent spikes targeting the excitatory population. The response of the network to this input is shown in
Figure 6 in comparison with the prediction of the Markovian formalism. The excitatory population activity
raises and immediately entrains a raise of the inhibitory population. The analytical description captures
well the order of magnitude of the deflection, it only slightly underestimates the peak value (Figure 6B).
But the numerical simulations also show a marked hyperpolarization after the stimulation, the return to the
baseline level happens only ∼ 200-300 ms after the end of the stimulus, and not immediately as predicted
by the Markovian framework. Here this strong hyperpolarization is the result of the strong spike-frequency
adaptation current that remains as a consequence of the high activity evoked by the stimulus.
In the
Markovian there is no memory of the previous activity and therefore this phenomena can not be accounted
for. This typically illustrates a limitation of the analytical description provided here. Note that this is not a
fundamental limitation of the Markovian formalism, it is a limitation of this version of the formalism, that
contains only variables related to the instantaneous activity (see Discussion).
[Figure 7 about here.]
To study more precisely the temporal validity of the formalism, we modulated the network activity by
sinusoidal input and compared the response predicted by the analytical description.
The numerical simulations showed a marked resonance at ∼50Hz. Given the relatively high strength
(compared to the external input) of the excitatory-inhibitory loop, the network is close to a bifurcation toward
oscillations that are typically in the gamma range (Brunel and Wang, 2003). A sinusoidal input therefore
amplifies those frequencies (Ledoux and Brunel, 2011). Because the individual excitatory and inhibitory
post-synaptic currents approximately match each other, the theoretical study of Brunel and Wang (2003)
would predict oscillations at 50-60Hz (the bifurcation would be achieved by reducing τe), thus compatible
with the present observation.
More importantly, the main insight of this analysis is to show that the network can track very fast
temporal variations in the input, even at time scales smaller than the integration time constant of the single
neurons (van Vreeswijk and Sompolinsky, 1996). Recurrent neural networks globally behave as low-pass filters
12
(though see Ledoux and Brunel (2011) for a detailed treatment of the appearance of resonances), but with a
high cutoff frequency compared to the frequency content of thalamic input for classical artificial stimuli (e.g.
in the visual system: drifting gratings, supra-10ms flashes, etc. . . ).
Leaving apart the failure of capturing the network resonance (that is linked to this special configuration of
synaptic parameters), we conclude that in the frequency range that will be used in the following (f<50-100Hz)
the description of the formalism gives a relatively accurate description of the network response in the sense
that it accurately predicts that there should not be a frequency filtering within this range. Again, in vivo
experiments in awake mice suggested that V1 cortical networks had a cut-off frequency above this range
(∼100Hz in Reinhold et al. (2015)).
Thus, by comparing numerical simulations of network dynamics and the Markovian formalism, we showed
that, despite some discrepancies, this analytical framework describes both the spontaneous activity and the
response in the [0,100]Hz range of a sparsely connected recurrent network of distinct excitatory and inhibitory
cells.
5 Discussion
In the present study, we investigated a mean-field model of networks with different electrophysiological
properties, described using the AdEx model with conductance-based synapses. We found that the Markovian
formalism proposed in El Boustani and Destexhe (2009) was able to describe the steady-state and temporal
dynamics of such networks. Though this formalism was shown to be a relatively accurate description of
the response simulated in numerical networks, we also showed the limits of this formalism. The relative
complexity of the theoretical problem should be stressed: our model includes non-linear phenomena such as
an voltage-dependent activation curve for spike emission or spike frequency. The proposed semi-analytical
appraoch thus offers a convenient description for theoretical models where an exact analytical treatment
would not be achievable.
Unlike previous studies (Brunel, 2000; Vogels and Abbott, 2005; Kumar et al., 2008; El Boustani
and Destexhe, 2009), we considered networks of non-linear integrate-and-fire neurons with asymmetric
electrophysiological properties between excitatory and inhibitory cells. This type of network is more realistic
because it includes the adaptation properties of excitatory cells, and the fact that inhibitory cells are more
excitable and fire at higher rates. We could demonstrate the relative accuracy of the markovian formalism
(with the semi-analytical approach) in a situation including this increased complexity. The mean-field model
obtained was able to predict the level of spontaneous activity of the network, as well as its response to
external time-varying inputs.
This versatile theoretical description of the local cortical network could be improved. For example the
strong hyperpolarization of population activity after a transient rise (see Figure 6B) was shown to be missed
by the mean-field formalism. Indeed, this version does not have a memory of the previous activity levels
and thus can not account for the effect of the long-lasting spike-frequency adaptation mechanism that has
13
been strongly activated by the activity evoked by the stimulus. One could design another version of the
Markovian formalism to capture such adaptation-mediated effects. Instead of accounting for adaptation
within the transfer function (i.e. accounting only for its stationary effects), one can introduce a new variable
with a dependency on time and activity: a "population adaptation current", that can directly be derived
from the equation of the AdExp model. Investigating the accuracy of such theoretical descriptions should be
the focus of future work.
We conclude by proposing the present model as a good candidate for modeling VSDi data. Not only the
present mean-field framework gives access to the mean voltage and its time evolution, but it could easily be
extended to model VSDi signals. The present model represents a local population of cortical excitatory and
inhibitory neurons, and thus can be thought to represent a "pixel" of the VSDi. The full VSDi model could
be obtained by embedding the present local population description within a spatial model, and yield 1-D or
2-D ensembles of such pixels, and thereby model VSDi recordings in mammalian neocortex. This is the focus
of current investigations (Destexhe et al., 2015).
Acknowledgments
Research supported by the CNRS, the ICODE excellence network, and the European Community (Human
Brain Project, H2020-720270). Y.Z. was supported by fellowships from the Initiative d'Excellence Paris-Saclay
and the Fondation pour la Recherche Médicale (FDT 20150532751).
6 References
Amit DJ, Brunel N (1997) Model of global spontaneous activity and local structured activity during delay periods in
the cerebral cortex. Cerebral Cortex 7:237 -- 252.
Arieli a, Sterkin a, Grinvald a, Aertsen A, An JH (1996) Dynamics of ongoing activity: explanation of the large
variability in evoked cortical responses. Science (New York, N.Y.) 273:1868 -- 71.
Ascoli GAG, Alonso-Nanclares L, Anderson SA, Barrionuevo G, Benavides-Piccione R, Burkhalter A, Buzsáki G,
Cauli B, Defelipe J, Fairén A, Others (2008) Petilla terminology: nomenclature of features of GABAergic interneurons
of the cerebral cortex. Nature Reviews . . . 9:557 -- 568.
Berger T, Borgdorff A, Crochet S, Neubauer FB, Lefort S, Fauvet B, Ferezou I, Carleton A, Lüscher HR, Petersen
CCH (2007) Combined voltage and calcium epifluorescence imaging in vitro and in vivo reveals subthreshold and
suprathreshold dynamics of mouse barrel cortex. Journal of neurophysiology 97:3751 -- 3762.
Brette R, Gerstner W (2005) Adaptive exponential integrate-and-fire model as an effective description of neuronal
activity. Journal of neurophysiology pp. 3637 -- 3642.
Brunel N (2000) Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. Journal of
computational neuroscience 8:183 -- 208.
Brunel N, Hakim V (1999) Fast global oscillations in networks of integrate-and-fire neurons with low firing rates.
Neural computation 11:1621 -- 1671.
14
Brunel N, Wang XJ (2003) What determines the frequency of fast network oscillations with irregular neural
discharges? I. Synaptic dynamics and excitation-inhibition balance. Journal of neurophysiology 90:415 -- 430.
Chemla S, Chavane F (2010) A biophysical cortical column model to study the multi-component origin of the VSDI
signal. NeuroImage 53:420 -- 438.
Civillico EF, Contreras D (2012) Spatiotemporal properties of sensory responses in vivo are strongly dependent on
network context. Frontiers in systems neuroscience 6:25.
Contreras D, Llinas R (2001) Voltage-sensitive dye imaging of neocortical spatiotemporal dynamics to afferent
activation frequency. The Journal of neuroscience : the official journal of the Society for Neuroscience 21:9403 -- 9413.
Daley DJ, Vere-Jones D (2007) An introduction to the theory of point processes: volume II: general theory and
structure, Vol. 2 Springer Science & Business Media.
Destexhe A, Rudolph M, Paré D (2003) The high-conductance state of neocortical neurons in vivo. Nature Reviews
Neuroscience 4:739 -- 751.
Destexhe A, Zerlaut Y, Reynaud A, Chemla S, Chavane F (2015) Conductance-Based Interactions Predict The
Suppressive Effect Of Interacting Propagating Waves In Awake Monkey Visual Cortex. Society for Neuroscience,
conference abstract .
El Boustani S, Destexhe A (2009) A master equation formalism for macroscopic modeling of asynchronous irregular
activity states. Neural computation 21:46 -- 100.
Ferezou I, Bolea S, Petersen CCH (2006) Visualizing the Cortical Representation of Whisker Touch: Voltage-Sensitive
Dye Imaging in Freely Moving Mice. Neuron 50:617 -- 629.
Goodman DFM, Brette R (2009) The brian simulator. Frontiers in Neuroscience 3:192 -- 197.
Kuhn A, Aertsen A, Rotter S (2004) Neuronal integration of synaptic input in the fluctuation-driven regime. The
Journal of neuroscience : the official journal of the Society for Neuroscience 24:2345 -- 56.
Kumar A, Schrader S, Aertsen A, Rotter S (2008) The high-conductance state of cortical networks. Neural
Computation 20:1 -- 43.
Latham PE, Richmond BJ, Nelson PG, Nirenberg S (2000) Intrinsic Dynamics in Neuronal Networks. I. Theory. J
Neurophysiol 83:808 -- 827.
Ledoux E, Brunel N (2011) Dynamics of networks of excitatory and inhibitory neurons in response to time-dependent
inputs. Frontiers in computational neuroscience 5:25.
Markram H, Muller E, Ramaswamy S, Reimann Mea (2015) Reconstruction and Simulation of Neocortical Microcir-
cuitry. Cell 163:456 -- 492.
Markram H, Toledo-Rodriguez M, Wang Y, Gupta A, Silberberg G, Wu C (2004) Interneurons of the neocortical
inhibitory system. Nature reviews. Neuroscience 5:793 -- 807.
McCormick DA, Connors BW, Lighthall JW, Prince Da (1985) Comparative electrophysiology of pyramidal and
sparsely spiny stellate neurons of the neocortex. Journal of neurophysiology 54:782 -- 806.
Papoulis A (1991) Probability, random variables and stochastic processes McGraw-Hill.
Petersen CCH, Sakmann B (2001) Functionally Independent Columns of Rat Somatosensory Barrel Cortex
Revealed with Voltage-Sensitive Dye Imaging. The Journal of neuroscience : the official journal of the Society for
Neuroscience 21:8435 -- 8446.
15
Peyrache A, Dehghani N, Eskandar EN, Madsen JR, Anderson WS, Donoghue Ja, Hochberg LR, Halgren E, Cash SS,
Destexhe A (2012) Spatiotemporal dynamics of neocortical excitation and inhibition during human sleep. Proceedings
of the National Academy of Sciences of the United States of America 109:1731 -- 6.
Reinhold K, Lien AD, Scanziani M (2015) Distinct recurrent versus afferent dynamics in cortical visual processing.
Nature Neuroscience 18.
Renart A, Brunel N, Wang XJ (2004) Mean-field theory of irregularly spiking neuronal populations and working
memory in recurrent cortical networks. Computational neuroscience: A comprehensive approach pp. 431 -- 490.
Reynaud A, Masson GS, Chavane F (2012) Dynamics of local input normalization result from balanced short- and
long-range intracortical interactions in area V1. The Journal of neuroscience : the official journal of the Society for
Neuroscience 32:12558 -- 69.
Steriade M, Timofeev I, Grenier F (2001) Natural waking and sleep states: a view from inside neocortical neurons.
Journal of neurophysiology 85:1969 -- 1985.
van Vreeswijk C, Sompolinsky H (1996) Chaos in neuronal networks with balanced excitatory and inhibitory activity.
Science (New York, N.Y.) 274:1724 -- 6.
Vogels TP, Abbott LF (2005) Signal propagation and logic gating in networks of integrate-and-fire neurons. The
Journal of neuroscience 25:10786 -- 10795.
Zerlaut Y, Telenczuk B, Deleuze C, Bal T, Ouanounou G, Destexhe A (2016) Heterogeneous firing response of mice
layer V pyramidal neurons in the fluctuation-driven regime. The Journal of Physiology 594:3791 -- 808.
16
Figure 1: Schematic of the local network architecture. The network is made of Ne = (1 − g) Ntot
excitatory and Ni = g Ntot inhibitory neurons. All excitatory connections (afferent and recurrent) onto a
neuron corresponds to Ke = (1 − g) Ntot synapses of weight Qe. All inhibitory connections (afferent and
recurrent) onto a neuron corresponds to Ki = g Ntot synapses of weight Qi
17
Figure 2: Single cell models of the excitatory and inhibitory populations. Top: response to a
current step of 200pA lasting 300ms. Bottom: transfer function of the single cell, i.e. output firing rate as a
function of the excitatory (x-axis) and inhibitory (color-coded) presynaptic release frequencies. Note that the
range of the excitatory and frequencies assumes numbers of synapses (Ke=40 and Ki=10 for the excitation
and inhibition respectively). (A) Excitatory cells. Note the presence of spike-frequency adaptation and
subthreshold adaptation. (B) Inhibitory cells. Note the very narrow spike initiation dynamics (ka=0.5mV).
Also, note the steepest relation to excitation (with respect to the excitatory cell) at various inhibitory levels
as a result of the increased excitability as a result of the increased excitability of the inhibitory cell (with
respect to the excitatory cell).
18
Figure 3: Using the analytical description to look for a stable configuration of spontaneous
network activity. Phase space of the dynamical system resulting from the first order of the markovian
description, shown for two levels of external excitatory drive νdrive
. The lines represents trajectories resulting
from different initial conditions. The vector field correspond to the time-evolution operator (the arrows
represent the direction in the two-dimensional space and the color codes for the norm of the vector). (A)
Phase space in the absence of an external drive νdrive
=0Hz, the stable fixed point of the dynamics correspond
to the quiescent network state νe = νi=0Hz. (A) Phase space with an external drive νdrive
=4Hz, the stable
fixed point of the dynamics correspond now corresponds to an active state with asymmetric activity levels:
νe=1.6Hz and νi=8.9Hz (round marker).
e
e
e
19
Figure 4: Numerical simulations of the dynamics of a recurrent network of 10000 neurons (see
parameters in Table 1). Note that all plots have the same x-axis: time. (A) Sample of the spiking
activity of 500 neurons (green, 400 excitatory and red, 100 inhibitory). (B) Population activity (i.e. spiking
activity sampled in 5ms time bins across the population) of the excitatory (green) and inhibitory (red)
sub-populations. We also show the applied external drive (νdrive
(t), black line), note the slow linear increase
=4Hz and try to reduce the initial synchronization that would result from an abrupt onset. (C)
to reach νdrive
Membrane potential (top) and conductances (bottom, excitatory in green and inhibitory in red) time courses
of three randomly chosen inhibitory neurons. (D) Membrane potential and conductances time courses of
three randomly chosen excitatory neurons.
e
e
20
Figure 5: Mean field prediction of the stationary activity. Those quantities are evaluated after
discarding the initial 500ms transient. (A) Gaussian predictions of the population activities (filled curve)
compared to those observed in numerical simulations (empty bars). (B) Mean of the membrane potential
and conductances time courses. Evaluated over 3 cells for the numerical simulations (empty bars, mean and
standard deviation). (C) Standard deviation of membrane potential and conductances time courses.
21
Figure 6: Network response to a time-varying input and associated prediction of the Markovian
formalism. For all plots, the x-axis corresponds to time. Shown after 500ms of initial stimulation. (A)
Sample of the spiking activity of 500 neurons (green, 400 excitatory and red, 100 inhibitory). (B) Population
activity (in 5ms bins) of the excitatory (green) and inhibitory (red) sub-populations. Superimposed is
the mean and standard deviation over time predicted by the Markovian formalism. We also show the
(t), dotted line). (C) Membrane potential time courses of three excitatory
applied external stimulation (νaf f
cells (green, top) and three inhibitory cells (red, bottom) with the prediction of the mean and standard
deviation in time. (D) Conductance time courses of the six cells in C with the predictions of the fluctuations
superimposed.
e
22
Figure 7: Limitations of the Markovian description in the frequency domain. Response of the
network (numerical simulation and analytical description) to sinusoidal stimulation of the form νaf f
e =
5Hz(cid:0)1− cos(2πf(t− t0))(cid:1)/2. The stimulation was set on at t0=500ms. The response was fitted by a function
of the form ν(t) = A(cid:0)1 − cos(2πf(t − t0) − φ)(cid:1)/2. (A) Amplitude of the sinusoidal response (A in the fitted
response) for various frequencies. (B) Phase shift of the sinusoidal response (φ in the fitted response) for
various frequencies.
23
|
1907.02351 | 1 | 1907 | 2019-07-04T12:14:30 | Evaluation of the criticality of in vitro neuronal networks: Toward an assessment of computational capacity | [
"q-bio.NC",
"cs.ET"
] | Novel computing hardwares are necessary to keep up with today's increasing demand for data storage and processing power. In this research project, we turn to the brain for inspiration to develop novel computing substrates that are self-learning, scalable, energy-efficient, and fault-tolerant. The overarching aim of this work is to develop computational models that are able to reproduce target behaviors observed in in vitro neuronal networks. These models will be ultimately be used to aid in the realization of these behaviors in a more engineerable substrate: an array of nanomagnets. The target behaviors will be identified by analyzing electrophysiological recordings of the neuronal networks. Preliminary analysis has been performed to identify when a network is in a critical state based on the size distribution of network-wide avalanches of activity, and the results of this analysis are reported here. This classification of critical versus non-critical networks is valuable in identifying networks that can be expected to perform well on computational tasks, as criticality is widely considered to be the state in which a system is best suited for computation. This type of analysis is expected to enable the identification of networks that are well-suited for computation and the classification of networks as perturbed or healthy. | q-bio.NC | q-bio | Evaluation of the criticality of
in vitro neuronal networks:
Toward an assessment of computational capacity
9
1
0
2
l
u
J
4
]
.
C
N
o
i
b
-
q
[
1
v
1
5
3
2
0
.
7
0
9
1
:
v
i
X
r
a
Kristine Heiney∗†§, Vibeke Devold Valderhaug‡, Ioanna Sandvig‡,
Axel Sandvig‡, Gunnar Tufte†, Hugo Lewi Hammer∗, and Stefano Nichele∗
∗Department of Computer Science, Oslo Metropolitan University, Oslo, Norway
†Department of Computer Science, Norwegian University of Science and Technology, Trondheim, Norway
‡Department of Neuromedicine and Movement Science, Norwegian University of Science and Technology, Trondheim, Norway
§Email: [email protected]
Abstract -- Novel computing hardwares are necessary to keep
up with today's increasing demand for data storage and pro-
cessing power. In this research project, we turn to the brain
for inspiration to develop novel computing substrates that are
self-learning, scalable, energy-efficient, and fault-tolerant. The
overarching aim of this work is to develop computational models
that are able to reproduce target behaviors observed in in vitro
neuronal networks. These models will be ultimately be used to
aid in the realization of these behaviors in a more engineerable
substrate: an array of nanomagnets. The target behaviors will
be identified by analyzing electrophysiological recordings of the
neuronal networks. Preliminary analysis has been performed
to identify when a network is in a critical state based on the
size distribution of network-wide avalanches of activity, and the
results of this analysis are reported here. This classification of
critical versus non-critical networks is valuable in identifying
networks that can be expected to perform well on computational
tasks, as criticality is widely considered to be the state in which
a system is best suited for computation. This type of analysis
is expected to enable the identification of networks that are
well-suited for computation and the classification of networks
as perturbed or healthy.
I. INTRODUCTION
Current computing technology is based on the von Neu-
mann architecture, in which tasks are performed sequentially
and control, processing, and memory are each allocated to
structurally distinct components. With this architecture, con-
ventional computers struggle to cope with the rising demand
for data processing and storage. Furthermore, although recent
advancements in machine learning technology have conferred
great advantages to our data handling capabilities, processing
continues to be performed on conventional hardware that
has no inherent learning capabilities and thus requires huge
amounts of training data, computational time, and computing
power.
To continue to fulfill the rapidly growing computing de-
mands of the modern day, it will be necessary to develop
novel physical computing architectures that are self-learning,
scalable, energy-efficient, and fault-tolerant. The use of self-
organizing substrates showing an inherent capacity for infor-
mation transmission, storage, and modification [1] would bring
computation into the physical domain, enabling improved
efficiency through the direct exploitation of material and
physical processes for computation [2, 3]. Some key properties
of self-organizing systems that make them well-suited for
computational tasks include their lack of centralized control
and their adaptive response to changes in their environment
[4]. Such systems are composed of many autonomous units
that interact with each other and the environment through a set
of local rules to give rise to organized emergent behaviors at a
macroscopic scale. This type of spontaneous pattern formation
is fairly common in nature, and there has been recent interest
in determining how to develop interaction rules to generate
various desired emergent behaviors [5], including those geared
toward computation. In addition, it has been demonstrated
that self-organizing substrates can be used as computational
reservoirs by training a readout layer to map the output of the
physical system to a target problem [6].
The brain is an excellent example of a self-organizing
system; it shows a remarkable capacity for computation with
very little energy consumption and no centralized control, and
scientists and engineers have long looked to the structure and
behavior of the brain for inspiration. Neurons grown in vitro
self-organize into networks that show complex patterns of
spiking activity, which can be analyzed to gain insight into the
network's capacity for information storage and transmission.
This behavior indicates that in vitro neuronal networks may
serve as a suitable computational reservoir [7] and could also
provide insights into the characteristics and dynamics desired
for more engineerable substrates.
The aim of the present research project is to construct
computational models that are able to reproduce desired be-
haviors observed in electrophysiological data recorded from
engineered neuronal networks. These models will provide
insight into the behavior of the neurons and enable us to
reproduce it in other substrates. The computational capabilities
of the models and different physical substrates developed from
the models will be explored and their dynamics characterized.
This work is part of a project entitled Self-Organizing Com-
putational substRATES (SOCRATES) [8], which aims to take
inspiration from the behavior of in vitro neuronal networks
toward the development of novel self-organizing computing
hardwares based in nanomagnetic substrates.
In addition to providing an avenue for the development of
novel computational hardwares, the developed models are also
expected to provide insight into the functionality of neuronal
networks in healthy and perturbed conditions, where typi-
cal perturbations include chemical manipulation or electrical
stimulation. The dynamics of perturbed neuronal networks
will also be modeled using the developed framework and
their computational capabilities and dynamics characterized.
On the basis of this modeling, strategies of interfacing with
perturbed networks to recover their dynamics will be explored.
The behavior of perturbed networks and their capacity for
recovery will also provide insight
into the robustness of
the computational capabilities of engineered self-organizing
substrates against analogous damage or perturbation.
The remainder of this paper is organized as follows. Section
II presents an analysis method that will be used in this research
project to assess the criticality of in vitro neuronal networks
toward identifying networks that may be considered well-
suited for computation. Results from the preliminary analysis
of electrophysiological data recorded from a neuronal network
cultured on a microelectrode array (MEA) are presented and
discussed in this section. A brief overview of the long-term
plan for this research project is then given in Section III.
Section IV concludes the paper.
II. NEURONAL AVALANCHE ANALYSIS
A. Background
An analysis method based on the size distribution of neu-
ronal avalanches was applied to the analysis of an in vitro
neuronal network in this study (see Fig. 1a for an example of
such an in vitro network); this method is based on previous
analysis performed on cortical networks [9, 10]. The aim of
this method is to determine whether a given neuronal network
is in the critical state, which is presumed to be beneficial
for the network in terms of its capacity to store information
and perform computation. A system in the critical state rests
at the boundary between two qualitatively different types of
behavior. In the subcritical phase, a system shows highly
ordered behavior characterized as static or oscillating between
very few distinct states, whereas in the supercritical or chaotic
phase,
the system shows highly unpredictable, essentially
random behavior. Near the transition point between these two
regimes, the system is poised to effectively respond to a wide
range of inputs as well as store and transmit information,
making it ideal in terms of the capacity a system has for
computation [1].
It has been proposed that the brain self-organizes into a
critical state to optimize its computational properties;
the
foundations and theorized functional benefits of this behavior
have been reviewed in recent articles [11, 12]. As first defined
by Beggs and Plenz [13], a neuronal avalanche is any number
of consecutive time bins in which at
least one spike is
recorded, bounded before and after by time bins containing
2
(a)
(b)
Fig. 1: (a) Microscope image of a neuronal network cultured
on an MEA for illustrative purposes. Image taken by Ola
Huse Ramstad. (b) Definition of a neuronal avalanche. Each
dot represents a spike recorded by one of the electrodes (Ch.
1 -- 4). A time bin is active when it contains at
least one
spike and empty when there are no spikes. An avalanche is
defined as a sequence of consecutive active time bins preceded
and followed by empty bins, and the size is the number of
electrodes active during the avalanche.
no activity, as shown in Fig. 1b. In their study, Beggs and
Plenz [13] demonstrated that the size and duration of neuronal
avalanches follows a power law, indicating that the propagation
of activity in the cortex is in the critical state [14]. It has been
further demonstrated that criticality is established by a balance
between excitation and inhibition and that cortical networks
at criticality show greater dynamic range and information
capacity and transmission than networks functioning outside
of criticality [15, 16]. Studies on the spontaneous activity
of dissociated cortical networks have indicated that
these
networks tend to self-organize into the critical state over
the course of their maturation, after first showing an early
subcritical followed by an intermediate supercritical phase
[17, 18].
the present
As a long-term goal of
research project,
avalanche size distribution analysis will be applied to record-
ings obtained from different in vitro neuronal networks to
assess the criticality of the networks. Emerging network dy-
namics in healthy and perturbed conditions will be studied,
characterized, and classified. In the preliminary results pre-
sented here, the development of a single unperturbed neuronal
network as it matures is reported.
To the authors' knowledge, this work represents the first
time avalanche analysis has been applied to neurons derived
from human induced pluripotent stem cells (iPSCs). Further-
more, this type of analysis has also not yet been applied to the
characterization of the neuronal network dynamics in in vitro
disease models, which will be the focus of future work.
B. Methods
The neuronal network assessed here was prepared as fol-
lows. Human iPSCs (ChiPSC18, Takara Bioscience) were
reprogrammed using a protocol for midbrain dopaminergic
neurons adapted from previous studies [19 -- 21]. Reprogram-
ming was concluded on day 16, at which point the cells were
left to mature. The spontaneous electrophysiological activity of
the network was recorded using a 60-electrode MEA together
with the corresponding in vitro recording system (MEA2100-
System, Multi Channel Systems) and software (Multi Channel
Experimenter, Multi Channel Systems). Recordings of 6 min
were taken starting after three weeks of maturation (day in
vitro (DIV) 21), starting from the date at which the repro-
gramming was concluded, and a total of 18 recordings taken
over the period from DIV 21 to DIV 56 were analyzed.
Avalanches were detected according to the method de-
scribed by Beggs and Plenz [13]. Briefly, events were detected
using thresholding on the data after applying a bandpass filter
with a pass band of 300 Hz to 3 kHz 1. The spikes were then
binned into time bins equal to the average inter-event interval
(IEI), which is the time between events recorded across all
electrodes, and avalanches were detected as any number of
consecutive active time bins (bins containing at least one spike)
bounded before and after by empty time bins. The size of an
avalanche is defined as the number of electrodes that were
active during the avalanche.
A power law was then fitted to the avalanche size distri-
bution data using a least-squares fitting followed by nonlinear
regression with the result from the least-squares fitting for the
initial parameter values. This power law takes the form
p(s) ∝ s−α,
(1)
where s is the avalanche size, p(s) is the probability of an
avalanche having size s, and α is the power of the fitted power
law. The fit was applied over the size range of s = 2 to 59
electrodes, following previous works [9, 10]. The goodness
of fit was computed following Clauset et al. [22]. Synthetic
datasets were generated from the fitted distribution, and their
Kolmogorov -- Smirnov (KS) distances from the theoretical dis-
tribution were compared to the empirical KS distance. The
fitting was rejected if the fraction p of synthetic KS distances
that were lower than the empirical KS distance was greater
than 0.1 (p > 0.1) 2.
C. Results and discussion
The neuronal avalanche size distribution of a single in vitro
neuronal network was observed as the network matured. In
this preliminary work, no rigorous analysis was yet applied
to classify network as super- or subcritical; rather, only the
goodness of fit of the size distribution to a power law was
evaluated to assess whether the network was in a critical state
during each analyzed recording. Preliminary classification of
non-critical cases was performed by visual inspection.
The fitting results indicate that the network was already be
in the critical state at DIV 21 (Fig. 2a) and remained as such
until DIV 42, with some brief deviations or periods of low
activity. Many of these deviations from criticality were during
the early recording period, and the network was stably in the
critical state between DIVs 36 and 42. The mean power of
the fitted power law distributions in the recordings where the
network was in the critical state was α = 1.84 ± 0.07, which
is higher than the value of 3/2 reported by Beggs and Plenz
[13]. From DIV 44 until the final recording on DIV 56, the
network was no longer in the critical state, and a preliminary
visual assessment of the avalanche size distributions indicate
that activity progressed to supercritical (Fig. 2b) and finally
subcritical (Fig. 2c) during the later recordings. Supercritical
behavior is characterized by a bimodal distribution, with an
initially low slope in the log -- log plot followed by a peak in
the number of larger avalanches, whereas subcritical behavior
is characterized by exponential decay, with few or no large
avalanches; although no fittings were performed to rigorously
assess whether the networks were in either of these states,
the plots in Fig. 2 appear to be consistent with this type of
behavior.
The observed network did not ultimately settle into a critical
state in the considered timeframe, though it did pass through
a period of relatively stable critical behavior. The deviation
from criticality is likely due to the different cell types that
arise during the reprogramming of iPSCs, particularly as the
proliferation of these cells causes the composition of the
culture to change over time. When differentiating iPSCs into a
target cell type, it is impossible to avoid having other types of
cells of the same lineage; for example, glial cells are inevitably
present in iPSC-derived neuronal cultures. This is different
from networks that have been assessed in previous studies on
self-organized criticality in neuronal networks, as these have
focused solely on primary cortical networks, which can be
prepared with greater homogeneity. The heterogeneous and
time-varying cellular composition likely produces changes to
the signalling environment of the neurons, which may tem-
porarily push the network away from criticality. Additionally,
the neurons assessed here were dopaminergic neurons, which
are likely to show a different course of maturation in terms
1Code for spike detection is available at https://github.com/SocratesNFR/
MCSspikedetection.
2Code for avalanche detection and goodness of fit evaluation is available
at https://github.com/SocratesNFR/avalanche.
3
a power law distribution, further analysis will be performed
to classify the state as sub- or supercritical; these cases are
known to show exponential and bimodal size distributions,
respectively, as described previously. This analysis will also
be applied to networks that have been perturbed chemically or
electrically to characterize how their behavior deviates from
that of unperturbed networks and gain insights into how the
perturbation may interrupt normal function.
III. PLAN FOR FUTURE RESEARCH
This work represents a first step in a larger research project,
which will be described briefly here. The plan for this research
project is divided into four stages. In the first stage, a data
analysis framework will be developed, with the avalanche
analysis method described here constituting a crucial part of
this framework. The framework involves methods of extracting
meaningful features from electrophysiological data recorded
from in vitro neuronal networks. Such features include con-
ventional parameters considered in electrophysiological data
analysis, such as the mean firing rate, as well as more complex
measures, such as entropy and measures of connectivity. The
connectivity of the engineered networks will also be modeled
using graph theory approaches. The avalanche method pre-
sented here represents a useful tool for classifying networks
as critical or non-critical. Other methods of classification and
clustering of networks will also be explored.
The second stage of the project involves the construction of
computing models, such as cellular automata (CAs), random
Boolean networks (RBNs), and recurrent neural networks
(RNNs), that show behavior similar to that of the neuronal
networks [23, 24]. The data analysis framework developed in
the first phase will be used as a method of capturing the target
behavior to be reproduced in the models, and this framework
will be continually refined as we improve our understanding
of the important aspects of neuronal behavior that contribute
to their computational capabilities. These computing models
are developed using evolutionary algorithms with appropriate
fitness functions defined on the basis of the target behavior.
Important features of the models, such as their input and output
mappings and number of states, will be explored, and the
dynamics of the models will be characterized.
The third phase involves the use of the developed models
and the in vitro neuronal networks as reservoirs to perform
computational and classification tasks as a proof-of-concept
using reservoir computing. The models from the second stage
will be refined based on their performance as computing
reservoirs.
The final stage consists of the exploration of the application
of the models developed in the second stage to the study
of engineered neuronal networks under perturbed conditions
mimicking pathologies related to the central nervous system
(CNS). Networks that have had their synaptic function per-
turbed will be modeled and analyzed using the developed
methods to characterize how their behavior differs from that
of unperturbed networks. Methods of interfacing with the per-
(a)
(b)
(c)
Fig. 2: Probability distribution functions for three representa-
tive cases. (a) DIV 21: The fitting indicates the network is
in a critical state, with α = 1.85 (p = 0.13). The power-
law fitting result is shown as a red line. (b) DIV 51: The
network appears to be in a supercritical state with a bimodal
distribution. (c) DIV 52: The network appears to be in a
subcritical state with an exponential distribution. In (b) and (c),
the dashed red lines correspond to α = 1.85 for comparison
with the distribution shown in (a).
of criticality than cortical neurons. It is possible that these
types of networks show more complex oscillatory behaviors
as they mature, or they may eventually settle into a critical
state given enough time. Further work is necessary to capture
the expected time course of the development of the criticality
of such networks.
In future work, this analytical framework will be applied
to recordings from additional unperturbed networks to gain a
better understanding of the typical progression of networks as
they mature in vitro. In cases where the data do not follow
4
turbed networks to restore their dynamics to the unperturbed
state will then be explored.
IV. CONCLUSION
The aim of this research project is to extract meaningful
behaviors and features from electrophysiological data recorded
from in vitro neuronal networks and construct models that
reproduce these behaviors toward the eventual realization of
novel computing substrates based in nanomagnetic materials.
This paper reported the application of an avalanche size
distribution analysis to electrophysiological data, representing
a first step in the development of an analytical framework
to extract target behaviors from such data. The preliminary
results reported here demonstrate emerging behavior that does
not settle into criticality within the investigated time frame;
further work is needed to better characterize the time course
of the development of criticality in the networks studied in this
work and characterize how this affects the network's suitability
for computation.
With this type of analysis, it can be determined if a network
is in a critical state, which gives an indication of its suitability
for use in computational tasks. In addition to the computational
applications of this analysis, it is also expected to be useful in
distinguishing healthy and perturbed networks and to provide
insight into how different diseases affect neuronal connectivity
and communication, which will be the target of future work.
ACKNOWLEDGEMENTS
This work was conducted as part of the SOCRATES project,
which is partially funded by the Norwegian Research Coun-
cil (NFR) through their IKTPLUSS research and innovation
action on information and communication technologies under
the project agreement 270961.
REFERENCES
[1] C. G. Langton, "Computation at
the edge of chaos:
Phase transitions and emergent computation," Physica D,
vol. 40, pp. 12 -- 37, 1990.
[2] S. Stepney, S. Rasmussen, and M. Amos, Computational
Matter, 1st ed. Springer Publishing Company, Incorpo-
rated, 2018.
[3] J. H. Jensen, E. Folven, and G. Tufte, "Computation in
artificial spin ice," The 2018 Conference on Artificial
Life: A Hybrid of the European Conference on Artificial
Life (ECAL) and the International Conference on the
Synthesis and Simulation of Living Systems (ALIFE),
no. 30, pp. 15 -- 22, 2018.
[4] F. Heylighen, "The science of self-organization and adap-
tivity," in in: Knowledge Management, Organizational
Intelligence and Learning, and Complexity, in: The En-
cyclopedia of Life Support Systems, EOLSS. Publishers
Co. Ltd, 1999, pp. 253 -- 280.
[5] R. Doursat, H. Sayama, and O. Michel, "A review of mor-
phogenetic engineering," Natural Computing, vol. 12, pp.
517 -- 535, 2013.
5
[6] B. Schrauwen, D. Verstraeten, and J. Van Campenhout,
"An overview of reservoir computing: theory, applica-
tions and implementations," in Proceedings of the 15th
European Symposium on Artificial Neural Networks,
2007, pp. 471 -- 482.
[7] P. Aaser, M. Knudsen, O. Huse Ramstad, R. van de
Wijdeven, S. Nichele, I. Sandvig, G. Tufte, U. S. Bauer,
Ø. Halaas, S. Hendseth et al., "Towards making a cyborg:
A closed-loop reservoir-neuro system," in Proceedings of
the European Conference on Artificial Life 2017. MIT
Press, 2017.
[8] (2018) Socrates: Self-organizing computational sub-
https://www.ntnu.edu/
[Online]. Available:
strates.
socrates
[9] V. Pasquale, P. Massobrio, L. L. Bologna, M. Chiap-
palone, and S. Martinoia, "Self-organization and neu-
ronal avalanches in networks of dissociated cortical neu-
rons," Neuroscience, vol. 153, pp. 1354 -- 1369, 2008.
[10] P. Massobrio, V. Pasquale, and S. Martinoia, "Self-
organized criticality in cortical assemblies occurs in con-
current scale-free and small-world networks," Scientific
Reports, 2015.
[11] W. L. Shew and D. Plenz, "The functional benefits of
criticality in the cortex," The Neuroscientist, vol. 19,
no. 1, pp. 88 -- 100, 2013.
[12] J. Hesse and T. Gross, "Self-organized criticality as a
fundamental propertie of neural systems," Frontiers in
Systems Neuroscience, vol. 8, 2014.
[13] J. M. Beggs and D. Plenz, "Neuronal avalanches in neo-
cortical circuits," The Journal of Neuroscience, vol. 23,
no. 35, pp. 11 167 -- 11 177, 2003.
[14] P. Bak, C. Tang, and K. Wiesenfeld, "Self-organized
criticality: An explanation of the 1/f noise," Physical
Review Letters, vol. 59, pp. 381 -- 384, Jul 1987.
[15] W. L. Shew, H. Yang, T. Petermann, R. Roy, and
D. Plenz, "Neuronal avalanches imply maximum dy-
namic range in cortical networks at criticality," Journal of
Neuroscience, vol. 29, no. 49, pp. 15 595 -- 15 600, 2009.
[16] W. L. Shew, H. Yang, S. Yu, R. Roy, and D. Plenz,
"Information capacity and transmission are maximized
in balanced cortical networks with neuronal avalanches,"
Journal of Neuroscience, vol. 31, no. 1, pp. 55 -- 63, 2011.
[17] C. Tetzlaff, S. Okujeni, U. Egert, W org otter, and
M. Butz, "Self-organized criticality in developing neu-
ronal networks," PLoS Computational Biology, vol. 6,
no. 12, 2010.
[18] Y. Yada, T. Mita, A. Sanada, R. Yano, D. J. Bakkum,
A. Hierlemann, and H. Takahashi, "Development of neu-
ral population activity toward self-organized criticality,"
Neuroscience, pp. 55 -- 65, 2017.
[19] A. Kirkeby, J. Nelander, and M. Parmar, "Generating
regionalized neuronal cells from pluripotency, a step-by-
step protocol," Frontiers in Cellular Neuroscience, vol. 6,
p. 64, 2012.
[20] A. Kirkeby, S. Nolbrant, K. Tiklova, A. Heuer, N. Kee,
T. Cardoso, D. R. Ottosson, M. J. Lelos, P. Rifes,
[21] D. Doi, B. Samata, M. Katsukawa, T. Kikuchi,
A. Morizane, Y. Ono, K. Sekiguchi, M. Nakagawa,
M. Parmar, and J. Takahashi, "Isolation of human in-
duced pluripotent stem cell-derived dopaminergic pro-
genitors by cell sorting for successful transplantation,"
Stem Cell Reports, vol. 2, no. 3, pp. 337 -- 350, 2014.
[22] A. Clauset, C. R. Shalizi, and M. E. J. Newman,
"Power-law distributions in empirical data," SIAM Re-
view, vol. 51, no. 4, pp. 661 -- 703, 2009.
S. B. Dunnett, S. Grealish, T. Perlmann, and M. Parmar,
"Predictive markers guide differentiation to improve graft
outcome in clinical translation of hesc-based therapy for
parkinson's disease," Cell Stem Cell, vol. 20, no. 1, pp.
135 -- 148, 2017.
[23] S. Nichele and A. Molund, "Deep learning with cel-
lular automaton-based reservoir computing," Complex
Systems, vol. 26, 2017.
[24] S. Nichele and M. S. Gundersen, "Reservoir computing
using nonuniform binary cellular automata," Complex
Systems, vol. 26, 2017.
6
|
1410.6769 | 2 | 1410 | 2015-01-31T17:59:34 | Dynamical criticality in the collective activity of a population of retinal neurons | [
"q-bio.NC",
"cond-mat.dis-nn",
"cond-mat.stat-mech"
] | Recent experimental results based on multi-electrode and imaging techniques have reinvigorated the idea that large neural networks operate near a critical point, between order and disorder. However, evidence for criticality has relied on the definition of arbitrary order parameters, or on models that do not address the dynamical nature of network activity. Here we introduce a novel approach to assess criticality that overcomes these limitations, while encompassing and generalizing previous criteria. We find a simple model to describe the global activity of large populations of ganglion cells in the rat retina, and show that their statistics are poised near a critical point. Taking into account the temporal dynamics of the activity greatly enhances the evidence for criticality, revealing it where previous methods would not. The approach is general and could be used in other biological networks. | q-bio.NC | q-bio |
Dynamical criticality in the collective activity of a population of retinal neurons
Thierry Mora,1 St´ephane Deny,2 and Olivier Marre2
1Laboratoire de physique statistique, ´Ecole normale sup´erieure,
CNRS and UPMC, 24 rue Lhomond, 75005 Paris, France
2Institut de la Vision, INSERM and UMPC, 17 rue Moreau, 75012 Paris, France
(Dated: October 11, 2018)
Recent experimental results based on multi-electrode and imaging techniques have reinvigorated
the idea that large neural networks operate near a critical point, between order and disorder [1, 2].
However, evidence for criticality has relied on the definition of arbitrary order parameters, or on
models that do not address the dynamical nature of network activity. Here we introduce a novel
approach to assess criticality that overcomes these limitations, while encompassing and generalizing
previous criteria. We find a simple model to describe the global activity of large populations of
ganglion cells in the rat retina, and show that their statistics are poised near a critical point. Taking
into account the temporal dynamics of the activity greatly enhances the evidence for criticality,
revealing it where previous methods would not. The approach is general and could be used in other
biological networks.
Complex brain functions usually involve large num-
bers of neurons interacting in diverse ways and spanning
a wide range of time and length scales. At first sight,
systems of inanimate matter seem to enjoy more regular
properties, but they may also display complex and het-
erogeneous behaviors when in a critical state, which cor-
responds to special points of the parameter space. Think-
ing about the brain as a system near a critical point has
been an attractive idea, which has gained attention after
the suggestion that such critical states could be achieved
in a self-organized manner, without fine-tuning [3], but
also the proposal that operating near a critical point
could be beneficial for computation [4].
Despite considerable work on the foundations of a the-
ory of critical neural networks (see [5, 6] for recent exam-
ples), the validation of these ideas by experimental data
has proven difficult, largely because it requires to mea-
sure the detailed activity of large populations of neurons.
Recent progress has been made possible by the advance of
multi-electrode or imaging techniques, which have helped
detect signatures of criticality in a variety of neural con-
texts. Two lines of empirical evidence, rooted in different
approaches to critical systems, have been followed, albeit
with little intersection.
In line with the original ideas
of self-organised criticality and branching processes, the
statistics of neural avalanches in cortical layers has been
shown to display power-law statistics [7 -- 10]. This ob-
servation is indicative of the critical nature of the sys-
tem's dynamics, but it relies on arbitrary choices, such
as the number of units considered, the minimal silence
time to call the end of an avalanche, or the definition of
a neural event itself. The stability exponents of the neu-
ral dynamics, which become positive at the transition
to chaos, have also been used as signatures of critical-
ity [11]. This criterion relies on a continuous description
of neural activity, which is inappropriate for codes rely-
ing on combinations of spikes and silences. Both these
approaches address the dynamical aspect of criticality.
They require the definition of an ad hoc order parameter
(avalanche size, firing rates), which may not be the most
relevant one for neural activity. A second line of enquiry,
which focuses on the thermodynamic aspect of critical-
ity, has been to study the frequency of combinations of
spikes and silences in a neural population as a statisti-
cal mechanics problem, and explore its properties in the
thermodynamic limit [12, 13], using non-parametric sig-
natures such as the divergence of the specific heat to
demonstrate critical behaviour [14, 15]. These analyses
have however been restricted to the simultaneous distri-
bution of neural activity, with no regard to its dynami-
cal properties, which may be strongly out of equilibrium
and may contain important clues about critical behavior.
Because of their respective limitations, neither of these
approaches gives us a coherent picture for assessing and
understanding all aspects of criticality.
In this paper we overcome these limitations by intro-
ducing a framework for analysing the critical dynamics of
neural networks. We apply a thermodynamic approach
to the population's spiking activity over long periods,
treating time as an extra dimension. We propose a gener-
alized, time-normalized specific heat of spike trains as an
indicator of critical dynamics. The approach accounts for
the combinatorial nature of the code, and does not rely
on the choice of an order parameter. It reduces to the
usual notion of dynamical criticality through the stabil-
ity exponents of the dynamics when the number of spikes
can be approximated as a continuous variable. It is also
equivalent to the thermodynamic criticality of [13, 15]
when time correlations are ignored. We apply our crite-
rion to a dense population of ganglion cells recorded in
the rat retina. We will show that the dynamics of this
population are close to a critical point, where the specific
heat diverges. This divergence appears to be much more
pronounced once the temporal dynamics are taken into
account.
To describe the discrete spiking activity of a population
of N neurons, we divide time into small windows of length
∆t, and assign a binary variable σi;t = 1 if neuron i has
spiked at least once within window t, and 0 otherwise. ∆t
must be small enough so that two spikes are unlikely to
2
where δE = E −(cid:104)E(cid:105) denotes fluctuations from the mean
energy, and (cid:104)·(cid:105)β denotes averages taken under probability
law Pβ (see Appendix B). The specific heat has a clear
biological interpretation in terms of the spike train statis-
tics: it is the normalized variance of the surprise of neural
spike trains, Var(log P )/N L.
It quantifies the breadth
of codeword utilization: c(β) = 0 means that all uti-
lized codewords have uniform usage probability, whereas
a large c(β) means that the code is balanced between
a few frequent codewords and many more rare code-
words [2]. We included the normalization N L because
the variance of the surprise is expected to be an exten-
sive quantity scaling linearly with the system size, taken
both across neurons and time. Thus, in the limit where
spiking events σi,t are independent or weakly correlated,
c(β) should converge to a finite value as N and L → ∞.
For example, if all spiking events were independent with
the same spiking probabibility p in each time window, we
would have c(β) = β2(pq)β(log p− log q)2/(pβ + qβ) with
q = 1 − p, for all N and L (see Appendix B). However,
if the system is strongly interacting (between neurons,
across time, or both) the specific heat may diverge for a
certain critical value of the control parameters. Treating
time windows and neurons on equal footing allows us to
address both the many-body nature of the problem and
its critical dynamics with a single criterion. Since this
criterion is based on the surprise, which follows directly
from the probabilistic nature of the process, it does not
require us to choose an order parameter (spike rates, size
of avalanche, etc.).
Using the divergence of this specific heat as a diag-
nostic tool for criticality generalizes previous approaches.
Firstly, in the limit L = 1, where codewords are simul-
taneous combinations of neurons and silences, with no
regards to the dynamics, we recover the static thermo-
dynamic approach of [15]. Secondly, the method is con-
sistent with the notion of dynamical criticality based on
stability exponents. Let us assume that the dynamics is
well described by a single projection of the spikes onto a
i=1 σi,t, and linearized
continuous variable, e.g. Kt =(cid:80)N
to a Gaussian, Markovian dynamics
Kt+1 = aKt + b + Gaussian noise,
(3)
where the stability exponent of the dynamics is log(a) <
0. This system is critical for a ∼ 1; above the transition,
the linearized dynamics breaks down as the system be-
comes chaotic. The specific heat of this model at β = 1,
(cid:18)
c(β = 1) ∼
N − (cid:104)K(cid:105)
(cid:104)K(cid:105)
log
(cid:19)2 Var(K)
1 + a
1 − a
,
(4)
N
diverges at the critical point a = 1 (see Appendix
C). Lastly, the approach can detect criticality in sim-
ple models of neural avalanches. Consider the spik-
ing model proposed in [7], where a neuron i spikes
at time t + 1 in response to a pre-synaptic neuron j
spiking at time t with probability pij (Appendix D).
FIG. 1: The model captures the global dynamics
of the network.
(a) Predicted versus observed con-
nected correlation functions C3 = P (Kt, Kt+1, Kt+2) −
P (Kt+2)P (Kt+1)P (Kt) between the total number of spik-
ing neurons in three consecutive time windows of length
∆t = 10ms, for a subnetwork of N = 61 neurons. (b) and
(c) Model prediction for the size and duration of avalanches,
with different temporal ranges v, for the same subnetwork of
N = 61 neurons. An avalanche is a series of non-silent 10 ms
windows, ended by a silent window. While a model of inde-
pendent spikewords (v = 0) is a poor predictor of avalanche
statistics, including time correlations over a few time windows
greatly improves the prediction. (d) The distribution of the
number of spiking neurons in a window ∆t = 10ms (black
curve) is exactly fitted by the model, by construction. By
contrast, it is not well predicted by a Gaussian model (red
curve).
occur in the same window. In the following we will take
∆t = 10ms. The probability of a given multi-neuron
spike train between t = 1 and t = L, or generalized
"codeword" {σi,t}, can formally be written in a Boltzman
form:
Pβ({σi,t}) =
1
Z(β)
e−βE({σi,t}),
(1)
where Z(β) is a normalization constant. By analogy to
equilibrium statistical mechanics, E is interpreted as the
energy of the spike train. In information-theoretic terms,
the surprise of the spike train is related to its energy
through − log P = βE + log Z(β). β is an adjustable
control parameter equivalent to an inverse temperature,
set to 1 by convention to describe the observed spike
statistics. Its function is to study the parameter space of
models in the vicinity of the actual system at β = 1, and
thus assess its proximity to a critical state.
One possible indicator for detecting a critical point is
the specific heat [2, 16], defined in our formalism as:
c(β) =
1
N T
β
∂
∂β
β2 ∂
∂β
log Z(β)
β
=
β2
N L
(cid:104)δE2(cid:105)β,
(2)
−0.00200.0020.004−0.00200.0020.004C3,observedC3,predicteda.0204010−1010−810−610−410−2100NumberofspikesKProbabilityd. GaussianmodelObserved10010110210310−510−410−310−210−1Avalanchesize(numberofspikes)Probabilityb. observedv=0v=1v=2v=3v=4v=510110210310−510−410−310−210−1Avalancheduration(ms)Probabilityc. observedv=0v=1v=2v=3v=4v=5 3
ter ω = (1/N )(cid:80)
This model is parametrized by the branching parame-
ij pij, which controls the spread of neu-
ral avalanches. At the critical point ω = 1 the system
exhibits avalanches with power-law statistics. We esti-
mated the specific heat c(β = 1) of that model numer-
ically, and found it to diverge with the system size pre-
cisely at the critical value of the branching parameter
ω = 1 (Fig. S1).
In sum, the specific heat, when de-
fined on the temporal statistics of spike trains, allows
us to detect dynamical critical transitions, without hav-
ing to know the order parameter or the definition of an
avalanche.
Our goal is to apply our criterion to the spiking activity
of a dense population of N = 185 retinal ganglion cells
in the rat retina [17], stimulated by films of randomly
moving bars and binned with ∆t = 10ms (see Methods).
However, to carry out our analysis we first need to learn
a probabilistic law P ({σi;t}) from the spike trains, which
in itself can be a daunting task. We do so by employing
the principle of maximum entropy [12, 18, 19], which con-
sists in finding the least constrained distribution of spike
trains (i.e. of maximum entropy −(cid:80) P log P ) consistent
E = −(cid:88)
h(Kt) −(cid:88)
v(cid:88)
with a few selected observables of the data (see Appendix
E). In [13] the global network activity of the salaman-
der retina was modeled by constraining the distribution
P (K) of the total number of spikes in the population (see
also [20]). The inferred model was shown to be near a
critical point. However that choice of constraints did not
address the dynamical nature of the spike trains. To do
that while making as few additionnal assumptions as pos-
sible, we also constrain a dynamical quantity -- the joint
distribution of Kt at two different times Pu(Kt, Kt+u).
This leads to a family of time translation invariant mod-
els of the form in Eq. 1 with:
Ju(Kt, Kt+u),
(5)
t
t
u=1
where v is the model's temporal range -- the larger the v,
the more accurate the model. Applying the maximum en-
tropy principle to trajectories rather than instantaneous
states is sometimes also refered to as the maximum cal-
iber method [21]. The model is learned by fitting the
parameters h(K) and Ju(K, K(cid:48)) to the data using the
technique of transfer matrices (see Appendices G and H).
We find that a temporal range of v = 4 suffices to account
for the temporal correlations of K (see Appendix I and
Figs. S2 and S3).
The obtained model reproduces key dynamical features
of the data. Fig. 1a compares data and model for the
joint distribution of the numbers of spikes in three con-
secutive time windows, showing excellent agreement de-
spite this observable not being fitted by the model. More
importantly, the model predicts well the distributions of
size and duration of neural avalanches, defined as con-
tinuous epochs of K > 0, as shown in Fig. 1b and c for
a subset of N = 61 neurons. The agreement extends
over seconds, way beyond the model's temporal range of
FIG. 2: Divergence of the specific heat of spike trains.
(a) Specific heat c(β) of spike trains of the entire population
(N = 185), as a function of the temperature 1/β, for an
increasing temporal range v. Temperature β = 1 corresponds
to the observed statistics of spike trains. The curve with
v = 4, which fully accounts for the dynamics of the spike
trains, shows a markedly higher peak than that obtained from
the statistics of instantaneous codewords (v = 0). (b) Specific
heat of spike trains for subnetworks of increasing sizes N , for
v = 4. Each point is averaged over 100 random subnetworks
for N ≤ 50, and shows one representative network for N = 61
and 97. The error bars show standard deviations. The peak
increases with network size, indicating a divergence in the
thermodynamic limit.
v × 10ms = 50ms. Although we will not use avalanche
statistics to discuss criticality in this paper, as is often
done [7 -- 10], the success of our model in predicting them
demonstrates its ability to capture complex, collective
dynamical behaviour.
Simplifying the model further does not capture impor-
tant statistics of the data. We could make a continuous
approximation for Kt and constrain only the first two
moments of the distributions. This approximation would
yield an autoregressive model generalizing Eq. 3 (see Ap-
pendix F). However, the statistics of such models would
all be Gaussian, in plain contradition with the observed
distribution of spikes P (K), see Fig. 1d. Since the tail
of that distribution is related to the collective properties
of the population [13], it is important to account for it
fully, and our model is the simplest one that does that.
0.90.9511.051.105101520Temperature1/βSpecificheatc(β)a. v=0v=1v=2v=3v=40.90.9511.051.105101520Temperature1/βSpecificheatc(β)b. N=5N=10N=20N=30N=40N=50N=61N=97N=185 Confident that our model gives a precise account of the
temporal dynamics of the global network activity, we can
use it to estimate its specific heat. Fig. 2A represents
the specific heat of the learned models (Eq. 5) for all
N = 185 neurons as a function of the temperature 1/β,
for different choices of the temporal range v. The special
case v = 0, in which time correlations are ignored, shows
only a moderate peak in specific heat, and far from β = 1.
By contrast, including time correlations (v > 0) greatly
enhances the peak, which rapidly approaches β = 1 as
the temporal range v is increased. Fig. 2B shows how
the peak in specific heat behaves for random subgroups
of neurons of increasing size, for v = 4. Similarly, the
peak becomes larger, sharper and closer to β = 1 as the
network size grows. These are striking results, if we recall
that all these curves would fall on top of each other for
independent (or weakly correlated) spiking events. The
unusal scalings of Fig. 2 suggest that the system is indeed
close to a critical point. But they also show that both the
collective behaviour of the population and the temporal
correlations play a crucial role in revealing the critical
properties of the network. In fact, the convergence of the
peak of the specific heat towards β = 1 is only apparent
when time correlations are taken into account (v > 0),
as illustrated by Fig. 3A. This is in contrast with the
results of [13], which found signatures of criticality even
for v = 0, although this apparent disagreement may be
attributed to differences between species (the rat having
much higher average firing rates in their retinal ganglion
cells than the salamander).
Although the peak of the specific heat is a some-
what abstract quantity, the fact that it increases and
approaches β ∼ 1 implies that the normalized variance
of the surprise c(β = 1) = Var(log P )/N L also increases
with the system size, as shown in Fig. 3B. These variances
are extremely high compared to that we would obtain if
all spiking events were independent, cinde(β = 1) = 0.38,
indicating a very wide spectrum of codeword usage.
For the sake of simplicity and tractability, we have here
only modelled the global network activity of the retina.
Although these models capture important features of the
dynamics (Fig. 1), more detailed models accounting for
the full temporal cross-correlations between individual
neurons [22, 23] could provide us with a more precise
description of the spiking dynamics, and better approx-
imations to the specific heat curves.
In principle our
inference procedure may also depend on the choice of
window size ∆t. We repeated the analysis for windows
of 5ms, and found the same results, with an excellent
agreement between models that have a different ∆t but
the same temporal range v×∆t expressed in seconds (see
Appendix J and Figs. S4 and S5).
We have introduced a framework for studying the col-
lective dynamics of a population of neurons. This for-
malism provides us with a non-parametric criterion for
detecting the proximity to a critical state, whether this
criticality stems from strong collective effects in the pop-
ulation, from critical dynamics at the edge of chaos, or
4
FIG. 3: Finite-size scaling. (a) Position of the peak 1/β in
specific heat (see Fig. 2) as a function of network size N , for
increasing time ranges v. Accounting for the dynanics of spike
trains (v > 0) gives peaks that are much closer to the tem-
perature of real spike trains (β = 1) than for instantaneous
spikewords (v = 0). (b) The specific heat of real spike trains,
c(β = 1), is equal to the variance of the surprise per neuron
and per unit time, Var(log P )/N L. This variance increases
with the system size N and with the temporal range v. Note
that the v = 5 curves are very close to the v = 4 ones up to
N = 61 (above which they are not calculated).
from both, thus generalizing previous approaches. When
we apply our approach to large-scale recordings in the
retina, we find that the population dynamics are very
close to a critical state. Compared to the static thermo-
dynamic approach of [13, 15], which focused on the statis-
tics of instantaneous codewords, the peak in specific heat
that we find is 10 times larger, and much closer to the sys-
tem's actual temperature of 1. Our results suggest that
although simultaneous correlations between neurons are
an important marker of near-critical behavior, account-
ing for their dynamical component greatly enhances our
confidence and understanding of it.
The idea that biological systems may operate near a
critical point is not restricted to the case of neurons
[2], with evidence in systems as diverse as the cochlea
[24], immune repertoires [25], natural images [26], ani-
mal flocks [27 -- 29] or the regulation of genes in early fly
developpment [30], to name but just a few, and we ex-
pect our approach to be useful when both the collective
behaviour and the dynamics play an important role.
We thank William Bialek and Gasper Tkacik for help-
ful comments on the manuscript. This work was sup-
ported by the DEFI-SENS 2013 program from the Centre
National de la Recherche Scientifique, by the Agence Na-
tionale pour la Recherche (ANR: OPTIMA), and by the
French State program "Investissements d'Avenir" man-
aged by the Agence Nationale de la Recherche [LIFE-
SENSES: ANR-10-LABX-65]. SD was supported by a
PhD fellowship from the region Ile-de-France.
Methods
Retinal recordings. Recordings were performed on the
Long-Evans adult rat. In brief, animals were euthanized
according to institutional animal care standards. The
010020011.051.11.151.2NetworksizeNPeaktemperaturea. v=0v=1v=2v=3v=4v=5010020005101520NetworksizeNVar(surprise)/NLb. v=0v=1v=2v=3v=4v=5 retina was isolated from the eye under dim illumination
and transferred as quickly as possible into oxygenated
AMES medium. The retina was then lowered with the
ganglion cell side against a multi-electrode array whose
electrodes were spaced by 60 microns, as previously de-
scribed [17]. Raw voltage traces were digitized and
stored for off-line analysis using a 252-channel preampli-
fier (MultiChannel Systems, Germany). The recordings
were sorted using custom spike sorting software devel-
oped specifically for these arrays [17]. We extracted the
activity from 185 neurons with satisfying standard tests
of stability and limited number of refractory period vio-
lations.
Visual stimulation. Our stimulus was composed of
several black bars moving randomly on a gray back-
ground. The trajectory was a random walk with a restor-
ing force to keep the bar close to the array (see Appendix
A for details). The stimulus was displayed using a Digi-
tal Mirror Device and focused on the photoreceptor plane
using standard optics. The analysed data corresponds to
one hour (L = 360, 000 with ∆t = 10ms) of recordings.
5
[1] Chialvo DR (2010) Emergent complex neural dynamics.
[19] Jaynes ET (1957) Information theory and statistical me-
Nature Physics 6:744 -- 750.
[2] Mora T, Bialek W (2011) Are biological systems poised
at criticality? J Stat Phys 144:268 -- 302.
[3] Bak P, Tang C, Wiesenfeld K (1988) Self-organized crit-
icality. Phys Rev, A 38:364 -- 374.
[4] Bertschinger N, Natschlager T (2004) Real-time compu-
tation at the edge of chaos in recurrent neural networks.
Neural Comput 16:1413 -- 36.
[5] Levina A, Herrmann JM, Geisel T (2007) Dynamical
synapses causing self-organized criticality in neural net-
works. Nat Phys 3:857 -- 860.
[6] Magnasco MO, Piro O, Cecchi GA (2009) Self-tuned
critical anti-hebbian networks. Phys Rev Lett 102:258102.
[7] Beggs JM, Plenz D (2003) Neuronal avalanches in neo-
cortical circuits. J Neurosci 23:11167 -- 77.
[8] Petermann T, et al. (2009) Spontaneous cortical activity
in awake monkeys composed of neuronal avalanches. Proc
Natl Acad Sci USA 106:15921 -- 6.
[9] Friedman N, et al. (2012) Universal critical dynamics in
high resolution neuronal avalanche data. Phys Rev Lett
108:208102.
[10] Tagliazucchi E, Balenzuela P, Fraiman D, Chialvo DR
(2012) Criticality in large-scale brain fmri dynamics un-
veiled by a novel point process analysis. Front. Physio.
3:15.
[11] Solovey G, Miller KJ, Ojemann JG, Magnasco MO, Cec-
chi GA (2012) Self-regulated dynamical criticality in hu-
man ecog. Front. Integr. Neurosci. 6:1 -- 9.
[12] Schneidman E, Berry MJ, Segev R, Bialek W (2006)
Weak pairwise correlations imply strongly correlated net-
work states in a neural population. Nature 440:1007 -- 12.
[13] Tkacik G, et al. (2013) The simplest maximum entropy
model for collective behavior in a neural network. J. Stat.
Mech. 2013:P03011.
[14] Tkacik G, Schneidman E, Berry MJ, Bialek W (2006)
Ising models for networks of real neurons. arXiv :q --
bio/0611072.
[15] Tkacik G, et al. (2014) Thermodynamics for a network
of neurons: Signatures of criticality. arXiv :1407.5946.
[16] Chaikin PM, Lubensky TC (1995) Principles of Con-
(Cambridge University Press,
densed Matter Physics
Cambridge).
[17] Marre O, et al. (2012) Mapping a complete neural pop-
ulation in the retina. Journal of Neuroscience 32:14859 --
14873.
[18] Jaynes ET (1957) Information theory and statistical me-
chanics. Physical Review 106:620.
chanics. ii. Physical Review 108:171.
[20] Okun M, et al.
(2012) Population rate dynamics and
multineuron firing patterns in sensory cortex. J Neurosci
32:17108 -- 19.
[21] Press´e S, Ghosh K, Lee J, Dill KA (2013) Principles
of maximum entropy and maximum caliber in statistical
physics. Reviews of Modern Physics 85:1115 -- 1141.
[22] Marre O, Boustani SE, Fr´egnac Y, Destexhe A (2009)
Prediction of spatiotemporal patterns of neural activity
from pairwise correlations. Phys Rev Lett 102:138101.
[23] Vasquez JC, Marre O, Palacios AG, Ii MJB, Cessac B
(2012) Gibbs distribution analysis of temporal correla-
tions structure in retina ganglion cells. Journal of Phys-
iology - Paris 106:120 -- 127.
[24] Egu´ıluz VM, Ospeck M, Choe Y, Hudspeth AJ, Mag-
nasco MO (2000) Essential nonlinearities in hearing.
Phys Rev Lett 84:5232 -- 5.
[25] Mora T, Walczak AM, Bialek W, Callan CG (2010) Max-
imum entropy models for antibody diversity. Proc Natl
Acad Sci USA 107:5405 -- 10.
[26] Stephens GJ, Mora T, Tkacik G, Bialek W (2013) Statis-
tical thermodynamics of natural images. Phys Rev Lett
110:018701.
[27] Bialek W, et al.
(2014) Social interactions dominate
speed control in poising natural flocks near criticality.
Proc Natl Acad Sci USA.
[28] Attanasi A, et al. (2013) Wild swarms of midges linger
at the edge of an ordering phase transition. arXiv cond-
mat.stat-mech.
[29] Cavagna A, et al. (2014) Dynamical maximum entropy
approach to flocking. Phys. Rev. E 89:042707.
[30] Krotov D, Dubuis JO, Gregor T, Bialek W (2014)
Morphogenesis at criticality. Proc Natl Acad Sci USA
111:3683 -- 8.
[31] Weigt M, White RA, Szurmant H, Hoch JA, Hwa T
(2009) Identification of direct residue contacts in protein-
protein interaction by message passing. Proc Natl Acad
Sci USA 106:67 -- 72.
Appendix A: Visual stimulation
Our stimulus was composed of two black bars moving
randomly on a gray background. Each bar was animated
by a brownian motion, with additional feedback force to
stay above the array, and repulsive forces so that they
do not overlap. The bars stay within an area that covers
the whole recording array. The amplitude of the bar
trajectories allowed them to sweep the whole recording
zone. The trajectories of the bars x1 and x2 are described
by the following equations:
(cid:18)
= − v1
+ sign(x1 − x2)
τ
(cid:18)
0(x1 − µ1) + σ W1(t)
−ω2
= − v2
+ sign(x2 − x1)
τ
0(x2 − µ2) + σ W2(t)
−ω2
R
x1 − x2
R
x2 − x1
(cid:19)6
(cid:19)6
dv1
dt
dv2
dt
(A1)
(A2)
where W1(t) and W2(t) are two Gaussian white noises of
unit amplitude, µ2 − µ1 = 600µm is the shift between
the means, ω0 = 1.04 Hz and τ = 16.7 ms. The width of
one bar is 100µm. The stimulus was displayed using a
Digital Mirror Device and focused on the photoreceptor
plane using standard optics.
Appendix B: Thermodynamics of spike trains
Let us start with the probability distribution for entire
spike trains {σi,t}, i = 1, . . . , N , t = 1, . . . , L. By analogy
with Boltzmann law we can write this probability as:
1
Z
P ({σi,t}) =
e−E({σi,t}),
(B1)
where E({σi,t}) and − log Z are defined up to a com-
mon constant. The surprise − log P ({σi,t}) is equal to
E({σi,t}) + log Z. Note that considering the statistics
of entire spike trains over time allows for a well-defined
∆t → 0 limit, with the concomitant scaling L ∼ 1/∆t, by
contrast to the static thermodynamic approach (L = 1)
where this limit tends to the all-silent state with proba-
bility one.
The probability distribution in Eq. B1 will produce
typical spike trains with the same statistics as the exper-
iment. To explore this model across a line in parameter
space, we can generalize Eq. B1 to an arbitrary fictious
temperature:
Pβ({σi,t}) =
1
Z(β)
e−βE({σi,t}).
(B2)
While Pβ=1 describes "typical" spike trains with the
same statistics as the experiment, this generalized dis-
tribution allows us to explore atypical spike trains of low
or high energy (accessed by high and low β), or equiva-
lently of high and low surprises.
The free energy is defined as F (β) = −β−1 log Z(β).
The Shannon entropy of Pβ,
Pβ({σi,t}) log Pβ({σi,t}),
(B3)
S(β) = − (cid:88)
{σi,t}
6
can be calculated as S(β) = ∂F/∂β = β−1((cid:104)E(cid:105)β−F (β)),
where (cid:104)·(cid:105)β denotes an average taken over spike trains with
probability law Pβ. This last relation is better known in
the form F = E − T S, with T = β−1 is temperature.
The heat capacity is defined as:
C(β) = T
∂S
∂T
= −β
∂S
∂β
= β2((cid:104)E2(cid:105)β − (cid:104)E(cid:105)2
β).
(B4)
In statistical physics it is an extensive quantity, meaning
that it scales with the system size N L. The specific heat
c(β) = C(β)/N L is the heat capacity normalized by the
system size.
Let us consider a simple example, where each neuron
spikes with probabily pi = ri∆t in each time window
(where ri is its spike rate), independently of the other
neurons and of its own spiking history. In the limit ∆t →
0 these are just Poisson neurons. The probability of a
given spike train factorizes over neurons and over time,
and reads:
Pβ({σi,t}) =
1
pβσi,t
i
qβ(1−σi,t)
i
,
(B5)
L(cid:89)
N(cid:89)
zi(β)
where qi = 1 − pi and zi(β) = pβ
can be calculated from Eq. B4:
t=1
i=1
i + qβ
i . The specific heat
N(cid:88)
i=1
c(β) =
1
N
β2(piqi)β(log pi − log qi)2/(pβ
i + qβ
i ). (B6)
This expression has no divergence as a function of β. For
small uniform spiking probability pi = p (cid:28) 1, the specific
heat at the natural temperature is also small: c(β =
1) ∼ p(log p)2. In that same limit, the peak in specific
heat is reached at high temperatures, βc ∼ −α/ log p,
where α ≈ 2.2 is solution of the irrational equation α =
2(1 + e−α); the value of the peak does not depend on p,
and is c(βc) ∼ α(α − 2) ≈ 0.48.
Appendix C: Thermodynamics of a simple
auto-regressive model
a time window. Let us call Kt =(cid:80)N
We consider a simple case where we assume that the
neural population is well described by a continuous pa-
rameter describing the total number of spiking neurons in
i=1 σi,t that number.
Its mean is (cid:104)K(cid:105) = rN , where r is the average spike rate
of each cell per time window. We denote Kt = rN + xt.
Assuming that K and N are large, we can treat xt as
a continuous variable, and model it by a simple Markov
dynamics, or auto-regressive model:
xt+1 = axt + t,
(C1)
with t a Gaussian noise of mean zero and variance σ2.
xt is of mean zero, and its auto-correlation function of xt
reads:
(cid:104)xtxt(cid:48)(cid:105) =
t(cid:105)at−t(cid:48) ≡ f rN at−t(cid:48), (C2)
1 − a2 at−t(cid:48) = (cid:104)x2
σ2
where f = Var(K)/(cid:104)K(cid:105) is the Fano factor of the number
of spiking neurons. f = 1 when the distribution of K
is Poisson. When a → 1, the system becomes critical in
the traditional dynamical sense, with a diverging corre-
lation time −1/ log(a). This is the "stability parameter"
obtained from an auto-regressive model [11].
For each K, the probability of a given spiking pattern
is uniform:
P (σ1, . . . , σNK) =
i σi, K)
.
(C3)
δ((cid:80)
(cid:0)N
(cid:1)
K
Assuming that the system is stationary at t = 1, the
probability of a whole spike train of duration L is thus
given by:
log P ({σi,t}) = − L
2
log(2πσ2) − x2
2f rN
log(2πf rN )
− 1
2
1
− EK − Eσ,
L(cid:88)
t=1
(xt+1 − axt)2
(C4)
(C5)
with
EK = − 1
2σ2
L(cid:88)
Eσ =
[log Γ(N + 1) − log Γ(rN + xt + 1)
t=1
where we have replaced (cid:0)N
(cid:1) = (cid:0) N
− log Γ((1 − r)N − xt + 1)] .
(cid:1) by its expres-
(C6)
K
rN +xt
sion in terms of Gamma functions Γ(x). The term −EK,
combined with the first term on the right-hand side of
Eq. C4, corresponds to the Gaussian distribution of t
with replacement using Eq. C1. The term −Eσ corre-
sponds the conditional distribution in Eq. C3. The sec-
ond and third terms on the right-hand side correspond to
the Gaussian distribution of x1, of zero mean and vari-
ance f rN = Var(K).
We expand Eσ by assuming that x (cid:28) N , using Stir-
ling's formula, and obtain at leading order:
Eσ ≈(cid:88)
t
(cid:18) N r + xt
(cid:19)
N H
N
(cid:20)
≈(cid:88)
t
1 − r
r
N H(r) + log
xt
,
(cid:21)
(C7)
where H(x) = −x log(x)− (1− x) log(1− x) is the binary
entropy.
If we neglect terms containing the initial condition x1,
the total surprise is, up to a constant, equal to EK + Eσ.
Its variance, also called heat capacity by analogy with
statistical mechanics, is given by
C(β = 1) = ((cid:104)E2
K(cid:105) − (cid:104)EK(cid:105)2) + ((cid:104)E2
σ(cid:105) − (cid:104)Eσ(cid:105)2),
(C8)
as the cross-correlation term involves the third moments
of xt and thus is zero. A calculation using Gaussian
integration rules gives, at leading order in the limit L →
∞:
(cid:104)E2
K(cid:105) − (cid:104)EK(cid:105)2 =
L
2
.
(C9)
On the other hand we obtain:
(cid:18)
σ(cid:105) − (cid:104)Eσ(cid:105)2 = N L
(cid:104)E2
1 − r
r
log
7
(cid:19)2
f r
1 + a
1 − a
.
(C10)
Both variances scale linearly with L. This is consistent
with the extensivity of the heat capacity: the average
surprise scales linearly with L, and its variance does as
well. But only the second part of the variance scales
linearly with N . Thus in the limit N , L → ∞,
C(β = 1)
1 + a
1 − a
.
=
f r
log
N L
c(β = 1) =
(C11)
The variance of the surprise diverges as a → 1, i.e. as
the system becomes critical in the usual dynamical sense.
When the Fano factor f = Var(K)/(cid:104)K(cid:105) diverges with
N , the specific heat c(β = 1) diverges as well. This is
the case when fluctuations of K are of the same order
of magnitude as K itself, e.g. Var(K) ∼ K 2 and thus
f ∼ K ∼ rN , as was observed in the salamander retina
[13].
(cid:18)
(cid:19)2
1 − r
r
Appendix D: Thermodynamics of a model of neural
avalanches
t
i=1
(D1)
P ({σi,t}) =
pi(t)σi,t[1 − pi(t)]1−σi,t
We now study a simple model of spiking dynamics that
is known to display critical avalanche statistics [7]. We
will show that applying our specific heat criterion allows
us to detect the critical point. In this model, neuron i
spikes at time t if it receives signal from at least one other
neuron j, which happens with probability pij, provided
that that neuron has spiked at time t−1. The probability
for a spike train can be written as:
N(cid:89)
(cid:89)
where pi(t) = 1−(cid:81)
be easily calculated as E = − log(P ) =(cid:80)
j(1− pij)σi,t−1 is the probability that
t = −(cid:88)
neuron i spikes at time t. The energy of this process can
The parameter ω = (1/N )(cid:80)
t t, with
σi,t log pi(t) − (1 − σi,t) log[1 − pi(t)]
ij pij quantifies the prob-
ability that a spike generates another spike at the next
time step. When ω < 1, the spiking activity goes
extinct, while when ω > 1, it explodes exponentially.
Around ω ∼ 1, the system is critical and exhibits neural
avalanches with power-law statistics [7]. Since the all-
silent state is absorbing, in the simulation we further as-
sume that when the system goes into the all-silent state,
one random neuron (out of N ) is made to spike to restart
the activity.
Taking the L → ∞ limit, the specific heat is just esti-
(D2)
i
mated numerically from simulations as
c(β = 1) =
(cid:104)δ2
t(cid:105) +
1
N
2
N
(cid:104)δtδt+u(cid:105),
(D3)
(cid:88)
u≥1
8
(cid:35)
L−u(cid:88)
(cid:88)
t=1
{σi,t}
where λa are Lagrange multipliers that must be adjusted
to satisfy Eq. E2, and Z is a normalization constant.
There are many ways to choose the set of observ-
ables Oa, and just as many resulting models. Here for
simplicity we assume that the system is in a station-
ary state, so that the statistics of spike trains is time-
invariant. This implies that the observables will be time
averaged. Our choice of observables are the joint dis-
tributions of the number of spiking neurons at different
times, Pu(Kt, Kt+u), for u = 1, . . . , v, defined as:
Pu(K, K(cid:48)) =
1
L − u
δK,KtδK(cid:48),Kt+u P ({σi,t}),
(E4)
where δa,b = 1 if a = b and 0 otherwise. The correspond-
ing model of maximum entropy is:
(cid:34)(cid:88)
(cid:88)
v(cid:88)
P ({σi,t}) =
1
Z
exp
h(Kt) +
Ju(Kt, Kt+u)
t
t
u=1
(E5)
where h(K) and Ju(K, K(cid:48)) are the Lagrange multipliers
λa associated with the constraints on Pu(K, K(cid:48)). Intro-
ducing h(K) is not necessary, because Ju(K, K(cid:48)) suffices
to enforce the constraints on the marginals, but doing
so allows us to formally separate first-order from second-
order terms, at the cost of redundancy. As a result, the
definition of the model in Eq. E5 allows for some freedom
in the definition of the parameters. Indeed the distribu-
tion is unchanged upon the transformations:
Ju(K, K(cid:48)) → Ju(K, K(cid:48)) + (K),
h(K) → h(K) − (K),
(E6)
and likewise for the second argument of Ju. This degen-
eracy can be lifted by imposing the following relations:
(cid:88)
(cid:88)
(cid:88)
K
K(cid:48)
P (K)h(K) = 0,
K
P (K)Ju(K, K(cid:48)) = 0 for all K(cid:48),
P (K(cid:48))Ju(K, K(cid:48)) = 0 for all K.
(E7)
(E8)
(E9)
Note that this choice of parametrization does not affect
the model distribution itself.
It is merely a choice of
convention, which ensures that the energy terms h and
J are balanced around 0.
In practice it is enough to study the model
for
(K1, . . . , KL), the distribution of which is:
(cid:18) N
(cid:19)(cid:19)
Kt
h(Kt) + log
(cid:18)
exp
1
Z
P (K1, . . . , KL) =
(cid:34)(cid:88)
(cid:35)
where the binomial factors(cid:0) N
Ju(Kt, Kt+u)
(cid:88)
v(cid:88)
+
t
u=1
,
t
Kt
(cid:1) counts the spiking pat-
(E10)
terns (σ1,t, . . . , σN,t) having Kt spiking cells among N .
FIG. S1: Specific heat of a simple model of neural
avalanches. The specific heat c(β = 1), or variance of the
surprise Var(log P )/N L, is plotted as a function of the branch-
ing parameter ω in a simple model of neural avalanches, for
increasing network sizes N . The specific heat gets increas-
ingly peaked as the network size grows, and the peak gets
closer to the critical value branching parameter ω = 1.
where δt = t − (cid:104)t(cid:105). Fig. S1 shows the specific heat
as function of the branching parameter ω, for increasing
network sizes N . The specific heat peaks close to ω = 1.
The peak diverges and gets closer to 1 as the system
size is increased. This demonstrates that our criterion
for criticality based on the specific heat can help detect
a critical transition in this simple model. Note that, in
doing so, we have not had to define what an avalanche
is. Instead, we have solely relied on the thermodynamic
properties of the spike train statistics.
Appendix E: Maximum entropy modeling
We want to infer a model for the probability of a
entire multi-neuron spike train {σi,t}, i = 1, . . . , N ,
t = 1, . . . , L. The principle of maximum entropy allows
us to infer an approximation of that probability from
measurable observables. We look for a model distribu-
tion P ({σi,t}) that has maximum entropy:
P ({σi,t}) log P ({σi,t})
− (cid:88)
(E1)
{σi,t}
the constraint
under
it agrees with the ex-
pected value of a few chosen observables O1({σi,t}),
O2({σi,t}), . . ., estimated from the data:
that
(cid:104)Oa(cid:105)data =
Oa({σi,t})P ({σi,t}), for all a.
(E2)
(cid:88)
{σi,t}
The technique Lagrange multipliers gives us the form of
such a distribution:
P ({σi,t}) =
1
Z
exp
λaOa({σi,t})
,
(E3)
(cid:34)(cid:88)
a
(cid:35)
0.911.11.21.302004006008001000120014001600BranchingparameterωSpecificheatc(β=1) N=5N=10N=15N=30N=60N=125N=250N=500N=750N=1000 Appendix F: Gaussian approximation
9
v(cid:88)
u=1
It is possible to further simplify the maximum entropy
model by treating K as a continuous variable and con-
straint only its first and second moments (cid:104)Kt(cid:105), (cid:104)KtKt+u(cid:105).
Using Eq. E3, these constraints lead to a Gaussian dis-
tribution for the number of spiking neurons:
(cid:34)
− 1
2
(cid:88)
v(cid:88)
1
Z
(cid:35)
exp
P (K1, . . . , KL) =
where xt = Kt−(cid:104)K(cid:105) as before. This process is equivalent
to a generalized auto-regressive model:
xtAuxt+u
, (F1)
u=0
t
xt =
γuxt−u + t,
(F2)
with t a Gaussian variable of zero mean and covariance
(cid:104)tt(cid:48)(cid:105) = σ2δtt(cid:48) and the correspondance:
(cid:32)
(cid:33)
v(cid:88)
γu − (cid:88)
1 +
γ2
u
u=1
u(cid:48)−u(cid:48)(cid:48)=u
A0 =
1
σ2
Au = − 2
σ2
.
(F3)
(F4)
γu(cid:48)γu(cid:48)(cid:48)
This class of models generalizes Eq. C1. They predict a
Gaussian distribution for the number of spiking neurons,
in contradiction with experimental observations.
Appendix G: Model solution
The fitting problem of the maximum entropy distribu-
tion reduces to finding the parameters h(K), Ju(K, K(cid:48))
so that the distribution in Eq. E10 agrees with the ex-
periments on the values of the marginal probabilities
Pu(K, K(cid:48)) (for all K, K(cid:48)). Data estimates are simply
obtained from the frequency of (Kt, Kt+u) pairs in the
recordings. The model prediction, defined by Eq. E4, re-
quires to sum over all possible trajectories of Kt, which, if
done with brute force, would be prohibitively long. How-
ever, it is possible to perform these sums using the tech-
nique of transfer matrices, which requires much less com-
putational power. This technique is commonly used to
solve one-dimensional problems in statistical mechanics.
It is also known in computer science as an instance of
dynamic programming.
We start by assuming that the trajectory (K1, . . . , KL)
is an vth order Markov process (this assumption will
be verified later). We define the super variable Xt =
(Kt, Kt+1, . . . , Kt+v−1), and rewrite Eq. E10 as:
P ({Xt}) =
H(Xt) +
W (Xt, Xt+1)
(cid:35)
(cid:34)(cid:88)
v−1(cid:89)
t
1
Z
exp
×(cid:89)
t
u=1
(cid:88)
t
,X (u+1)
t−1
δX (u)
t
FIG. S2: Temporal correlations. Mutual information be-
tween Kt and Kt+u as a function of u × ∆t (∆t = 10 ms), for
all N = 185 neurons. The mutual information quantifies the
correlation between two quantities. The model prediction for
different v is compared to the data. The agreement is good
for v = 3 and 4. The gray curve shows the direct information
between different times [31], which quantifies the strengh of
interaction between t and t + u, within the v = 4 model.
Value
FIG. S3:
coupling parameters
Ju(Kt, Kt+u). The x and y axes represent Kt and Kt+u,
respectively. The model was fittted with v = 4 and all
N = 185 neurons.
the
of
is the uth component of Xt, i.e. Kt+u−1, and
(cid:18) N
(cid:19)(cid:21)
Kt+u−1
h(Kt+u−1) + log
1
v − (u − u(cid:48))
Ju−u(cid:48)(Kt+u−1, Kt+u(cid:48)−1).
(G2)
where X (u)
with
t
H(Xt) =
1
v
(cid:20)
v(cid:88)
v(cid:88)
u=1
+
u(cid:48)<u=1
and
(G1)
W (Xt, Xt+1) = Jv(Kt, Kt+v).
(G3)
05010015020001020304050Time(ms)MI(Kt,Kt+u)(bits/sec) observedv=1v=2v=3v=4directinfou = 1 20401020304050−100102030u = 2 20401020304050−100102030u = 3 20401020304050−100102030u = 4 20401020304050−100102030 If Kt is vth-order Markovian, then the super-variable Xt
is Markovian:
P ({Xt}) = P (X1)
P (XtXt−1).
(G4)
L(cid:89)
t=2
The conditional distribution can be written in the form:
P (XtXt−1) =
eH(Xt)+gt→(Xt)−gt−1→ (Xt−1)+W (Xt−1,Xt)
and Z =(cid:81)L
where gt→ is a function that will be specified by normal-
ization (see below). This identification can be verified
by replacing Eq. G5 into Eq. G4 and comparing with
Eq. G1, with gL→(XL) = 0,
P (X1) =
1
z1
eH(X1)+g1→(X1),
(G6)
t=1 zt→. Thus, Xt is indeed Markovian, and
Kt is vth order Markovian.
The parameter to be learned is the function g→(X). In
general that function depends on t, but here we assume
that it is constant because of stationarity. This assump-
tion is only valid in the bulk (i.e.
for t far away from
both 1 and L). The normalization condition
P (XtXt−1) = 1,
(G7)
1
zt→
× v−1(cid:89)
u=1
(cid:88)
Xt
(cid:34)
(cid:88)
δX (u)
t
,X (u+1)
t−1
,
(G5)
which gives the self-consistent equation:
10
simply gives the eigenvalue z. This procedure takes a
computational time of order (Kmax+1)v+1, which is large
but manageable for small enough v.
The same reasoning can be repeated by writing the
Markov dynamics of Xt backward in time:
P (XtXt+1) =
eH(Xt)+g←(Xt)−g←(Xt+1)+W (Xt,Xt+1)
1
z
× v−1(cid:89)
(cid:34)
u=1
(cid:88)
δX (u+1)
t
,X (u)
t+1
,
v−1(cid:89)
(cid:35)
(G9)
eg←(X(cid:48)).
eg←(X) =
1
z
eH(X(cid:48))+W (X(cid:48),X)
X(cid:48)
u=1
δX(cid:48)(u+1),X (u)
(G10)
The only difference with Eq. G8 is the exchange of X
and X(cid:48). Thus, g→ and g← may be different for general
time-irreversible processes. The eigenvalue z remains un-
changed, however, because the right and left eigenvalues
of a matrix are the same.
Armed with g→ and g←, we can now calculate all
marginals. Using the Markovian nature of the sequence:
P ({Xt(cid:48)}) =P (Xt)P (X1, . . . , Xt−1Xt)P (Xt+1, . . . , XLXt)
t(cid:89)
t(cid:48)=2
=P (Xt)
P (Xt(cid:48)−1Xt(cid:48))
L−1(cid:89)
t(cid:48)=t
P (Xt(cid:48)+1Xt(cid:48)),
(G11)
which must hold for all Xt−1, gives the following self-
consistent equation for g→(X):
v−1(cid:89)
(cid:35)
eg→(X) =
1
z
eH(X(cid:48))+W (X,X(cid:48))
δX(cid:48)(u),X (u+1)
eg→(X(cid:48)),
X(cid:48)
u=1
(G8)
where we have replaced Xt−1 by X and Xt by X(cid:48) to
ease notations (but also because these are dummy vari-
ables). We can view this equation as an eigenvalue prob-
lem: eg→(X) is the eigenvector of the matrix defined in
the bracket (called the transfer matrix), associated with
its largest eigenvalue z. This equation can be solved by
simply iterating Eq. G8, and normalizing eg→(X) after
X g→(X)). Af-
ter convergence, that normalization constant at each step
each iteration (by e.g. maxX g→(X) or(cid:80)
and replacing with Eqs. G5,G9 and G1, we get:
P (Xt) =
1
zt
eg→(Xt)+g←(Xt)+H(Xt)
(G12)
and
P (Xt, Xt+1) = P (Kt, . . . , Kt+v)
=
1
ztzt+1
eg→(Xt+1)+g←(Xt)+H(Xt)+H(Xt+1)+W (Xt,Xt+1).
(G13)
We can also calculate pairwise marginals between Kt at
arbitrary time differences by using the following recur-
sion, for u > v:
P (Kt, Kt+u+1, . . . , Kt+u+v) =
(cid:88)
Kt+u
P (Kt, Kt+u, . . . , Kt+u+v−1)P (Kt+u+vKt+u, . . . , Kt+u+v−1).
(G14)
starting with u = 0:
This whole procedure can be performed at an arbitrary
P (Kt, Kt+1, . . . , Kt+v) = P (Xt, Xt+1).
(G15)
11
inverse temperature β. The energy of a given spike train
is, according to Eq. E5:
E = −(cid:88)
h(Kt) −(cid:88)
v(cid:88)
Ju(Kt, Kt+u),
(G16)
t
t
u=1
and thus at temperature 1/β the distribution of spike
trains reads:
Pβ({σi,t}) =
1
Z(β)
e−βE({Kt}),
(G17)
where Z(β) enforces normalization. The distribution
Pβ(K1, . . . , Kt) is given by Eq. E10 with the substitu-
tions:
h(K) → βh(K),
Ju(K, K(cid:48)) → βJu(K, K(cid:48)).
(G18)
(G19)
All the results of the procedure, z(β), g→(X; β) and
g←(X; β) thus depend on β. The free energy F (β) =
−β−1 log Z(β) can be calculated per unit time through
f (β) ≡ F (β)/N L = −β−1 log z(β)/N . The average en-
ergy (Eq. G16) per unit time is given by:
(β) ≡(cid:104)E(cid:105)β
N L
− 1
N
(cid:88)
(cid:88)
Kt
= − 1
N
v(cid:88)
u=1
Kt,Kt+u
h(Kt)Pβ(Kt)
Ju(Kt, Kt+u)Pβ;u(Kt, Kt+u)
(G20)
and the entropy per unit time by s(β) ≡ S(β)/N L =
β(β) + log z(β)/N . The specific heat c(β) = −β∂s/∂β
is obtained by numerical derivation.
The technique of transfer matrices can also be ex-
tended to calculate the statistics of avalanches. Two dis-
tributions can be calculated: that of the duration of the
avalanche, and that of the number of spikes in it. An
avalanche starts at t if Kt−1 = 0 and Kt > 0. It ends
after (cid:96) steps if Kt+(cid:96) = 0, and Kt(cid:48) > 0 for all t(cid:48) such that
t ≤ t(cid:48) < t + (cid:96). The probability Q(cid:96) for an avalanche to
last at least (cid:96) steps, and have Kt+(cid:96), . . . , Kt+(cid:96)+v−1 spiking
neurons at the v subsequent step is given recursively by:
(cid:88)
Kt+(cid:96)−1>0
Q(cid:96)(Kt+(cid:96), . . . , Kt+(cid:96)+v−1) =
with initialization (cid:96) = 0:
Q(cid:96)(Kt+(cid:96)−1, . . . , Kt+(cid:96)+v−2)P (Kt+(cid:96)Kt+(cid:96)−1, . . . , Kt+(cid:96)+v−2)
(G21)
Q(cid:96)(Kt, . . . , Kt+v−1) =
P (Kt−1 = 0, Kt, . . . , Kt+v−1)
P (Kt−1 = 0)
=
P (Xt−1, Xt)
P (Kt−1 = 0)
.
Then the probability that the avalanche lasts (cid:96) steps is calculated through:
(G22)
(G23)
(cid:88)
P(cid:96) =
Q(cid:96)(Kt+(cid:96) = 0, Kt+(cid:96)+1, . . . , Kt+(cid:96)+v−1).
Restricting to non-zero avalanches, the distribution is given by P(cid:96)/(1 − P(cid:96)=0).
Kt+(cid:96)+1,...,Kt+(cid:96)+v−1
(cid:88)
The distribution of the number of spiking events in the avalanche can be calculated in a similar way, although at
a higher computational cost. We define R(cid:96)(Kt+(cid:96), . . . , Kt+(cid:96)+v−1; n) as the probability that an avalanche has lasted at
least (cid:96) steps, has accumulated n spiking events during these steps, and has (Kt+(cid:96), . . . , Kt+(cid:96)+v−1) spiking cells in the
v time windows following the (cid:96)th step. Then the following recursion holds:
R(cid:96)(Kt+(cid:96), . . . , Kt+(cid:96)+v−1; n) =
R(cid:96)(Kt+(cid:96)−1, . . . , Kt+(cid:96)+v−2; n− Kt+(cid:96)−1)P (Kt+(cid:96)Kt+(cid:96)−1, . . . , Kt+(cid:96)+v−2). (G24)
The initialization at (cid:96) = 0 simply reads:
Kt+(cid:96)−1>0
R(cid:96)(Kt, . . . , Kt+v−1, n) =
P (Kt−1 = 0, Kt, . . . , Kt+v−1)
P (Kt−1 = 0)
δn,0
(G25)
As before the joint distribution P(cid:96),n for the size and duration of avalanches is obtained by summing over
Kt+(cid:96)+1, . . . , Kt+(cid:96)+v−1 as in Eq. G23, and restricting to non-zero avalanches ((cid:96) > 0).
Appendix H: Model learning
ties for a given set of parameters h(K) and Ju(K, K). We
The procedure described in the previous section allows
us to calculate the marginals and thermodynamic quanti-
12
FIG. S4: Effect of window size on the specific heat. Same as figure 2 of the main text, for a window size ∆t = 5 ms.
(a) Specific heat c(β) of spike trains of the entire population (N = 185), as a function of temperature 1/β, for an increasing
temporal range v. (b) Comparison with the curves obtained for ∆t = 10 ms, with the same temporal range v × ∆t expressed
in seconds (=10 ms for cyan curves, 20 ms for the red curves). Solid lines are for ∆t = 5 ms, and dashed line for ∆t = 10 ms.
(c) Specific heat of spike trains of subnetworks of increasing sizes N , for v = 4. Each point is averaged over 100 random
subnetworks for N ≤ 50, and shows one representative network for N = 61 and 97. The error bars show standard deviations.
want to solve the inverse problem, which is to find these
parameters for a given set of marginals Pu(Kt, Kt+u). To
do this we implement the following iteration:
FIG. S5: Effect of window size on finite-size scaling.
Same as figure 3 of the main text, for a window size ∆ = 5
ms. (a) Position of the peak 1/β in specific heat as a function
of network size N , for an increasing temporal range v. (b)
Normalized variance of the surprise as a function of N , for an
increasing temporal range v.
h(K) ← h(K) + [Pdata(K) − Pmodel(K)]
Ju(Kt, Kt+u) ← Ju(Kt, Kt+u) + [Pdata(Kt, Kt+u) − Pmodel(Kt, Kt+u)] ,
after which we enforce our constraints (Eqs. E7, E8, E9) by:
h(K) ← h(K) +
(cid:34)(cid:88)
v(cid:88)
h(K) ← h(K) −(cid:88)
Ju(K, K(cid:48)) ← Ju(K, K(cid:48)) −(cid:88)
K(cid:48)
u=1
K(cid:48)
K(cid:48)(cid:48)
P (K(cid:48))Ju(K, K(cid:48)) +
(cid:88)
P (K(cid:48)(cid:48))Ju(K(cid:48)(cid:48), K(cid:48)) −(cid:88)
K(cid:48)
P (K(cid:48))h(K(cid:48))
(cid:35)
P (K(cid:48))Ju(K(cid:48), K)
P (K(cid:48)(cid:48))Ju(K, K(cid:48)(cid:48)) +
(cid:88)
K(cid:48)(cid:48),K(cid:48)(cid:48)(cid:48)
K(cid:48)(cid:48)
(H1)
(H2)
(H3)
(H4)
P (K(cid:48)(cid:48))P (K(cid:48)(cid:48)(cid:48))Ju(K(cid:48)(cid:48), K(cid:48)(cid:48)(cid:48))(H5)
Note that only the first step H1,H2 actually modifies the
model. At each step, Pmodel must be re-calculated from
the new set of parameters (h, Ju).
We initialize the algorithm by setting Ju = 0 for u > 1.
This corresponds to the case v = 1, for which h(K) and
J1(K, K(cid:48)) can be deduced directly from P (XtXt−1).
0.90.9511.051.10246810121416Temperature1/βSpecificheatc(β)a. v=0v=1v=2v=3v=40.9511.0502468101214161/βc(β)b. v×∆t=10ms,∆t=5msv×∆t=10ms,∆t=10msv×∆t=20ms,∆t=5msv×∆t=20ms,∆t=10ms0.90.9511.051.10510151/βc(β)c. N=5N=10N=20N=30N=40N=50N=61N=97N=185010020011.051.11.151.2NetworksizeNPeaktemperaturea. v=0v=1v=2v=3v=4v=501002000510NetworksizeNVar(surprise)/NLb. v=0v=1v=2v=3v=4v=5 This procedure is equivalent to a gradient descent al-
gorithm on the log-likelihood [2], and therefore is guaran-
teed to converge to the solution provided that is small
enough.
Appendix I: Inferred parameters, and choice of v
To assess the performance of the model, we can ask
how well it predicts the correlations of K at different
times. The mutual information, defined as:
MI(Kt, Kt+u) =
Pu(K, K(cid:48)) log
Pu(K, K(cid:48))
P (K)P (K(cid:48))
,
(I1)
(cid:88)
K,K(cid:48)
is a non-parametric measure of these correlations. Fig. S2
shows this mutual information estimated from the data,
as well as its prediction for models with different v. Note
that by construction, the agreement is perfect for u ≤ v.
The v = 3 and v = 4 model predictions are fairly good
even for larger u, indicating that a larger v would not
improve the model prediction much.
The inferred Ju(Kt, Kt+u) are represented in Fig. S3
for v = 4 and N = 185. They become smaller as u in-
creases, indicating that the effective interactions between
time windows decay with the time difference. This can
be quantified using the Direct Information, which mea-
sures the strength of interaction between two variables in
a complex interaction network [31]. The direct pairwise
distribution is defined as:
u (K, K(cid:48)) = eJu(K,K(cid:48))+φ(K)+φ(cid:48)(K(cid:48)),
P dir
(I2)
13
are
so
K P dir
(cid:80)
where
φ(K)
and
φ(K(cid:48))
u (K, K(cid:48)) = P (K(cid:48)), and (cid:80)
that
u (K, K(cid:48)) =
P (K). This distribution corresponds to the effect that
Kt and Kt+u would have on each other if they were
not interacting with Kt(cid:48) at other times t(cid:48). The direct
information is then defined as the mutual information in
this pairwise distribution:
chosen
K(cid:48) P dir
DI(Kt, Kt+u) =
u (K, K(cid:48)) log
P dir
u (K, K(cid:48))
P dir
P (K)P (K(cid:48))
. (I3)
(cid:88)
K,K(cid:48)
This quantity is represented in gray in Fig. S2, and
shows a sharp decay as a function of u, a further indica-
tion that v = 4 is sufficient.
Appendix J: Effect of the window size
Both the thermodynamic approach and the model used
to describe spike trains depend on the window size ∆t.
We repeated the analysis with a shorter window size of
∆t = 5 ms. The results are shown in Figs. S4 and S5.
In the limit of small window sizes, we expect that mod-
els with different ∆t, but with the same temporal range in
seconds, v×∆t, should yield similar predictions. Fig. S4b
shows that this is indeed the case. This indicates that the
results of our analysis do not depend much on the choice
of window size.
|
1912.05050 | 1 | 1912 | 2019-12-10T23:49:31 | Suppression of hypersynchronous network activity in cultured cortical neurons using an ultrasoft silicone scaffold | [
"q-bio.NC",
"physics.bio-ph"
] | The spontaneous activity pattern of cortical neurons in dissociated culture is characterized by burst firing that is highly synchronized among a wide population of cells. The degree of synchrony, however, is excessively higher than that in cortical tissues. Here, we employed polydimethylsiloxane (PDMS) elastomers to establish a novel system for culturing neurons on a scaffold with an elastic modulus resembling brain tissue, and investigated the effect of the scaffold's elasticity on network activity patterns in cultured rat cortical neurons. Using whole-cell patch clamp to assess the scaffold effect on the development of synaptic connections, we found that the amplitude of excitatory postsynaptic current, as well as the frequency of spontaneous transmissions, was reduced in neuronal networks grown on an ultrasoft PDMS with an elastic modulus of 0.5 kPa. Furthermore, the ultrasoft scaffold was found to suppress neural correlations in the spontaneous activity of the cultured neuronal network. The dose of GsMTx-4, an antagonist of stretch-activated cation channels (SACs), required to reduce the generation of the events below 1.0 event/min on PDMS substrates was lower than that for neurons on a glass substrate. This suggests that the difference in the baseline level of SAC activation is a molecular mechanism underlying the alteration in neuronal network activity depending on scaffold stiffness. Our results demonstrate the potential application of PDMS with biomimetic elasticity as cell-culture scaffold for bridging the in vivo-in vitro gap in neuronal systems. | q-bio.NC | q-bio | Suppression of hypersynchronous network activity in cultured
cortical neurons using an ultrasoft silicone scaffold
Takuma Sumia, Hideaki Yamamoto*b, and Ayumi Hirano-Iwataab
aResearch Institute of Electrical Communication, Tohoku University, 2-1-1 Katahira, Aoba-ku,
Sendai 980-8577, Japan. E-mail: [email protected]
bWPI-Advanced Institute for Materials Research (WPI-AIMR), Tohoku University, 2-1-1
Katahira, Aoba-ku, Sendai 980-8577, Japan
1
Abstract
The spontaneous activity pattern of cortical neurons in dissociated culture is
characterized by burst firing that is highly synchronized among a wide population of
cells. The degree of synchrony, however, is excessively higher than that in cortical
tissues. Here, we employed polydimethylsiloxane (PDMS) elastomers to establish a novel
system for culturing neurons on a scaffold with an elastic modulus resembling brain tissue, and
investigated the effect of the scaffold's elasticity on network activity patterns in cultured rat
cortical neurons. Using whole-cell patch clamp to assess the scaffold effect on the development
of synaptic connections, we found that the amplitude of excitatory postsynaptic current, as well
as the frequency of spontaneous transmissions, was reduced in neuronal networks grown on an
ultrasoft PDMS with an elastic modulus of 0.5 kPa. Furthermore, the ultrasoft scaffold was
found to suppress neural correlations in the spontaneous activity of the cultured neuronal
network. The dose of GsMTx-4, an antagonist of stretch-activated cation channels (SACs),
required to reduce the generation of the events below 1.0 event/min on PDMS substrates was
lower than that for neurons on a glass substrate. This suggests that the difference in the baseline
level of SAC activation is a molecular mechanism underlying the alteration in neuronal network
activity depending on scaffold stiffness. Our results demonstrate the potential application of
PDMS with biomimetic elasticity as cell-culture scaffold for bridging the in vivo-in vitro gap in
neuronal systems.
2
Main text
1. Introduction
In vitro modelling of in vivo multicellular functions is essential in biology and medicine not
only for basic studies but also for applied research, such as the screening of candidate molecules
in drug development.1,2 In fields such as cardiology and oncology, cultured-cell models have
been established and are used in disease modelling and toxicity assays.1,3 However, in
neuroscience, cortical and hippocampal neurons
in dissociated culture generate a
non-physiological activity characterized by globally synchronized burst firing, often referred to
as 'network bursts'.4-7 This activity pattern is significantly different from that observed in an
animals' cortex or hippocampus, which is highly complex both spatially and temporally.8,9 Such
complexity in neural activity is important, as it underlies the computational capacity of the
neuronal networks.10,11
Several approaches have been taken to suppress the globally synchronized bursting in
cultured neuronal networks. For instance, it has been shown that the synchronized bursts are
inhibited and the complexity in the spontaneous activity is upregulated by growing cultured
neurons on micropatterned surfaces to induce a network architecture such as those observed in
the in vivo networks.12 The role of external inputs in shaping the spontaneous dynamics of the
cultured neural networks has also been investigated both experimentally and computationally,
showing that chronic application of external stimulus that resembles thalamic input decorrelates
cortical neuronal network activity.13-15 Furthermore, pharmacological blockade of an
AMPA-type glutamate receptor with CNQX at a dose below its IC50 reduces the spatial extent
of the burst spreading,5 possibly through a reduction in the excitatory synaptic strength that is
excessively strong in cultured neurons as compared to the in vivo cortex.16-18
3
Another major difference between the in vitro and in vivo neuronal networks is the
mechanical property of their scaffolds. Cultured neurons are usually grown on a polystyrene or
glass substrate, whose elastic moduli, E, are in the order of GPa.19,20 In contrast, the brain is the
softest tissue in an animals' body, with an E below 1 kPa.21 Several studies on non-neuronal
cells have pointed to the importance of culturing cells on a scaffold with biomimetic elasticity.
For instance, mesenchymal stem cells commit to the lineage specified by scaffold elasticity.22
Furthermore, the expression of chondrocyte phenotype is stabilized when cultured on a scaffold
with an E of 5.4 kPa, similar to that of the in vivo environment.23 Based on these observations,
we hypothesized that the non-physiological synchronized bursting in cultured neuronal
networks could be suppressed by growing neurons on a biomimetic scaffold.
In this work, we established a biomimetic culture platform using polydimethylsiloxane
(PDMS) that is as soft as brain tissue (i.e. E ~ 0.5 kPa). PDMS is a well-established
biocompatible material, whose elasticity can be tuned in a wide range, from ~0.1 kPa to tens of
MPa by choosing the precursors and changing their mixing ratio.24,25 It also offers several
advantages over more commonly used materials (e.g. polyacrylamide), such as being
compatible with surface modification techniques, being electrically insulating, and having a
long shelf life.26 Primary rat cortical neurons were cultured on the PDMS substrate, and the
effect of the scaffold's stiffness on synaptic strength and the complexity of the neuronal
network activity was assessed using whole-cell patch-clamp recording and fluorescent calcium
imaging, respectively. We show that the excitatory synapses are weakened on the softer
substrates and that the neuronal correlation in spontaneous network activity is significantly
reduced on the PDMS substrate with an E ~ 0.5 kPa. The underlying molecular mechanism
responsible for the stiffness-dependent modulation on spontaneous network activity is
pharmacologically explored by blocking stretch-activated cation channels (SACs).
4
2. Experimental
2.1 Mechanical characterization of the PDMS
PDMS was prepared using Sylgard 184 (Dow Corning; mixing ratio = 50:1) and Sylgard 527
(Dow Corning; mixing ratio = 5:4). For each PDMS, 200 g of the mixtures were poured in a
glass petri dish (diameter, 90 mm; height, 60 mm), degassed in a vacuum chamber, and cured in
an oven (AS-ONE SONW-450S) for two days at 80 oC.
The elastic modulus of the PDMS was determined by the spherical indentation method
(Fig. 1a) following Zhang et al.27,28 Briefly, a chromium steel ball of 3.175-mm radius (R) was
attached onto the load cell of the Instron 5943 Universal Testing System. The depth
(δ)-indentation load (P) curves were measured (Fig. 1b), and the elastic moduli, E, were
determined by fitting the load curves to the following equation:
𝑃 =
16
9
𝐸√𝑅𝛿δ (1 − 0.15
δ
𝑅
).
(1)
2.2 PDMS substrates for neuronal culture
Glass coverslips (Matsunami C018001; diameter, 18 mm; thickness 0.17 mm) were first cleaned
by sonication in 99.5% ethanol and rinsed two times in Milli-Q grade water. After a thorough
mixing of the two PDMS components and subsequent degassing, 100 μL of the mixture was
drop casted on the coverslip. PDMS was then cured in an oven for 11 h at 80 oC.
2.3 Contact angle measurement
The hydrophilicity of the surfaces was characterized by measuring the water contact angle.
Using the LSE-B100 equipment (NiCK Corporation, Japan), a 0.5-μL water droplet was
5
dropped onto the substrate and was imaged from the side. The contact angle of the droplet was
measured using the i2win software (NiCK Corporation, Japan). Three samples were prepared
for each condition, and measurements were performed at three different positions for each
sample.
2.4 Cell culture
For cell culturing, the PDMS substrate was first treated in air plasma (Yamato PM-100) for 10 s
and was sterilized under UV light for 60 min. The surface of the PDMS was then coated with
poly-D-lysine
(PDL; Sigma P-0899) by
floating
the sample upside-down on a
phosphate-buffered saline (Gibco 14190-144) containing 50 μg/mL PDL overnight. The sample
was then rinsed two times in sterilized water and dried in air inside a laminar flow hood. One
day prior to cell plating, the sample was immersed in the plating medium [minimum essential
medium (Gibco 11095-080) + 5% foetal bovine serum + 0.6% D-glucose] and stored in a CO2
incubator (37 ºC). Glass coverslips without the PDMS layer were used in control experiments.
These were prepared by cleaning coverslips in ethanol and water, treating the surface with air
plasma (60 s), UV-sterilization (60 min), and subsequent coating with PDL (overnight).
Rat cortical neurons from 18-d old embryos were used in our experiments. All
procedures were approved by the Center for Laboratory Animal Research, Tohoku University
(approval number: 2017AmA-001-1). After dissection of the cortical tissues and cell dispersion,
the cells were plated on the samples immersed in the plating medium. After a 3 h incubation, the
medium was changed to Neurobasal medium [Neurobasal (Gibco 21103-049) + 2% B-27
supplement (Gibco 17504-044) + 1% GlutaMAX-I (Gibco 3505-061)]. Half of the medium was
replaced with fresh Neurobasal medium at 4 and 8 days of the culture.
6
2.5 Electrophysiology
Whole-cell patch-clamp recordings (HEKA EPC-10) were performed on neurons at 14−18 DIV
under the voltage-clamp mode (holding potential, -70 mV). Signals were sampled at 20 kHz and
filtered with 10 kHz and 2.9 kHz Bessel filters. Recordings were performed at room temperature.
The intracellular solution contained: 146.3 mM KCl, 0.6 mM MgCl2, 4 mM ATP-Mg, 0.3 mM
GTP-Na, 5 U/mL creatine phosphokinase, 12 mM phosphocreatine, 1 mM EGTA, and 17.8 mM
HEPES (pH 7.4). The extracellular solution for the recording contained: 140 mM NaCl, 2.4 mM
KCl, 10 mM HEPES, 10 mM glucose, 2 mM CaCl2, and 1 mM MgCl2 (pH 7.4).18 The GABAA
receptor antagonist, bicuculline (Sigma 14343; 10 μM), was added to the extracellular solution
to block inhibitory synaptic transmission. The membrane resistance was ~30 MΩ, and the
synaptic currents with amplitude of 10−150 pA were analysed using a custom code written in
MATLAB (Mathworks).
2.6 Fluorescent calcium imaging
Cultured neurons were loaded with a fluorescence calcium indicator Cal-520 AM (AAT
Bioquest).12 The cells were first rinsed in HEPES-buffered saline (HBS) containing 128 mM
NaCl, 4 mM KCl, 1 mM CaCl2, 1 mM MgCl2, 10 mM D-glucose, 10 mM HEPES, and 45 mM
sucrose, and subsequently incubating in HBS containing 2 μM Cal-520 AM for 30 min at 37 °C.
The cells were then rinsed in fresh HBS and were imaged on an inverted microscope (IX83,
Olympus) equipped with a 20× objective lens (numerical aperture, 0.70), a light-emitting diode
light source (Lambda HPX, Sutter Instrument), a scientific complementary metal-oxide
semiconductor camera (Zyla 4.2, Andor), and an incubation chamber (Tokai Hit). All recordings
were performed at 14−18 DIV, while incubating in HBS at 37 °C. In some experiments,
GsMTx-4 (Peptide Institute 4393-s) was added to the HBS to inhibit SACs.29 Each recording
7
was performed for 10 min at a frame rate of 10 Hz.
3. Results and discussion
3.1 Material properties of silicone scaffolds
The elastic scaffolds for neuronal culture were prepared with two types of PDMS, i.e. Sylgard
184 mixed at a ratio of 50:1 (hereafter referred to as 'soft') and Sylgard 527 mixed at a ratio of
5:4 (hereafter referred to as 'ultrasoft'). We first prepared the PDMS in glass petri dishes and
determined their elastic moduli by the spherical indentation method27,28 (Fig. 1). The elastic
moduli of soft and ultrasoft PDMS were determined to be 13.6 ± 1.1 kPa (mean ± S.D.; n = 4)
and 0.5 ± 0.03 kPa (n = 5), respectively (Fig. 1c). The values are in good agreement with
previous studies,24,27 and the elastic modulus of the ultrasoft PDMS was nearly equal to that of
brain tissue.21
We next evaluated the wettability of the PDMS surface by measuring water contact
angles. Neurons require the scaffold surface to be coated with cationic molecules, such as PDL.
However, the strong hydrophobicity of as-prepared PDMS prevents the molecules from stably
adsorbing on the surface.30 Therefore, the samples were exposed to air plasma for a designated
amount of time, which hydrophilizes the PDMS surface by substituting methyl groups with
hydroxyl groups.31 The changes in water contact angle θ of the soft and ultrasoft PDMS upon
the plasma treatment are shown in Fig. 2a. Prior to the plasma treatment, the PDMS surface was
hydrophobic, and θ were measured to be 132.3 and 128.0 for the soft and ultrasoft PDMS,
respectively. The hydrophilicities of samples increased with the plasma exposure time. The
hydrophilized surface was finally coated with PDL, and rat cortical neurons were cultured on
the substrates. Plain glass coverslips coated with PDL were used as controls. For the cell-culture
8
experiment, samples exposed to the plasma for 10 s were used in order to minimize the effect of
surface vitrification and cracking.31,32 As shown in Fig. 2a, θ for the soft and ultrasoft PDMS
immediately after the 10 s plasma treatment were significantly different. The values of θ for the
two scaffolds were found to converge after the PDL and the subsequent immersion in the
neuronal plating medium (Fig. 2b). Representative micrographs of the rat cortical neurons
cultured on the soft and ultrasoft PDMS are shown in Figs. 2c -- e. The cell bodies of the neurons
were well spread, and the neurites uniformly covered the entire surface. In order to compensate
for the difference in cell affinity between glass and PDMS, initial plating density was increased
1.5-fold for the two PDMS scaffolds to achieve a constant attachment density of ~950
cells/mm2 (Fig. 2f).
3.2 Reduction of excitatory synaptic currents on ultrasoft scaffolds
Previous work has shown that the amplitude of excitatory postsynaptic current (EPSC) in
hippocampal neurons cultured on Sylgard 184 with E = 457 kPa was significantly higher than
that of neurons on Sylgard 184 with E = 46 kPa.27 To investigate whether a further reduction of
substrate stiffness to mimic that of the brain tissue (E ~ 0.5 kPa) influences the synaptic
strengths, we compared the amplitude and frequency of spontaneous EPSC (sEPSC) in neuronal
networks grown on the soft (E = 14 kPa) and ultrasoft (E = 0.5 kPa) PDMS. sEPSC was
recorded from cultured cortical neurons at 14−18 DIV under whole-cell patch clamp. To inhibit
spontaneous inhibitory transmissions, a GABAA receptor blocker, bicuculline (10 μM), was
added to the extracellular solution during recording
Representative traces from neurons cultured on glass, soft PDMS, and ultrasoft PDMS
are shown in Figs. 3a -- c, respectively. The amplitude of sEPSC observed in the neurons on soft
substrates was 15% lower than those on glass substrates [soft: 23.5 ± 4.1 pA (mean ± S.D.; n =
9
13), glass: 27.8 ± 7.0 pA (n = 11)]. sEPSC amplitude in neurons on ultrasoft substrates was
further reduced from those on soft substrates and was approximately 30% lower than those on
glass substrates [ultrasoft: 20.4 ± 2.3 pA (n = 12)]. In addition, the frequency of sEPSC from the
neurons on soft and ultrasoft substrates was significantly lower than that on glass substrates
(ultrasoft: 8.7 ± 3.3 Hz, soft: 9.7 ± 3.3 Hz, glass: 13.3 ± 5.6 Hz). These data are summarized in
Figs. 3d and 3e. These results indicate that ultrasoft substrates that resemble the elastic moduli
of brain tissues suppress the excitatory synaptic strength in cultured cortical neurons. The
molecular mechanisms underlying the observations are further investigated and discussed in section
3.4.
3.3 Suppression of neural synchrony on ultrasoft scaffolds
Next, fluorescence calcium imaging was used to quantify the difference in the spontaneous
firing patterns of neuronal networks on respective substrates. Representative traces of relative
fluorescence intensity (ΔF/Fo) from single neurons are shown in Figs. 4a -- c. On the glass surface,
the peak amplitude of the calcium transients was 0.42 ± 0.01, and the rate was 9.7 ± 0.2
events/min (n = 500). Both the peak amplitude and the event rate were significantly reduced on
the soft PDMS (0.37 ± 0.01 and 7.1 ± 0.3 events/min, respectively; n = 500). On the ultrasoft
substrates, both the amplitude and rate were further reduced as compared to the soft substrate
and the control (0.27 ± 0.01 and 5.5 ± 0.2 events/min, respectively; n = 500). The reduction is
likely to be caused by the reduction in the excitatory synaptic strength. These data are
summarized in Figs. 4d and 4e.
In order to analyse the degree of neural correlations in the spontaneous activity, we
evaluated the correlation coefficient, rij, between neurons i and j, as:
𝑟𝑖𝑗 =
∑ (𝑓𝑖(𝑡)−𝑓𝑖 )(𝑓𝑗(𝑡)−𝑓𝑗 )
𝑡
√∑ (𝑓𝑖(𝑡)−𝑓𝑖 )2
𝑡
2
√∑ (𝑓𝑗(𝑡)−𝑓𝑗 )
𝑡
,
(2)
10
where fi(t) is the relative fluorescence intensity of cell i at time t, and the overline represents
time average. Then, we compared their mean, 𝑟 = (∑
𝑖,𝑗(𝑖≠𝑗)
𝑟𝑖𝑗
)/𝑁2, where N (= 50) is the total
number of analysed neurons on respective substrates. Although no significant difference in 𝑟
was observed between glass and soft substrates, the value was significantly lower in the
neuronal network grown on the ultrasoft scaffold (Fig. 4f). These results show that excessive
neural synchronization was suppressed by reducing the scaffold stiffness to 0.5 kPa.
The results obtained in this work are in agreement with the previous study, which
showed that a stiff PDMS substrate with E = 457 kPa increased hippocampal neuronal network
activity as compared to a PDMS substrate with E = 46 kPa.27 However, no discernible change in
network synchrony was observed within the range of the elasticities investigated by the previous
study. In the present study, we found that the non-physiological bursting activity is suppressed,
and the mean correlation coefficient significantly decreases when the elastic modulus of the
scaffold is further reduced to 0.5 kPa. Thus, Sylgard 527 is a promising scaffold for suppressing
the hypersynchrony in neuronal culture.
3.4 Molecular mechanism of the scaffold effect
The above results show that the ultrasoft scaffold weakens the excitatory synaptic strength and
reduces the synchrony in the neuronal network activity. We hypothesized that SACs, whose
activity is downregulated on softer substrates,33 would be the underlying molecular mechanism
and investigated the effect of its pharmacological blockade on the neuronal network activity.
GsMTx-4 is a selective antagonist for SACs with an equilibrium constant of
approximately 500 nM.29,34 We first investigated the effect of reducing SAC activity in neurons
on glass substrates. Bath application of GsMTx-4 at a concentration of 250 nM was found to
reduce the peak amplitude and the rate of spontaneous calcium transients [0.30 ± 0.01 and 4.2 ±
11
0.1 events/min (mean ± S.E.M.), respectively; Fig. 5]. When GsMTx-4 was applied at a higher
concentration of 500 nM, the rate was further reduced to 0.24 ± 0.01 events/min (Fig. 5b), while
the peak amplitude did not significantly vary from the value observed at 250 nM (Fig. 5a).
These results indicate that the fraction of active SACs in the neuronal plasma membrane plays a
key role in the generation of spontaneous bursting events and the size of individual events.
We next examined the impact of GsMTx-4 application on cortical neurons grown on
the PDMS substrates. Application of GsMTx-4 at a concentration of 250 nM reduced the rate of
spontaneous calcium transients down to 0.62 ± 0.06 and 0.42 ± 0.02 events/min on the soft and
ultrasoft substrates, respectively (Fig. 5b). Therefore, the dose of GsMTx-4 required to reduce
the spontaneous occurrence of the calcium transients below 1.0 event/min was lower than that
for the neurons on the glass substrate. This suggests that the difference in the baseline level of
SAC activation is a molecular mechanism that contributes to the alteration in neuronal network
activity depending on scaffold stiffness.
Penn et al.35 previously showed that synchronized network activity in cultured
hippocampal neurons decreased with extracellular calcium concentration, which was discussed
to be caused by a reduction in presynaptic vesicle release probability. Considering that SACs
permeate calcium ions,36 the decrease in SAC activation could underlie the reduction in sEPSC
amplitude and frequency, and neuronal synchrony on ultrasoft substrates.35,37 Another
possibility is that the influx of sodium ions through SACs36 could directly enhance neuronal
excitability independent of the modulation of synaptic strength (e.g. through facilitation of
action potential generation). Finally, a mechanism independent of SACs could also have a role.
A recent study reported that stiff substrates increase the number of synapses and reduce
voltage-dependent Mg2+ blockade in N-methyl-D-aspartate receptors, which lead to higher
postsynaptic activity in cultured hippocampal neurons.38 Figure 6 summarizes the above
12
discussion concerning
the underlying molecular mechanisms for
the suppression of
hypersynchrony on the ultrasoft substrate.
4. Conclusions
We established a protocol for culturing primary cortical neurons on an ultrasoft PDMS gel that
mimics the elasticity of brain tissues and investigated the impact of the biomimetic scaffold on
synaptic strength and spontaneous activity patterns. Our study showed that the ultrasoft
substrate reduces the amplitude of sEPSCs (Fig. 3) that are excessively strong in the in vitro
cultures. This led to significant reduction in the peak fluorescence amplitude and event rate of
spontaneous network bursts on the ultrasoft substrate as compared to the glass substrate (Fig. 4).
No significant difference in the correlation of neuronal network activity was observed on the
scaffolds with E > 13.5 kPa. In contrast, this value was significantly lower for the neuronal
network grown on the scaffold with E = 0.5 kPa (Fig. 4f), a stiffness similar to that of brain
tissue. This is the first evidence that the ultrasoft scaffold with biomimetic elasticity effectively
suppresses the hypersynchrony in the spontaneous network activity. A difference in the baseline
activation of SACs underlie these stiffness-dependent changes in synaptic transmission and
neuronal network activity.
Understanding of cellular mechanosensitivity has advanced rapidly since Engler et
al.22 found in 2006 that mesenchymal stem cells commit to the lineage specified by scaffold
elasticity. The ultrasoft PDMS scaffold offers a mechanically biomimetic culture platform that
is beneficial in suppressing the synchronous bursting in neuronal cultures. Moreover, it is a
useful platform to study the influence of mechanical cues on neuronal network development.
Further work is necessary to fully suppress the synchronized bursting in neuronal cultures. This
13
could be accomplished by integrating cell micropatterning technology with ultrasoft scaffolds or
by adding external noise to fill in for functional interactions between brain regions.12,13
Conflicts of interest
There are no conflicts to declare.
Acknowledgements
We acknowledge Prof. Hisashi Kino and Prof. Tetsu Tanaka of Tohoku University for the
mechanical analysis of PDMS. This work was supported by the Japan Society for the Promotion
of Science (Kakenhi Grant No. 18H03325) and by the Japan Science and Technology Agency
(PRESTO: JPMJPR18MB and CREST: JPMJCR14F3).
14
References
[1] K. H. Benam, S. Dauth, B. Hassell, A. Herland, A. Jain, K.-J. Jang, K. Karalis, H. J. Kim, L.
MacQueen, R. Mahmoodian, S. Musah, Y. Torisawa, A. D. van der Meer, R. Villenave, M.
Yadid, K. K. Parker and D. E. Ingber, Annu. Rev. Pathol. Mech. Dis., 2015, 10, 195-262.
[2] G. Quadranto, J. Brown and P. Arlotta, Nat. Med., 2016, 22, 1220-1228.
[3] A. Skardal, S. V. Murphy, M. Devarasetty, I. Mead, H-W. Kang, Y-J. Seol, Y. S. Zhang,
S.-R. Shin, L. Zhao, J. Aleman, A. R. Hall, T. D. Shupe, A. Kleensang, M. R. Dokmeci, S. J.
Lee, J. D. Jackson, J. J. Yoo, T. Hartung, A. Khademhosseini, S. Soker, C. E. Bishop and A.
Atala, Sci. Rep., 2017, 7, 8837.
[4] D. A. Wagenaar, J. Pine and S. M. Potter, BMC Neurosci., 2006, 7:11, 1271-2202.
[5] J. Soriano, M. Rodríguez Martínez, T. Tlusty and E. Moses, Proc. Natl. Acad. Sci. U.S.A.,
2008, 105, 13758-13763.
[6] J. G. Orlandi, J. Soriano, E. Alvarez-Lacalle, S. Teller and J. Casademunt, Nat. Phys., 2013
9, 582-590.
[7] H. Yamamoto, S. Kubota, Y. Chida, M. Morita, S. Moriya, H. Akima, S. Sato, A.
Hirano-Iwata, T. Tanii and M. Niwano, Phys. Rev. E, 2016, 94, 012407.
[8] P. Golshani, J. T. Goncalves, S. Khosahkhoo, R. Monstany, S. Smirnakis and C.
Portera-Cailliau, J. Neurosci., 2009, 29, 10890-10899.
[9] J. K. Miller, I. Ayzenshtat, L. Carrillo-Reid and R. Yuste, Proc. Natl. Acad. Sci. U.S.A., 8,
E4053-E4061 (2014).
[10] E. M. Izhikevich, Neural Comput., 2006, 18, 245-282.
[11] H. Ju, M. R. Dranias, G. Banumurthy and A. M. J. VanDongen, J. Neurosci., 2015, 35,
4040-4051.
[12] H. Yamamoto, S. Moriya, K. Ide, T. Hayakawa, H. Akima, S. Sato, S. Kubota, T. Tanii, M.
15
Niwano, S. Teller, J. Soriano and A. Hirano-Iwata, Sci. Adv., 2018, 4, eaau4914.
[13] K. P. Dockendorf, I. Park, P. He, J. C. Princípe and T. B. DeMarse, BioSystems, 2009, 95,
90-97.
[14] E. M. Izhikevich, J. A. Gally and G. M. Edelman, Cereb. Cortex, 2004, 14, 933-944.
[15] J. Zierenberg, J. Wilting and V. Priesemann, Phys. Rev. X, 2018, 8, 031018.
[16] A. K. Vogt, L. Lauer, W. Knoll and A. Offenhäusser, Biotechnol. Prog., 2003, 19,
1562-1568.
[17] S. Song, P. J. Sjöström, M. Reigl, S. Nelson and D. B. Chklovskii, PLoS Biol., 2005, 3,
e68.
[18] H. Yamamoto, R. Matsumura, H. Takaoki, S. Katsurabayashi, A. Hirano-Iwata and M.
Niwano, Appl. Phys. Lett., 2016, 109, 043703.
[19] G. V. Lubarsky, M. R. Davidson and R. H. Bradley, Surf. Sci., 2004, 558, 135-144.
[20] N. Soga, J. Non-Cryst. Solids, 1985, 73, 305-313.
[21] N. D. Leipzig and M. S. Shoichet, Biomaterials, 2009, 30, 6867-6878.
[22] A. J. Engler, S. Sen, H. L. Sweeney and D. E. Discher, Cell, 2006, 126, 677-689.
[23] T. Zhang, T. Gong, J. Xie, S. Lin, Y. Liu, T. Zhou and Y. Lin, ACS Appl. Mater. Interfaces,
2016, 8, 22884-22891.
[24] C. Moraes, J. M. Labuz, Y. Shao, J. Fu and S. Takayama, Lab Chip, 2015, 15, 3760-3765.
[25] M. P. Wolf, G. B. Salieb-Beugelaar and P. Hunziker, Prog. Polym. Sci., 2018, 83, 97-134.
[26] H. Yamamoto, L. Grob, T. Sumi, K. Oiwa, A. Hirano-Iwata and B. Wolfrum, Adv. Biosys.,
2019, 3, 1900130.
[27] Q.-Y. Zhang, Y.-Y. Zhang, J. Xie, C.-X. Li, W.-Y. Chen, B.-L. Liu, X.-a. Wu, X.-N. Li, B.
Huo, L.-H. Jiang and H.-C. Zhao, Sci. Rep., 2014, 4, 6215.
[28] M. G. Zhang, Y.-P. Cao, G.-Y. Li and X.-Q. Feng, Biomech. Model. Mechanobiol., 2014, 13,
16
1-11.
[29] T. M. Suchyna, J. H. Johnson, K. Hamer, J. F. Leykam, D. A. Gage, H. F. Clemo, C. M.
Baumgarten and F. Sachs, J. Gen. Physlol., 2000, 115, 583-598.
[30] K. Kang, I. S. Choi and Y. Nam, Biomaterials, 2011, 32, 6374-6380.
[31] N. Y. Adly, H. Hassani, A.Q. Tran, M. Balski, A. Yakuschenko, A. Offenhäusser, D. Mayer
and B. Wolfrum, Soft Matter, 2017 13, 6297-6303.
[32] D. MacNearnet, B. Mak, G. Ongo, T. E. Kennedy and D. Juncker, Langmuir, 2016, 32,
13525-13533.
[33] M. M. Pathak, J. L. Nourse, T. Tran, J. Hwe, J. Arulmoli, D. T. T. Le, E. Bernardis, L. A.
Flanagan and F. Tombola, Proc. Natl. Acad. Sci. U.S.A., 2014, 111, 16148-16153.
[34] C. L. Bowman, P. A. Gottlieb, T. M. Suchyna, Y. K. Murphy and F. Sachs, Toxicon, 2007,
49, 249-270.
[35] Y. Penn, M. Segal and E. Moses, Proc. Natl. Acad. Sci. U.S.A., 2016, 113, 3341-3346.
[36] F. B. Kalapesi, J. C. H. Tan and M. T. Coroneo, Clin. Exp. Optom., 2005, 33, 210-217.
[37] N. R. Hardingham, N. J. Bannister, J. C. A. Read, K. D. Fox, G. E. Hardingham and J. J. B.
Jack, J. Neurosci., 2006, 26, 6337-6345.
[38] Y. Yu, S. Liu, X. Wu, Z. Yu, Y. Xu, W. Zhao, I. Zavodnik, J. Zheng, C. Li and H. Zhao, ACS
Biomater. Sci. Eng, 2019, 5, 3475-3482.
17
Figures
Figure 1
Figure 2
18
Figure 3
19
Figure 4
20
Figure 5
Figure 6
21
Figure captions
Fig. 1. Mechanical properties of PDMS. (a) Schematic illustration of the spherical indentation
apparatus. (b) Load-displacement curves for soft (left) and ultrasoft (right) PDMS. Open circles
represent the measured data, and the solid curve the fit with Eq. (1) (r = 0.9999 for both
samples). For the data points, every 50th point is plotted for clarity. (d) Measured elastic moduli
of soft and ultrasoft PDMS. Error bars, S.D.
Fig. 2 Culturing primary neurons on PDMS. (a) Change in water contact angles of soft and
ultrasoft PDMS upon exposure to air plasma. (b) Water contact angles measured after plasma
irradiation for 10 s, after coating with PDL, and after immersion in the plating medium
overnight. The surfaces of both samples were superhydrophilic after the immersion in the
plating medium, and thus the data are plotted as 0o. (c -- e) Primary cortical neurons cultured on
(c) glass, (d) soft, and (e) ultrasoft scaffolds. Scale bars, 50 μm. (f) Average cell densities on the
glass, soft, and ultrasoft substrates. Error bars, S.D.
Fig. 3. Effects of elastic modulus on sEPSC. (a -- c) Representative recordings of spontaneous
EPSCs on (a) glass, (b) soft, and (c) ultrasoft scaffolds. (d and e) The mean values of the
amplitude (d) and frequency (e) of sEPSCs on respective surfaces. Error bars, S.D. * p < 0.05;
** p < 0.01.
Fig. 4. Impact of substrate stiffness on network activity of cultured cortical neurons. (a -- c)
Fluorescence intensity traces of representative neurons on (a) glass, (b) soft, and (c) ultrasoft
scaffolds. Fluorescence micrographs are shown on the right. Scale bars, 100 μm. (d and e)
Average peak amplitudes (d) and frequency of bursting events (e) on respective substrates. (f)
22
Mean correlation coefficient (mean CC) of neural activity on respective substrates. Error bars,
S.E.M. * p < 0.05; ** p < 0.01; *** p < 0.001.
Fig. 5. Impact of the pharmacological blockade of SAC on neuronal network activity. (a and b)
Average peak amplitudes (a) and rate of bursting events (b) at various concentrations of
GsMTx-4 on respective substrates. Error bars, S.E.M. * p < 0.05; ** p < 0.01; *** p < 0.001.
Fig. 6. Diagram summarizing the present findings and the mechanisms underlying the
suppression of hypersynchronous neuronal network activity on soft scaffolds.
23
|
1007.3122 | 2 | 1007 | 2013-01-30T19:34:49 | Robust short-term memory without synaptic learning | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.AO"
] | Short-term memory in the brain cannot in general be explained the way long-term memory can -- as a gradual modification of synaptic weights -- since it takes place too quickly. Theories based on some form of cellular bistability, however, do not seem able to account for the fact that noisy neurons can collectively store information in a robust manner. We show how a sufficiently clustered network of simple model neurons can be instantly induced into metastable states capable of retaining information for a short time (a few seconds). The mechanism is robust to different network topologies and kinds of neural model. This could constitute a viable means available to the brain for sensory and/or short-term memory with no need of synaptic learning. Relevant phenomena described by neurobiology and psychology, such as local synchronization of synaptic inputs and power-law statistics of forgetting avalanches, emerge naturally from this mechanism, and we suggest possible experiments to test its viability in more biological settings. | q-bio.NC | q-bio |
Robust Short-Term Memory without Synaptic
Learning
Samuel Johnson1,2,
∗, J. Marro3, and Joaqu´ın J. Torres3
1Department of Mathematics, Imperial College London, SW7 2AZ, United Kingdom.
2Oxford Centre for Integrative Systems Biology, and Department of Physics,
University of Oxford, Clarendon Lab, OX1 3QU, United Kingdom.
3Departamento de Electromagnetismo y F´ısica de la Materia, and
Institute Carlos I for Theoretical and Computational Physics,
University of Granada, 18071 Granada, Spain.
Abstract
Short-term memory in the brain cannot in general be explained the
way long-term memory can -- as a gradual modification of synaptic weights
-- since it takes place too quickly. Theories based on some form of cel-
lular bistability, however, do not seem able to account for the fact that
noisy neurons can collectively store information in a robust manner. We
show how a sufficiently clustered network of simple model neurons can be
instantly induced into metastable states capable of retaining information
for a short time (a few seconds). The mechanism is robust to different
network topologies and kinds of neural model. This could constitute a vi-
able means available to the brain for sensory and/or short-term memory
with no need of synaptic learning. Relevant phenomena described by neu-
robiology and psychology, such as local synchronization of synaptic inputs
and power-law statistics of forgetting avalanches, emerge naturally from
this mechanism, and we suggest possible experiments to test its viability
in more biological settings.
Keywords: Working memory; sensory memory; power-law forgetting; local
synchronization in neural networks.
Author summary
Whenever an image is flashed briefly before your eyes, or you hear a sudden
sound, you are usually able to recall the information presented for a few seconds
thereafter. In fact, it is most vivid at first but fades gradually. According to our
current understanding of neural networks, memories are stored by strengthening
and weakening the appropriate connections (synapses) between neurons. But
these biochemical processes take place on a timescale of minutes. Therefore,
when it came to understanding such short-term memory tasks, it seemed that
either each neuron had an individual memory that could work fast enough (not
very robust), or we could only actually remember things that had been stored in
∗[email protected]
1
our brains previously (not very credible). Here we suggest a simple mechanism
-- Cluster Reverberation -- whereby neurons with no individual memory can
nonetheless store completely novel information almost instantly and maintain it
for a few seconds, thanks to a preponderance of local connections in the network.
If the brain were indeed using this mechanism, it might explain the statistics of
forgetting as well as some recent neurobiological findings.
Introduction
Slow but sure, or fast and fleeting?
Memory -- the storage and retrieval of information by the brain -- is probably
nowadays one of the best understood of all the collective phenomena to emerge
in that most complex of systems. Thanks to a gradual modification of synap-
tic weights (the interaction strengths with which neurons signal to one other)
particular patterns of firing and non-firing cells become energetically favourable
and so systems evolve towards these attractors according to a mechanism known
as Associative Memory [1, 2, 3, 4]. In nature, these synaptic modifications occur
via the biochemical processes of long-term potentiation (LTP) and depression
(LTD) [5, 6]. However, some memory processes take place on timescales of sec-
onds or less and in many instances cannot be accounted for by LTP and LTD
[7], since these require at least minutes to be effected [8, 9]. For example, visual
stimuli are recalled in great detail for up to about one second after exposure
(iconic memory); similarly, acoustic information seems to linger for three or
four seconds (echoic memory) [10, 11]. In fact, it appears that the brain actu-
ally holds and continually updates a kind of buffer in which sensory information
regarding its surroundings is maintained (sensory memory) [12]. This is eas-
ily observed by simply closing one's eyes and recalling what was last seen, or
thinking about a sound after it has finished. Another instance is the capability
referred to as working memory [7, 13]: just as a computer requires RAM for its
calculations despite having a hard drive for long-term storage, the brain must
continually store and delete information to perform almost any cognitive task.
We shall here use short-term memory to describe the brain's ability to store
information on a timescale of seconds or less.
Evidence that short-term memory is related to sensory information while
long-term memory is more conceptual can be found in psychology. For instance,
a sequence of similar sounding letters is more difficult to retain for a short time
than one of phonetically distinct ones, while this has no bearing on long-term
memory, for which semantics seems to play the main role [14, 15]; and the way
many of us think about certain concepts, such as chess, geometry or music,
is apparently quite sensorial: we imagine positions, surfaces or notes as they
would look or sound. Most theories of short-term memory -- which almost
always focus on working memory -- make use of some form of previously stored
information (i.e., of synaptic learning) and so can account for labelling tasks,
such as remembering a particular series of digits or a known word,1 but not
for the instant recall of novel information [17, 18, 19]. An interesting exception
1This method can also be used to represent a continuous variable, such as the value of an
angle or the length of an object, because concepts such as angle and length are in some sense
already "known" at the time of the stimulus [16].
2
is the mechanism proposed by Chialvo et al.
[20] which allows for arbitrary
patterns of activity to be temporarily retained thanks to the refractory times of
neurons.
Attempts to deal with novel information have been made by proposing mech-
anisms of cellular bistability: neurons are assumed to retain the state they
are placed in (such as firing or not firing) for some period of time thereafter
[21, 22, 23]. Although there may indeed be subcellular processes leading to a
certain bistability, the main problem with short-term memory depending exclu-
sively on such a mechanism is that if each neuron must act independently of
the rest the patterns will not be robust to random fluctuations [7] -- and the
behaviour of individual neurons is known to be quite noisy [24].
It is worth
pointing out that one of the strengths of Associative Memory is that the be-
haviour of a given neuron depends on many neighbours and not just on itself,
which means that robust global recall can emerge despite random fluctuations
at an individual level.
Harnessing network structure
Something that, at least until recently, most neural-network models have failed
to take into account is the structure of the network -- its topology -- it often being
assumed that synapses are placed among the neurons completely at random,
or even that all neurons are connected to all the rest. Although relatively
little is yet known about the architecture of the brain at the level of neurons
and synapses, experiments have shown that it is heterogeneous (some neurons
have very many more synapses than others), clustered (two neurons have a
higher chance of being connected if they share neighbours than if not) and
highly modular (there are groups, or modules, with neurons forming synapses
preferentially to those in the same module) [25, 26]. We show here that it suffices
to use a more realistic network topology, in particular one that is modular and/or
clustered, for a randomly chosen pattern of activity the system is placed in to
be metastable. This means that novel information can be instantly stored and
retained for a short period of time in the absence of both synaptic learning and
cellular bistability. The only requisite is that the patterns be coarse grained
versions of the usual patterns -- that is, whereas it is often assumed that each
neuron in some way represents one bit of information, we shall allocate a bit to
a small group or neurons. (This does not, of course, mean that memories are
expected to be encoded as bitmaps. In fact, we are not making any assumptions
regarding neural coding.)
The mechanism, which we call Cluster Reverberation (CR), is very simple.
If neurons in a group are more densely connected to each other than to the rest
of the network, either because they form a module or because the network is
significantly clustered, they will tend to retain the activity of the group: when
they are all initially firing, they each continue to receive many action poten-
tials and so go on firing, whereas if they start off silent, there is not usually
enough input current from the outside to set them off. (This is similar to the
're-entrant' activity exhibited by excitable elements [27].) The fact that each
neuron's state depends on its neighbours confers to the mechanism a certain
robustness to random fluctuations. This robustness is particularly important
for biological neurons, which as mentioned are quite noisy. Furthermore, not
only does the limited duration of short-term memory states emerge naturally
3
from this mechanism (even in the absence of interference from new stimuli) but
this natural forgetting follows power-law statistics, as has been observed exper-
imentally [28, 29, 30].
It is also coherent with recent observations of locally
synchronized neural activity in vivo [31], and of clustering in both synaptic in-
puts [32] and plasticity [33] during development. The viability of this mechanism
in a more realistic setting could perhaps be put to the test by growing modular
and/or clustered networks in vitro and carrying out similar experiments as we
do here in simulation [34, 35] (see Discussion).
Results
The simplest neurons on modular networks
Consider a network of N model neurons, with activities si = ±1. The topology
is given by the adjacency matrix aij = {1, 0}, each element representing the
existence or absence of a synapse from neuron j to neuron i (a need not be
symmetric). In this kind of model -- a network of what are often referred to as
Amari-Hopfield neurons -- each edge usually has a synaptic weight associated,
ωij ∈ R, which serves to store information [1, 2, 3, 4]. However, since our
objective is to show how this can be achieved without synaptic learning, we
shall here consider these to have all the same value: ωij = ω > 0 ∀i, j. Neurons
are updated in parallel (Little dynamics) at each time step, according to the
stochastic transition rule
P (si → ±1) =
1
2(cid:20)± tanh(cid:18) hi
T (cid:19) + 1(cid:21) ,
(1)
where hi = ωPj aijsj is the field at neuron i, and T is a stochasticity parameter
called temperature. This dynamics can be derived by considering coupled binary
elements in a thermal bath, the transition rule stemming from energy differences
between states [3, 4, 36].
We shall consider the network defined by a to be made up of M distinct
modules. To achieve this, we can first construct M separate random directed
networks, each with n = N/M nodes and mean degree (mean number of neigh-
bours) hki. Then we evaluate each edge aij = 1 and, with probability λ, elim-
inate it (aij → 0), to be substituted for another edge between the original
(postsynaptic) neuron i and a new (presynaptic) neuron l chosen at random
from among any of those in other modules (ail → 1). We do not allow self-edges
(although they can occur in reality) since these could be regarded as equiv-
alent to a form of cellular bistability. Note that this protocol does not alter
the number of presynaptic neighbours of each node, kin
the number of postsynaptic neurons, kout
λ can be seen as a measure of modularity of the partition considered, since it
coincides with the expected value of the proportion of edges that link different
modules [37]. In particular, λ = 0 defines a network of disconnected modules,
while λ = 1 − M −1 yields a random network in which this partition has no
modularity. If λ ∈ (1 − M −1, 1), the partition is less than randomly modular --
i.e., it is quasi-multipartite (or multipartite if λ = 1).
i = Pj aij , although
i = Pj aji, can vary. The parameter
4
Cluster Reverberation
A memory pattern, in the form of a given configuration of activities, {ξi = ±1},
can be stored in this system with no need of prior learning. (The system will
recall the pattern perfectly when si = ξi, ∀i.) Imagine a pattern such that the
activities of all n neurons found in any module are the same -- i.e., ξi = ξµ(i),
where the index µ(i) denotes the module that neuron i belongs to. The system
can be induced into this configuration through the application of an appropriate
stimulus: the field of each neuron will be altered for just one time step according
to
hi → hi + δξµ(i),∀i,
where the factor δ is the intensity of the stimulus (see Fig. 1). This mechanism
for dynamically storing information will work for values of parameters such that
the system is sensitive to the stimulus, acquiring the desired configuration, yet
also able to retain it for some interval of time thereafter (a similar setting is
considered, for instance, in Ref. [38]).
Figure 1: Diagram of a modular network composed of four five-neuron clusters.
The four circles enclosed by the dashed line represent the stimulus: each is
connected to a particular module, which adopts the input state (red or blue) and
retains it after the stimulus has disappeared thanks to Cluster Reverberation.
The two configurations of minimum energy of the system are si = 1 ∀i and
si = −1 ∀i (see the next section for a more detailed discussion on energy).
However, the energy is locally minimized for any configuration in which each
module comprises either all active or all inactive neurons (that is, for configu-
rations si = dµ(i) ∀i, with dµ(i) = ±1 a binary variable specific to the whole
module µ(i) that neuron i belongs to). These are the configurations that we
shall use to store information. We define the mean activity of each module,
mµ ≡ hsiii∈µ, which is a mesoscopic variable, as well as the global mean ac-
tivity, m ≡ hsii∀i (these magnitudes change with time, but, where possible, we
shall avoid writing the time dependence explicitly for clarity; h·ix stands for an
5
average over x). The mean activity in a neural network model is usually taken
to represent the mean firing rate measured in experiments [39]. The extent to
which the network, at a given time, retains the pattern {ξi} with which it was
stimulated is measured with the overlap parameter mstim ≡ hξisiii = hξµmµiµ.
Ideally, the system should be capable of reacting immediately to a stimulus by
adopting the right configuration, yet also be able to retain it for long enough
to use the information once the stimulus has disappeared. A measure of perfor-
mance for such a task is therefore
η ≡
1
τ
t0+τ
Xt=t0+1
mstim(t),
where t0 is the time at which the stimulus is received and τ is the period of time
we are interested in (η ≤ 1) [38]. If the intensity of the stimulus, δ, is very
large, then the system will always adopt the right pattern perfectly and η will
only depend on how well it can then retain it. In this case, the best network will
be one that is made up of mutually disconnected modules (λ = 0). However,
since the stimulus in a real brain can be expected to arrive via a relatively small
number of axons, either from another part of the brain or directly from sensory
cells, it might be more realistic to assume that δ is of a similar order as the
input a typical neuron receives from its neighbours, hhi ∼ ωhki.
Figure 2 shows the mean performance obtained in Monte Carlo (MC) simula-
tions when the network is repeatedly stimulated with different randomly gener-
ated patterns. For low enough values of λ and stimuli of intensity δ & ωhki, the
system can capture and successfully retain any pattern it is "shown" for some
period of time, even though this pattern was in no way previously learned. For
less intense stimuli (δ < ωhki), performance is nonmonotonic with modularity:
there exists an optimal value of λ at which the system is sensitive to stimuli yet
still able to retain new patterns quite well.
Just as some degree of structural (quenched) noise, given by λ, can improve
performance by increasing sensitivity, so too the dynamical (annealed) noise set
by T can have a similar effect. This apparent stochastic resonance is analysed
in Analysis: The effect of noise.
Energy and topology
Each pair of neurons contributes a configurational energy eij = − 1
2 ω(aij +
aji)sisj [4]; that is, if there is an edge from i to j and they have opposite
activities, the energy is increased in 1
2 ω, whereas it is decreased by the same
amount if their activities are equal. Given a configuration, we can obtain its
associated energy by summing over all pairs. To study how the system relaxes
from the metastable states (i.e., how it "forgets" the information stored) we shall
be interested in configurations with x neurons that have s = +1 (and N − x
neurons with s = −1), chosen in such a way that one module at most, say µ,
has neurons in both states simultaneously. Therefore, x = nρ + z, where ρ is
the number of modules with all their neurons in the positive state and z is the
number of neurons with positive sign in module µ. We can write m = (2x−1)/N
and mµ = (2z − 1)/n. The total configurational energy of the system will be
eij =
E =Xij
ω(L↑↓ − hkiN ),
1
2
6
1
1
1
1
1
1
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
η
η
η
η
η
η
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
P
P
P
P
P
P
1
1
1
m
m
m
i
i
i
t
t
t
s
s
s
0.5
0.5
0.5
m
m
m
0
0
0
500
500
500
Time (MCS)
Time (MCS)
Time (MCS)
3500
3500
3500
δ=8.5
δ=8.5
δ=8.5
δ=9
δ=9
δ=10
λ=0.0
λ=0.0
λ=0.0
λ=0.25
λ=0.25
λ=0.5
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Figure 2: Performance η against λ for networks of the sort described in the main
text with M = 160 modules of n = 10 neurons each, hki = 9, obtained from
Monte Carlo (MC) simulations; patterns are shown with intensities δ = 8.5, 9
and 10, and performance is computed evey 200 time steps, preceding the next
random stimulus; T = 0.02 (error bars represent standard deviations; lines --
splines -- are drawn as a guide to the eye). Inset: typical time series of mstim
(i.e., the overlap with whichever pattern was last shown) for λ = 0.5 (bad
performance), 0 (intermediate), and 0.25 (optimal); with δ = ωhki = 9.
where L↑↓ is the number of edges linking nodes with opposite activities. By
simply counting over expected numbers of edges, we can obtain the expected
value of L↑↓ (which amounts to a mean-field approximation), yielding:
E
ωhki
+
N
2
=
λn
N − n{ρ[n − z + n(M − ρ − 1)]
z(n − z)
n − 1
.
+(M − ρ − 1)(z + nρ)} + (1 − λ)
(2)
Figure 3 shows the mean-field configurational energy curves for various values
of the modularity on a small modular network. The local minima (metastable
states) are the configurations used to store patterns. It should be noted that
the mapping x → m is highly degenerate: there are CM
mM patterns with mean
activity m that all have the same energy.
Forgetting avalanches
In obtaining the energy we have assumed that the number of synapses rewired
from a given module is always equal to its expected value: ν = hkinλ. However,
since each edge is evaluated with probability λ, ν will in fact vary somewhat
from one module to another, being approximately Poisson distributed with mean
hνi = hkinλ. The depth of the energy well corresponding to a given module is
then, neglecting all but the last term in Eq. (2) and approximating n − 1 ≃ n,
∆E ≃ 1
4 ω(nhki − ν). The typical escape time τ from an energy well of depth
7
M=5
M=5
M=5
M=5
λ=0.2
λ=0.2
λ=0.2
λ=0.2
λ=0.4
λ=0.4
λ=0.4
λ=0.6
λ=0.6
λ=0.8
0
0
0
0
-0.2
-0.2
-0.2
-0.2
-0.4
-0.4
-0.4
-0.4
-0.6
-0.6
-0.6
-0.6
-0.8
-0.8
-0.8
-0.8
)
)
)
)
N
N
N
N
k
k
k
k
ω
ω
ω
ω
(
(
(
(
/
/
/
/
E
E
E
E
2
2
2
2
-1
-1
-1
-1
-1
-1
-1
-1
-0.5
-0.5
-0.5
-0.5
0
0
0
0
m
m
m
m
0.5
0.5
0.5
0.5
1
1
1
1
Figure 3: Configurational energy of a network made up of M = 5 modules of
n = 10 neurons each, according to Eq. (2), for various values of λ (increasing
from bottom to top). The minima correspond to situations such that all neurons
within any given module have the same sign.
∆E at temperature T is τ ∼ e∆E/T [40]. Using Stirling's approximation [n! ∼
√2πn(n/e)n] in the Poisson distribution over ν and expressing it in terms of τ ,
we find that the escape times are distributed according to
where
β(τ ) = 1 +
P (τ ) ∼(cid:18)1 −
4T
ωnhki
ωnhki"1 + ln
4T
ωnhki ln τ!# .
λnhki
1 − 4T
2
ln τ(cid:19)− 3
τ −β(τ ),
(3)
(4)
Therefore, at low temperatures, P (τ ) will behave approximately like a power
law. Note also that the size of the network, N , does not appear in Eqs. (3) and
(4). Rather, T scales with n, which could be small even in the thermodynamic
limit (N → ∞).
The left panel of Fig. 4 shows the distribution of time intervals between
events in which the overlap mµ of at least one module µ changes sign. The
power-law-like behaviour is apparent, and justifies talking about forgetting avalanches
-- since there are cascades of many forgetting events interspersed with long pe-
riods of metastability. This is very similar to the behaviour observed in other
nonequilibrium settings in which power-law statistics arise from the convolution
of exponentials, such as demagnetization processes [41] or Griffiths phases on
networks [42].
It is known from experimental psychology that forgetting in humans is indeed
quite well described by power laws [28, 29, 30] -- although most experiments to
date seem to refer to slightly longer timescales than we are interested in here.
The right panel of Fig. 4 shows the value of the exponent β(τ ) as a function of
τ . Although for low temperatures it is almost constant over many decades of τ
-- approximating a pure power law -- for any finite T there will always be a τ
such that the denominator in the logarithm of Eq. (4) approaches zero and β
diverges, signifying a truncation of the distribution.
8
)
)
τ
τ
(
(
p
p
100
100
10-1
10-1
10-2
10-2
10-3
10-3
10-4
10-4
101
101
102
102
104
104
105
105
103
103
τ
τ
)
)
)
τ
τ
τ
(
(
(
β
β
β
2
2
2
1.8
1.8
1.8
1.6
1.6
1.6
1.4
1.4
1.4
1.2
1.2
1.2
1
1
1
T=1
T=2
T=3
101
101
101
103
103
103
τ
τ
τ
105
105
105
107
107
107
Figure 4: Left panel: distribution of escape times τ , as defined in the main
text, for λ = 0.25 and T = 2, from MC simulations. Slope is for β = 1.35.
Other parameters as in Fig. 2. Right panel: exponent β of the quasi-power-law
distribution p(τ ) as given by Eq. (4) for temperatures T = 1, 2 and 3 (from
bottom to top).
Note that we have considered the information stored in a pattern to be lost
once the system evolves to any other energy minimum. However, this new pat-
tern will be highly correlated with the original one, and it might be reasonable
to assume that the system has to escape from a large number of energy minima,
L, before the information can be considered to have been entirely forgotten. The
i=0 τi, where τi are independently drawn from Eq (3).
If L is sufficiently large, the distribution of times τ sum will tend to a L´evy dis-
tribution [43]. In practice, these different broad-tailed distributions [power-law,
L´evy, or as given by Eq. (3)] are likely to be indistinguishable experimentally
unless it is possible to observe over many orders of magnitude.
time for this is τ sum =PL
Clustered networks
Although we have illustrated how the mechanism of Cluster Reverberation works
on a modular network, it is not actually necessary for the topology to have this
characteristic -- only for the patterns to be in some way "coarse-grained," as
described, and that each region of the network encoding one bit have a small
enough parameter λ, defined as the proportion of synapses to other regions. For
instance, for the famous Watts-Strogatz small-world model [44] -- a ring of N
nodes, each initially connected to its k nearest neighbours before a proportion
p of the edges are randomly rewired -- we have λ ≃ p (which is not surprising
considering the resemblance between this model and the modular network used
above). More precisely, the expected modularity of a randomly imposed box of
n neurons is
λ = p −
n − 1
N − 1
p +
n (cid:18) k
1 − p
4 −
1
2(cid:19) ,
the second term on the right accounting for the edges rewired to the same box,
and the third to the edges not rewired but sufficiently close to the border to
connect with a different box.
Perhaps a more realistic model of clustered network would be a random
9
network embedded in d-dimensional Euclidean space. For this we shall use the
scheme laid out by Rozenfeld et al.
[45], which consists simply in allocating
each node to a site on a d-torus and then, given a particular degree sequence,
placing edges to the nearest nodes possible -- thereby attempting to minimize
total edge length. For a scale-free degree sequence (i.e., a set {ki} drawn from a
degree distribution p(k) ∼ k−γ) according to some exponent γ, then, as shown
in Analysis: Effective modularity of clustered networks, such a network has a
modularity
1
λ ≃
d(γ − 2) − 1hd(γ − 2)l−1 − l−d(γ−2)i ,
(5)
where l is the linear size of the boxes considered. It is interesting that even in
this scenario, where the boxes of neurons which are to receive the same stimulus
are chosen at random with no consideration for the underlying topology, these
boxes need not have very many neurons for λ to be quite low (as long as the
degree distribution is not too heterogeneous).
δ=3.5
δ=3.5
δ=3.5
δ=3.5
δ=4
δ=4
δ=4
δ=5
δ=5
δ=10
1
1
1
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
η
η
η
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
P
P
P
γ=2
γ=2
γ=2
γ=3
γ=3
γ=4
1
1
1
0.5
0.5
0.5
0
0
0
500
500
500
Time (MCS)
Time (MCS)
Time (MCS)
3000
3000
3000
m
m
m
i
i
i
t
t
t
s
s
s
m
m
m
0
0
0
0
0
0
0
0
0
0
0
0
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
4
4
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
SF exponent γ
5
5
5
5
5
5
5
5
5
5
5
5
6
6
6
6
6
6
6
6
6
6
6
6
Figure 5: Performance η against exponent γ for scale-free networks, embedded
on a 2D lattice, with patterns of M = 16 modules of n = 100 neurons each,
hki = 4 and N = 1600; patterns are shown with intensities δ = 3.5, 4, 5 and
10, and T = 0.01 (error bars represent standard deviations; lines -- splines -- are
drawn as a guide to the eye). Inset: typical time series for γ = 2, 3, and 4, with
δ = 5.
Carrying out the same repeated stimulation test as on the modular networks
in Fig. 2, we find a similar behaviour for the scale-free embedded networks. This
is shown in Fig. 5, where for high enough intensity of stimuli δ and scale-free
exponent γ, performance can, as in the modular case, be η ≃ 1. We should
point out that for good performance on these networks we require more neurons
for each bit of information than on modular networks with the same λ (in Fig.
5 we use n = 100, as opposed to n = 10 in Fig. 2). However, that we should be
10
able to obtain good results for such diverse network topologies underlines that
the mechanism of Cluster Reverberation is robust and not dependent on some
very specific architecture.
Spiking neurons
In the usual spirit of determining the minimal ingredients for a mechanism to
function we have, up until now, used the simplest model neurons able to ex-
hibit CR. This approach makes for a good illustration of the main idea and
allows for a certain amount of analytical understanding of the underlying phe-
nomena. However, before CR can be considered as a plausible candidate for
helping to explain short-term memory, we must check that it is compatible with
more realistic neural models. For this we examine the behaviour of the popular
Integrate-and-Fire (IF) model neurons -- often referred to as spiking neurons --
in the same kind of setting as described above for the simpler Amari-Hopfield
neurons. In the IF model, each neuron is characterized at time t by a membrane
potential V (t), described by the differential equation
τm
dV (t)
dt
= −V (t) + RmIin(t),
(6)
where τm and Rm are, respectively, the membrane time constant and resistance,
and Iin(t) = Isyn(t)+Ist(t)+Iext(t); the term Isyn(t) =Pj I j
syn(t) is the synap-
tic current generated by the arrival of Action Potentials (AP) from the neuron's
presynaptic neighbours, Ist(t) is the current generated by the presentation of
a particular external stimulus to the network and Iext(t) = I0 + √τmDξ(t) is
an additional noisy external current. Here I0 and D are constants and ξ(t) is
a Gaussian noise of mean hξ(t)i = 0 and autocorrelation hξ(t)ξ(t′)i = δ(t − t′).
Each synaptic contribution to the total synaptic current is modelled as I j
syn(t) =
Ayj(t), where yj(t) represents the fraction of neurotransmitters in the synaptic
cleft, which follows the dynamics [46]
dyj(t)
dt
= −
yj(t)
τin
+ U δ(t − tj
sp).
(7)
Here, tj
sp is the time at which an AP arrives at synapse j, inducing the re-
lease of a fraction U of neurotransmitters, and τin is the typical time-scale for
neurotransmitter inactivation. Whenever V surpasses a given threshold θ, the
neuron fires an AP to all its postsynaptic neighbours and V is reset to zero,
then undergoing a refractory time τref before again becoming susceptible to in-
put. Because the parameters and variables of this model represent measurable
physiological quantities, it is possible to use it to make quantitative -- albeit
tentative -- predictions about the timescales on which CR might be expected to
be effective in a real neural system.
Figure 6 is a raster plot of a modular network of IF neurons. The system
performs a short-term memory task akin to the one previously described for the
Amari-Hopfield neural network: the neurons belonging to clusters that corre-
spond to ones in a random pattern are stimulated, for 10 ms, with an intensity
Istim, while the the remaining neurons receive an opposite stimulus, −Istim.
We then allow the system to evolve for 500 ms, before choosing a new random
pattern and stimulating again. In such tests, the neurons in positively stimu-
lated clusters usually begin to oscillate in synchrony, while the rest remain silent
11
Figure 6: Raster plot, obtained from MC simulations, of a network of 1000
integrate-and-fire (IF) neurons wired up (as described in the main text) in
groups of 50, with a rewiring probability λ = 0.02. Every 500 ms, a new
pattern is shown for 10 ms with an intensity Istim = 500 pA (plotted in blue).
Parameters for the neurons are A = 42.5 pA, θ = 8 mV, τref = 5 ms, τin = 3
ms, U = 0.02, Rm = 0.1 GΩ and τm = 10 ms, which are all within the physio-
logical range; and the external noisy current is modelled with I0 = 15 pA and
D = 10 pA ms−1/2.
(save for occasional individual APs caused by noise). However, since this is a
metastable state, with time active clusters can suddenly go mostly silent, or the
neurons in silent clusters begin spontaneously to fire in synchrony. Thus, the
information is gradually lost, as in the case with simpler neurons.
To gauge how well the system is performing the task, we look at each cluster
µ for the last 100 ms before the next stimulus and assign a value mµ = 1 to
its mean activity if it is active, and mµ = −1 if it is silent. We then define the
performance as:
η =
mµξµ.
(8)
1
M Xµ
In Fig. 7 we show the values of η obtained in MC simulations against λ. Using
different values of Istim we observe a similar behaviour to that of Fig. 2. In
particular, for Istim ≃ 100 pA, we have the interesting nonmonotonic behaviour
in which performance benefits from a certain degree of rewiring. While, for the
sake of illustration, in Fig. 6 we only show the evolution of the system for 500 ms
after stimulation, in Fig. 7 we wait for five seconds. Although the model is too
simple, and the network too small, to make quantitative predictions about the
brain, it is nevertheless promising that with physiologically realistic parameters
we observe high performance (η ≃ 1) over several seconds, since this is the
12
Istim=98
Istim=98
Istim=98
Istim=100
Istim=100
Istim=200
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.01
0.02
0.03
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.02
0.03
0.03
0.03
0.03
0.03
0.03
0.03
0.03
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
Rewiring prob. λ
0.04
0.04
0.04
0.04
0.04
0.04
0.04
0.04
0.04
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
0.05
Figure 7: Performance η against rewiring λ for modular networks of IF neurons,
as obtained from MC simulations. The network is periodically stimulated with a
new random pattern for 10 ms with an intensity Istim = 98 pA (green squares),
100 pA (red circles) and 200 pA (blue triangles) (error bars represent standard
deviations; lines -- splines -- are drawn as a guide to the eye). The system evolves
in the absence of stimuli for 5000 ms and performance, η, is computed according
to Eq. (8). (An interval of 5 seconds corresponds roughly to the timescale on
which short-term memory operates in the brain.) Other parameters are as in
Fig. 6.
timescale on which short-term memory operates in humans.
Discussion
Cluster Reverberation may be a means available to neural systems for perform-
ing certain short-term tasks, such as sensory memory or working memory. To
the best of our knowledge, it is the first mechanism proposed to use network
properties with no need of synaptic learning. All that is required is for the
underlying network to be highly clustered or modular, and for small groups of
neurons in some sense to store one bit of information, as opposed to a con-
ventional view which assumes one bit per neuron. Considering the enormous
number of neurons in the brain, and the fact that real neurons are possibly
too noisy to store information individually anyway, these hypotheses do not
seem far-fetched. The mechanism is furthermore consistent with what is known
about the structure of biological neural networks, with experiments that have
revealed power-law statistics of forgetting, and with recent observation of locally
synchronized synaptic activity.
For the sake of illustration, we have focused here on the simplest model
13
neurons that are able to exhibit the behaviour of interest. However, we have
shown how the mechanism can also work with the slightly more sophisticated
Integrate-and-Fire neurons, and there is no reason to believe that it would not
also be viable with more realistic models, or even actual cells. Although CR
comes about thanks to the high modularity of small groups of neurons, we have
shown how robust it is to the details of the topology by carrying out simulations
on clustered networks with no explicitly built-in modularity. And this setting
suggests an interesting point. If an initially homogeneous (i.e., neither modular
nor clustered) area of brain tissue were repeatedly stimulated with different
patterns in the same way as we have done in our simulations, then synaptic
plasticity mechanisms (LTP and LTD) might be expected to alter the network
structure in such a way that synapses within each of the imposed modules
would all tend to become strengthened, while inter-module synapses would vary
their weights in accordance with the details of the patterns being shown [47].
The result would be a modular structure conducive to efficient CR for arbitrary
patterns, with simultaneous Hebbian learning in the inter-synapses of the actual
patterns shown. In this way, the same network might be capable of both short-
term and long-term memory, explaining, perhaps, why our brains can indeed
store completely novel information but usually with a certain bias in favour of
what we are expecting to perceive.
Although we have not gone into the question of neural coding, there would
seem to be an intrinsic difference between semantic storage of information --
used for long-term memory and probably useful for certain working-memory
tasks that require the labelling of previously learned information -- and sensory
storage, for which some mechanism such as the one proposed here must store
novel information immediately -- in a similar but more efficient way to how
the retina retains the pigmentation left by an image it was recently exposed
to. If novel sensory information were held for long enough in metastable states,
Hebbian learning (either in the same or other areas of the brain) could take place
and the information be stored thereafter indefinitely. This might constitute the
essence of concentrating so as to memorise a recent stimulus.
Finally, we should mention that CR could work in conjunction with other
mechanisms, such as processes leading to cellular bistability, making these more
robust to noise and augmenting their efficacy. Whether CR would work for bio-
logical neural systems could perhaps be put to the test by growing such modular
networks in vitro, stimulating appropriately, and observing the duration of the
metastable states [34, 35]. In vivo recordings of neural activity during short-term
memory tasks, together with a mapping of the underlying synaptic connections,
might be used to ascertain whether the brain could indeed harness this mech-
anism. For this it must be borne in mind that the neurons forming a module
need not find themselves close together in metric space, and that effective mod-
ularity might come about via stronger intra- than inter-connexions, instead of
simply through a higher density of synapses within the clusters. We hope that
observations and experiments such as these will be carried out and eventually
reveal something more about the basis of this puzzling emergent property of the
brain's known as thought.
14
Analysis
The effect of noise
On a random network (λ = 1 − M −1), the Amari-Hopfield model described
in the main text has a second order phase transition with temperature, T , at
Tc = ωk [4]. This can be seen by considering the mean-field equation for the
overlap at the steady state, m = tanh(ωkm/T ), where we have substituted
hi = ωPj aijsi → ωkm in Eq. (1). For T < Tc, the paramagnetic solution
m = 0 becomes unstable, and ferromagnetic solutions m 6= 0 appear [36]. This
result also holds for the modular networks described in the main text. However,
that the global overlap m is different form zero does not mean that the short-
term memory configurations we are interested in are stable. In fact, we know
they are metastable for any T > 0 (see Section Energy and topology), but we can
set an upper bound on the temperature at which these states can be maintained
even for a short time by considering again the mean-field equation for such a
configuration. For a neuron in module µ, hi → ωk[(1 − λ)mµ + λm]. For
patterns with mean activity zero (m = 0), states mµ 6= 0 will be unstable if
T > (1 − λ)ωk ≤ Tc.
λ=0.0
λ=0.0
λ=0.0
λ=0.0
λ=0.1
λ=0.1
λ=0.1
λ=0.2
λ=0.2
λ=0.3
η
η
η
η
η
η
η
η
η
η
η
η
e
e
e
e
e
e
e
e
e
e
e
e
c
c
c
c
c
c
c
c
c
c
c
c
n
n
n
n
n
n
n
n
n
n
n
n
a
a
a
a
a
a
a
a
a
a
a
a
m
m
m
m
m
m
m
m
m
m
m
m
r
r
r
r
r
r
r
r
r
r
r
r
o
o
o
o
o
o
o
o
o
o
o
o
f
f
f
f
f
f
f
f
f
f
f
f
r
r
r
r
r
r
r
r
r
r
r
r
e
e
e
e
e
e
e
e
e
e
e
e
P
P
P
P
P
P
P
P
P
P
P
P
1
1
1
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
1
1
1
1
1
1
1
1
1
1
1
2
2
2
2
2
2
2
2
2
2
2
2
3
3
3
3
3
3
3
3
3
3
3
3
4
4
4
4
4
4
4
4
4
4
4
4
5
5
5
5
5
5
5
5
5
5
5
5
6
6
6
6
6
6
6
6
6
6
6
6
7
7
7
7
7
7
7
7
7
7
7
7
8
8
8
8
8
8
8
8
8
8
8
8
9
9
9
9
9
9
9
9
9
9
9
9
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Temperature T
Figure 8: Performance η against T for the Hopfield-Amari networks described
in the main text, obtained from MC simulations, for values of the rewiring
λ = 0.0, 0.1, 0.2 and 0.3, and stimulus δ = 8.5. All other parameters as in Fig.
2. (Error bars represent standard deviations; lines -- splines -- are drawn as a
guide to the eye).
As we saw from Fig. 2, for stimuli δ . ωhki, the system does not always
leave whichever meatastable state it is in to go perfectly to the pattern shown.
A degree of "structural noise" (λ > 0) can lead to a better response. In the same
way, the dynamical noise set by T can improve performance. Figure 8 shows
how performance varies with T for different values of λ. Due to the trade-off
between sensitivity to stimuli and stability of the memory states, there is in
general an optimum level of noise at which the system performs best. This
15
dynamics can be interpreted as a kind of stochastic resonance, with the stimuli
playing the part of the periodic forcing typically seen in such systems [48]. Both
the dynamic (annealed) noise, T , and the structural (quenched) noise, λ, serve
to increase the sensitivity of the system to stimuli.
It is interesting to observe in Fig. 8 that whereas highly modular networks
(λ ≃ 0) are most robust to T , for no values of parameters do they exhibit as
good performance as the less modular networks when T is relatively low.
Effective modularity of clustered networks
We wish to estimate λ, the proportion of edges that cross the boundaries of a
box of linear size l placed randomly on a network embedded in d-dimensional
space according to the scheme laid out in Ref. [45]. The number of nodes within
a radius r is n(r) = Adrd, with Ad a constant. We shall therefore assume a node
with degree k to have edges to all nodes up to a distance r(k) = (k/Ad)1/d, and
none beyond (note that this is not necessarily always feasible in practice). To
estimate λ, we shall first calculate the probability that a randomly chosen edge
have length x. The chance that the edge belong to a node with degree k is
π(k) ∼ kp(k) (where p(k) is the degree distribution). The proportion of edges
that have length x among those belonging to a node with degree k is ν(xk) =
dAdxd−1/k if Adxd < k, and 0 otherwise. Considering, for example, scale-free
networks (as in Ref. [45]), so that the degree distribution is p(k) ∼ k−γ in some
interval k ∈ [k0, kmax], and integrating over p(k), we have the distribution of
lengths,
P (x) = (Const.)Z kmax
max(k0,Axd)
π(k)ν(kx)dk = d(γ − 2)x−[d(γ−2)+1],
where we have assumed, for simplicity, that the network is sufficiently sparse
that max(k0, Axd) = Axd, ∀x ≥ 1, and where we have normalised for the
interval 1 ≤ x < ∞; strictly, x ≤ (kmax/A)1/d, but we shall also ignore this
effect. Next we need the probability that an edge of length x fall between two
compartments of linear size l. This depends on the geometry of the situation as
well as dimensionality; however, a first approximation which is independent of
such considerations is
Pout(x) = min(cid:16)1,
x
l(cid:17) .
We can now estimate the modularity λ as
λ =Z ∞
1
Pout(x)P (x)dx =
1
d(γ − 2) − 1hd(γ − 2)l−1 − l−d(γ−2)i .
Figure 9 compares this expression with the value obtained numerically after
averaging over many network realizations, and shows that it is fairly good --
considering the approximations used for its derivation.
Acknowledgements
Many thanks to Jorge F. Mejias, Sebastiano de Franciscis, Miguel A. Munoz,
Sabine Hilfiker, Peter E. Latham, Ole Paulsen and Nick S. Jones for useful
16
λ
λ
λ
λ
λ
λ
λ
λ
λ
1
1
1
1
1
1
1
1
1
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.8
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.6
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.4
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0.2
0
0
0
0
0
0
0
0
0
l=2
l=4
l=8
2
2
2
2
2
2
2
2
2
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
2.5
3
3
3
3
3
3
3
3
3
3.5
3.5
3.5
3.5
3.5
3.5
3.5
3.5
3.5
γ
γ
γ
γ
γ
γ
γ
γ
γ
4
4
4
4
4
4
4
4
4
4.5
4.5
4.5
4.5
4.5
4.5
4.5
4.5
4.5
5
5
5
5
5
5
5
5
5
Figure 9: Proportion of outgoing edges, λ, from boxes of linear size l against
exponent γ for scale-free networks embedded on 2D lattices. Lines from Eq. (5)
and symbols (with error bars representing standard deviations) from simulations
with hki = 4 and N = 1600.
comments and suggestions. This work was supported by Junta de Andaluc´ıa
projects FQM-01505 and P09-FQM4682, by the joint Spanish Research Min-
istry (MEC) and the European Budget for the Regional Development (FEDER)
project FIS2009-08451, and by the Granada Research of Excellence Initiative
on Bio-Health (GREIB) traslational project GREIB.PT 2011 19 of the Spanish
Science and Innovation Ministry (MICINN) "Campus of International Excel-
lence." S.J. is grateful for financial support from the Oxford Centre for Inte-
grative Systems Biology, and from the European Commission under the Marie
Curie Intra-European Fellowship Programme PIEF-GA-2010-276454.
17
References
[1] Hebb, D.O., The Organization of Behavior, Wiley, New York, 1949.
[2] S. Amari, "Characteristics of random nets of analog neuron-like elements",
IEEE Trans. Syst. Man. Cybern., 2 643 -- 657 (1972).
[3] J.J. Hopfield, "Neural networks and physical systems with emergent col-
lective computational abilities", Proc. Natl. Acad. Sci. USA 79 2554 -- 8
(1982).
[4] Amit, D.J.,Modeling Brain Function, Cambridge Univ. Press, Cambridge,
1989.
[5] A. Gruart, M.D. Munoz, and J.M. Delgado-Garc´ıa, "Involvement of the
CA3-CA1 synapse in the acquisition of associative learning in behaving
mice", J. Neurosci. 26, 1077 -- 87 (2006).
[6] M. De Roo, P. Klauser, P. Mendez, L. Poglia, and D. Muller, "Activity-
dependent PSD formation and stabilization of newly formed spines in hip-
pocampal slice cultures", Cerebral Cortex 18, 151 (2008).
[7] D. Durstewitz, J.K. Seamans, and T.J. Sejnowski, "Neurocomputational
models of working memory", Nature Neuroscience 3, 1184 -- 91 (2000).
[8] K.S. Lee, F. Schottler, M. Oliver, and G. Lynch, "Brief bursts of high-
frequency stimulation produce two types of structural change in rat hip-
pocampus", J. Neurophysiol. 44, 247 (1980).
[9] A.Y. Klintsova and W.T. Greenough, "Synaptic plasticity in cortical sys-
tems", Current Opinion in Neurobiology 9, 203 (1999).
[10] G.A. Sperling, "The information available in brief visual persentation",
Psychological Monographs: General and Applied 74, 1 -- 30 (1960).
[11] N. Cowan, "On Short And Long Auditory Stores", Psychological Bulletin
96, 341 -- 70 (1984).
[12] Baddeley, A. D., Essentials of Human Memory, Psychology Press, London,
1999.
[13] A. Baddeley, "Working memory: looking back and looking forward", Na-
ture Reviews Neuroscience 4, 829-39 (2003).
[14] R. Conrad, "Acoustic confusion in immediate memory", B. J. Psychol. 55,
75 -- 84 (1964).
[15] R. Conrad, "Information, acoustic confusion and memory span", B. J. Psy-
chol. 55, 429 -- 432 (1964).
[16] X.-J. Wang, "Synaptic reverbaration underlying mnemonic presistent ac-
tivity", TRENDS in Neurosci. 24, 455 -- 63 (2001).
[17] O. Barak and M. Tsodyks, "Persistent activity in neural networks with
dynamic synapses", PLoS Comput. Biol. 3(2): e35 (2007).
18
[18] Y. Roudi and P.E. Latham, "A balanced memory network", PLoS Comput.
Biol. 3(9): e141 (2007).
[19] G. Mongillo, O. Barak, and M. Tsodyks, "Synaptic Theory of Working
Memory", Science 319, 1543 -- 1546 (2008).
[20] D.R. Chialvo, G.A. Cecchi, and M.O. Magnasco, "Noise-induced memory
in extended excitable systems", Phys. Rev. E 61 5654 -- 7 (2000).
[21] M. Camperi and X.-J. Wang, "A model of visuospatial working memory
in prefrontal cortex: recurrent network and cellular bistability", J. Comp.
Neurosci. 5, 383 -- 405 (1998).
[22] J.-N. Teramae and T. Fukai, "A cellular mechanism for graded persistent
activity in a model neuron and its implications for working memory", J.
Comput. Neurosci. 18, 105 -- 21 (2005).
[23] E. Tarnow, "Short Term Memory May Be the Depletion of the Readily
Releasable Pool of Presynaptic Neurotransmitter Vesicles", Cognitive Neu-
rodynamics 3, 263 -- 9, (2008).
[24] A. Compte, C. Constantinidis,
J. Tegn´er, S. Raghavachari, S.
Raghavachari, M.V. Chafee, P.S. Goldman-Rakic, and X.-J. Wang, "Tem-
porally irregular mnemonic persistent activity in prefrontal neurons of mon-
keys during a delayed response task", J. Neurophysiol. 90 3441-54 (2003).
[25] O. Sporns, D.R. Chialvo, M. Kaiser and C.C. Hilgetag, "Organization,
development and function of complex brain networks", Trends Cogn. Sci.
8 418 -- 25 (2004).
[26] S. Johnson, J. Marro, and J.J. Torres, "Evolving networks and the devel-
opment of neural systems", J. Stat. Mech. (2010) P03003.
[27] "Self-organized synchronous oscillations in a network of excitable cells cou-
pled by gap junctions", T.J. Lewis and J. Rinzel, Comput. Neural Syst. 11
299 -- 320 (2000).
[28] J.T. Wixted and E.B. Ebbesen, "On the form of forgetting", Psychological
Science 2 409 -- 15 (1991).
[29] J.T. Wixted and E.B. Ebbesen, "Genuine power curves in forgetting: A
quantitative analysis of individual subject forgetting functions", Memory
& Cognition 25, 731 -- 9 (1997).
[30] S. Sikstrom, "Forgetting curves:
implications for connectionist models",
Cognitive Psychology 45, 95 -- 152 (2002).
[31] N. Takahashi, K. Kitamura Naoki Matsuo, M. Mayford, M. Kano, N. Mat-
suki and Y. Ikegaya "Locally Synchronized Synaptic Inputs"Science 335,
353 -- 6 (2012).
[32] T. Kleindienst, J. Winnubst, C. Roth-Alpermann, T. Bonhoeffer and C.
Lohmann, "Activity-Dependent Clustering of Functional Synaptic Inputs
on Developing Hippocampal Dendrites", Neuron 72, 1012 (2011).
19
[33] H. Makino and R. Malinow, "Compartmentalized versus Global Synaptic
Plasticity on Dendrites Controlled by Experience", Neuron 72, 1001 (2011).
[34] M.M. Kohl, O.A. Shipton, R.M. Deacon, J.N. Rawlins, K. Deisseroth, and
O. Paulsen, "Hemisphere-specific optogenetic stimulation reveals left-right
asymmetry of hippocampal plasticity", Nat. Neurosci. 14, 1413 -- 5 (2011)
[35] M. Shein-Idelson, E. Ben-Jacob, and Y. Hanein, "Engineered neuronal cir-
cuits: A new platform for studying the role of modular topology", Front.
Neuroeng. 4, 00010 (2011)
[36] D. Fraiman, P. Balenzuela, J Foss, D. Chialvo, "Ising-like dynamics in
large-scale functional brain networks", Phys. Rev. E, 79, 61922 -- 31 (2009).
[37] M.E.J. Newman, "The structure and function of complex networks", SIAM
Reviews 45, 167 (2003).
[38] S. Johnson, J. Marro, and J.J. Torres, "Functional optimization in complex
excitable networks", EPL 83, 46006 (2008).
[39] J.J. Torres, and P. Varona, "Modeling Biological Neural Networks", Hand-
book of Natural Computing, Springer-Verlag, Berlin, 2010.
[40] Levine, R.D., Molecular Reaction Dynamics, Cambridge University Press,
Cambridge, 2005.
[41] P.I. Hurtado, J. Marro, and P.L. Garrido, "Demagnetization via nucleation
of the nonequilibrium metastable phase in a model of disorder", J. Stat.
Phys. 133, 29-58 (2008).
[42] M.A. Munoz, R. Juhasz, C. Castellano, G. Odor, "Griffiths phases in com-
plex networks", Phys. Rev. Lett. 105, 128701 (2010).
[43] Gnedenko, B.V, and Kolmogorov, A. N., Limit Distributions for Sums of
Independent Random Variables, Cambridge, Addison-Wesley, 1954.
[44] D.J. Watts and S.H. Strogatz, "Collective dynamics of 'small-world' net-
works" Nature 395, 440 -- 2 (1998)
[45] A.F. Rozenfeld, R. Cohen, D. ben-Avraham, and S. Havlin, "Scale-free
networks on lattices", Phys. Rev. Lett. 89, 218701 (2002).
[46] M. V. Tsodyks and H. Markram, "The neural code between neocortical
pyramidal neurons depends on neurotransmitter release probability", Proc.
Natl. Acad. Sci. USA 94, 719 (1997).
[47] K.I. van Aerde, E.O. Mann, C.B. Canto, T.S. Heistek, K. Linkenkaer-
Hansen, M. van der Roest, A.B. Mulder, O. Paulsen, A.B. Brussaard, and
H.D. Mansvelder, "Flexible spike timing of layer 5 neurons during dynamic
beta-oscillation shifts in rat prefrontal cortex" J. Physiol. 587, 5177 -- 96
(2009).
[48] R. Benzi, A. Sutera and A. Vulpiani, "The mechanism of stochastic reso-
nance", J. Phys. A 14, 453 -- 457 (1981).
20
|
1610.05654 | 2 | 1610 | 2016-10-24T11:22:52 | The infochemical core | [
"q-bio.NC",
"cs.CL"
] | Vocalizations and less often gestures have been the object of linguistic research over decades. However, the development of a general theory of communication with human language as a particular case requires a clear understanding of the organization of communication through other means. Infochemicals are chemical compounds that carry information and are employed by small organisms that cannot emit acoustic signals of optimal frequency to achieve successful communication. Here the distribution of infochemicals across species is investigated when they are ranked by their degree or the number of species with which it is associated (because they produce or they are sensitive to it). The quality of the fit of different functions to the dependency between degree and rank is evaluated with a penalty for the number of parameters of the function. Surprisingly, a double Zipf (a Zipf distribution with two regimes with a different exponent each) is the model yielding the best fit although it is the function with the largest number of parameters. This suggests that the world wide repertoire of infochemicals contains a chemical nucleus shared by many species and reminiscent of the core vocabularies found for human language in dictionaries or large corpora. | q-bio.NC | q-bio | The infochemical core
Antoni Hernández‐Fernández1,2,*, Ramon Ferrer‐i‐Cancho1
(1) Complexity and Quantitative Linguistics Lab, LARCA Research Group. Departament de
Ciències de la Computació, Universitat Politècnica de Catalunya. Barcelona (Catalonia), Spain.
(2) Institut de Ciències de l'Educació, Universitat Politècnica de Catalunya. Barcelona
(Catalonia), Spain.
*Author for correspondence ([email protected]).
Keywords: chemical communication, pheromones, infochemicals, semiochemicals, Zipf's law.
*Address correspondence to: Antoni Hernández-Fernández, Institut de Ciències de l'Educació, Universitat
Politècnica de Catalunya, Campus Nord, Edifici Vèrtex, Plaça Eusebi Güell, 6, 08034 Barcelona (Catalonia), Spain.
Tel: +34 649 00 51 53. Email: [email protected]
ABSTRACT
Vocalizations, and less often gestures, have been the object of linguistic research for decades.
However, the development of a general theory of communication with human language as a
particular case requires a clear understanding of the organization of communication through
other means. Infochemicals are chemical compounds that carry information and are employed
by small organisms that cannot emit acoustic signals of an optimal frequency to achieve
successful communication. Here, we investigate the distribution of infochemicals across
species when they are ranked by their degree or the number of species with which they are
associated (because they produce them or are sensitive to them). We evaluate the quality of
the fit of different functions to the dependency between degree and rank by means of a
penalty for the number of parameters of the function. Surprisingly, a double Zipf (a Zipf
distribution with two regimes, each with a different exponent) is the model yielding the best fit
although it is the function with the largest number of parameters. This suggests that the
2
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
worldwide repertoire of infochemicals contains a core which is shared by many species and is
reminiscent of the core vocabularies found for human language in dictionaries or large
corpora.
1. INTRODUCTION
Quantitative linguistics is a discipline with a tremendous capacity to explore connections
between human language and other natural systems. The key is that a hypothetical connection
between language and certain natural systems can be investigated simply by looking at certain
statistical properties for which only minimal assumptions are required. For instance, the
tendency of more frequent words to be shorter (Zipf, 1935, 1949), is also found in DNA
sequences (Naranan & Balasubrahmanyan, 2000; Ferrer‐i‐Cancho et al., 2013a) and the
behaviour of non‐human species (see Ferrer‐i‐Cancho et al., 2013b for a review). A non‐trivial
version of Menzerath's law, i.e. the tendency of the size of the parts of a linguistic construct to
decrease as its number of parts increases (Menzerath, 1954; Altmann, 1980), is found in
genomes at different levels of analysis (Wilde & Schwibbe, 1989; Li, 2012; Ferrer‐i‐Cancho et
al., 2013a). In general, the depth of a connection of the kind described above depends on
various factors. One factor is the existence of conceptual similarities between both at some
level of abstraction (Ferrer‐i‐Cancho et al., 2013a). For instance, the striking conceptual
similarities between human words and codons (Bel‐Enguix & Giménez‐López, 2011) are
reinforced by other similarities arising from statistical analyses (Naranan & Balasubrahmanyan,
2000; Balasubrahmanyan & Naranan, 2000; Naranan, 2011).
Another factor is the number of statistical properties that coincide simultaneously: the
likelihood of a non‐trivial connection cannot decrease as the number of shared statistical
regularities increases (Rao et al., 2012; Ferrer‐i‐Cancho & McCowan, 2012). For instance, the
various statistical similarities between dolphin whistles and human words (McCowan et al.,
1999; Ferrer‐i‐Cancho & McCowan, 2009; Ferrer‐i‐Cancho & McCowan, 2012) suggest that
THE INFOCHEMICAL CORE
3
dolphin whistles may have a communicative function resembling that of human words.
However, this is not the only reason why quantitative linguistics has a tremendous potential to
bridge the gap between different fields. Due to its minimal assumptions, quantitative
linguistics is in a privileged position for selecting candidates for extraterrestrial forms of
intelligence (Doyle et al., 2011) or shedding light on the linguistic nature of undeciphered
scripts (Rao, 2010).
Contrary to what many believe, quantitative linguistics is more than data analysis or the art of
collecting statistical curiosities. A less well‐known contribution of quantitative linguistics going
back at least to the foundation of modern quantitative linguistics by G. K. Zipf during the first
half of the twentieth century (Zipf, 1949), are abstract principles that can explain the
recurrence of certain statistical patterns, not only in language but beyond, to ultimately
establish laws of language or human behaviour. G. K. Zipf's idea of a minimum equation
defining the cost of a set of tools has recently been interpreted as a precursor of the notion of
mean code length in information theory (Ferrer‐i‐Cancho et al., 2013b). Compression, i.e. the
minimization of that mean code length has been used to shed light on the origins of the law of
brevity, not only in human language but in other species too (Ferrer‐i‐Cancho et al., 2013b),
and DNA sequences (Naranan & Balasubrahmanyan, 2000). A compromise between entropy
minimization and mutual information minimization, inspired by Zipf's conflict between
unification and diversification (Zipf, 1949), has been used to explain the recurrence of "Zipf's
law" patterning in languages (Ferrer‐i‐Cancho, 2005; Prokopenko et al., 2010; Dickman et al.,
2011). This explanation (a) is abstract enough to be valid for DNA (Ji, 1999) and other natural
systems where Zipf's‐law‐like patterning has been found (e.g. Searls, 2002; McCowan et al.,
1999) and (b) challenges the view that the solutions of information theory do not resemble
those of natural languages (Christiansen & Chater, 2015).
4
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Human vocalizations, in the form of speech or written language (Akmajian et al., 1997), and
less intensively gestures (Sandler & Lillo‐Martin, 2006), have been objects of research from
standard linguistics and neighbouring disciplines for decades. However, the development of a
general theory of communication with human language as a particular case requires a clear
understanding of the organization of information transfer through other means. While
language is believed to be uniquely human (Hauser et al., 2002; Pinker, 2003), a statistically
rigorous test of the statement is lacking. While researchers struggle to date the origins of
language in modern humans or ancestors, bacteria started linguistic communication millions
and millions of years before a vocalizing multicellular organism appeared on Earth (Ben Jacob
et al., 2004). Thus, chemical communication might be the oldest communication system
produced in our planet. The same applies to writing, a milestone in the development of
communicative skills (in humans), being as it is a persistent method of communication that
helps one to detach oneself from the here and now. However, bacteria pioneered the
exchange of documents, i.e. plasmids, on our planet (Head, 1999/2000).
Language research is anthropocentric: it is based on the assumption that putting human
language and human biology at the centre provides a sufficient level of abstraction for solving
the puzzle of the origins of language (Hurford, 2012; Fitch, 2010). However, this human‐
centred vision has a certain flexibility: it allows one to include existing or extinct species that
are close in phylogeny or to jump further to species exhibiting, like ourselves, complex vocal
behaviour such as songs. This view is challenged again by the non‐vocal and non‐gestural
linguistic communication of brainless unicellular organisms through chemical compounds:
bacteria (Ben Jacob et al., 2004). The challenge adds further support for the proposal of a new
paradigm for language research including biology and computer science (Bel‐Enguix &
Jiménez‐López, 2012).
THE INFOCHEMICAL CORE
5
Here we consider the particular case of chemical communication in multicellular organisms, a
domain that has received, to our knowledge, little attention by quantitative linguistics
research, with some exceptions (Doyle, 2009). The goal of this article is applying concepts and
tools from quantitative linguistics to shed light on the organization of infochemicals, chemical
compounds that carry information (Wyatt, 2003).
2. INFOCHEMICALS
The complexity of chemical communication probably requires a zoosemiotic approach (Maran,
Martinelli & Turovski, 2011; Riba, 1990), but at present we must settle for analysing signals
that can be detected and describe the communicative contexts in which they are emitted.
Chemicals provide information (cues) or act as precursors of communication (signals) or as
central elements of communication systems that probably evolved from non‐communicative
compounds with a phylogenetic pattern (Symonds & Elgar, 2004; Steiger et al., 2011, for a
review). Chemical signals constitute much of the language of life in the sea (Hay, 2009) and
also on dry land (Wyatt, 2003). We humans have a poor capacity to understand chemical
interactions, partly due to our rather stunted sense of smell, but our modern technology
allows us to explore the fascinating world of infochemicals. Infochemicals are usually divided
into two groups: pheromones and allelochemicals (Wyatt, 2010; Dicke & Sabelis, 1988;
Nordlund & Lewis, 1976). Pheromones are defined classically as semiochemicals involved in
intraspecific communication (Law & Regnier, 1971; Regnier & Law, 1968), substances secreted
to the outside by an organism (in contrast with hormones, secreted inside an organism) and
perceived by a second individual of the same species in which they release a specific reaction
(Karlson & Lüscher, 1959; Wyatt, 2003), in opposition to allelochemicals, which mediate
interspecies communications (Nordlund & Lewis, 1976) as happens between plants and
insects.
6
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
The modern division of infochemicals by function (Wyatt, 2010, for a general review)
distinguishes various classes of infochemicals, more or less appropriate according to the area
of research concerned (Dicke & Sabelis, 1988). Thus, for example, the classification chosen by
El‐Sayed for the Pherobase, a free database of infochemicals (El‐Sayed, 2012) distinguishes
between five behavioural functions (for alternative classifications see Nordlund & Lewis, 1976;
Dicke & Sabelis, 1988; Wyatt, 2010): pheromones, involved in intraspecific communication;
attractants, or infochemicals that cause aggregation of individuals, secreted by species or
synthesized by humans; allomones, or allelochemicals that benefit the sender; kairomones or
allelochemicals beneficial for the receiver; and finally synomones, which are allelochemicals
benefiting both signaller and receiver in mutualistic interactions.
The Pherobase (El‐Sayed, 2012) is a wide‐ranging database that incorporates a list of species
that produce, or are sensitive to, each
infochemical, as well as other biochemical
characteristics. Pheromones and allelochemicals are a way of transmitting information whose
main advantage is their ability to spread and persist in the environment in which they expand.
Their diffusion is conditioned to the predominant species' habitat and its activity (Okubo et al.,
2001), as had already been concluded by pioneer studies of infochemicals (Wilson, 1958;
Wilson & Bossert, 1963; Wilson, 1970).
***TABLE 1 NEAR HERE***
Here, the degree of an infochemical is defined as the number of species that are associated
with it, because they produce it or are sensitive to it, according to the Pherobase (El‐Sayed,
2012). The infochemical with the highest degree has rank 1, the second has rank 2, and so on
(Table 1). The number of functions that an infochemical serves for a given species is irrelevant
for our notion of degree. For instance, if an infochemical is associated with only one species,
the infochemical will have degree one regardless of the number of functions served.
THE INFOCHEMICAL CORE
7
In this article, the fit of different functions to the rank distribution of infochemical associations
in nature is studied. The list of functions considered is summarized in Table 2. It is found that
the function providing the best balance between the exactness of the fit and the number of
parameters is a double Zipf (a power law with two different exponents) although it is the
function with the largest number of parameters. This suggests that infochemicals have a core
repertoire analogous to the core vocabulary found in human language (Ferrer‐i‐Cancho & Solé,
2001; Petersen et al., 2012; Gerlach & Altmann, 2013; Cocho et al., 2015).
***TABLE 2 NEAR HERE***
3. MATERIALS AND METHODS
f(r) is defined as the degree of rank r, n as the maximum rank (r = 1,2,…,n) and T as the sum of
all the degrees, i.e.
T
where n is the size of the repertoire.
n
r
1
rf
,)(
(1)
Various two‐parameter models such as Zipf's function, Beta function, Yule function or
Menzerath‐Altmann function (see Table 2) can be fitted to rank‐frequency data (Li et al.,
2010). Here, Li et al.'s (2010) model‐selection methodology is adopted to study rank‐degree
data in infochemicals. A two regime distribution Zipf distribution (Ferrer i Cancho & Solé, 2001)
is added to the list of functions explored by Li et al. (2010).
3.1. Materials
The degree and the classification of each infochemical come from the Pherobase (El‐Sayed
2012) which is freely available at http://www.pherobase.com/. A study primarily concerned
8
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
with compounds which exist in nature and which regulate communication without human
intervention should discard attractants synthesized in human laboratories. Therefore, two
kinds of analyses are considered: one concentrating on the pheromones and allelochemicals of
the Pherobase (El‐Sayed, 2012) and another on the whole database, including attractants
synthesized by humans. The whole database comprises a repertoire of n=1686 chemical
compounds and T=17633 species‐infochemical associations. A summary of the elementary
features of the more frequent elements of the dataset is shown in Table 1.
3.2. Methods
The methodology for fitting functions to the dependency between rank and degree is
borrowed from Li et al.'s for the dependency between the rank of a word and its frequency (Li
et al., 2010). Baayen's (2008) methodology is used to compute the breakpoint parameter of
the double Zipf (parameter r* in Table 2). Li et al.'s (2010) methodology consists of a linear
regression of the target function on a double logarithmic transformation, i.e. log(r) versus
log(fr), and then using Akaike's information criterion, a combination of likelihood to evaluate
the quality of the fit and a penalty for the number of parameters used. Table 2 summarizes the
functions that are fitted to the rank distribution of infochemicals and the corresponding
double logarithmic transformation that is used for linear regression.
SSE is defined as the sum of the squared differences between the logarithm of fr, the observed
degree of rank r, and the logarithm of Fr, expected degree of rank r, i.e.
SSE
n
r
1
w
r
(log(
f
)
log(
F
r
2))
r
(2)
w being a weight that is wr = 1 in unweighted regression and wr =1/r for weighted regression
following Li et al.'s (2010) methodology. The goal of the weighted regression is to give more
importance to low ranks.
THE INFOCHEMICAL CORE
The log‐likelihood L is defined (Venables & Ripley, 1999; Li et al., 2010) as
nCL
2
log
SSE
n
n
2
9
(3)
where n is the maximum rank and C is an additive constant that depends on the model (see
Table 2). For model selection, the Akaike Information Criterion (AIC) is used as in Li et al.
(2010). As Li et al. (2010) point out, any constant term of Eq. 3 will be cancelled when two
models are compared and then the AIC defined by Akaike (1974) is:
AIC
2
L
2
nK
log
SSE
n
K
,2
(4)
where K is the number of free parameters (Table 2) in the model under consideration.
The AIC difference of a model, ∆, is defined as the difference between the AIC of the model
and that of the model with smallest AIC (Li et al., 2010). Then, trivially ∆=0 for the best model.
The relationship between AIC and SSE defined in equation 3 is still true for weighted regression
as Li et al. (2010) demonstrate, but the correlation‐based regression of R2 for weighted
regression in logarithmic scale has a specific definition (see Li et al. (2010) for further details).
To fit the double Zipf, a function not considered by Li et al. (2010), Baayen's (2008, pp.234‐
239) technique is used to compute the breakpoint automatically. That breakpoint is the rank r*
that defines the boundary between two consecutive simple Zipf distributions. The optimal
breakpoint r* is the rank that minimizes the SSE, which is obtained by exhaustively exploring
all the possible values of r*, that is fitting the double Zipf with a given r* that varies between 1
and n, and keeping the r* yielding the smallest SSE.
10
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
***FIGURE 1 NEAR HERE***
Fig. 1 indicates that the breakpoint obtained for the Pherobase is not a local optimum. The R
statistical package is used to analyse all the data (see R online manuals at http://www.r‐
project.org/ and Baayen (2008)). Table 2 summarizes the list of functions fitted to the rank
distribution of degree.
4. RESULTS
Table 3 and Table 4 summarize the fitting of different functions to the empirical relationship
between the degree of an infochemical and its rank according to the Pherobase (El‐Sayed
2012), including or excluding attractants, respectively. Both the correlation coefficient (ρ) and
AIC differences (∆) suggest that the double Zipf is the function providing the best fit, both in
the weighted and unweighted regression, regardless of whether attractants are included or
not. All the other functions considered (simple Zipf, Beta, Yule and Menzerath‐Altmann) are far
from yielding the best fit: the second best function, both in unweighted and weighted
regression for the whole database, is Yule function with AIC differences ∆=967.5 and ∆=1450,
respectively (Table 3) and similar results for the analysis excluding attractants (Table 4). AIC
differences of the order of a thousand are normally considered sufficient to discard the model
under consideration (Burham & Anderson, 2002).
***TABLE 3 AND TABLE 4 NEAR HERE***
Figure 2 shows the best fit of the double Zipf equation, with and without weights, for the
whole database (Fig 2 A and C) and without attractants (Fig 2 B and D), respectively. When all
infochemicals are considered, the breakpoint is r*= 180 for unweighted, and r*=79 for
weighted regression. When attractants are excluded, the breakpoint is r*=151 for unweighted,
and r*=61 for weighted regression (see Table 5). The value of the breakpoint parameter of the
THE INFOCHEMICAL CORE
11
double Zipf suggests that infochemicals are divided into two groups: a core of the order of one
hundred infochemicals and the remainder.
***TABLE 5 AND FIGURE 2 NEAR HERE***
5. DISCUSSION
This double Zipf distribution of ranks (in the degree of an infochemical as a function of its rank
rank) is also found in the rank distribution of words, i.e. the frequency of occurrence of a word
(in tokens) as a function of its rank (Ferrer‐i‐Cancho & Solé, 2001; Petersen et al., 2012;
Gerlach & Altmann, 2013; Cocho et al., 2015), where the breakpoint defines the boundary
between a vocabulary core, i.e. a finite vocabulary of semantically versatile word types, and a
potentially infinite peripheral vocabulary.
The connection may not be obvious at first glance: our target has not been the frequency of
occurrence of an infochemical in nature but its degree, namely, the number of species that
produce it or are sensitive to each infochemical. However, word frequency is connected
lawfully with another linguistic variable, namely word polytextuality, which can be defined as
the number of texts in a corpus where the target word appears at least once (Köhler, 1986).
The metaphor that words types are infochemicals and texts are the infochemicals from the
environment that the members of a species have produced or been sensitive to, and the
positive correlation between frequency and polytextuality (Köhler, 1986), suggest that the
infochemicals in the high degree regime (the low ranks) form a core repertoire analogous to
the core lexicon found in the high frequency regime of human language (Ferrer‐i‐Cancho &
Solé, 2001), while the infochemicals in the low degree regime would form a peripheral
repertoire analogous to the peripheral lexicon found in the lower frequency domain of words
(Ferrer‐i‐Cancho & Solé, 2001).
12
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Cores have also been investigated in cognitive networks (Baronchelli et al., 2013). A network
core is defined by Baronchelli et al. (2013) as "a powerful subset of the network because of the
high frequency of occurrence of its nodes, their importance for the existence of remainder of
nodes, or the fact that it is both densely connected and central (in terms of graph distance)". A
network analysis of cross‐referencing between dictionary entries has shown that dictionaries
have a core consisting of about 10% of words (typically with a concrete meaning and acquired
early), from which other words can be defined (Picard et al., 2009). In our case, we are
analysing a bipartite graph of infochemical‐species associations (one partition for infochemical
and another partition for species). Table 5 summarizes the proportion of infochemical types
within the chemical core taking the breakpoint of the double Zipf as the boundary. The
percentage of infochemicals that belong to the core repertoire is about 10% according to
unweighted regression, about the same percentage of word types in the grounding kernel of a
dictionary (Picard et al., 2009). The 10% of the core both in dictionaries and infochemicals
might be simply anecdotal and should be explored further. The percentage of infochemical‐
species associations within the chemical kernel is two‐thirds (11755 infochemicals in the core
repertoire over 17633 in total, just 66.66%) of the total, but this percentage varies according to
the kind of function and increases to more than seventy percent for attractants and allomones
(Table 6). Similarly, the core vocabulary of words identified by means of a double Zipf by
Ferrer‐i‐Cancho & Solé (2001) is responsible for 85% of word tokens in a large English
multiauthor corpus. Although infochemical degrees and word frequencies are not fully
comparable, the high percentages of associations/tokens in the core highlights the importance
of cores.
***TABLE 6 NEAR HERE***
Pheromones and kairomones are just a little below sixty percent of infochemical associations
in the chemical kernel (Figure 3) and both have a greater presence than attractants and
THE INFOCHEMICAL CORE
13
allomones (in percentage) in the peripheral chemical repertoire, with synomones not
appearing because their number is very low (Figure 3). The presence of pheromones increases
in the peripheral chemical repertoire probably as a response to a major necessity of
communication specificity and to reduce communication interferences with other species. The
information carried by general attractants of the infochemical core may need to be completed
and detailed by peripheral, perhaps species‐specific, pheromones to avoid confusion.
***FIGURE 3 NEAR HERE***
The chemical kernel constitutes a core of infochemicals shared by many species (Table 5). We
hypothesize that the design of such a core could be driven by the economy of its compounds
from the perspective of the sender (e.g. ease of production of the compound) and
communicative efficiency in a given environment from the perspective of the receiver.
Communicative efficiency is determined by various factors such as ease of detection or
persistence of the chemical compound. The heterogeneity of the ecosystems of the species
included in the database probably limits the cases of communicative interference between
species. Furthermore, to avoid the confusion that might arise in an ecosystem with different
species using the same infochemicals, species may adopt different diffusion strategies and
exploit the zoosemiotic communication context to reduce confusion (Maran, Martinelli &
Turovsky, 2011; Riba, 1990).
The most frequently analyzed infochemicals (Table 2), by definition on the core chemical
repertoire, include, for instance, some isomers of tetradecenyl acetate that are well‐known sex
pheromones (Shorey, 1976; Chapman, 1998) and are attractive to all of the males in a large
group of different closely‐related species. As already pointed out by Shorey (1976) in his
classical review, other chemicals present
in relatively small quantities enhance the
14
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
attractiveness of the pheromone for males of the correct species and, at the same time, inhibit
attraction of males of wrong species.
In fact, each species has its own chemical communication system, with a finite set of
associated infochemicals that it can emit, Ve, and another set to which it responds, Vr. Actually,
Ve≠Vr since there are semiochemicals (e.g., synthesized compounds) that a species will not
emit, but will be attracted to. Using set theory, the chemicals that constitute a chemical
communication system is Ve U Vr, while the elements emitted and also detectable are Ve ∩ Vr.
In this Universe of chemical communication there are species that are sensitive to a shared
chemical repertoire, emitted by themselves or not. It is a complex Universe because there are
substances that, within the same species, can attract females and not males or vice versa.
The analysis of non‐human communication systems may be able to provide insights into the
efficiency of signalling systems that might otherwise be inaccessible (Doyle, 2009). Keeping a
certain distance, parallels between our study and other linguistic phenomena, e.g. shared
elements between different languages, can be established. Above we have considered the
metaphor of a species as a text of related infochemicals. Here we regard the infochemicals
with which a species is related as a language. There are infochemicals shared in the
communication systems of different species thus, arguably, species may speak "different
chemical languages" but do have some communicative elements in common (not necessarily
with the same meaning). Similarly, different human languages can share word elements such
as phonemes or syllables in their linguistic systems. Some languages not only share those
building blocks of words but also word themselves. In a more complex way, languages
sometimes share elements like words of basic vocabulary, e.g. a Swadesh list, especially if they
are related languages. From an evolutionary perspective, related languages tend to maintain
diachronically basic words in the same or practically identical word form (Wang & Wang, 2004)
THE INFOCHEMICAL CORE
15
just as different species can be sensitive to very similar -or the same- semiochemicals when
they come from the same phylum (Symonds & Elgar, 2004).
The ubiquity and variety of pheromones can be explained by natural selection (Wyatt, 2009;
Hauser, 1996). From a Zipfian perspective (Zipf, 1949), the existence of a peripheral chemical
repertoire (the infochemicals outside of the core) could be a consequence of the need to
diversify species' communication possibilities, which would be particularly useful in a noisy
channel. If the principle of least effort leads all species to emit similar chemicals compounds
from the core because of their high availability, ease of production or utility in different
environments (Okubo et al., 2001), then interference and confusion emerges as inevitable,
especially in very rich terrestrial ecosystems like tropical forests (Basset et al., 2012). Thus,
diversification is opposed to unification (Ferrer‐i‐Cancho & Solé, 2003; Zipf, 1949) also in
chemical communication systems. From a Darwinian evolution perspective, it is expected that
an optimal occupation of the chemical channel in ecosystems arises over time. In sum, the
existence of both a chemical core and a periphery could be a natural consequence that
communication has to solve two different problems: efficient coding of information and
transmission (Ferrer‐i‐Cancho et al., 2013).
Although linguistics is mainly concerned with the vocal modality, interest in gestural language
through conventional sign language or spontaneous gestures has been increasing over time
(Goldin‐Meadow et al., 2008; Sandler & Lillo‐Martin, 2006). The chemical modality might be
the next frontier to be explored intensively. Quantitative linguistics now has the challenge of
understanding the mechanisms underlying the emergence of two regimes in words and
infochemicals and identifying the distinctive features of those cores (Gerlach & Altmann, 2013;
Cocho et al., 2015). Throughout this long research, we hope to answer a very important
question: what are the principles of organization shared between human language and
16
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
chemical communication? Understanding both as evolutionary and complex adaptative
systems might be crucial (Beckner et al., 2009; Bel‐Enguix & Jiménez‐López, 2012).
The quantitative exploration of the Pherobase is just beginning and our analysis has focused on
the large scale. Future research should pay attention to specific ecological niches or a concrete
phylum. The generosity of those who share data on infochemicals is helping to explore
connections between apparently distant domains, which will become increasingly more
common in twenty‐first century science.
ACKNOWLEDGEMENTS
We thank G. Bel‐Enguix for helpful comments and discussions. We are grateful to A.M. El‐
Sayed and all the people who have made the Pherobase® possible and to Jordi Las for his help
with the programming of the data acquisition. RFC and AHF wish to express their gratitude for
support in the form of (a) the grant 2014SGR 335 890 (MACDA) from AGAUR (Generalitat de
Catalunya) and (b) the APCOM project (TIN2014‐57226‐P) from MINECO (Ministerio de
Economia y Competitividad). Additionally, RFC is grateful to the Universitat Politècnica de
Catalunya for support in the form of the grant "Iniciació i reincorporació a la recerca" and to
the Spanish Ministry of Science and Innovation for the grants BASMATI (TIN2011‐27479‐C04‐
03) and OpenMT‐2 (TIN2009‐14675‐C03).
THE INFOCHEMICAL CORE
REFERENCES
17
Akmajian, A., Demers, R.A., Farmer, A.K. & Harnish, R. (1997). Linguistics. An introduction to
language and communication. Cambridge, MA: MIT Press, 2nd edition.
Akaike, H. (1974). A new look at statistical model identification. IEEE Transactions on
Automatic Control, 19, 716–722.
Altmann, G. (1980). Prolegomena to Menzerath's law. Glottometrika 2, 1–10.
Baayen, R.H. (2008). Analyzing Linguistic Data. A Practical Introduction to Statistics Using R.
Cambridge: Cambridge University Press.
Balasubrahmanyan, V.K. & Naranan, S. (2000). Information Theory and Algorithmic Complexity:
Applications to Linguistic Discourses and DNA Sequences as Complex Systems Part II:
Complexity of DNA sequences, analogy with linguistic discourses. Journal of Quantitative
Linguistics, 7(2), 153‐183.
Baronchelli, A., Ferrer‐i‐Cancho, R., Pastor‐Satorras, R., Chater, N. & Christiansen, M.H. (2013).
Networks in cognitive science. Trends in Cognitive Sciences, 17 (7) 348‐360.
Basset, Y., et al. (2012). Arthropod diversity in a tropical forest. Science, 338, pp. 1481‐1484.
Beckner, C., et al. (2009). Language is a complex adaptive system. Language Learning, 59 (s1),
1‐26.
Bel‐Enguix, G. & Jiménez‐López, M. D. (2011). Genetic code and verbal language: syntactic and
interdisciplinary
semantic analogies.
paradigms. Bel‐Enguix, G., Dahl, V., Jiménez‐López, M.D., (Eds); IOS Press: Amsterdam. pp.
85‐103.
In: Biology, computation and
linguistics. New
Bel‐Enguix, G. & Jiménez‐López, M.D. (2012). Biocomputing: an insight from linguistics. Natural
Computing, 11, 131‐139.
Ben Jacob, E., Becker, I., Shapira, Y. & Levine, H. (2004). Bacterial linguistic communication and
social intelligence, Trends in Microbiology, 12 (8), 366–372.
Burnham, K.P. & Anderson, D.R. (2002). Model selection and multimodel inference. A practical
information‐theoretic approach. 2nd edition. New York: Springer.
Byers, J.A. (2006). Pheromone component patterns of moth evolution revealed by computer
analysis of the Pherolist. Journal of Animal Ecology 75, 399–407.
Chapman, R.F. (1998). The insects. Structure and function, 4th edition. Cambridge: Cambridge
University Press.
Christiansen, M. & Chater, N. (2015). The now‐or‐never bottleneck: a fundamental constraint
on language. Brain and Behavioral Sciences, in press.
18
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Cocho, G., Flores, J., Gershenson, C., Pineda, C. & Sánchez, S. (2015). Rank diversity of
languages: generic behavior in computational linguistics. PLoS ONE, 10 (4), e0121898.
Dicke, M. & Sabelis, M.W. (1988). Infochemicals terminology: Based on cost‐benefit analysis
rather than origin of compounds? Functional Ecology, 2, 131‐139.
Dickman, R., Moloney, N.R. & Altmann, E.G. (2012). Analysis of an information‐theoretic model
for communication. Journal of Statistical Mechanics: Theory and Experiment, P12022.
Doyle, L.R., McCowan, B., Johnston, S., & Hanser, S.F. (2011). Information theory, animal
communications, and the search for extraterrestrial intelligence. Acta Astronomica, 68,
406–417.
El‐Sayed, A.M. (2012). The Pherobase: Database of Pheromones and Semiochemicals.
http://www.pherobase.com.
Ferrer‐i‐Cancho, R. (2005). Zipf's law from a communicative phase transition. European
Physical Journal B, 47, 449‐457.
Ferrer‐i‐Cancho, R. & Solé, R.V. (2003). Least effort and the origins of scaling in human
language. Proceedings of the National Academy of Sciences, 100, 788‐791.
Ferrer‐i‐Cancho, R. & Solé, R.V. (2001). Two regimes in the frequency of words and the origin
of complex lexicons: Zipf's law revisited. Journal of Quantitative Linguistics, 8, 165‐173.
Ferrer‐i‐Cancho, R., Forns, N., Hernández‐Fernández, A., Bel‐Enguix, G. & Baixeries, J. (2013a).
The challenges of statistical patterns of language: the case of Menzerath's law in genomes.
Complexity, 18 (3), 11–17.
Ferrer‐i‐Cancho, R., Hernández‐Fernández, A., Lusseau, D., Agoramoorthy, G., Hsu, M. J. &
Semple, S. (2013b). Compression as a universal principle of animal behavior. Cognitive
Science, 37 (8), 1565–1578.
Ferrer‐i‐Cancho, R. & McCowan, B. (2009). A law of word meaning in dolphin whistle types.
Entropy, 11 (4), 688‐701.
Ferrer‐i‐Cancho, R. & McCowan, B. (2012). The span of correlations in dolphin whistle
sequences. Journal of Statistical Mechanics, P06002.
Fitch, W. T. (2010). The evolution of language. Cambridge: Cambridge University Press.
Gerlach, M. & Altmann, E.G. (2013). Stochastic model for the vocabulary growth in natural
languages. Physical Review X, 3, 021006. Greenhill, S.J., Blust, R. & Gray, R.D. (2008). The
Austronesian Basic Vocabulary Database: From bioinformatics to lexomics. Evolutionary
Bioinformatics Online 4, 271–283.
Goldin‐Meadow, S., So, W.‐C., Ozyurek, A., & Mylander, C. (2008). The natural order of events:
how speakers of different languages represent events nonverbally. Proceedings of the
National Academy of Sciences, 105 (27), 9163‐9168.
THE INFOCHEMICAL CORE
19
Hauser, M.D. (1996). The Evolution of Communication. Cambridge, Massachusetts: MIT Press.
Hauser, M.D., Chomsky, N., Fitch, W.T. (2002). The faculty of language: what is it, who has it,
and how dit it evolve? Science, 298, 1569–1579.
Hay, M. (2009). Marine chemical ecology: chemical signals and cues, structure marine
populations, communities, and ecosystems. Annual Review of Marine Sciences, 1, 193–212.
Head, T. (1999/2000). Communication by documents in communities of organisms. Millenium
III (4), 33‐42.
Hurford, J. (2011). The origins of grammar. Language in the light of evolution. Oxford: Oxford
University Press.
Ji, H. (1999). The Linguistics of DNA: Word, Sentences, Grammar, Phonetics and Semantics.
Annals of the New York Academy of Sciences, 870, 411‐417.
Karlson, P. & Lüscher, M.(1959). 'Pheromones': a new term for a class of biologically active
substances. Nature, 183, 55–56.
Köhler, R. (1986). Zur linguistischen Synergetik. Struktur und Dynamik der Lexik. Bochum:
Brockmeyer.
Law, R.H. & Regnier, F.E. (1971). Pheromones. Annual Review of Biochemistry, 40, 533–548.
Li, W., Miramontes, P. & Cocho, G. (2010). Fitting ranked linguistic data with two‐parameter
functions. Entropy, 12, 1743‐1764.
Li, W. (2012). Menzerath's law at the gene‐exon level in the human genome. Complexity, 17,
49‐53.
Pinker, S. (2003). Language as an adaptation to the cognitive niche. In: Language evolution:
states of the art, M.H. Christiansen & S. Kirby (Eds), pp. 16‐37. New York: Oxford University
Press.
Maran, T., Martinelli, D. & Turovsky, A. (Eds) (2011). Readings in zoosemiotics. Göttingen: De
Gruyter‐Mouton.
McCowan, B., Hanser, S.F. & Doyle, L.R. (1999). Quantitative tools for comparing animal
information theory applied to bottlenose dolphin whistle
communication systems:
repertoires. Animal Behaviour, 57, 409‐419.
Menzerath, P. (1954). Die Architektonik des deutschen Wortschatzes. Dümmler: Bonn.
Naranan, S. & Balasubrahmanyan, V.K. (2000). Information Theory and Algorithmic Complexity:
Applications to Linguistic Discourses and DNA Sequences as Complex Systems Part I:
Efficiency of the Genetic Code of DNA. Journal of Quantitative Linguistics, 7 (2), 129‐151.
Naranan, S. (2011). Historical Linguistics and Evolutionary Genetics. Based on Symbol
Frequencies in Tamil Texts and DNA Sequences. Journal of Quantitative Linguistics, 18(4),
359‐380.
20
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Nordlund, D.A. & Lewis, W.J. (1976). Terminology of chemical releasing stimuli in intraspecific
and interspecific interactions. Journal of Chemical Ecology, 2, 211–220.
Okubo, A., Armstrong, R.A., & Yen, J. (2001). Diffusion of "Smell" and "Taste": Chemical
Communication. In Diffusion and Ecological Problems, 2nd edition, ed. Okubo, A. & Levin,
S.A. pp. 107–126. New‐York: Springer‐Verlag.
Petersen, A.M., Tenenbaum, J.N., Havlin, S., Stanley, H.E. & Perc, M. (2012). Languages cool as
they expand: Allometric scaling and the decreasing need for new words. Scientific Reports,
2, 943.
Picard, O., Blondin‐Masse, A., Harnad, S., Marcotte, O., Chicoisne, G. & Gargouri, Y. (2009).
Hierarchies in Dictionary Definition Space. 23rd Annual Conference on Neural Information
Processing Systems (NIPS): Workshop on Analyzing Networks and Learning.
Prokopenko, M., Ay, N., Obst, O., & Polani, D., (2010). Phase transitions in least‐effort
communications. Journal of Statistical Mechanics: Theory and Experiment, 11025.
Rao, R. (2010). Probabilistic analysis of an ancient undeciphered script. IEEE Computer, 43 (4),
76‐80.
Rao, R.P.N., Yadav, N., Vahia, M.N., Joglekar, H., Adhikari, R. & Mahadevan, I. (2012). Entropy,
the Indus script, and language: a reply to R. Sproat. Computational Linguistics, 36 (4), 795‐
805.
Regnier, F.E. & Law, R.H. (1968). Insect Pheromones. Journal of Lipid Research, 9, 541–551.
Riba, C. (1990). La comunicación animal. Un enfoque zoosemiótico. Barcelona: Anthropos.
Sandler, W. & Lillo‐Martin, D. (2006). Sign language and linguistic universals. Cambridge:
Cambridge Universty Press.
Searls, D.B. (2002). The language of genes. Nature, 420, 211‐217.
Semple, S., Hsu, M.J. & Agoramoorthy, G. (2010). Efficiency of coding in macaque vocal
communication. Biology Letters, 6, 469‐471.
Shorey, H.H. (1976). Animal Communication by Pheromones. New York: Academic Press.
Steiger, S., Schmitt, T. & Schaefer, H.M. (2011). The origin and dynamic evolution of chemical
information transfer. Proceedings of the Royal Society of London B, 278, 970–979.
Symonds, M.R.E. & Elgar, M.A. (2004). The mode of pheromone evolution: evidence from bark
beetles. Proceedings of the Royal Society of London B, 271, 839–846.
Wang, F. & Wang, S.Y. (2004). Basic Words and Language Evolution. Language and Linguistics,
5(3),643‐662.
Wilde, J. & Schwibbe, M.H. (1989). Organizationsformen von Erbinformation im Hinblick auf
die Menzerathsche Regel. In Das Menzerathsche Gesetz in informationsverarbeitenden
Systemen. G. Altmann & M.H. Schwibbe (Eds); Olms: Hildesheim, 1989. pp. 92‐107.
THE INFOCHEMICAL CORE
21
Wilson, E.O. & Bossert, W.H. (1963). Chemical communication among animals. Recent Progress
in Hormone Research, 19, 673–716.
Wilson, E.O. (1970). Chemical communication within animal species. In Chemical ecology, 9,
ed. Sondheimer, E. & Simeone J. B. pp. 133–155. New York: Academic Press.
Wilson, E.O. (1958). A chemical release of alarm and digging behavior
in the ant
Pogonomyrmex badius (Latreille). Psyche 65, 41‐51.
Wyatt, T.D. (2009). Fifty years of pheromones. Nature, 457, 262–263.
Wyatt, T.D. (2010). Pheromones and signature mixtures: defining species‐wide signals and
variable cues for individuality in both invertebrates and vertebrates. Journal of Comparative
Physiology A, 196,685–700.
Wyatt, T.D. (2003). Pheromones and Animal Behaviour. Cambridge: Cambridge University
Press.
Zipf, G.K. (1935). The psycho‐biology of language: an introduction to dynamic philology.
Cambridge, MA: MIT Press.
Zipf, G.K. (1949). Human behaviour and the principle of least effort. Cambridge, MA: Addison‐
Wesley.
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
22
Table 1: List of the ten infochemicals with the highest degree (in number of species) according
to the Pherobase (El‐Sayed 2012). fr is the degree of the infochemical with the r‐th largest
degree. Each chemical compound can serve different biological functions.
Infochemicals
(Z)‐9‐Tetradecenyl acetate
(Z)‐11‐Tetradecenyl acetate
(Z)‐11‐Hexadecenyl acetate
2,6,6‐Trimethylbicyclo[3,1,1]hept‐2‐ene
(Z)‐7‐Dodecenyl acetate
(E)‐11‐Tetradecenyl acetate
Carbon dioxide
Ethanol
(Z)‐11‐Hexadecenal
1‐Octen‐3‐ol
r
1
2
3
4
5
6
7
8
9
10
fr
342
309
278
245
235
229
222
221
185
183
THE INFOCHEMICAL CORE
23
Table 2: List of functions used in this article (simplified and adapted from Li et al., 2010 and
enriched with the double Zipf) and the parameters fitted, i.e. a, b and normalization constants
(C and C'). In the double logarithmic transformation for linear regression, Co is the independent
term and C1 is the coefficient that multiplies log r. r* is the rank of the breakpoint in the double
Zipf function and defines two regimes, a 1st regime from r=1 to r=r* and a 2nd regime for r>r*.
MODEL
FREE
PARAMETERS
(K)
ORIGINAL FUNCTION
(r=rank, Fr=degree,
n=repertoire size)
Zipf
Beta
Yule
Menzerath‐
Altmann
Double Zipf
1
2
2
2
3
DOUBLE LOGARITHMIC
TRANSFORMATION
log Fr = C0 + C1 log r
log Fr = C0 + C1 log r +C2log(n+1‐r)
log Fr = C0 + C1 log r +C4r
F
r
C
r
a
F
r
nC
(
b
r
)
1
r
a
F
r
r
bC
·
r
a
erCF
·
r
b
·
ra
/
log Fr = C0 + C1 log r + C3/r
1st regime: log Fr = C01 + C11 log r
2nd regime: log Fr = C02 + C12 log r
1st regime:
F
r
2nd regime:
F
r
C
r
a
;
'
'
C
r
a
24
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Figure 1: The sum of squared errors of the double Zipf, SSE, versus the breakpoint rank, r*. A
and B are the plots for the whole Pherobase, unweighted and weighted respectively. C and D
are analogous to A and B excluding attractants synthesized by humans, unweighted and
weighted, respectively. It can be seen that the breakpoint r* obtained corresponds to a global
minimum of deviance in all cases. Only in C do we find another local minimum for very large
ranks that is not relevant due to its comparatively large SSE value with regard to the global
0,5
0,4
0,3
0,2
0,1
0,0
E
S
S
A
1200
1400
1600
1800
0
200
400
600
800
1000
r*
800
1000
r*
B
1200
1400
1600
1800
D
0
200
400
600
800
r*
1000
1200
1400
1600
minimum.
35
30
25
E
20
S
S
15
10
5
0
26
24
22
20
E
18
S
S
16
14
12
10
8
200
400
600
E
S
S
0,5
0,4
0,3
0,2
0,1
0,0
C
0
200
400
600
800
r*
1000
1200
1400
1600
THE INFOCHEMICAL CORE
25
Figure 2: Degree (number of species that that are associated with each infochemical) versus
the infochemical rank in double logarithmic scale (white circles) versus the best fit of the
double Zipf model (solid line). A and B correspond to the whole database while attractants are
excluded for C and D. A and C are unweighted while B and D are weighted. In each subplot, the
rank breakpoint (r*), the Zipf's exponent for the first and second regime (C11 and C12,
respectively) are shown.
C11= -0.58
C12 = -1.43
r*=151
10
100
1000
rank
C11= -0.30
C12 = -1.27
r*=61
10
100
1000
rank
i
)
s
e
c
e
p
s
n
i
(
e
e
r
g
e
d
100 B
10
1
1
100 D
i
)
s
e
c
e
p
s
n
i
(
e
e
r
g
e
d
10
1
1
A
100
10
i
)
s
e
c
e
p
s
n
i
(
e
e
r
g
e
d
1
1
C
100
10
i
)
s
e
c
e
p
s
n
i
(
e
e
r
g
e
d
1
1
C11= -0.66
C12 = -1.61
r*=180
10
100
1000
rank
C11= -0.40
C12 = -1.43
r*=79
10
100
1000
rank
26
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Table 3: Summary of the results of the fitting of the ensemble of functions to the whole
database (El‐Sayed, 2012). For every target function, the coefficients giving the best fit are
shown (the meaning of each coefficient is explained in Table 1). ∆ is the difference between
the AIC of the target function and that of the function giving the lowest AIC), SSE is the sum of
squared errors and ρ is the correlation coefficient (R2
corr,log and R2
corr,log,weight , for unweighted
and weighted fit, respectively, following Li et al.'s (2010) notation).
Unweighted
Weighted
Coefficients
∆
SSE
ρ
Coefficients
∆
SSE
ρ
C0 =4.12± 0.02
C1 = -1.29±0.01
C0 =3.59 ± 0.05
C1= -1.21± 0.01
C2 = 0.051±0.004
C0 =3.36± 0.02
C1 = -0.90± 0.01
C4=-39·10-5 ±1·10-5
C0 =4.41± 0.02
C1 = -1.39± 0.01
C3 = -2.9±0.1
C01 =1.43± 0.01
C11 = -0.66± 0.01
C12 = -1.61±0.01
2074.4
31.24
0.94
1939.9
28.81
0.95
967.5
16.18
0.97
1354.7
20.36
0.96
0
9.11
0.98
C0 =2.84± 0.01
C1 = -0.74±0.01
C0 =-0.17± 0.08
C1= -0.625±0.005
C2 = 0.39±0.01
C0 =2.66± 0.01
C1 = -0.49± 0.03
C4 =-92·10-5 ±1·10-5
C0 = 3.40± 0.01
C1 = -0.98± 0.01
C3 = -0.95±0.02
C01 =1.83± 0.01
C11 = -0.40± 0.01
C12 = -1.43±0.01
4425.2
0.47
0.90
3426.4
0.26
0.95
1450.5
0.08
0.98
2756.8
0.17
0.96
0
0.034
0.99
Function
Zipf
Beta
Yule
Menzerath
-Altmann
Double
Zipf
THE INFOCHEMICAL CORE
27
Table 4: Summary of the fitting of the ensemble of functions to the Pherobase (El‐Sayed, 2012)
excluding attractants. The format is the same as in Table 3.
Function
Zipf
Beta
Yule
Menzerath
-Altmann
Double
Zipf
Unweighted
Weighted
Coefficients
∆
SSE
ρ
Coefficients
∆
SSE
ρ
C0 =3.68± 0.02
C1 = -1.17±0.01
C0 =3.34 ± 0.05
C1= -1.12± 0.01
C2 = 0.033±0.004
C0 =3.08± 0.02
C1 = -0.85± 0.01
C4=-35·10-5 ±1·10-5
C0 =3.95± 0.02
C1 = -1.26± 0.01
C3 = -2.6±0.1
C01 =1.32± 0.01
C11 = -0.58± 0.01
C12 = -1.43±0.01
1601.5
24.50
0.94
1543.6
23.58
0.94
881.3
15.41
0.96
924.9
15.85
0.96
0
8.76
0.98
C0 =2.52± 0.01
C1 = -0.66±0.01
C0 =-0.26± 0.08
C1= -0.55±0.01
C2 = 0.37±0.01
C0 =2.34± 0.01
C1 = -0.42± 0.01
C4 =-95·10-5 ±1·10-5
C0 = 3.09± 0.01
C1 = -0.98± 0.01
C3 = -0.95±0.02
C01 =1.73± 0.01
C11 = -0.30± 0.01
C12 = -1.27±0.01
3845.3
0.44
0.88
3030.3
0.26
0.93
1448.2
0.1
0.98
2107.5
0.15
0.96
0
0.038
0.99
28
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
Table 5: Percentage of infochemical types both in the core and peripheral chemical repertoire
considering the breakpoint r* of the two regime distribution as the boundary between the two
(El‐Sayed, 2012). T is the number of infochemical‐species associations (the total sum of
degrees) and n is the repertoire size (in infochemical types).
All infochemicals
All infochemicals without
attractants
Core chemical repertoire (1st
regime)
Peripheral chemical repertoire
(2nd regime)
Breakpoint (r*)
Total
associations (T)
infochemical‐species
Repertoire (n)
Unweighted
Weighted
Unweighted
Weighted
10.68%
4.69%
9.69%
3.91%
89.32%
95.31%
90.31%
96.09%
180
17633
1686
79
17633
1686
151
11380
1560
61
11380
1560
THE INFOCHEMICAL CORE
29
Table 6: Number of infochemical‐species associations that are within the "Core Chemical
Repertoire", and in the "Peripheral Chemical Repertoire", by kind of infochemical and for the
whole database (percentages are shown in parentheses and are relative to the total in the
right‐most column). The breakpoint r* of the double Zipf function in unweighted regression
defines the boundary between both repertoires.
1506
588
(12.81%)
(41.11%)
Attractants Allomones Kairomones Pheromones Synomones TOTAL
11755
(100%)
5878
(100%)
17633
(100%)
(35.46%)
(24.16%)
(60.62%)
8568
(48.59%)
(10.00%)
2094
(11.88%)
5005
(42.57%)
(<0.01%)
(5.14%)
712
(4.04%)
1
5
6
410
(3.49%)
302
(0.09%)
4833
1420
6253
3563
(0.03%)
Core repertoire
Peripheral repertoire
TOTAL
HERNÁNDEZ-FERNÁNDEZ & FERRER-I-CANCHO
30
Figure 3: Percentage of infochemical associations within the "Core Chemical Repertoire", and
in the "Peripheral Chemical Repertoire" by kind of infochemical. Synomones cannot be seen
due to their very low proportion. Data borrowed from Table 6.
t
n
e
c
r
e
P
100
80
60
40
20
0
Core repertoire
Peripheral repertoire
Total
Infochemicals
Attractants
Allomones
Kairomones
Pheromones
|
1603.00500 | 1 | 1603 | 2016-03-01T21:43:14 | BluePyOpt: Leveraging open source software and cloud infrastructure to optimise model parameters in neuroscience | [
"q-bio.NC"
] | At many scales in neuroscience, appropriate mathematical models take the form of complex dynamical systems. Parametrising such models to conform to the multitude of available experimental constraints is a global nonlinear optimisation problem with a complex fitness landscape, requiring numerical techniques to find suitable approximate solutions. Stochastic optimisation approaches, such as evolutionary algorithms, have been shown to be effective, but often the setting up of such optimisations and the choice of a specific search algorithm and its parameters is non-trivial, requiring domain-specific expertise. Here we describe BluePyOpt, a Python package targeted at the broad neuroscience community to simplify this task. BluePyOpt is an extensible framework for data-driven model parameter optimisation that wraps and standardises several existing open-source tools. It simplifies the task of creating and sharing these optimisations, and the associated techniques and knowledge. This is achieved by abstracting the optimisation and evaluation tasks into various reusable and flexible discrete elements according to established best-practices. Further, BluePyOpt provides methods for setting up both small- and large-scale optimisations on a variety of platforms, ranging from laptops to Linux clusters and cloud-based compute infrastructures. The versatility of the BluePyOpt framework is demonstrated by working through three representative neuroscience specific use cases. | q-bio.NC | q-bio | BluePyOpt: Leveraging open source software and cloud
infrastructure to optimise model parameters in neuroscience
Werner Van Geit 1
Jean-Denis Courcol 1, Eilif Muller 1, Felix Schurmann 1, Idan Segev 3
∗, Michael Gevaert 1, Giuseppe Chindemi 1, Christian Rossert 1,
4 and Henry
,
,
6
1
0
2
r
a
M
1
]
.
C
N
o
i
b
-
q
[
1
v
0
0
5
0
0
.
3
0
6
1
:
v
i
X
r
a
Markram 1
,
2
,
∗
1Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne (EPFL) Biotech
2Laboratory of Neural Microcircuitry, Brain Mind Institute, ´Ecole Polytechnique
Campus, Geneva, Switzerland
F´ed´erale de Lausanne, Lausanne, Switzerland
3Department of Neurobiology, Alexander Silberman Institute of Life Sciences, The
Hebrew University of Jerusalem, Jerusalem, Israel
4The Edmond and Lily Safra Centre for Brain Sciences, The Hebrew University of
Jerusalem, Jerusalem, Israel
∗ [email protected], [email protected]
March 3, 2016
Abstract
At many scales in neuroscience, appropriate mathematical models take the form of com-
plex dynamical systems. Parametrising such models to conform to the multitude of available
experimental constraints is a global nonlinear optimisation problem with a complex fitness
landscape, requiring numerical techniques to find suitable approximate solutions. Stochastic
optimisation approaches, such as evolutionary algorithms, have been shown to be effective, but
often the setting up of such optimisations and the choice of a specific search algorithm and its
parameters is non-trivial, requiring domain-specific expertise. Here we describe BluePyOpt, a
Python package targeted at the broad neuroscience community to simplify this task. BluePy-
Opt is an extensible framework for data-driven model parameter optimisation that wraps and
standardises several existing open-source tools.
It simplifies the task of creating and shar-
ing these optimisations, and the associated techniques and knowledge. This is achieved by
abstracting the optimisation and evaluation tasks into various reusable and flexible discrete
elements according to established best-practices. Further, BluePyOpt provides methods for
setting up both small- and large-scale optimisations on a variety of platforms, ranging from
laptops to Linux clusters and cloud-based compute infrastructures. The versatility of the
BluePyOpt framework is demonstrated by working through three representative neuroscience
specific use cases.
1
1 Introduction
Advances in experimental neuroscience are bringing an increasing volume and variety of data,
and inspiring the development of larger and more detailed models (Izhikevich and Edelman, 2008;
Merolla et al., 2014; Markram et al., 2015; Eliasmith et al., 2016). While experimental constraints
are usually available for the emergent behaviours of such models, it is unfortunately commonplace
that many model parameters remain inaccessible to experimental techniques. The problem of infer-
ring or searching for model parameters that match model behaviours to experimental constraints
constitutes an inverse problem (Tarantola, 2016), for which analytical solutions rarely exist for com-
plex dynamical systems, i.e. most mathematical models in neuroscience. Historically, such parame-
ter searches were done by hand tuning, but the advent of increasingly powerful computing resources
has brought automated search algorithms that can find suitable parameters (Bhalla and Bower,
1993; Vanier and Bower, 1999; Achard and De Schutter, 2006; Gurkiewicz and Korngreen, 2007;
Druckmann et al., 2007; Van Geit et al., 2007, 2008; Huys and Paninski, 2009; Taylor et al., 2009;
Hay et al., 2011; Bahl et al., 2012; Svensson et al., 2012; Friedrich et al., 2014; Pozzorini et al.,
2015; Stefanou et al., 2016). While many varieties of search algorithms have been described and
explored in the literature (Vanier and Bower, 1999; Van Geit et al., 2008; Svensson et al., 2012),
stochastic optimisation approaches, such as simulated annealing and evolutionary algorithms, have
been shown to be particularly effective strategies for such parameter searches (Vanier and Bower,
1999; Druckmann et al., 2007; Gurkiewicz and Korngreen, 2007; Svensson et al., 2012). Never-
theless, picking the right type of stochastic algorithm and setting it up correctly remains a
non-trivial task requiring domain-specific expertise, and could be model and constraint specific
(Van Geit et al., 2008).
With the aim of bringing widely applicable and state-of-the-art automated parameter search
algorithms and techniques to the broad neuroscience community, we describe here a Python-based
open-source optimisation framework, BluePyOpt, which is available on Github (see (Blue Brain Project,
2016)), and is designed taking into account model optimisation experience accumulated during the
Blue Brain Project (Druckmann et al., 2007; Hay et al., 2011; Markram et al., 2015; Ramaswamy et al.,
2015) and the ramp-up phase of the Human Brain Project. The general purpose high-level pro-
gramming language Python was chosen for developing BluePyOpt, so as to contribute to, and
also leverage from the growing scientific and neuroscientific software ecosystem (Oliphant, 2007;
Muller et al., 2015), including state-of-the-art search algorithm implementations, modelling and
data access tools.
At its core, BluePyOpt is a framework providing a conceptual scaffolding in the form of an
object-oriented application programming interface or API for constructing optimization problems
according to established best-practices, while leveraging existing search algorithms and modelling
simulators transparently "under the hood". For common optimisation tasks, the user configures
the optimisation by writing a short Python script using the BluePyOpt API. For more advanced
use cases, the user is free to extend the API for their own needs, potentially contributing these
extensions back to the core library. The latter is important for BluePyOpt APIs to remain broadly
applicable and state-of-the-art, as best-practices develop for specific problem domains, mirroring
the evolution that has occured for neuron model optimization strategies (Bhalla and Bower, 1993;
Hay et al., 2011).
Depending on the complexity of the model to be optimised, BluePyOpt optimisations can
require significant computing resources. The systems available to neuroscientists in the commu-
2
nity can be very heterogeneous, and it is often difficult for users to set up the required software.
BluePyOpt therefore also provides a novel cloud configuration mechanism to automate setting up
the required environment on a local machine, cluster system, or cloud service such as Amazon Web
Services.
To begin, this technical report provides an overview of the conceptual framework and open-
source technologies used by BluePyOpt, followed by a presentation of the software architecture
and API of BluePyOpt. Next, three concrete use cases are elaborated in detail, showing how the
BluePyOpt APIs, concepts and techniques can be put to use by potential users. The first use case is
an introductory example demonstrating the optimisation of a single compartmental neuron model
with two Hodgkin-Huxley ion channels. The second use case shows a BluePyOpt-based state-of-
the-art optimisation of a morphologically detailed thick-tufted layer 5 pyramidal cell model of the
type used in a recent in silico reconstruction of a neocortical microcircuit (Markram et al., 2015).
The third use case demonstrates the broad applicability of BluePyOpt, showing how it can also be
used to optimise parameters of synaptic plasticity models.
2 Concepts
The BluePyOpt framework provides a powerful tool to optimise models in the field of neuroscience,
by combining several established Python-based open-source software technologies. In particular,
BluePyOpt leverages libraries providing optimisation algorithms, parallelisation, compute environ-
ment setup, and experimental data analysis. For numerical evaluation of neuroscientific models,
many open-source simulators with Python bindings are available for the user to chose from. The
common bridge allowing BluePyOpt to integrate these various softwares is the Python program-
ming language, which has seen considerable uptake and a rapidly growing domain-specific software
ecosystem in the neuroscience modelling community in recent years (Muller et al., 2015). Python
is recognized as a programming language which is fun and easy to learn, yet also attractive to
experts, meaning that novice and advanced programmers alike can easily use BluePyOpt, and
contribute solutions to neuroscientific optimisation problems back to the community.
BluePyOpt was developed using an object oriented programming model. Figure 1 shows an
overview of the class hierarchy of BluePyOpt. In its essence, the BluePyOpt object model defines
the Optimisation class which applies a search algorithm to an Evaluator class. Both are abstract
classes, meaning they define the object model, but not the implementation. Taking advantage of
Pythonic duck typing, the user can then choose from a menu of implementations, derived classes, or
easily define their own implementations to meet their specific needs. This design makes BluePyOpt
highly versatile, while keeping the API complexity to a minimum. The choice of algorithm and
evaluator is up to the user, but many are already provided for various use cases. For many common
use cases, these are the only classes users are required to instantiate.
For neuron model optimizations in particular, BluePyOpt provides further classes to sup-
port feature-based multi-objective optimizations using NEURON, as shown in Figure 1. Classes
CellModel, Morphology, Mechanisms, Protocol, Stimuli, Recordings, Location are specific to setting
up neuron models and assessing their input-output properties. Other classes Objectives and eFea-
ture are more generally applicable, with derived classes for specific use cases, e.g. eFELFeature
provides features extracted from voltage traces using the open-source eFEL library discussed below.
They define features and objectives for feature-based multi-objective optimization, a best-in-class
stochastic optimization strategy (Druckmann et al., 2007, 2011; Hay et al., 2011). We generally
3
Abstract classes
Optimisation
Evaluator
Model
Morphology
Mechanisms
Protocols
Stimuli
Derived classes
DEAPOptimisation
External tools
DEAP
ModelEvaluator, GraupnerBrunelEvaluator
CellModel, NetworkModel, SynapseModel
NrnFileMorphology
NrnModMechanism
SequenceProtocol, SweepProtocol
NrnCurrentPlayStimulus, NrnSquarePulse
Recordings
CompRecording
Simulator
Parameters
Objectives
eFeature
ObjectivesCalculator
NrnSimulator
Neuron
NrnRangeParameter, NrnSectionParameter
SingletonObjective, WeightedSumObjective
eFELFeature
eFEL
Figure 1.Hierarchy of the most important classes in BluePyOpt. Ephys abstraction layer in blue.
recommend it as the first algorithm to try for a given problem domain. For example, the third
example for the optimisation of synaptic plasticity models also employs this strategy.
In the sub-sections to follow, an overview is provided for the various software components and
the manner in which BluePyOpt integrates them.
2.1 Optimisation algorithms
Multiobjective evolutionary algorithms have been shown to perform well to optimise parameters
of biophysically detailed multicompartmental neuron models (Druckmann et al., 2007; Hay et al.,
2011). To provide optimisation algorithms, BluePyOpt relies on a mature Python library, Dis-
tributed Evolutionary Algorithms in Python (DEAP), which implements a range of such algo-
rithms (Fortin et al., 2012). The advantage of using this library is that it provides many useful
features out of the box, and it is mature, actively maintained and well documented. DEAP provides
many popular algorithms, such as Non-dominated Sorting Genetic Algorithm-II (Deb et al., 2002),
Covariance Matrix Adaptation Evolution Strategy (Hansen and Ostermeier, 2001), and Particle
Swarm Optimisation (Kennedy and Eberhart, 1995). Moreover, due to its extensible design, im-
plementing new search algorithms in DEAP is straight-forward. Historically, the Blue Brain Project
has used a C implementation of the Indicator Based Evolutionary Algorithm IBEA to optimise the
parameters of biophysically detailed neuron models (Zitzler and Kunzli, 2004; Bleuler et al., 2003;
Markram et al., 2015), as this has been shown to have excellent convergence properties for these
problems (Schmucker, 2010). Case in point, we implemented a version of IBEA for the DEAP
framework, so this algorithm is consequently available to be used in BluePyOpt.
4
Moreover, DEAP is highly versatile, whereby most central members of its class hierarchy, such
as individuals and operators, are fully customizable with user defined implementations. Classes
are provided to keep track of the Pareto Front or the Hall-of-Fame of individuals during evolution.
Population statistics can be recorded in a logbook, and the genealogy between individuals can
be saved, analysed and visualised. In addition, checkpointing can be implemented in DEAP by
storing the algorithm's state in a Python pickle file for any generation, as described in DEAP's
documentation (DEAP Project, 2016).
Although the use cases below use DEAP as a library to implement the search algorithm, it is
worth noting that BluePyOpt abstracts the concept of a search algorithm. As such, it is entirely
possible to implement algorithms that are independent of DEAP, or that use other third-party
libraries.
2.2 Simulators
To define a BluePyOpt optimisation, the user must provide an evaluation function which maps
model parameters to a fitness score. It can be a single Python function that maps the parameters
to objectives by solving a set of equations, or a function that uses an external simulator to eval-
uate a complex model under multiple scenarios. For the latter, the only requirement BluePyOpt
imposes is that it can interact with the external simulator from within Python. Often, this inter-
action is implemented through Python modules provided by the user's neuroscientific simulator of
choice, as is the case for many simulators in common use, including NEURON (Hines et al., 2009),
NEST (Eppler et al., 2009; Zaytsev and Morrison, 2014), PyNN (Davison et al., 2009), BRIAN
(Goodman and Brette, 2009), STEPS (Wils and De Schutter, 2009), and MOOSE (Ray and Bhalla,
2008). Otherwise, communication through shell commands and input/output files is also possible,
so long as an interface can be provided as a Python class.
2.3 Feature Extraction
For an evaluation function to compute a fitness score from simulator output, the resulting traces
must be compared against experimental constraints. Voltage recordings obtained from patch clamp
experiments are an example of experimental data that can be used as a constraint for neuron
models. From such recordings the neuroscientist can deduce many interesting values, like the input
resistance of the neuron, the action potential characteristics, firing frequency etc. To standardise
the way these values are measured, the Blue Brain Project has released the Electrophysiology
Feature Extract Library (eFEL) (Blue Brain Project, 2015), also as open-source software. The
core of this library is written in C++, and a Python wrapper is provided. BluePyOpt can interact
with eFEL to compute a variety of features of the voltage response of neuron models. A fitness
score can then be computed by some distance metric comparing the resulting model features to
their experimental counterparts. As we will see for the last example in this article, a similar
approach can also be taken for other optimisation problem domains.
2.4 Parallelisation
Optimisations of the parameters of an evaluation function typically require the execution of this
function repeatedly. For a given optimisation integration step, such executions are often in the
hundreds (scaling e.g. with evolutionary algorithm population size), are compute bound, and are
5
essentially independent, making them ripe for parallelisation. Parallelisation of the optimisation
can be performed in several ways. DEAP provides an easy way to evaluate individuals in a
population on several cores in parallel. The user need merely provide an implementation of a
map function. In its simplest form, this function can be the Python serial map in the standard
library, or the parallel map function in the multiprocessing module to leverage local hardware
threads. To parallelise over a large cluster machine, the DEAP developers encourage the use of
the SCOOP (Hold-Geoffroy et al., 2014) map function. SCOOP is a library that builds on top
of ZeroMQ (ZeroMQ Project, 2007), which provides a socket communication layer to distribute
the computation over several computers. Other map functions and technologies can be used like
MPI4Py (Dalcn et al., 2005) or iPython ipyparallel package (P´erez and Granger, 2007). Moreover,
parallelisation doesn't necessarily have to happen at the population level. Inside the evaluation of
individuals, map functions can also be used to parallelise over stimulus protocols, feature types,
etc., however for the problem examples presented here, such an approach wouldn't make good use
of anything more than 10 to 20 cores.
2.5 Cloud
To increase the throughput of optimisations, multiple computers can be used to parallelise the work.
Such a group of computers can be composed of machines in a cluster, or they can be obtained from
a cloud provider like Amazon Web Services, Rackspace Public Cloud, Microsoft Azure, Google
Compute Engine, or the Neuroscience Gateway portal (Sivagnanam et al., 2013).
These and other cloud providers allow for precise allocation of numbers of machines and their
storage, compute power and memory. Depending on the needs and resources of an individual or
organization, trade-offs can be made on how much to spend versus how fast the results are needed.
Setting up a cluster or cloud environment with the correct software requirements is often
complicated and error prone: Each environment has to be exactly the same, and scripts and data
need to be available in the same locations. To ease the burden of this configuration, BluePyOpt
includes Ansible (Red Hat, Inc., 2012) configuration scripts for setting up a local test environment
(on one machine, using Vagrant (HashiCorp, 2010)), for setting up a cluster with a shared file
system, or for provisioning and setting up an Amazon Web Services cluster.
Ansible is open-source software that allows for reproducible environments to be created and
configured from simple textual descriptions called 'Playbooks'. These Playbooks encapsulate the
discrete steps needed to create an environment, and offer extra tools to simplify things like package
management, user creation and key distribution. Furthermore, when a Playbook is changed and
run against an already existing environment, only the changes necessary will be applied. Finally,
Ansible has the advantage over other systems, like Puppet (Puppet Labs, 2005) and Chef (Chef,
2009), that nothing except a Python interpreter needs to be installed on the target machine and all
environment discovery and configuration is performed through SSH from the machine on which An-
sible is run. This decentralized system means that a user can use Ansible to setup an environment
in their home directory on a cluster, without intervention from the system administrators.
3 Software Architecture
The BluePyOpt software architecture follows an object oriented programming model, whereby the
various concepts of the software are modularised into cleanly seperated and well defined classes
6
Optimisation
Evaluator
Evaluator: evaluator
run() - > results
Parameter
bool: frozen
value
bounds
instantiate(Simulator)
Objective
value
ObjectivesCalculator
Objectives: objectives
Model: model
Protocol[ ]: protocols
Stimulus[ ]: stimuli
ObjectivesCalculator: obj_calc
parameters
objectives
eval(Parameter [ ]) -> Objective []
Model
Parameter []: parameters
Mechanism []: mechanisms
Morphology: morphology
instantiate(Simulator)
Simulator
python_module
simulator_settings
Protocol
Stimulus []
Recordings []
run(Model, Simulator) -> Response []
Stimulus
Location: location
instantiate(Simulator)
Recording
variable
Location: location
instantiate(Simulator)
Location
location_speci! ers
calc(Response []) -> Objectives
run()
instantiate(Simulator) -> handle
Figure 2: General structure of most important classes. Every box represents a class. In every box
the top panel is the name, the middle panel the most important fields and the bottom panel the
most important methods. Ephys abstraction layer in blue.
which interact, as defined in a class hierarchy (Figure 1) and object model (Figure 2) show to
define the program control flow, as shown in Figure 3. In what follows, the role of each class and
how it relates to and interacts with other classes in the hierarchy is described.
3.1 Optimisation abstraction layer
At the highest level of abstraction, the BluePyOpt API contains the classes Optimisation and
Evaluator (Figure 2). An Evaluator object defines an evaluation function that maps Parameters
to Objectives. The Optimisation object accepts the Evaluator as input, and runs a search algorithm
on it to find the parameter values that generate the best objectives.
The task of the search algorithm is to find the parameter values that correspond to the best
objective values. Defining 'best' is left to the specific implementation. As in the use cases below,
this could be a weighted sum of the objectives or a multiobjective front in a multidimensional
space.
The Optimisation class allows the user to control the settings of the search algorithm. In case
of IBEA, this could be the number of individuals in the population, the mutation probabilities,
etc.
3.2 EPhys model abstraction layer
On a different level of abstraction, we have classes that are tailored for electrophysiology (ephys)
experiments and can be used inside the Evaluator. The ephys model layer provides an abstraction
to the simulator, so that the person performing the optimisation doesn't have to have knowledge
of the intricate details of the simulator.
A Protocol is applied to a Model for a certain set of Parameters, generating a Response. An
ObjectivesCalculator is then used to calculate the Objectives based on the Response of the Model.
All these classes are part of the bluepyopt.ephys package.
7
1. run()
14. results
Optimisation
2. evaluate(Parameters)
13. Objectives
Evaluator
12. Objectives
11. calc(Responses)
ObjectiveCalculator
3. run(Parameters)
10. Responses
Protocols
9. Recordings
4. freeze(Parameters)
5. instantiate()
6. instantiate()
7. instantiate()
8. run()
Model
Stimuli
Recordings
Simulator
Figure 3: Graph representing control flow in BluePyOpt. Ordering is clarified by the numbers.
Arrow labels that contain parentheses represent function calls, the other labels data being returned.
This figure is meant to give a high level description of the control flow, not all function calls and
intermediate objects are included. Ephys abstraction layer in blue.
3.2.1 Model
By making a Model an abstract class, we give users the ability to use our software for a broad
range of use cases. A Protocol can attach Stimuli and Recordings to a Model. When the Simulator
is then run, a Response is generated for each of the Stimuli for a given set of Model parameter
values.
Examples of broad subclasses are a NetworkModel, CellModel and SynapseModel. Specific
subclasses can be made for different simulators, or assuming some level of similarity, the same
model object can know how to instantiate itself in different simulators. In the future, functionality
could be added to import/export the model configuration from/to standard description languages
like NeuroML (Cannon et al., 2014) or NineML (Raikov et al., 2011).
Particular parameters of a Model can be in a frozen state. This means that their value is fixed,
and won't be considered for optimisation. This concept can be useful in multi-stage optimisation
in which subgroups of a model are optimised in a sequential fashion.
Another advantage of this abstraction is that a Model is a standalone entity that can be run
outside of the Optimisation and have exactly the same Protocols applied to it, generating exactly
the same Response. One can also apply extra Protocols to assess how well the model generalises,
or to perfom a sensitivity analysis.
3.2.2 Simulator
Every model simulator should have a subclass of Simulator. Objects of this type will be passed
on to objects that are simulator aware, like the Model and Stimuli when their instantiate method
is invoked. This architecture allows e.g. the same model object to be run in different simulators.
Examples of functionality this class can provide are links to the Python module related to the
simulator, the enabling of variable time step integration, etc. Simulators also have run() method
that starts the simulation.
8
3.2.3 Protocol
A Protocol is an object that elicits a Response from a model. In its simplest form it represents,
for example, a step current clamp stimulus, but more complicated versions are possible, such as
stimulating a set of cells in a network with an elaborate protocol and recording the response. A
Protocol can also contain sub-protocols, providing a powerful mechanism to reuse components.
3.2.4 Stimulus, Recording and Response
The Stimulus and Recording objects, which are part of a Protocol are applied to a model and
are aware of the simulator used. Subclasses of Stimulus are concepts like current/voltage clamp,
synaptic activation, etc. Both of these classes accept a Location specifier. Several Recording objects
can be combined in a Response which can be analysed by an ObjectiveCalculator.
3.2.5 Location
Specifying the location on a neuron morphology of a recording, stimulus or parameter in a simulator
can be complicated. Therefore we created an abstract class Location. As arguments the constructor
accepts the location specification, e.g. in NEURON this could be a sectionlist name and an index
of the section, or it could point to a section at a certain distance from the soma. Upon request, the
object will return a reference to the object at the specified location, this could e.g be a NEURON
section or compartment. At a location, a variable can be set or recorded by a Parameter or
Recording, respectively.
3.2.6 ObjectivesCalculator, eFeature
The ObjectivesCalculator takes the Response of a Model and calculates the objective values from
it. When using ephys recordings, one can use the eFEL library to extract eFeatures. Examples
of these eFeatures are spike amplitudes, steady state voltages, etc. The values of these eFeatures
can then be compared with experimental data values, and a score can be calculated based on the
difference between model and experiment.
4 Example Use Cases
To provide hands on experience how real-world optimisations can be developed using the BluePy-
Opt API, this section provides step-by-step guides for three use-cases. The first is a single
compartmental neuron model optimisation, the second is an optimisation of a state-of-the-art
morphologically detailed neuron model, and the third is an optimisation of a synaptic plasticity
model. All examples to follow assume NEURON default units, i.e. ms, mV, nA, µF cm−2, etc.
(Carnevale and Hines, 2016).
4.1 Single compartmental model
The first use case shows how to set up an optimisation of single compartmental neuron model with
two free parameters: The maximal conductances of the sodium and potassium Hodgkin-Huxley
ion channels. This example serves as an introduction for the user to the programming concepts in
BluePyOpt. It uses the NEURON simulator as backend.
9
First we need to import the top-level bluepyopt module. This example will also use BluePy-
Opt's electrophysiology features, so we also need to import the bluepyopt.ephys subpackage.
import bluepyopt as bpop
import bluepyopt.ephys as ephys
Next we load a morphology from a file. By default a morphology in NEURON has the fol-
lowing sectionlists: somatic, axonal, apical and basal. We create a Location object (specifically, a
NrnSecListLocation object) that points to the somatic sectionlist. This object will be used later
to specify where mechanisms are to be added etc.
morph = ephys.morphologies.NrnFileMorphology('simple.swc')
somatic_loc = ephys.locations.NrnSeclistLocation('somatic', seclist_name='somatic')
Now we can add ion channels to this morphology. First we add the default NEURON Hodgkin-
Huxley mechanism to the soma, as follows.
hh_mech = ephys.mechanisms.NrnMODMechanism(
name='hh',
prefix='hh',
locations=[somatic_loc])
The name argument can be chosen by the user, and should be unique across mechanisms.
The prefix argument string should correspond to the SUFFIX field in the NEURON NMODL
description file (Carnevale and Hines, 2006) of the channel. The locations argument specifies which
sections the mechanism are to be added to.
Next we need to specify the parameters of the model. A parameter can be in two states: frozen
and not-frozen. When a parameter is frozen it has an exact known value, otherwise it has well-
defined bounds but the exact value is not known yet. The parameter for the capacitance of the
soma will be a frozen value.
cm_param = ephys.parameters.NrnSectionParameter(
name='cm',
param_name='cm',
value=1.0,
locations=[somatic_loc],
frozen=True)
The two parameters that represent the maximal conductance of the sodium and potassium
channels are to be optimised, and are therefore specified as frozen=False, i.e. not-frozen, and
bounds for each are provided with the bounds argument.
gnabar_param = ephys.parameters.NrnSectionParameter(
name='gnabar_hh',
param_name='gnabar_hh',
locations=[somatic_loc],
bounds=[0.05, 0.125],
frozen=False)
gkbar_param = ephys.parameters.NrnSectionParameter(
name='gkbar_hh',
param_name='gkbar_hh',
10
bounds=[0.01, 0.075],
locations=[somatic_loc],
frozen=False)
To create the cell template, we pass all these objects to the constructor of the model.
simple_cell = ephys.cellmodels.NrnCellModel(
name='simple_cell',
morph=morph,
mechs=[hh_mech],
params=[cm_param, gnabar_param, gkbar_param])
To optimise the parameters of the cell, we further need to create a CellEvaluator object. This
object needs to know which protocols to inject, which parameters to optimise, and how to compute
a score, so we'll first create objects that define these aspects.
A protocol consists of a set of stimuli and a set of responses (i.e. recordings). These responses
will later be used to calculate the score of the specific model parameter values. In this example, we
will specify two stimuli, two square current pulses delivered at the soma with different amplitudes.
To this end, we first need to create a location object for the soma.
soma_loc = ephys.locations.NrnSeclistLocation(
name='soma',
seclist_name='somatic',
sec_index=0,
comp_x=0.5)
For each step in the protocol, we add a stimulus (NrnSquarePulse) and a recording (CompRe-
cording) in the soma.
sweep_protocols = {}
for protocol_name, amplitude in [('step1', 0.01), ('step2', 0.05)]:
stim = ephys.stimuli.NrnSquarePulse(
step_amplitude=amplitude,
step_delay=100,
step_duration=50,
location=soma_loc,
total_duration=200)
rec = ephys.recordings.CompRecording(
name='%s.soma.v' % protocol_name,
location=soma_loc,
variable='v')
protocol = ephys.protocols.SweepProtocol(protocol_name, [stim], [rec])
sweep_protocols[protocol.name] = protocol
The step_amplitude argument of the NrnSquarePulse specifies the amplitude of the current
pulse, and step_delay, step_duration, and total_duration specify the start time, length and
total simulation time. Finally, we create a combined protocol that encapsulates both current pulse
protocols.
twostep_protocol = ephys.protocols.SequenceProtocol('twostep', protocols=sweep_protocols)
11
Now to compute the model score that will be used by the optimisation algorithm, we define
objective objects. For this example, our objective is to match the eFEL "Spikecount" feature to
specified values for both current injection amplitudes. In this case, we will create one objective
per feature.
efel_feature_means = {'step1': {'Spikecount': 1}, 'step2': {'Spikecount': 5}}
objectives = []
for protocol_name, protocol in protocols.iteritems():
stim_start = protocol.stimuli[0].step_delay
stim_end = stim_start + protocol.stimuli[0].step_duration
for efel_feature_name, mean in
efel_feature_means[protocol_name].iteritems():
feature_name = '%s.%s' % (protocol_name, efel_feature_name)
feature = ephys.efeatures.eFELFeature(
feature_name,
efel_feature_name=efel_feature_name,
recording_names={'': '%s.soma.v' % protocol_name},
stim_start=stim_start,
stim_end=stim_end,
exp_mean=mean,
exp_std=0.05 * mean)
objective = ephys.objectives.SingletonObjective(
feature_name,
feature)
objectives.append(objective)
We then pass these objective definitions to a ObjectivesCalculator object, calculate the total
scores from a protocol response.
obj_calc = ephys.scorecalculators.ObjectivesCalculator(objectives)
Finally, we can combine everything together into a CellEvaluator. The CellEvaluator construc-
tor has a field param names which contains the (ordered) list of names of the parameters that are
used as input (and will be fitted later on).
cell_evaluator = ephys.evaluators.CellEvaluator(
cell_model=simple_cell,
param_names=['gnabar_hh', 'gkbar_hh'],
fitness_protocols=protocols,
fitness_calculator=obj_calc)
Now that we have a cell template and an evaluator for this cell, the Optimisation object can
be created and run.
optimisation = bpop.optimisations.DEAPOptimisation(
evaluator=cell_evaluator,
offspring_size = 100)
final_pop, hall_of_fame, logs, hist = optimisation.run(max_ngen=10)
12
After a short time, the optimisation returns us the final population, the hall of fame, a logbook
and an object containing the history of the population during the execution of the algorithm.
Figure 4 shows the results in a graphical form.
4.2 Neocortical pyramidal cell
Our second use case is a more complex example demonstrating the optimisation of a morphologi-
cally detailed model of a thick-tufted layer 5 pyramidal cell (L5PC) from the neocortex (Fig. 5a).
This example uses a BluePyOpt port of the state-of-the-art methods for the optimisation of the
L5PC model described in Markram et al. (2015). The original model is available online from the
Neocortical Microcircuit Collaboration Portal (Ramaswamy et al., 2015). Due to its complexity,
we will not describe the complete optimisation script here. The full code is available from the
BluePyOpt website. What we will do here is highlight the particularities of this model compared
to the introductory single compartmental model optimisation. As a first validation and point of
reference, we ran the BluePyOpt model with its original parameter values from (Ramaswamy et al.,
2015), as shown in Figure 5b.
For clarity, the code for setting the parameters, objective and optimisation algorithm is par-
titioned into separate modules. Configuration values are stored and read from JavaScript Object
Notation JSON files.
4.2.1 Parameters
Evidently, the parameters of this model, as shown in Table 1, far exceed in number those of the
single compartmental use case. The parameters marked as frozen are kept a constant throughout
the optimisation. The parameters to be optimised are the maximal conductances of the ion channels
and two values related to the calcium dynamics. The location of the parameters is based on
sectionlist names, whereby sections are automatically assigned to the somatic, axonal, apical and
basal sectionlists by NEURON when it loads a morphology.
An important aspect of this neuron model is the non-uniform distribution of certain ion channel
conductances. For example, the h-channel conductance is specified to increase exponentially with
distance from the soma (Kole et al., 2006), as follows.
soma_loc = ephys.locations.NrnSeclistLocation(
seclist_name='somatic',
seclist_index=0,
seg_x=0.5)
exponential_scaler = ephys.parameterscalers.NrnDistanceScaler(
origin=soma_loc,
distribution='(-0.8696 + 2.087*math.exp(({distance})*0.0031))*{value}')
parameter = ephys.parameters.NrnRangeParameter(
name='gIhbar_Ih.apical',
param_name='gIhbar_Ih'
value_scaler=exponential_scaler,
value=8e-5,
frozen=True,
locations=[apical_loc]))
13
)
V
m
(
e
g
a
t
l
o
V
)
V
m
(
e
g
a
t
l
o
V
60
40
20
0
−20
−40
−60
−80
0
60
40
20
0
−20
−40
−60
−80
0
250
200
150
100
50
j
s
e
v
i
t
c
e
b
o
f
o
m
u
S
50
50
100
Time (ms)
150
200
100
Time (ms)
150
200
population average
population minimum
population standard deviation
0
0
2
4
6
Generation #
8
10
(a) Top plots: In light blue, voltage traces recording during the two different step current injection for all
the individuals found that have objectives sum equal to zero. In dark blue, an example of ones of these
individuals. The target objectives of Step1 and Step2 were 1 and 5 action potentials respectively. Bottom
plot: Evolution of minimal objectives sum during the 10 generations of the evolutionary algorithm.
0.10
0.08
0.06
0.04
0.02
r
a
b
k
g
r
e
t
e
m
a
r
a
P
0.00
0.00
5.6
4.8
4.0
3.2
2.4
1.6
0.8
0.0
0.05
0.10
0.15
0.20
Parameter gnabar
j
)
1
+
s
e
v
i
t
c
e
b
o
f
o
m
u
s
(
g
o
l
(b) Triangular grid plot of the parameter space. Every point of the grid is a point where the algorithm
evaluated an individual. X- and Y-axis represent the values of the sodium and potassium maximal con-
ductance respectively (units S cm−2). The color represents the average natural logarithm of the objectives
sum of every triangle's three points. Circles represent the solutions with an objectives sum of 0.
Figure 4: Results of the single compartmental model optimisation.
14
v
.
a
m
o
s
.
1
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
2
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
3
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
1
d
n
e
d
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
v
.
2
d
n
e
d
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
0
0
0
0
0
0
500
1000
1500
2000
2500
3000
500
1000
1500
2000
2500
3000
500
1000
1500
2000
2500
3000
100
200
300
400
500
600
100
200
300
400
500
600
100
200
300
Time (ms)
400
500
600
(a) Morphological reconstruction of L5PC used
in the model obtained from the NMC portal
(Ramaswamy et al., 2015).
(b) Voltage traces recorded in soma and den-
drites (dend1 660 µm, dend2 800 µm from soma
in apical trunk).
s
e
v
i
t
c
e
b
O
j
bAP.soma.Spikecount
bAP.soma.AP_width
bAP.soma.AP_height
bAP.dend2.AP_amplitude_from_voltagebase
bAP.dend1.AP_amplitude_from_voltagebase
Step3.soma.time_to_first_spike
Step3.soma.mean_frequency
Step3.soma.doublet_ISI
Step3.soma.adaptation_index2
Step3.soma.ISI_CV
Step3.soma.AP_width
Step3.soma.AP_height
Step3.soma.AHP_slow_time
Step3.soma.AHP_depth_abs_slow
Step3.soma.AHP_depth_abs
Step2.soma.time_to_first_spike
Step2.soma.mean_frequency
Step2.soma.doublet_ISI
Step2.soma.adaptation_index2
Step2.soma.ISI_CV
Step2.soma.AP_width
Step2.soma.AP_height
Step2.soma.AHP_slow_time
Step2.soma.AHP_depth_abs_slow
Step2.soma.AHP_depth_abs
Step1.soma.time_to_first_spike
Step1.soma.mean_frequency
Step1.soma.doublet_ISI
Step1.soma.adaptation_index2
Step1.soma.ISI_CV
Step1.soma.AP_width
Step1.soma.AP_height
Step1.soma.AHP_slow_time
Step1.soma.AHP_depth_abs_slow
Step1.soma.AHP_depth_abs
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Objective value (# std)
(c) Objective scores for the model calculated
based on experimental mean and standard de-
viation.
Figure 5: L5PC model as simulated by BluePyOpt with parameter values from (Markram et al.,
2015)
15
Table 1. List of parameters for L5PC example. Optimised parameters with bounds are in upper
part of the table, lower part lists the frozen parameters with their value. The parameter with exp
distribution is scaled for every morphological segment with the equation −0.8696 + 2.087 · e0.0031·d
with d the distance of the segment to the soma.
Location Mechanism
Parameter name Distribution Units
Lower bound Upper bound
0.04
0.04
0.001
4
4
1
0.1
0.1
2
0.001
0.01
0.05
1000
1
1
0.1
0.001
0.01
0.05
1000
apical
apical
apical
axonal
axonal
axonal
axonal
axonal
axonal
axonal
axonal
axonal
axonal
somatic
somatic
somatic
somatic
somatic
somatic
somatic
NaTs2 t
SKv3 1
Im
NaTa t
Nap Et2
K Pst
K Tst
SK E2
SKv3 1
Ca HVA
Ca LVAst
CaDynamics E2
CaDynamics E2
NaTs2 t
SKv3 1
SK E2
Ca HVA
Ca LVAst
CaDynamics E2
CaDynamics E2
gNaTs2 tbar
gSKv3 1bar
gImbar
gNaTa tbar
gNap Et2bar
gK Pstbar
gK Tstbar
gSK E2bar
gSKv3 1bar
gCa HVAbar
gCa LVAstbar
gamma
decay
gNaTs2 tbar
gSKv3 1bar
gSK E2bar
gCa HVAbar
gCa LVAstbar
gamma
decay
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
ms
S cm−2
S cm−2
S cm−2
S cm−2
S cm−2
ms
Location Mechanism
Parameter name Distribution Units
global
global
all
all
all
all
apical
apical
apical
somatic
somatic
basal
axonal
axonal
basal
apical
somatic
Ih
Ih
Ih
v init
celsius
g pas
e pas
cm
Ra
ena
ek
cm
ena
ek
cm
ena
ek
gIhbar
gIhbar
gIhbar
mV
◦C
S cm−2
mV
µF cm−2
Ω cm
mV
mV
µF cm−2
mV
mV
µF cm−2
mV
mV
S cm−2
S cm−2
S cm−2
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
uniform
exp
uniform
0
0
0
0
0
0
0
0
0
0
0
0.0005
20
0
0
0
0
0
0.0005
20
Value
-65
34
3e-05
-75
1
100
50
-85
2
50
-85
2
50
-85
8e-05
8e-05
8e-05
16
4.2.2 Protocols
During the optimisation, the model is evaluated using three square current step stimuli applied
and recorded at the soma. For these protocols, a holding current is also applied during the entire
stimulus, the amplitude of which is the same as was used in the in vitro experiments to keep the
cell at a standardised membrane voltage before the step current injection.
Another stimulus protocol checks for a backpropagating action potential (bAP ) by stimulating
the soma with a very short pulse, and measuring the height and width of the bAP at a location of
660 µm and 800 µm from the soma in the apical dendrite. It is specified as follows.
for loc_name, loc_distance in [('dendloc1', 660), ('dendloc2', 800)]:
loc = ephys.locations.NrnSomaDistanceCompLocation(
name=loc_name,
soma_distance=loc_distance)
recording = nrpel.recordings.CompRecording(
name='bAP.%s.v' % (loc_name),
location=loc)
4.2.3 Objectives
For each of the four stimuli defined above, a set of eFeatures is calculated (Table 2). These are then
compared with the same features extracted from experimental data. As described in Markram et al.
(2015), experiments were performed that applied these and other protocols to L5PCs in vitro. For
these cells, the same eFeatures were extracted, and the mean µexp and standard deviation σexp
calculated. The bAP target values are extracted from Larkum et al. (2001).
For every feature value fmodel, one objective value is calculated:
objective = (cid:12)(cid:12)(cid:12)(cid:12)
µexp − fmodel
σexp
(cid:12)(cid:12)(cid:12)(cid:12)
4.2.4 Optimisation
For the optimisation of this cell model we needed significantly more computing resources. The
goal was to find a solution that has objective values that are only several standard deviation
away from the experimental mean. For this we ran 200 generations with an offspring size of the
genetic algorithm of 100 individuals. The evaluation of these 100 individuals was parallelised over
50 CPU cores using SCOOP, and took a couple of hours to run. Figure 6 shows the results of
the optimisation, and Figure 7 shows a comparison of the optimised model to its reference under
Gaussian noise current injection (not used during the optimisation).
4.3 Spike-timing dependent plasticity (STDP) model
The BluePyOpt framework was designed to be versatile and broadly applicable to a wide range of
neuroscientific optimisation problems. In this use case, we demonstrate this versatility by using
BluePyOpt to optimise the parameters of a calcium-based STDP model (Graupner and Brunel,
2012) to summary statistics from in vitro experiments (Nevian and Sakmann, 2006). That is, we
show how to fit the model to literature data, commonly reported just as mean and SEM of the
amount of potentiation (depression) induced by one or more stimulation protocols.
17
Table 2. List of eFeatures for L5PC example. Locations dend1 and dend2 are respectively
660 µm and 800 µm from the soma in the apical trunk. Depending on the eFeature type the units
can be mV or ms.
Stimulus Location
eFeature
Mean
Std
Step1
soma
Step2
soma
Step3
soma
bAP
dend1
dend2
soma
AHP depth abs
AHP depth abs slow
AHP slow time
AP height
AP width
ISI CV
adaptation index2
doublet ISI
mean frequency
time to first spike
AHP depth abs
AHP depth abs slow
AHP slow time
AP height
AP width
ISI CV
adaptation index2
doublet ISI
mean frequency
time to first spike
AHP depth abs
AHP depth abs slow
AHP slow time
AP height
AP width
ISI CV
adaptation index2
doublet ISI
mean frequency
time to first spike
AP amplitude from voltagebase
AP amplitude from voltagebase
AP height
AP width
Spikecount
-60.3636
-61.1513
0.1599
25.0141
3.5312
0.109
0.0047
62.75
6
27.25
-59.9055
-60.2471
0.1676
27.1003
2.7917
0.0674
0.005
44.0
8.5
19.75
-57.0905
-61.1513
0.1968
19.7207
3.5347
0.0737
0.0055
22.75
17.5
10.5
45
36
25.0
2.0
1.0
2.3018
2.3385
0.0483
3.1463
0.8592
0.1217
0.0514
9.6667
1.2222
5.7222
1.8329
1.8972
0.0339
3.1463
0.7499
0.075
0.0067
7.1327
0.9796
2.8776
2.3427
2.3385
0.0112
3.7204
0.8788
0.0292
0.0015
4.14
0.8
1.36
10
9.33
5.0
0.5
0.01
18
v
.
a
m
o
s
.
1
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
2
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
3
p
e
t
S
)
V
m
(
e
g
a
t
l
o
V
v
.
1
d
n
e
d
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
v
.
2
d
n
e
d
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
v
.
a
m
o
s
.
P
A
b
)
V
m
(
e
g
a
t
l
o
V
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
40
20
0
−20
−40
−60
−80
0
0
0
0
0
0
500
1000
1500
2000
2500
3000
500
1000
1500
2000
2500
3000
500
1000
1500
2000
2500
3000
100
200
300
400
500
600
100
200
300
400
500
600
100
200
300
Time (ms)
400
500
600
s
e
v
i
t
c
e
b
O
j
bAP.soma.Spikecount
bAP.soma.AP_width
bAP.soma.AP_height
bAP.dend2.AP_amplitude_from_voltagebase
bAP.dend1.AP_amplitude_from_voltagebase
Step3.soma.time_to_first_spike
Step3.soma.mean_frequency
Step3.soma.doublet_ISI
Step3.soma.adaptation_index2
Step3.soma.ISI_CV
Step3.soma.AP_width
Step3.soma.AP_height
Step3.soma.AHP_slow_time
Step3.soma.AHP_depth_abs_slow
Step3.soma.AHP_depth_abs
Step2.soma.time_to_first_spike
Step2.soma.mean_frequency
Step2.soma.doublet_ISI
Step2.soma.adaptation_index2
Step2.soma.ISI_CV
Step2.soma.AP_width
Step2.soma.AP_height
Step2.soma.AHP_slow_time
Step2.soma.AHP_depth_abs_slow
Step2.soma.AHP_depth_abs
Step1.soma.time_to_first_spike
Step1.soma.mean_frequency
Step1.soma.doublet_ISI
Step1.soma.adaptation_index2
Step1.soma.ISI_CV
Step1.soma.AP_width
Step1.soma.AP_height
Step1.soma.AHP_slow_time
Step1.soma.AHP_depth_abs_slow
Step1.soma.AHP_depth_abs
(a) Similar to Fig. 5b, with the top ten objective
values found by BluePyOpt, and the best one
plotted darker
(b) Objective scores for the best objective values
found by BluePyOpt.
0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
Objective value (# std)
population average
population minimum
population standard deviation
l
e
u
a
v
r
e
t
e
m
a
r
a
P
g
o
l
5
0
−5
−10
−15
t
_
2
s
T
a
N
_
r
a
b
t
_
2
s
T
a
N
g
l
a
c
i
p
a
1
_
3
v
K
S
_
r
a
b
1
_
3
v
K
S
g
l
a
c
i
p
a
m
I
_
r
a
b
m
I
g
l
a
c
i
p
a
t
_
a
T
a
N
_
r
a
b
t
_
a
T
a
N
g
l
a
n
o
x
a
2
t
E
_
p
a
N
_
r
a
b
2
t
E
_
p
a
N
g
l
a
n
o
x
a
t
s
P
_
K
_
r
a
b
t
s
P
_
K
g
l
a
n
o
x
a
t
s
T
_
K
_
r
a
b
t
s
T
_
K
g
l
a
n
o
x
a
2
E
_
K
S
_
r
a
b
2
E
_
K
S
g
l
a
n
o
x
a
1
_
3
v
K
S
_
r
a
b
1
_
3
v
K
S
g
l
a
n
o
x
a
l
a
n
o
x
a
l
a
n
o
x
a
A
V
H
_
a
C
_
r
a
b
A
V
H
_
a
C
g
t
s
A
V
L
_
a
C
_
r
a
b
t
s
A
V
L
_
a
C
g
2
E
_
s
c
i
m
a
n
y
D
a
C
_
a
m
m
a
g
Parameters
l
a
n
o
x
a
2
E
_
s
c
i
m
a
n
y
D
a
C
_
y
a
c
e
d
l
a
n
o
x
a
t
_
2
s
T
a
N
_
r
a
b
t
_
2
s
T
a
N
g
c
i
t
a
m
o
s
1
_
3
v
K
S
_
r
a
b
1
_
3
v
K
S
g
c
i
t
a
m
o
s
2
E
_
K
S
_
r
a
b
2
E
_
K
S
g
c
i
t
a
m
o
s
A
V
H
_
a
C
_
r
a
b
A
V
H
_
a
C
g
c
i
t
a
m
o
s
t
s
A
V
L
_
a
C
_
r
a
b
t
s
A
V
L
_
a
C
g
c
i
t
a
m
o
s
2
E
_
s
c
i
m
a
n
y
D
a
C
_
a
m
m
a
g
c
i
t
a
m
o
s
2
E
_
s
c
i
m
a
n
y
D
a
C
_
y
a
c
e
d
c
i
t
a
m
o
s
8000
6000
4000
j
s
e
v
i
t
c
e
b
o
f
o
m
u
S
2000
0
0
20
40
60
80
100
Generation #
(c) Parameter diversity in final solution. Param-
eter values for best (blue crosses) and 10 best
individuals (black dots) and all individuals with
all objectives below 5 (grey dots).
(d) Evolution of the L5PC optimisation that
found the model in panel 6a. Plot shows the
minimal, maximal and average scores found in
the consecutive generations of the evolutionary
algorithm.
Figure 6: Results of optimising L5PC model using BluePyOpt.
19
Current
(nA)
1.5
1.0
0.5
0.0
−0.5
−1.0
−1.5
−2.0
0
Voltage
(mV)
−100
1000
2000
3000
Time (ms)
4000
5000
Figure 7: Comparison of L5PC model solutions found by BluePyOpt to reference model. Top:
Gaussian noise current injected in the models. Middle: Raster plot of model responses to noise
current injection. Bottom: Voltage response of the models to noise current injection.
In red,
model parameters from Markram et al. (2015) model, in light blue the best 10 individuals found
by BluePyOpt, in dark blue the best individual. Figure as in Pozzorini et al. (2015).
In the set of experiments performed by Nevian and Sakmann (2006), a presynaptic action
potential (AP) is paired with a burst of three post-synaptic APs to induce either long-term po-
tentiation (LTP) or long-term depression (LTD) of the postsynaptic neuron response. The time
difference ∆t between the presynaptic AP and the postsynaptic burst determines the direction of
change: A burst shortly preceding the presynaptic AP causes LTD, with a peak at ∆t = −50 ms;
conversely, a burst shortly after the presynaptic AP results in LTP, with a peak at ∆t = +10 ms
(Nevian and Sakmann, 2006).
The model proposed by Graupner and Brunel (2012) assumes bistable synapses, with plasticity
of their absolute efficacies governed by post-synaptic calcium dynamics. That is, each synapse
is either in an high-conductance state or a low-conductance state; potentiation and depression
translate then into driving a certain fraction of synapses from the low-conductance state to the
high-conductance state and vice versa; synapses switch from one state to another depending on
the time spent by post-synaptic calcium transients above a potentiation (depression) threshold.
Following Graupner and Brunel (2012), the model is described as
τ
dρ
dt
= −ρ(1 − ρ)(ρ⋆ − ρ) + γp(1 − ρ)Θ[c(t) − θp]
− γdρΘ[c(t) − θd] + Noise(t)
dc
dt
= −
c
τCa
+ Cpre X
δ(t − ti − D) + Cpost X
δ(t − tj)
i
j
(1)
(2)
20
where ρ is the absolute synaptic efficacy, ρ⋆ delimits the basins of attraction of the potentiated
and depressed state, γp (γd) is the potentiation (depression) rate, Θ is the Heaviside function,
θp (θd) is the potentiation (depression) threshold, Noise(t) is an activity dependent noise. The
postsynaptic calcium concentration is described by the process c, with time constant τCa. Cpre
is the calcium transient caused by a presynaptic spike with a delay D to account for the slow
activation of NMDARs, while Cpost is the calcium transient caused by a postsynaptic spike.
For periodic stimulation protocols, such as in Nevian and Sakmann (2006), the synaptic tran-
sition probability can be easily calculated analytically (Graupner and Brunel, 2012), allowing es-
timation of the amount of potentiation (depression) induced by the stimulation protocol without
actually running any neuron simulations. The amount of potentiation (depression) obtained with
different protocols in vitro become the objectives of the optimisation.
A small Python module stdputil calculating this model is available in the example section on
the BluePyOpt website. To optimise this model, only an Evaluator class has to be defined that
implements an evaluation function:
class GraupnerBrunelEvaluator(bpop.evaluators.Evaluator):
def __init__(self):
super(GraupnerBrunelEvaluator, self).__init__()
# Graupner-Brunel model parameters and boundaries
# From Graupner and Brunel (2012)
self.graup_params = [('tau_ca', 1e-3, 100e-3),
('C_pre', 0.1, 20.0),
('C_post', 0.1, 50.0),
('gamma_d', 5.0, 5000.0),
('gamma_p', 5.0, 2500.0),
('sigma', 0.35, 70.7),
('tau', 2.5, 2500.0),
('D', 0.0, 50e-3),
('b', 1.0, 100.0)]
self.params = [bpop.parameters.Parameter
(param_name, bounds=(min_bound, max_bound))
for param_name, min_bound, max_bound in self.
graup_params]
self.param_names = [param.name for param in self.params]
self.protocols, self.sg, self.stdev, self.stderr = \
stdputil.load_neviansakmann()
self.objectives = [bpop.objectives.Objective(protocol.prot_id)
for protocol in self.protocols]
def get_param_dict(self, param_values):
return gbParam(zip(self.param_names, param_values))
def compute_synaptic_gain_with_lists(self, param_values):
param_dict = self.get_param_dict(param_values)
21
syn_gain = [stdputil.protocol_outcome(protocol, param_dict) \
for protocol in self.protocols]
return syn_gain
def evaluate_with_lists(self, param_values):
param_dict = self.get_param_dict(param_values)
err = []
for protocol, sg, stderr in zip(self.protocols, self.sg, self.stderr):
res = stdputil.protocol_outcome(protocol, param_dict)
err.append(numpy.abs(sg - res) / stderr)
return err
With the evaluator defined, running the optimisation becomes as simple as:
evaluator = GraupnerBrunelEvaluator()
opt = bpop.optimisations.DEAPOptimisation(GraupnerBrunelEvaluator())
results = opt.run(max_ngen=200)
Figure 8 shows the results of the optimisation.
5 Discussion
BluePyOpt was designed to be a state-of-the-art tool for neuroscientific model parameter search
problems that is both easy to use for inexperienced users, and versatile and broadly applicable
for power users. Three example use cases were worked through in the text to demonstrate how
BluePyOpt serves each of these user communities.
From a software point of view, this dual goal was achieved by an object oriented architecture
which abstracts away the domain-specific complexities of search algorithms and simulators, while
allowing extension and modification of the implementation and settings of an optimisation. Python
was an ideal implementation language for such an architecture, with its very open and minimal
approach to extending existing implementations. Object oriented programming allows users to
define new subclasses of existing BluePyOpt API classes with different implementations. The duck
typing of Python allows parameters and objectives to have any kind of type, e.g. they don't have
to be floating point numbers. In extreme cases, function implementations can even be overwritten
at run time by monkey patching. These features of Python gives extreme flexibility to the user,
which will make BluePyOpt applicable to many use cases.
A common issue arising for users of optimisation software is the configuration of computing
infrastructure. The fact that BluePyOpt is coded in Python, an interpreted language, and provides
Ansible scripts for its installation, makes straightforward to run on diverse computing platforms.
This will give the user the flexibility to pick the computing infrastructure which best fits their
needs, be it their desktop computer, university cluster or temporarily rented cloud infrastructure,
such as offered by Amazon Web Services.
22
Best model
In vitro
2.5
2.0
1.5
1.0
e
g
n
a
h
c
e
d
u
t
i
l
p
m
a
P
S
P
E
0.5
−100
−80
−60
−40
−20
∆t (ms)
0
20
40
60
(a) Comparison between model and experimen-
tal results; the models match the available in
vitro data and predict the outcome of the miss-
ing points.
In light blue, models generated by
individuals having fitness values within one stan-
dard error of the mean from experimental in
vitro data.
In dark blue, best model, defined
as the closest to all experimental data points.
Experimental data from Nevian and Sakmann
(2006) digitized using Ankit Rohatgi (2015).
population average
population minimum
m
u
i
c
l
a
c
m
u
i
c
l
a
c
m
u
i
c
l
a
c
m
u
i
c
l
a
c
m
u
i
c
l
a
c
m
u
i
c
l
a
c
2
1
0
2
1
0
2
1
0
2
1
0
2
1
0
2
1
0
0.0
0.1
0.2
-90ms
θp
θd
-50ms
θp
θd
-30ms
θp
θd
-10ms
θp
θd
+10ms
θp
θd
+50ms
θp
θd
0.3
Time (s)
0.4
0.5
0.6
(b) Calcium transients generated by the best
model (panel a) for each stimulation protocol.
Potentiation and depression thresholds, θp and
θd respectively, are indicated by the dashed
lines.
60
50
40
30
20
10
j
s
e
v
i
t
c
e
b
o
f
o
m
u
S
0
0
50
100
Generation #
150
200
(c) Evolution of the STDP optimisation that
found the model in panel 8a. Minimal and aver-
age scores found in the consecutive generations
of the evolutionary algorithm.
Figure 8: Results of STDP fitting.
23
This present paper focuses on the use of BluePyOpt as an optimisation tool.
It is worth
noting that the application domain of BluePyOpt needn't remain limited to this. The ephys
model abstraction can also be used in validation, assessing generalisation, and parameter sensitivity
analyses. E.g. when applying a map function to an ephys model evaluation function which takes
as input a set of morphologies, one can measure how well the model generalises when applied to
different morphologies. The present paper expressly does not touch on issues of generalization
power, overfitting, or uniqueness of solution. It is worth now making a few points on the latter.
While BluePyOpt could successfully optimise the three examples, Figures 4b and 6c show a diversity
of solutions giving good fitness values. That is, for these neuron model optimisation problems, the
solutions found are non-unique. This is compatible with the observation that Nature itself also
utilises various and non-unique solutions to provide the required phenotype (Schulz et al., 2007;
Taylor et al., 2009). For other problems solutions could be unique, making BluePyOpt useful e.g.
for extracting parameters for models of synapse dynamics (Fuhrmann et al., 2002).
Of course BluePyOpt is not the only tool available to perform parameter optimisations in neu-
roscience (Druckmann et al., 2007; Van Geit et al., 2007; Bahl et al., 2012; Friedrich et al., 2014;
Carlson et al., 2015; Pozzorini et al., 2015). Some tools provide a Graphical User Interface (GUI),
other tools are written in other languages, or use different types of evaluation functions or search
algorithms. We explicitly didn't make a detailed comparison between BluePyOpt and other tools
because many of these tools are developed for specific and non-overlapping applications, making
a systematic comparison difficult. This suggests perhaps BluePyOpt's greatest strength, its broad
applicability relative to previous approaches.
While BluePyOpt significantly reduces the domain specific knowledge required to employ pa-
rameter optimisation strategies, some thought from the user in setting up their problem is still
required. For example, BluePyOpt does in principle allow brute force optimisation of all param-
eters of the L5PC model example, including channel kinetics parameters and passive properties,
but such an approach would almost certainly be unsuccessful. Moreover, when it comes to as-
sessing fitness of models, care and experience is also required to avoid the optimisation getting
caught in local minima, or cannibalizing one objective for another. For neuron models for example,
feature-based approaches coupled with multi-objective optimisation strategies have proven espe-
cially effective (Druckmann et al., 2007). Indeed, even the stimuli and features themselves can be
optimised on theoretical grounds to improve parameter optimisation outcomes (Druckmann et al.,
2011). For these reasons, an important companion of BluePyOpt will be a growing library of
working optimisation examples developed by domain experts for a variety of common use cases,
to help inexperienced users quickly adopt a working strategy most closely related to their specific
needs.
As these examples library grows, so too will the capabilities of BluePyOpt evolve. Some im-
provements planned for the future include the following:
Support for multi-stage optimisations allowing for example the passive properties of a
neuron to optimised in a first stage, prior to optimising the full-active dendritic parameters
in a second phase
Embedded optimisation allowing for example an optimisation of a "current at rheobase"
feature requiring threshold detection during the optimisation using e.g. a binary search. Also,
for integrate-and-fire models such as the adapting exponential integrate-and-fire (Brette and Gerstner,
2005), a hybrid of a global stochastic search and local gradient descent has been shown to be
24
a competitive approach (Jolivet et al., 2008)
Fast pre-evaluation of models to exclude clearly bad parameters before computation time
is wasted on them
Support for evaluation time-outs to protect against optimisations getting stuck in long
evaluations, for example when using NEURON's CVODE solver, which can occasionally get
stuck at excessively high resolutions.
Support for explicit units to make optimisation scripts more readable, and sharing with
others less error prone.
Although parameter optimisations can require appreciable computing resources, the ability
to share the code of an optimisation through a light-weight script or ipython notebook using
BluePyOpt will improve reproducibility in the field. It allows for neuroscientists to exchange code
In the future,
and knowledge about search algorithms that perform well for particular models.
making it possible for users to read and write model descriptions from community standards
(Cannon et al., 2014; Raikov et al., 2011), could further ease the process of plugging in a model
into a BluePyOpt optimisation. By providing the neuroscientific community with BluePyOpt,
an open source tool to optimise model parameters in Python which is powerful, easy to use and
broadly applicable, we hope to catalyse community uptake of state-of-the-art model optimisation
approaches, and encourage code sharing and collaboration.
Downloads
The source code of BluePyOpt, the example scripts and cloud installation scripts are available
on Github at https://github.com/BlueBrain/BluePyOpt, the former under the GNU Lesser
General Public License version 3 (LGPLv3), and the latter two under a BSD license.
Disclosure/Conflict-of-Interest Statement
The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.
Author Contributions
WVG, MG and JDC designed the software and contributed code. WVG, GC, MG and CR de-
signed the examples and contributed code. WVG, EM, MG and GC wrote the manuscript. All:
Conception and design, drafting and revising, and final approval.
Acknowledgments
We wish to thank Michael Graupner for his support with the implementation of the calcium-based
STDP model (Graupner and Brunel, 2012) and for fruitful discussions, and Elisabetta Iavarone for
testing the cloud installation functionality.
25
Funding: The work was supported by funding from the EPFL to the Laboratory of Neural
Microcircuitry (LNMC) and funding from the ETH Domain for the Blue Brain Project (BBP).
Additional support was provided by funding for the Human Brain Project from the European
Union Seventh Framework Program (FP7/2007- 2013) under grant agreement no. 604102 (HBP).
The BlueBrain IV BlueGene/Q and Linux cluster used as a development system for this work is
financed by ETH Board Funding to the Blue Brain Project as a National Research Infrastructure
and hosted at the Swiss National Supercomputing Center (CSCS).
References
Achard, P. and De Schutter, E. (2006). Complex parameter landscape for a complex neuron model.
PLoS Comput Biol 2, e94
Ankit Rohatgi (2015). Webplotdigitizer. [Online; accessed 26-February-2016; Version: 3.9]
Bahl, A., Stemmler, M. B., Herz, A. V., and Roth, A. (2012). Automated optimization of a reduced
layer 5 pyramidal cell model based on experimental data. Journal of Neuroscience Methods 210,
22 – 34. doi:http://dx.doi.org/10.1016/j.jneumeth.2012.04.006. Special Issue on Computational
Neuroscience
Bhalla, U. S. and Bower, J. M. (1993). Exploring parameter space in detailed single neuron models:
simulations of the mitral and granule cells of the olfactory bulb. Journal of Neurophysiology 69,
1948–1965
Bleuler, S., Laumanns, M., Thiele, L., and Zitzler, E. (2003). PISA - a platform and programming
language independent interface for search algorithms. In Evolutionary Multi-Criterion Optimiza-
tion (EMO 2003), eds. C. M. Fonseca, P. J. Fleming, E. Zitzler, K. Deb, and L. Thiele (Berlin:
Springer), Lecture Notes in Computer Science, 494 – 508
Blue Brain Project (2015). eFEL. https://github.com/BlueBrain/eFEL.
[Online; accessed
16-February-2016]
Blue Brain Project (2016). BluePyOpt. https://github.com/BlueBrain/BluePyOpt.
[Online;
accessed 16-February-2016]
Brette, R. and Gerstner, W. (2005). Adaptive exponential integrate-and-fire model as an effective
description of neuronal activity. Journal of Neurophysiology 94, 3637–3642. doi:10.1152/jn.00686.
2005
Cannon, R. C., Gleeson, P., Crook, S., Ganapathy, G., Marin, B., Piasini, E., et al. (2014). LEMS:
A language for expressing complex biological models in concise and hierarchical form and its use
in underpinning neuroml 2. Frontiers in Neuroinformatics 8. doi:10.3389/fninf.2014.00079
Carlson, K. D., Nageswaran, J. M., Dutt, N., and Krichmar, J. L. (2015). An efficient automated
parameter tuning framework for spiking neural networks. Neuromorphic Engineering Systems
and Applications , 168
Carnevale, N. T. and Hines, M. L. (2006). The NEURON Book (Cambridge University Press).
Cambridge Books Online
26
Carnevale, N. T. and Hines, M. L. (2016). Units used in neuron.
[Online; accessed 26-February-
2016]
Chef (2009). Open Source Chef. http://www.chef.io. [Online; accessed 16-February-2016]
Dalcn, L., Paz, R., and Storti, M. (2005). MPI for Python. Journal of Parallel and Distributed
Computing 65, 1108 – 1115. doi:http://dx.doi.org/10.1016/j.jpdc.2005.03.010
Davison, A. P., Brderle, D., Eppler, J. M., Kremkow, J., Muller, E., Pecevski, D., et al. (2009).
PyNN: a common interface for neuronal network simulators. Frontiers in Neuroinformatics 2.
doi:10.3389/neuro.11.011.2008
DEAP Project (2016). DEAP documentation. [Online; accessed 26-February-2016]
Deb, K., Pratap, A., Agarwal, S., and Meyarivan, T. (2002). A fast and elitist multiobjective
genetic algorithm: NSGA-II. Trans. Evol. Comp 6, 182–197. doi:10.1109/4235.996017
Druckmann, S., Banitt, Y., Gidon, A., Schurmann, F., Markram, H., and Segev, I. (2007). A novel
multiple objective optimization framework for constraining conductance-based neuron models
by experimental data. Frontiers in Neuroscience 1, 7–18. doi:10.3389/neuro.01.1.1.001.2007
Druckmann, S., Berger, T. K., Schrmann, F., Hill, S., Markram, H., and Segev, I. (2011). Effective
stimuli for constructing reliable neuron models. PLoS Comput Biol 7, 1–13. doi:10.1371/journal.
pcbi.1002133
Eliasmith, C., Gosmann, J., and Choo, X. (2016). Biospaun: A large-scale behaving brain model
with complex neurons. arXiv preprint arXiv:1602.05220
Eppler, J. M., Helias, M., Muller, E., Diesmann, M., and Gewaltig, M.-O. (2009). PyNEST: a
convenient interface to the NEST simulator. Frontiers in Neuroinformatics 2. doi:10.3389/neuro.
11.012.2008
Fortin, F.-A., De Rainville, F.-M., Gardner, M.-A., Parizeau, M., and Gagn´e, C. (2012). DEAP:
Evolutionary algorithms made easy. Journal of Machine Learning Research 13, 2171–2175
Friedrich, P., Vella, M., Gulys, A. I., Freund, T. F., and Kli, S. (2014). A flexible, interactive
software tool for fitting the parameters of neuronal models. Frontiers in Neuroinformatics 8.
doi:10.3389/fninf.2014.00063
Fuhrmann, G., Segev, I., Markram, H., and Tsodyks, M. (2002). Coding of temporal information
by activity-dependent synapses. Journal of Neurophysiology 87, 140–148
Goodman, D. F. M. and Brette, R. (2009). The Brian simulator. Frontiers in Neuroscience 3.
doi:10.3389/neuro.01.026.2009
Graupner, M. and Brunel, N. (2012). Calcium-based plasticity model explains sensitivity of synap-
tic changes to spike pattern, rate, and dendritic location. Proceedings of the National Academy
of Sciences 109, 3991–3996
Gurkiewicz, M. and Korngreen, A. (2007). A numerical approach to ion channel modelling using
whole-cell voltage-clamp recordings and a genetic algorithm. PLoS Comput Biol 3, e169
27
Hansen, N. and Ostermeier, A. (2001). Completely derandomized self-adaptation in evolution
strategies. Evol. Comput. 9, 159–195. doi:10.1162/106365601750190398
HashiCorp (2010). Vagrant. http://www.vagrantup.com/. [Online; accessed 16-February-2016]
Hay, E., Hill, S., Schurmann, F., Markram, H., and Segev, I. (2011). Models of neocortical layer
5b pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS
Comput Biol 7, e1002107
Hines, M., Davison, A. P., and Muller, E. (2009). NEURON and Python. Frontiers in Neuroin-
formatics 3. doi:10.3389/neuro.11.001.2009
Hold-Geoffroy, Y., Gagnon, O., and Parizeau, M. (2014). Once you SCOOP, no need to fork.
In Proceedings of the 2014 Annual Conference on Extreme Science and Engineering Discovery
Environment (ACM), 60
Huys, Q. J. and Paninski, L. (2009). Smoothing of, and parameter estimation from, noisy biophys-
ical recordings. PLoS Comput Biol 5, e1000379
Izhikevich, E. M. and Edelman, G. M. (2008). Large-scale model of mammalian thalamocortical
systems. Proceedings of the National Academy of Sciences 105, 3593–3598. doi:10.1073/pnas.
0712231105
Jolivet, R., Schurmann, F., Berger, T. K., Naud, R., Gerstner, W., and Roth, A. (2008). The
quantitative single-neuron modeling competition. Biological Cybernetics 99, 417–426. doi:10.
1007/s00422-008-0261-x
Kennedy, J. and Eberhart, R. (1995). Particle swarm optimization. In Neural Networks, 1995.
Proceedings., IEEE International Conference on. vol. 4, 1942–1948 vol.4. doi:10.1109/ICNN.
1995.488968
Kole, M. H., Hallermann, S., and Stuart, G. J. (2006). Single Ih channels in pyramidal neuron
dendrites: properties, distribution, and impact on action potential output. The Journal of
neuroscience 26, 1677–1687
Larkum, M. E., Zhu, J. J., and Sakmann, B. (2001). Dendritic mechanisms underlying the coupling
of the dendritic with the axonal action potential initiation zone of adult rat layer 5 pyramidal
neurons. The Journal of Physiology 533, 447–466. doi:10.1111/j.1469-7793.2001.0447a.x
Markram, H., Muller, E., Ramaswamy, S., Reimann, M. W., Abdellah, M., Sanchez, C. A., et al.
(2015). Reconstruction and simulation of neocortical microcircuitry. Cell 163, 456–492. doi:
10.1016/j.cell.2015.09.029
Merolla, P. A., Arthur, J. V., Alvarez-Icaza, R., Cassidy, A. S., Sawada, J., Akopyan, F., et al.
(2014). A million spiking-neuron integrated circuit with a scalable communication network and
interface. Science 345, 668–673. doi:10.1126/science.1254642
Muller, E., Bednar, J. A., Diesmann, M., Gewaltig, M.-O., Hines, M., and Davison, A. P. (2015).
Python in neuroscience. Frontiers in Neuroinformatics 9. doi:10.3389/fninf.2015.00011
Nevian, T. and Sakmann, B. (2006). Spine Ca2+ signaling in spike-timing-dependent plasticity.
The Journal of Neuroscience 26, 11001–11013
28
Oliphant, T. E. (2007). Python for scientific computing. Computing in Science & Engineering 9,
10–20. doi:http://dx.doi.org/10.1109/MCSE.2007.58
P´erez, F. and Granger, B. E. (2007).
IPython: a system for interactive scientific computing.
Computing in Science and Engineering 9, 21–29. doi:10.1109/MCSE.2007.53
Pozzorini, C., Mensi, S., Hagens, O., Naud, R., Koch, C., and Gerstner, W. (2015). Automated
high-throughput characterization of single neurons by means of simplified spiking models. PLoS
Comput Biol 11, e1004275. doi:10.1371/journal.pcbi.1004275
Puppet Labs (2005). Puppet. http://puppetlabs.com/. [Online; accessed 16-February-2016]
Raikov, I., Cannon, R., Clewley, R., Cornelis, H., Davison, A., De Schutter, E., et al. (2011).
NineML: the network interchange for neuroscience modeling language. BMC Neuroscience 12,
P330–P330. doi:10.1186/1471-2202-12-S1-P330
Ramaswamy, S., Courcol, J.-D., Abdellah, M., Adaszewski, S. R., Antille, N., Arsever, S., et al.
(2015). The neocortical microcircuit collaboration portal: a resource for rat somatosensory cortex.
Frontiers in neural circuits 9
Ray, S. and Bhalla, U. S. (2008). PyMOOSE: interoperable scripting in python for moose. Frontiers
in Neuroinformatics 2. doi:10.3389/neuro.11.006.2008
Red Hat, Inc. (2012). Ansible. http://www.ansible.com/. [Online; accessed 16-February-2016]
Schmucker, N. (2010). Advancing Automated Parameter Constraining on Parallel Architectures
for Neuroscientic Applications. Bachelor thesis, Osnabruck University
Schulz, D. J., Goaillard, J.-M., and Marder, E. E. (2007). Quantitative expression profiling of iden-
tified neurons reveals cell-specific constraints on highly variable levels of gene expression. Pro-
ceedings of the National Academy of Sciences 104, 13187–13191. doi:10.1073/pnas.0705827104
Sivagnanam, S., Majumdar, A., Yoshimoto, K., Astakhov, V., Bandrowski, A., Martone, M. E.,
et al. (2013). Introducing the neuroscience gateway. In IWSG (Citeseer)
Stefanou, S. S., Kastellakis, G., and Poirazi, P. (2016). Advanced Patch-Clamp Analysis for Neuro-
scientists (New York, NY: Springer New York), chap. Creating and Constraining Compartmental
Models of Neurons Using Experimental Data. 325–343. doi:10.1007/978-1-4939-3411-9 15
Svensson, C.-M., Coombes, S., and Peirce, J. W. (2012). Using evolutionary algorithms
for fitting high-dimensional models to neuronal data. Neuroinformatics 10, 199–218. doi:
10.1007/s12021-012-9140-7
Tarantola, A. (2016). Inverse problem. From MathWorld–A Wolfram Web Resource, created by
Eric W. Weisstein. [Online; accessed 26-February-2016]
Taylor, A. L., Goaillard, J.-M., and Marder, E. (2009). How multiple conductances determine
electrophysiological properties in a multicompartment model. The Journal of Neuroscience 29,
5573–5586
Van Geit, W., Achard, P., and De Schutter, E. (2007). Neurofitter: a parameter tuning package
for a wide range of electrophysiological neuron models. Frontiers in Neuroinformatics 1. doi:
10.3389/neuro.11.001.2007
29
Van Geit, W., De Schutter, E., and Achard, P. (2008). Automated neuron model optimization
techniques: a review. Biol Cybern 99, 241–251
Vanier, M. C. and Bower, J. M. (1999). A comparative survey of automated parameter-search
methods for compartmental neural models. J Comput Neurosci 7, 149–171
Wils, S. and De Schutter, E. (2009). STEPS: modeling and simulating complex reaction-diffusion
systems with python. Frontiers in Neuroinformatics 3. doi:10.3389/neuro.11.015.2009
Zaytsev, Y. V. and Morrison, A. (2014). CyNEST: a maintainable Cython-based interface for the
NEST simulator. Frontiers in Neuroinformatics 8. doi:10.3389/fninf.2014.00023
ZeroMQ Project (2007). ZeroMQ. http://zeromq.org. [Online; accessed 16-February-2016]
Zitzler, E. and Kunzli, S. (2004). Indicator-based selection in multiobjective search. In Parallel
Problem Solving from Nature (PPSN VIII), eds. X. Yao et al. (Berlin, Germany: Springer-Verlag),
832–842
30
|
1206.4812 | 2 | 1206 | 2013-06-11T16:11:01 | A biological gradient descent for prediction through a combination of STDP and homeostatic plasticity | [
"q-bio.NC",
"cs.NE",
"math.DS"
] | Identifying, formalizing and combining biological mechanisms which implement known brain functions, such as prediction, is a main aspect of current research in theoretical neuroscience. In this letter, the mechanisms of Spike Timing Dependent Plasticity (STDP) and homeostatic plasticity, combined in an original mathematical formalism, are shown to shape recurrent neural networks into predictors. Following a rigorous mathematical treatment, we prove that they implement the online gradient descent of a distance between the network activity and its stimuli. The convergence to an equilibrium, where the network can spontaneously reproduce or predict its stimuli, does not suffer from bifurcation issues usually encountered in learning in recurrent neural networks. | q-bio.NC | q-bio |
1
A biological gradient descent for prediction
through a combination of STDP and
homeostatic plasticity.
Mathieu N. Galtier1, Gilles Wainrib2
1 School of Engineering and Science, Jacobs University Bremen
gGmbH, College Ring 1, 28759 Bremen, Germany
2 Laboratoire Analyse G´eom´etrie et Applications, Universit´e
Paris 13, 99 avenue Jean-Baptiste Cl´ement, Villetaneuse, France
Abstract
Identifying, formalizing and combining biological mechanisms which imple-
ment known brain functions, such as prediction, is a main aspect of current
research in theoretical neuroscience. In this letter, the mechanisms of Spike
Timing Dependent Plasticity (STDP) and homeostatic plasticity, combined
in an original mathematical formalism, are shown to shape recurrent neural
networks into predictors. Following a rigorous mathematical treatment, we
prove that they implement the online gradient descent of a distance between
the network activity and its stimuli. The convergence to an equilibrium,
where the network can spontaneously reproduce or predict its stimuli, does
not suffer from bifurcation issues usually encountered in learning in recurrent
neural networks.
2
1
Introduction
One of the main functions of the brain is prediction [Bar, 2009]. This func-
tion is generally thought to rely on the idea that cortical regions learn a
model of the world and simulate it to generate predictions of future events
[Gilbert and Wilson, 2007, Schacter et al., 2008]. Several recent experimen-
tal findings support this view, showing in particular that the spontaneous
neuronal activity after presentation of a stimulus is correlated with the evoked
activity [Kenet et al., 2003], and that this similarity increases along develop-
ment and learning [Berkes et al., 2011]. Moreover, at the scale of neuronal
networks, prediction can also be seen as a general organization principle : it
has been argued [Rao et al., 1999, Clark, 2012] that the brain would contain
a hierarchy of predictive units which are able to predict their direct stimuli
or inputs, through the modification of the synaptic connections.
Understanding the mechanisms and principles underlying this prediction
function is a key challenge, not only from a neuroscience perspective but also
for machine learning where the number of applications requiring prediction
is significant.
In the field of machine learning,
recurrent neural networks have
been successfully proposed as candidates
for
these predictive units
[Williams and Zipser, 1989, Williams and Zipser, 1995, Pearlmutter, 1995,
Jaeger and Haas, 2004, Sussillo and Abbott, 2009]. In most cases, these al-
gorithms aim at creating a neural network that autonomously and sponta-
3
neously reproduces a given time series. The Bayesian approach is also use-
ful in designing predictors [Dayan et al., 1995, George and Hawkins, 2009]
and has also been mapped to neural networks [Deneve, 2008, Friston, 2010,
Bitzer and Kiebel, 2012]. However, apart from a rough conceptual equiva-
lence, this paper is devoid of Bayesian terminology and directly focuses on
neural networks.
In this framework, prediction is often achieved by mini-
mizing a distance between the activity of the neural network and the target
time series. Although neural networks were originally studied in a feedfor-
ward framework [Rosenblatt, 1958], the most efficient networks for prediction
shall involve recurrent connections giving the network some memory prop-
erties. So called gradient descent algorithms in recurrent neural network
[Mandic and Chambers, 2001] involve the learning of the entire connectiv-
ity matrix. They minimize the distance between a target trajectory and
the trajectory of the network. On the other hand, researchers in the field of
reservoir computing [Lukosevicius and Jaeger, 2009] only optimize some con-
nections in the network whereas the others are randomly drawn and fixed.
To do prediction, they minimize the "one-step ahead" error corresponding
to the distance between the network predictions and the next time step of
the target time series. Thus, these algorithms are derived to optimize an ac-
curacy criterion, with learning rules generally favoring prediction efficiency
over biological meaning.
In the field of neuroscience, these last years have seen many discover-
ies in the study of synaptic plasticity, in particular providing experimental
4
evidences and possible mechanisms for two major concepts in the current
biology of learning.
The first is the discovery of spike-timing dependent plasticity (STDP)
[Markram et al., 1997,
Bi and Poo, 1998,
Caporale and Dan, 2008,
Sjostrom and Gerstner, 2010].
It is a temporally asymmetric form of
Hebbian learning induced by temporal correlations between the spikes
of pre- and post-synaptic neurons. The general principle is that if a
neuron fires before (resp. after) another then the strength of the connec-
tion from the former to the latter will be increased (resp.
decreased).
The summation of all
these modification leads
to the strengthen-
ing of causality links between neurons. Although STDP is originally
based on spiking network,
it has several extensions or analogs for rate-
based networks (those used in machine learning) [Kempter et al., 1999,
Izhikevich and Desai, 2003, Pfister and Gerstner, 2006]. The functional role
of STDP is still discussed, for instance: reducing latency [Song et al., 2000],
optimizing information transfer [Hennequin et al., 2010],
invariant recog-
nition [Sprekeler et al., 2007]
and even learning
temporal patterns
[Gerstner et al., 1993,
Rao and Sejnowski, 2001,
Yoshioka et al., 2007]
(non exhaustive list).
Second,
the
notion
of
homeostatic
plasticity
[Miller, 1996,
Abbott and Nelson, 2000, Turrigiano and Nelson, 2004],
including mecha-
nisms such as synaptic scaling, has proved to be important to moderate the
growth of connection strength. In contrast to previously theory-motivated
5
normalization of the connectivity [Miller and MacKay, 1994, Oja, 1982],
there is a need of a biologically plausible means to prevent the connectivity
from exploding under the influence of shaping mechanisms like Hebbian
learning or STDP.
From a theoretical viewpoint, STDP and homeostatic plasticity are al-
most always studied independently. An extensive bottom-up numerical
analysis of the combination of such learning mechanisms, done by Tri-
esch and colleagues, has already lead to biologically relevant behaviors
[Lazar et al., 2007, Lazar et al., 2009, Zheng et al., 2013]. However, the
mathematical understanding of their combination in terms of functionality
still stands as an undocumented challenge to researchers.
This letter aims at bridging the gap between biological mechanisms and
machine learning regarding the issue of predictive neural networks. We rig-
orously show how a biologically-inspired learning rule, made of an original
combination of STDP mechanism and homeostatic plasticity, mimics the gra-
dient descent of a distance between the activity of the neural network and its
direct stimuli. This results in capturing the underlying dynamical behavior
of the stimuli into a recurrent neural network and therefore in designing a
biologically plausible predictive network.
The letter is organized as follows. In section 2, we construct a theoret-
ical learning rule designed for prediction, based on an appropriate gradient
descent method. Then, in section 3, we introduce a biologically inspired
learning rule, combining the concepts of STDP and homeostatic plasticity,
6
whose purpose is to mimic the theoretical learning rule. We discuss the var-
ious biological mechanisms which may be involved in this new learning rule.
Finally in section 4, we provide a mathematical justification of the link be-
tween the theoretical and the biologically inspired learning rule, based on the
key idea that STDP can be seen as a differential operator.
2 Theoretical learning rule for prediction
In a machine learning approach, we introduce here a procedure to design a
neural network which autonomously replays a target time series.
2.1 Set-up
We consider a recurrent neural network made of n neurons which is exposed
to a time dependent input u(t) ∈ Rn of the same dimension. Our aim is
to construct a learning rule which will enable the network to reproduce the
input's behavior.
Our approach is focused on learning the underlying dynamics of the input.
Therefore, we assume that u is generated by an arbitrary dynamical system:
u = ξ(u)
(1)
with ξ a smooth vector field from Rn to Rn. We also assume that the trajec-
tory of the inputs or stimuli is τ -periodic. The key mathematical assumptions
7
on the input u is in fact ergodicity, but we restrict our study to periodic in-
puts for simplicity. In particular, periodic inputs can be constructed by the
repetition of a given finite-time sample.
Although the following method virtually works with any network equa-
tions, we focus on a neural network composed of n neurons and governed
by
v = −lv + W.S(v)
(2)
where v ∈ Rn is a vector representing neuronal activity, W ∈ Rn×n is the
connectivity matrix, l ∈ R+ is a decay constant and S is an entry-wise
sigmoid function.
2.2 Gradient descent learning rule
The idea behind our learning rule is to find the best connectivity matrix W
which will minimize a distance between the two functions ξ(u) and −lu +
W.S(u). In this perspective, we define the following quantity:
Hu :=
1
2 Z τ
0 (cid:13)(cid:13)(cid:13) − lu(t) + W.S(cid:0)u(t)(cid:1) − ξ(cid:0)u(t)(cid:1)(cid:13)(cid:13)(cid:13)
2
dt
(3)
When Hu = 0, the vector fields of systems (1) and (2) are equal on the
trajectories of the inputs. This quantity may be viewed as a distance between
the two vector fields defining the dynamics of the inputs and of the neuronal
network along the trajectories of the inputs. One shall notice that it is similar
to classical gradient methods [Pearlmutter, 1995, Williams and Zipser, 1995,
8
Mandic and Chambers, 2001], except that the norms are applied to the flows
of inputs and neural network instead of their activity. Thus, it focuses more
specifically on the dynamical structure of the inputs. Moreover, it is possible
to show, using Girsanov's Theorem, that this definition coincides with the
concept of relative entropy between two diffusion processes, namely the ones
obtained by adding a standard Gaussian perturbation to both equations.
Therefore, we will call Hu the relative entropy.
In order to capture the dynamics of the inputs into the network, it is
natural to look for a learning rule minimizing this quantity. To this end, we
consider the gradient of this measure with respect to the connectivity matrix:
∇WHu = −hξ(u) · S(u)′ + lu · S(u)′ − W.S(u) · S(u)′i
(4)
where the component i, j of x.y′ ∈ Rn×n is {x · y′}ij = R τ
0 xi(s)yj(s)ds for
x, y functions from [0, τ [ to Rn. Equivalently, these functions can be seen as
semi continuous matrices in Rn×[0,τ ] and x′ is the transpose of x.
Thus, an algorithm implementing the gradient descent
W = −∇WHu is a
good candidate to minimize the relative entropy between inputs and spon-
taneous activity. Since Hu is quadratic in W, it follows that W → Hu is a
convex function, thus excluding situations with multiple local minima. More-
over, if S(u) · S(u)′ is invertible, one can compute directly the minimizing
connectivity as
W∗ = (cid:2)ξ(u) · S(u)′ + lu · S(u)′(cid:3).(cid:2)S(u) · S(u)′(cid:3)−1
9
(5)
Implementing a gradient descent based on equation (4) does not imme-
diately lead to a biologically relevant mechanism. First, it requires a direct
access to the inputs u, whereas synaptic plasticity mechanisms shall only rely
on the network activity v. Second, it is a batch learning algorithm which
requires an access to the entire history of the inputs. Third, it requires the
ability to compute ξ(u). Therefore, we will see in section 3 how to overcome
these issues combining biologically inspired synaptic plasticity mechanisms.
2.3 Example
1
0.8
0.6
0.4
0.2
0
Figure 1: Learning how to write the letter A. Time evolution of the input
movie (top row) and of the network activity after learning (bottom row).
Each pixel corresponds to a neuron. Parameters: Number of neurons n =
400; l = 1; S(x) = tanh(x).
10
In order to illustrate the idea that learning rule (4) enables the network
to learn dynamical features of the input, we have constructed the following
experiment. We present to the network an input movie displaying sequen-
tially the writing of the letter A (Figure 1 - top row). To each pixel we assign
one neuron, so that the input and the network share the same dimension.
This input movie is repeated periodically until the connectivity matrix of
the network, evolving under rule (4), stabilizes. Then the input is turned
off and we set the initial state of the network to a priming image showing
the bottom left part of letter A. The network evolving without input strik-
ingly reproduces the dynamical writing of letter A as displayed in Figure 1.
Thus, with this example we have illustrated the ability of the learning rule
we have derived from a theoretical principle to capture a dynamic input into
a connectivity matrix.
3 A biological learning rule
We now introduce a biological learning rule made of the combination of STDP
and homeostatic plasticity. Later in section 4, we show that this learning rule
minimizes Hu. Here, we first give a mathematical description of this learning
rule and, second, relate the different terms to biological mechanisms.
11
3.1 Mathematical description
Learning corresponds to a modification of the connectivity simultaneous to
the network activity evolution. The result is a coupled system of equations.
The learning rate ǫ is chosen to be small: ǫ ≪ 1, so that learning can be
considered slow compared to the evolution of the activity. The full online
learning system is
1
ǫ
v = −lv + W.S(v) + u
Wij = δ[¯vi, S(¯vj)]
}
STDP
{z
− Xk
WikS(¯vk)S(¯vj)
homeostatic plasticity
{z
}
where
and
¯v = lv − W.S(v) ∗ gl
δ[x, y] =
γ + l
2
x(y ∗ gγ) −
γ − l
2
(x ∗ gγ)y
(6)
(7)
where the notation ∗ denotes the convolution operator. The function gc
is defined as gc : t 7→ ce−ctH(t) with H the Heaviside function and for any
positive number c as shown in the left picture of Figure 2. As illustrated in the
right picture of Figure 2, the operator δ roughly corresponds to the classical
STDP window [Bi and Poo, 1998] (taking into account a y-axis symmetry
corresponding to the symmetric formalism we are using). The constant γ ∈
R+ is a time constant corresponding to the width of the STDP window used
for learning.
12
Figure 2: (left) Plots of the function gc for c ∈ {1/2, 1, 2}. (right) Plot of the
function ∆γ(t) + lΣγ(t) = γ+l
2 gγ(t) − γ−l
2 gγ(−t) for γ = 0.5 and l = 0.2. This
function corresponds to the operator δ as shown in section 4.1.2.
3.2 Biological mechanisms
3.2.1 An input estimate
The variable ¯v can be seen as a spatio-temporal differential variable which
approximates the inputs u. Although unsupervised learning rules are often
algebraic combinations of element-wise functions applied to the activity of
the network [Gerstner and Kistler, 2002], it is not precisely the case here.
Indeed, learning is based on the variable ¯v which corresponds to the sub-
traction of the temporally integrated synaptic drive W.S(v) ∗ gl from the
activity of the neurons lv. For each neuron, this variable takes into ac-
13
count the past of all the neurons which are then spatially averaged through
the connectivity to be subtracted from the current activity. This gives
a differential flavor to this variable which is reminiscent of former learn-
ing rules [Bienenstock et al., 1982, Sejnowski, 1977] for the temporal aspect
and [Miller and MacKay, 1994] for the spatial aspect. Note that this variable
is not strictly speaking local (i.e. the connection Wij needs the values of vk
to be updated), yet it is biologically plausible since the term (cid:0)W.S(v)(cid:1)i is
accessible for neuron i on its dendritic tree, which is a form of locality in a
broader sense.
3.2.2 An STDP mechanism
The first term δ[¯vi, S(¯vj)] in (6) can be related to STDP. The antisym-
metric part of this term is responsible for retrieving the drift ξ in equa-
tion (4). The symmetric part (corresponding to Hebbian learning) is re-
sponsible for retrieving the second term in (4). Thus,
it captures the
causality structure of the inputs which is a task generally attributed to
STDP [Sjostrom and Gerstner, 2010]. Beyond the simple similarity of func-
tional role, we believe a simplification of this term may shed light on the
deep link it has with STDP. The main difference between our setup and
STDP is that the former is based on a rate-based dynamics, whereas the
latter is based on a spiking dynamics.
In a pure spike framework, i.e the
activity is a sum of Diracs, the STDP can be seen as this simple learning
rule
Wij = δ[vi, vj] = γ+l
2 vi(vj ∗ gγ) − γ−l
2 (vi ∗ gγ) vj.
Indeed, the term
14
vi(vj ∗ gγ) is non-null only when the post-synaptic neuron i is firing and
then, via the factor vj ∗ gγ, it counts the number of preceding pre-synaptic
spikes that might have caused i's excitation and weight them by the de-
creasing exponential gγ. Thus, this term exactly accounts for the positive
part of the STDP curve. The negative term −(vi ∗ gγ)vj takes the opposite
perspective and accounts for the negative part of the STDP curve. A loose
extension of this rule to the case where the activity is smoothly evolving
leads to identifying the function δ to the STDP mechanism for rate-based
networks [Galtier, 2012, Izhikevich and Desai, 2003].
3.2.3 Homeostatic plasticity
The second term −Pk WikS(¯vk)S(¯vj) in (6) accounts for what is usu-
ally presented as homeostatic plasticity mechanisms. The previous STDP
term seems to be a powerful mechanism to shape the response of the net-
work. However, there is a need of a regulatory process to prevent from
uncontrolled growth of the network connectivity [Abbott and Nelson, 2000,
Turrigiano and Nelson, 2004, Miller, 1996]. It has been argued that STDP
could be self regulatory [Van Rossum et al., 2000, Song et al., 2000], but it
is not the case in our framework and an explicit balancing mechanism is
necessary to avoid the divergence of the system. This last term is the
only one with a negative sign and is multiplicative with respect to the
connectivity. Thus, according to [Abbott and Nelson, 2000], it is a reason-
able candidate for homeostasis. It has been argued [Turrigiano et al., 1998,
15
Kim et al., 2012] that homeostatic plasticity might keep the relative synap-
tic weights by dividing the connectivity with a common scaling fac-
tor, theoretically preventing from a possible information loss.
In con-
trast to these ad hoc re-normalizations often introduced in other learning
rules [Miller and MacKay, 1994, Oja, 1982], our relative entropy minimizing
learning rule thus introduces naturally an original form of homeostatic plas-
ticity.
Although we have separated the description of the various terms in (6),
our approach suggests that homeostasis may be seen, not necessarily as a
scaling term, but as a constitutive part of a learning principle, deeply entan-
gled [Turrigiano, 1999] with the other learning mechanisms.
3.3 Numerical application
Although the focus of the paper is on theory, we introduce a simple numer-
ical example to illustrate the predictive properties of the biological learn-
ing rule. More precisely, we investigate the question of retrieving the con-
nectivity of a neural network based on the observation of the time se-
ries of its activity. This is an inverse problem which is a usual challeng-
ing topic in computational neuroscience [Friston et al., 2003, Gal´an, 2008,
Potthast and beim Graben, 2009] since it may give access to large scale ef-
fective connectivities simply from the observation of a neuronal activity.
Here we address it in an elementary framework. The network generating
the activity patterns is referred as input network and evolves according to
16
u = −l0u + W0.S(u). For this example, the network is made of n = 3
neurons and its connectivity W0 is shown in Figure 3.a). These parameters
were chosen so that the activity is periodic as shown by the dashed curves in
Figure 3.c). Then, we simulate the entire system (6) with a decay constant
lnet and observe that its connectivity Wnet converges to W0.
As shown in the next section, it is necessary u ∗ gl ∗ gγ ≃ u in order to
approximate accurately the input's activity with the online learning rule (6).
Given that the time constant of the inputs is governed by l0, the previous
approximation holds only if γ, lnet ≫ l0. If this assumption is broken, e.g.
lnet = l0, then the final connectivity matrix is different from W0, see the
black dot-dashed curve in Figure 3.b). Indeed, in this case, the network only
learns to replay the slow variation of the inputs.
A method to recover the precise time course of the inputs consists in
artificially changing the time constants at different steps of an algorithm
described in the following. First, simulate the network equation (top equation
in (6)) with a constant lnet ≫ l0. Yet, the time constant in the learning
equation (bottom equation in (6)) is to be kept at l0, thus introducing a
hybrid model. In this framework, the connectivity converges exactly to W0
as shown in the grey dashed curve in Figure 3.b). After learning, simulate
the network equation with the learned connectivity and with a time constant
switched back to l0. This gives an activity as shown in the plain curves in
Figure 3.c).
17
4 Link between theoretical and biological
learning rules
In this part, we show how the biological learning rule (6) implements the
gradient descent based on equation (4). First, we introduce three techni-
cal results which are the mathematical cornerstones of the paper and then
combine them to obtain the desired result.
4.1 Three technical results
As mentioned previously, equation (4) is not biologically plausible for three
main reasons: (i) it requires a direct access to the inputs u, whereas synaptic
plasticity mechanisms shall only rely on the network activity v; (ii) it is a
batch learning algorithm which requires an access to the entire history of
the inputs; (iii) it requires the ability to compute ξ(u) (equal to the time-
derivative of the inputs according to equation (1)).
On the contrary, the biological learning rule (6) does not have these prob-
lems. First, it uses an estimate of the inputs, noted ¯v, which is based on
the activity variable only. Second, it is an online learning rule, i.e. it takes
input on the fly, and relies on a slow-fast mechanism to temporally average
the variables. Third, it approximates the computation of ξ(u) with a func-
tion δ inspired from STDP convolution. In the following, we mathematically
address these three points successively.
18
4.1.1
¯v is an input estimate
The online learning rule (6) is expressed by means of the activity of the neural
network v governed by the top equation in (6). However, to be comparable
to the gradient of the relative entropy (4), we first need to make explicit the
dependency on the inputs u. Therefore, we show that the network dynamics
induces a simple relation between ¯v and the inputs u: a simple computation
in the Fourier domain shows that the convolution between the temporal
operator d
dt + lId and gl results in lId. Applying this result to the neural
dynamics leads to reformulating the top equation in (6) as lv−W·S(v)∗gl =
u ∗ gl. We recognize the definition of the variable ¯v (which was originally
defined according to this relation) such that the network's dynamics of the
fast equation in (6) is equivalent to
¯v = u ∗ gl
(8)
4.1.2 Temporal averaging of the learning rule
To prove that (6) implements the gradient descent of the relative entropy,
we need to use a time-scale separation assumption, enabling the input to
reveal its dynamical structure through ergodicity.
Indeed, the fact that
learning is very slow compared to the activity v corresponds to the as-
sumption ǫ ≪ 1.
In this case, we can apply classical results of periodic
averaging [Sanders and Verhulst, 1985] (see also [Galtier and Wainrib, 2012]
in the context of neural networks) to show that the evolution of W is well-
approximated by
W = (cid:2)¯v ∗ ∆γ + l¯v ∗ Σγ − W.S(¯v)(cid:3) · S(¯v)′
2(cid:0)gγ(−t) − gγ(t)(cid:1) and Σγ : t 7→ gγ(−t)+gγ (t)
2
of Figure 2 shows a plot of the function ∆γ + lΣγ.
19
(9)
where ∆γ : t 7→ γ
. The right picture
4.1.3 STDP as a differential operator
Here, we prove the two following equalities
[x ∗ Σγ] · y′ = (x ∗ gγ) · (y ∗ gγ)′
and
[x ∗ ∆γ] · y′ = ( x ∗ gγ) · (y ∗ gγ)′
(10)
(11)
This second equation is the key mathematical mechanism that makes STDP
a good approximation of the temporal derivative of the inputs.
Both proofs consist in going in the Fourier domain, where convolutions
are turned into multiplications. We use the unitary, ordinary frequency con-
vention for the Fourier transform. Observe that the Fourier transform of gγ
is gγ(ξ) = γ
γ+2iπξ . And we define1 g′
γ : t 7→ gγ(−t) such that ∆γ = γ
2 (g′
γ − gγ)
and g′
γ(ξ) = gγ(−ξ).
1This notation is motivated by the following: the convolution with gγ can be written
as a matrix multiplication with a continuous Toeplitz matrix RR×R whose component ts
is gγ(t − s). In this framework, the convolution with g ′
γ corresponds to the multiplication
1. Let us show that [x ∗ Σγ] · y′ = (x ∗ gγ) · (y ∗ gγ)′
We proceed in two steps:
20
(a) Let us show that gγ +g′
γ
2 = gγ ∗ g′
γ
The Fourier transform of the convolution gγ ∗ g′
γ is the product
γ 2
γ
γ+2iπξ
γ
γ−2iπξ =
that the right hand side is the Fourier transform of t 7→ γ
γ 2+4π2ξ2 . From the usual Fourier tables we observe
2 e−γt.
This immediately leads to the result.
(b) Let us show that (x ∗ g′
γ) · y′ = x · (y ∗ gγ)′
Compute
−∞(cid:0)xi∗g′
γ(cid:1)(s)yj(s)ds = Z ∞
−∞Z +∞
−∞
xi(t)g′
γ(s−t)yj(s)dsdt
{(x∗g′
γ)·y′}ij = Z ∞
= Z ∞
−∞Z +∞
−∞
xi(t)gγ(t − s)yj(s)dsdt = {x · (y ∗ gγ)′}ij
Using the result (a) and applying result (b) to x ∗ gγ and y leads to the
result [x ∗ Σγ] · y′ = (x ∗ gγ) · (y ∗ gγ)′.
2. To prove equation (11), let us first show that x ∗ ∆γ = x ∗ gγ +g′
γ
2
.
The Fourier transform of the convolution x ∗ ∆γ is the product x ∆γ.
with the transpose of the previous Toeplitz operator.
Besides,
2
γ
∆γ = \g′
γ − gγ(ξ) =
γ
γ
−
γ − 2iπξ
γ
=
2iπξ
γ (cid:16)
= 2iπξ
γ + 2iπξ
γ
+
γ − 2iπξ(cid:17) =
γ + 2iπξ
21
2γ
γ2 + 4π2ξ2
2iπξ
γ (cid:0) \g′
γ + gγ(ξ)(cid:1)
Because dx
[g′
γ+gγ
x 2iπξ
2
dt (ξ) = x 2iπξ, taking the inverse Fourier transform of
(ξ) gives the intermediary result x ∗ ∆γ = x ∗ gγ+g′
γ
2
.
Using it and applying the first result to x and y leads to equation (11).
4.2 Main result
We prove here that the biological learning rule (6) implements the gradient
descent based on (4).
The temporally averaged version of the biological learning rule is given by
(9). We simultaneously use equations (8) and (11) on this formula to show
that the solution of the biological learning rule is well approximated by
W = (cid:2)ξ(u ∗ gl) ∗ gγ(cid:3) ·(cid:2)S(u ∗ gl) ∗ gγ(cid:3)′ + l(u ∗ gl ∗ gγ) ·(cid:0)S(u ∗ gl) ∗ gγ(cid:1)′
− W · S(u ∗ gl) · S(u ∗ gl)′
If u is slow enough, i.e. u ∗ gl ∗ gγ ≃ u, then this equation is precisely
the gradient descent of Hu. If u is too fast, then the network will only learn
to predict the slow variations of the inputs. Some fast-varying information
22
is lost in the averaging process, mainly because the network equation in (6)
acts as a relaxation equation which filters the activity. Note that this loss
of information is not necessarily a problem for the brain, since it may be ex-
tracting and treating information at different time scales [Kiebel et al., 2008].
Actually, the choice of the parameter l specifies the time scale under which
the inputs are observed.
4.3 Stability
Both theoretical and biological learning rules are assured to make the con-
nectivity converge to an equilibrium point whatever the initial condition. In
particular, this method does not suffer from the bifurcation issues often en-
countered in recurrent neural network learning [Doya, 1992]. In most cases,
problems arise when learning leads the network activity to a bifurcation. The
bifurcation suddenly changes the value of the quantity being minimized and
this intricate coupling leads to very slow convergence rates. The reason why
this method has such an unproblematic converging property is because the
relative entropy (as opposed to other energy functions traditionally used)
is independent from the network activity v. Besides, for any network pat-
tern v equation (8) ensures that ¯v is always equal to the convolved inputs.
These two arguments break the pathologic coupling preventing from getting
interesting convergence results.
From a mathematical perspective, the Krasovskii-Lasalle invariance prin-
ciple [Khalil and Grizzle, 1992] ensures that the theoretical
learning rule
23
W = −∇WHu converges to a equilibrium point.
Indeed, the relative en-
tropy acts as an energy or Lyapunov function.
If S(u).S(u)′ is invertible
then the equilibrium point is unique and defined by equation (5). If it is not
invertible then the equilibrium depends on the initial condition.
If the inputs are slow enough, it has been shown that the biological learning
is well approximated by the theoretical gradient descent. Therefore, the
former exhibits the same stable converging behavior as the latter. Thus, the
biological learning rule (6) is stable. In practice, we have never encountered
a diverging situation even when the inputs are not slow enough.
5 Discussion and conclusion
We have shown that a biological learning rule shapes a network into a pre-
dictor of its stimuli. This learning rule is made of a combination of a STDP
mechanism and homeostatic plasticity. After learning the network would
spontaneously predict and replay the stimuli it has previously been exposed
to. This was achieved by showing that this learning rule minimizes a quan-
tity analogous to the relative entropy (or Kullback-Leibler divergence) be-
tween the stimuli and the network activity as for other well-known algorithms
[Ackley et al., 1985, Hinton, 2009].
We believe this letter brings interesting arguments in the debate concern-
ing the functional role of STDP. We have shown that the antisymmetric part
of STDP can be seen as a differential operator. When its effect is moder-
24
ated by an appropriate scaling term, we argue that it could correspond to a
generic mechanism shaping neural networks into predictive units. This idea
is not new, but this letter may give it a stronger theoretical basis.
This study also gives central importance to the time constants character-
izing the network. Indeed, the fact the biological learning rule (6) implements
a gradient descent is rigorously true for slow inputs. Inputs are slow if they
are significantly slower than both the time window of the STDP and the de-
cay constant of the network. This means that sub-networks in the brain could
select the frequency of the information they are predicting. Given that the
brain may process information at different time scales [Kiebel et al., 2008],
this may be an interesting feature. Besides, note that the proposed biological
learning rule (6) is partly characterized by the activity decay parameter l.
This is surprising because it links the intrinsic dynamics of the neurons to
the learning processes corresponding to the synapse. Therefore, it may pro-
vide a direction to experimentally check this theory: is the symmetric part
of STDP (i.e. Hebbian learning) stronger between faster neurons?
One of the characteristic features of this research is the combination of
rate-based networks and the concept of spike timing dependent plasticity.
Obviously, this made impossible to consider the rigorous definition of STDP.
However, we have argued that the function δ in (7) is an alternative for-
mulation which is equivalent to the others in the spiking framework and
which can be trivially extended to analog networks. Besides, it does capture
the functional mechanism of the STDP in the rate-based framework: when
25
the activation of a neuron precedes (resp. follows) the activation of another
the strength of the synapse from the former to the latter is increased (resp.
decreased). Finally, we believe the theory presented in this paper gives sup-
port to this rate-based STDP since it appears to fill a gap between machine
learning and biology by implementing a differential operator.
This approach can be applied to other forms of network equations than
(2) such as the Wilson-Cowan or Kuramoto models, leading to different learn-
ing rules. In this perspective, learning can be seen as a projection of a given
arbitrary dynamical system to a versatile neuronal network model, and the
learning rule will depend on the chosen model. However, we shall remark that
any network equation with an additive structure -- intrinsic dynamics + com-
munication with other neurons -- as in (2) will lead to a similar structure for
the learning rule, with terms that may share similar biological interpretations
as developed above. A special case is the linear network v = −lv + W.v for
which various statistical methods to estimate the connectivity matrix have
been applied e.g.
in climate modeling [Penland and Sardeshmukh, 1995],
gene regulatory networks [Yeung et al., 2002] and spontaneous neuronal ac-
tivity [Gal´an, 2008]. In this simpler case, the biological learning rule in equa-
tion (6) would be 1
ǫ
Wij = δ[¯vi, ¯vj]−Pk Wik ¯vk ¯vj with ¯v = lv−W.v∗gl. The
method developed in this article can be used to extend the inverse modeling
approach previously developed in the linear case to models with non-linear
interactions.
One of the main restrictions of this approach is that the dimensionality
26
of the stimuli and that of the network have to be identical. Therefore, as
such, the accuracy of this biological mechanism does not match the state of
the art machine learning algorithms.This is not necessarily a big issue since
a high dimensional projections of the inputs may be used as preprocessing.
From a biological viewpoint and taking the example of vision, this would
correspond to the retino-cortical pathway which is not one to one and very
redundant. But this limitation also puts forward the necessity to study a
network containing hidden neurons in a similar fashion. Ongoing research
is revealing that the mathematical formalism is well suited to extend this
approach to the field of reservoir computing [Jaeger and Haas, 2004].
Acknowledgments
The authors thank Herbert Jaeger for his helpful comments on the
manuscript.
MNG was partially funded by the Amarsi project (FP7-ICT #24833),
the ERC advanced grant NerVi #227747, the r´egion PACA, France and the
IP project BrainScaleS #269921.
References
[Abbott and Nelson, 2000] Abbott, L. and Nelson, S. (2000). Synaptic plas-
ticity: taming the beast. Nature neuroscience, 3:1178 -- 1183.
27
[Ackley et al., 1985] Ackley, D., Hinton, G., and Sejnowski, T. (1985). A
learning algorithm for boltzmann machines. Cognitive science, 9(1):147 --
169.
[Bar, 2009] Bar, M. (2009). Predictions: a universal principle in the opera-
tion of the human brain. Philosophical Transactions of the Royal Society
B: Biological Sciences, 364(1521):1181 -- 1182.
[Berkes et al., 2011] Berkes, P., Orb´an, G., Lengyel, M., and Fiser, J. (2011).
Spontaneous cortical activity reveals hallmarks of an optimal internal
model of the environment. Science, 331(6013):83.
[Bi and Poo, 1998] Bi, G. and Poo, M. (1998).
Synaptic modifications
in cultured hippocampal neurons: dependence on spike timing, synap-
tic strength, and postsynaptic cell type. The Journal of Neuroscience,
18(24):10464 -- 10472.
[Bienenstock et al., 1982] Bienenstock, E., Cooper, L., and Munro, P.
(1982). Theory for the development of neuron selectivity: orientation
specificity and binocular interaction in visual cortex. The Journal of
Neuroscience, 2(1):32 -- 48.
[Bitzer and Kiebel, 2012] Bitzer, S. and Kiebel, S. (2012). Recognizing re-
current neural networks (rrnn): Bayesian inference for recurrent neural
networks. Biological cybernetics, pages 1 -- 17.
28
[Caporale and Dan, 2008] Caporale, N. and Dan, Y. (2008). Spike timing-
dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci.,
31:25 -- 46.
[Clark, 2012] Clark, A. (2012). Whatever next? predictive brains, situated
agents, and the future of cognitive science. Behav. Brain Sci.
[Dayan et al., 1995] Dayan, P., Hinton, G., Neal, R., and Zemel, R. (1995).
The helmholtz machine. Neural computation, 7(5):889 -- 904.
[Deneve, 2008] Deneve, S. (2008). Bayesian spiking neurons i:
inference.
Neural computation, 20(1):91 -- 117.
[Doya, 1992] Doya, K. (1992). Bifurcations in the learning of recurrent neural
networks. In Circuits and Systems, 1992. ISCAS'92. Proceedings., 1992
IEEE International Symposium on, volume 6, pages 2777 -- 2780. IEEE.
[Friston, 2010] Friston, K. (2010). The free-energy principle: a unified brain
theory? Nature Reviews Neuroscience, 11(2):127 -- 138.
[Friston et al., 2003] Friston, K., Harrison, L., and Penny, W. (2003). Dy-
namic causal modelling. Neuroimage, 19(4):1273 -- 1302.
[Gal´an, 2008] Gal´an, R. (2008). On how network architecture determines the
dominant patterns of spontaneous neural activity. PLoS One, 3(5):e2148.
29
[Galtier, 2012] Galtier, M. (2012). A mathematical approach to unsupervised
learning in recurrent neural networks. PhD thesis, Mines Paristech / IN-
RIA.
[Galtier and Wainrib, 2012] Galtier, M. and Wainrib, G. (2012). Multiscale
analysis of slow-fast neuronal learning models with noise. The Journal of
Mathematical Neuroscience, 2(1):13.
[George and Hawkins, 2009] George, D. and Hawkins, J. (2009). Towards
a mathematical theory of cortical micro-circuits. PLoS computational
biology, 5(10):e1000532.
[Gerstner and Kistler, 2002] Gerstner, W. and Kistler, W. (2002). Spiking
neuron models: Single neurons, populations, plasticity. Cambridge Univ
Pr.
[Gerstner et al., 1993] Gerstner, W., Ritz, R., and Van Hemmen, J. (1993).
Why spikes? hebbian learning and retrieval of time-resolved excitation
patterns. Biological cybernetics, 69(5):503 -- 515.
[Gilbert and Wilson, 2007] Gilbert, D. and Wilson, T. (2007). Prospection:
experiencing the future. Science, 317(5843):1351 -- 1354.
[Hennequin et al., 2010] Hennequin, G., Gerstner, W., and Pfister, J. (2010).
Stdp in adaptive neurons gives close-to-optimal information transmission.
Frontiers in Computational Neuroscience, 4.
30
[Hinton, 2009] Hinton, G. E. (2009). Deep belief networks. Scholarpedia,
4(4):5947.
[Izhikevich and Desai, 2003] Izhikevich, E. and Desai, N. (2003). Relating
stdp to bcm. Neural Computation, 15(7):1511 -- 1523.
[Jaeger and Haas, 2004] Jaeger, H. and Haas, H. (2004). Harnessing nonlin-
earity: Predicting chaotic systems and saving energy in wireless commu-
nication. Science, 304(5667):78 -- 80.
[Kempter et al., 1999] Kempter, R., Gerstner, W., and Van Hemmen, J.
(1999). Hebbian learning and spiking neurons. Physical Review E,
59(4):4498.
[Kenet et al., 2003] Kenet, T., Bibitchkov, D., Tsodyks, M., Grinvald, A.,
and Arieli, A. (2003). Spontaneously emerging cortical representations of
visual attributes. Nature, 425(6961):954 -- 956.
[Khalil and Grizzle, 1992] Khalil, H. and Grizzle, J. (1992). Nonlinear
systems. Macmillan Publishing Company New York.
[Kiebel et al., 2008] Kiebel, S., Daunizeau, J., and Friston, K. (2008). A
hierarchy of time-scales and the brain. PLoS computational biology,
4(11):e1000209.
[Kim et al., 2012] Kim, J., Tsien, R., and Alger, B. (2012). An improved
test for detecting multiplicative homeostatic synaptic scaling. PloS one,
7(5):e37364.
31
[Lazar et al., 2007] Lazar, A., Pipa, G., and Triesch, J. (2007). Fading mem-
ory and time series prediction in recurrent networks with different forms
of plasticity. Neural Networks, 20(3):312 -- 322.
[Lazar et al., 2009] Lazar, A., Pipa, G., and Triesch, J. (2009).
Sorn:
a self-organizing recurrent neural network. Frontiers in computational
neuroscience, 3.
[Lukosevicius and Jaeger, 2009] Lukosevicius, M. and Jaeger, H. (2009).
Reservoir computing approaches to recurrent neural network training.
Computer Science Review, 3(3):127 -- 149.
[Mandic and Chambers, 2001] Mandic, D. and Chambers,
J.
(2001).
Recurrent neural networks
for prediction:
Learning algorithms,
architectures and stability. John Wiley & Sons, Inc.
[Markram et al., 1997] Markram, H., Lubke, J., Frotscher, M., and Sak-
mann, B. (1997). Regulation of synaptic efficacy by coincidence of post-
synaptic aps and epsps. Science, 275(5297):213 -- 215.
[Miller, 1996] Miller, K. (1996). Synaptic economics: Competition and co-
operation in correlation-based synaptic plasticity. Neuron, 17:371 -- 374.
[Miller and MacKay, 1994] Miller, K. and MacKay, D. (1994). The role of
constraints in hebbian learning. Neural Computation, 6(1):100 -- 126.
[Oja, 1982] Oja, E. (1982). Simplified neuron model as a principal component
analyzer. Journal of mathematical biology, 15(3):267 -- 273.
32
[Pearlmutter, 1995] Pearlmutter, B. (1995). Gradient calculations for dy-
namic recurrent neural networks: A survey. Neural Networks, IEEE
Transactions on, 6(5):1212 -- 1228.
[Penland and Sardeshmukh, 1995] Penland, C. and Sardeshmukh, P. (1995).
The optimal growth of tropical sea surface temperature anomalies. Journal
of climate, 8(8):1999 -- 2024.
[Pfister and Gerstner, 2006] Pfister, J. and Gerstner, W. (2006). Triplets of
spikes in a model of spike timing-dependent plasticity. The Journal of
neuroscience, 26(38):9673 -- 9682.
[Potthast and beim Graben, 2009] Potthast, R. and beim Graben, P. (2009).
Inverse problems in neural field theory.
SIAM Journal on Applied
Dynamical Systems, 8:1405.
[Rao et al., 1999] Rao, R., Ballard, D., et al. (1999). Predictive coding in the
visual cortex: a functional interpretation of some extra-classical receptive-
field effects. Nature neuroscience, 2:79 -- 87.
[Rao and Sejnowski, 2001] Rao, R. and Sejnowski, T. (2001). Spike-timing-
dependent hebbian plasticity as temporal difference learning. Neural
computation, 13(10):2221 -- 2237.
[Rosenblatt, 1958] Rosenblatt, F. (1958). The perceptron: a probabilistic
model for information storage and organization in the brain. Psychological
review, 65(6):386.
33
[Sanders and Verhulst, 1985] Sanders, J. and Verhulst, F. (1985). Averaging
methods in nonlinear dynamical systems, volume 59. Springer.
[Schacter et al., 2008] Schacter, D., Addis, D., and Buckner, R. (2008).
Episodic simulation of future events. Annals of the New York Academy of
Sciences, 1124(1):39 -- 60.
[Sejnowski, 1977] Sejnowski, T. (1977). Statistical constraints on synaptic
plasticity. Journal of theoretical biology, 69(2):385.
[Sjostrom and Gerstner, 2010] Sjostrom, J. and Gerstner, W. (2010). Spike-
timing dependent plasticity. 5(2):1362.
[Song et al., 2000] Song, S., Miller, K., and Abbott, L. (2000). Competi-
tive hebbian learning through spike-timing-dependent synaptic plasticity.
Nature neuroscience, 3:919 -- 926.
[Sprekeler et al., 2007] Sprekeler, H., Michaelis, C., and Wiskott, L. (2007).
Slowness: An objective for spike-timing -- dependent plasticity?
PLoS
Computational Biology, 3(6):e112.
[Sussillo and Abbott, 2009] Sussillo, D. and Abbott, L. (2009). Generat-
ing coherent patterns of activity from chaotic neural networks. Neuron,
63(4):544 -- 557.
[Turrigiano, 1999] Turrigiano, G. (1999). Homeostatic plasticity in neuronal
networks: the more things change, the more they stay the same. Trends
in neurosciences, 22(5):221 -- 227.
34
[Turrigiano et al., 1998] Turrigiano, G., Leslie, K., Desai, N., Rutherford, L.,
and Nelson, S. (1998). Activity-dependent scaling of quantal amplitude in
neocortical neurons. NATURE, 391:893.
[Turrigiano and Nelson, 2004] Turrigiano, G. and Nelson, S. (2004). Home-
ostatic plasticity in the developing nervous system. Nature Reviews
Neuroscience, 5(2):97 -- 107.
[Van Rossum et al., 2000] Van Rossum, M., Bi, G., and Turrigiano, G.
(2000). Stable hebbian learning from spike timing-dependent plasticity.
The Journal of Neuroscience, 20(23):8812 -- 8821.
[Williams and Zipser, 1989] Williams, R. and Zipser, D. (1989). A learning
algorithm for continually running fully recurrent neural networks. Neural
computation, 1(2):270 -- 280.
[Williams and Zipser, 1995] Williams, R. and Zipser, D. (1995). Gradient-
based learning algorithms for recurrent networks and their computational
complexity. Back-propagation: Theory, architectures and applications,
pages 433 -- 486.
[Yeung et al., 2002] Yeung, M., Tegn´er, J., and Collins, J. (2002). Reverse
engineering gene networks using singular value decomposition and robust
regression. Proceedings of the National Academy of Sciences, 99(9):6163.
35
[Yoshioka et al., 2007] Yoshioka, M., Scarpetta, S., and Marinaro, M. (2007).
Spatiotemporal learning in analog neural networks using spike-timing-
dependent synaptic plasticity. Physical Review E, 75(5):051917.
[Zheng et al., 2013] Zheng, P., Dimitrakakis, C., and Triesch, J. (2013). Net-
work self-organization explains the statistics and dynamics of synaptic
connection strengths in cortex. PLoS Computational Biology, 9(1).
36
time
1 -5 -2
3 5 -1
-5 -2 1
b)
r
o
r
r
e
a)
c)
2
0
y
t
i
v
i
t
c
a
0
time
100
Figure 3: Retrieving the connectivity a) Connectivity matrix W0 of the
input network. b) Evolution of the difference between current connectiv-
ity and input connectivity through learning. The black dot-dashed curve
corresponds to the online learning rule (6) in the homogeneous case, i.e.
l = lnet = l0 in both equations in (6). The grey dashed curve corresponds
to the online learning system (6) in the hybrid framework, i.e.
l = l0 for
the learning equation and l = lnet ≫ l0 for the network equation. For this
simulation we chose lnet = 50. The black plain curve (superposed to the grey
dashed curve) corresponds to the batch relative entropy minimization (4).
c) The dashed curves correspond to the activity of the inputs. It is a three
dimensional input (i.e. n = 3) and the three different grey levels correspond
to the different dimensions. The plain curves correspond to the simulation of
the network (top equation in (6)) post learning, in the hybrid framework, and
without inputs. The parameters for these simulations are l0 = 1, ǫ = 0.01
and γ = 100.
|
1105.4705 | 1 | 1105 | 2011-05-24T08:22:36 | A Tutorial in Connectome Analysis: Topological and Spatial Features of Brain Networks | [
"q-bio.NC",
"cs.SI",
"physics.soc-ph"
] | High-throughput methods for yielding the set of connections in a neural system, the connectome, are now being developed. This tutorial describes ways to analyze the topological and spatial organization of the connectome at the macroscopic level of connectivity between brain regions as well as the microscopic level of connectivity between neurons. We will describe topological features at three different levels: the local scale of individual nodes, the regional scale of sets of nodes, and the global scale of the complete set of nodes in a network. Such features can be used to characterize components of a network and to compare different networks, e.g. the connectome of patients and control subjects for clinical studies. At the global scale, different types of networks can be distinguished and we will describe Erd\"os-R\'enyi random, scale-free, small-world, modular, and hierarchical archetypes of networks. Finally, the connectome also has a spatial organization and we describe methods for analyzing wiring lengths of neural systems. As an introduction for new researchers in the field of connectome analysis, we discuss the benefits and limitations of each analysis approach. | q-bio.NC | q-bio | "!#$%&'()*!(+!,&++-.%&/-!"+)*01(12!!
#&3&*&4(.)*!)+5!63)%()*!7-)%$'-1!&8!9')(+!:-%;&'<1
Running title: Principles of Connectome Analysis
Marcus Kaiser1, 2, 3
1 School of Computing Science, Newcastle University, UK
2 Institute of Neuroscience, Newcastle University, UK
3 Department of Brain and Cognitive Sciences, Seoul National University, South Korea!
Corresponding author:
Dr Marcus Kaiser
School of Computing Science
Newcastle University
Claremont Tower
Newcastle upon Tyne, NE1 7RU
United Kingdom,
E-Mail: [email protected] Phone: +44 191 222 8161
!
Fax: +44 191 222 8232
Abstract
High-throughput methods for yielding the set of connections in a neural system, the connectome, are
now being developed. This tutorial describes ways to analyze the topological and spatial organization
of the connectome at the macroscopic level of connectivity between brain regions as well as the
microscopic level of connectivity between neurons. We will describe topological features at three
different levels: the local scale of individual nodes, the regional scale of sets of nodes, and the global
scale of the complete set of nodes in a network. Such features can be used to characterize components
of a network and to compare different networks, e.g. the connectome of patients and control subjects
for clinical studies. At the global scale, different types of networks can be distinguished and we will
describe Erdös-Rényi random, scale-free, small-world, modular, and hierarchical archetypes of
networks. Finally, the connectome also has a spatial organization and we describe methods for
analyzing wiring lengths of neural systems. As an introduction for new researchers in the field of
connectome analysis, we discuss the benefits and limitations of each analysis approach.
Keywords: cortical networks; neural networks; neuronal networks; brain connectivity; connectome;
network analysis
!
"!
=>!?+%'&5$.%(&+!
The set of connections in neural systems, now called the connectome!(Sporns et al., 2005), has been
the focus of neuroanatomy for more than a hundred years (His, 1888; Ramón y Cajal, 1892). However,
it attracted recent interest due to the increasing availability of network information at the global
(Burns and Young, 2000; Felleman and van Essen, 1991; Scannell et al., 1995; Tuch et al., 2003) and
local level (Denk and Horstmann, 2004; Lichtman et al., 2008; Seung, 2009; White et al., 1986) as
well as the availability of network analysis tools that can elucidate the link between structure and
function of neural systems. Within the neuroanatomical network (structural connectivity), the
nonlinear dynamics of neurons and neuronal populations result in patterns of statistical dependencies
(functional connectivity) and causal interactions (effective connectivity), defining three major
modalities of complex neural systems (Sporns et al., 2004). How is the network structure related to its
function and what effect does changing network components have (Kaiser, 2007)? Since 1992
(Achacoso and Yamamoto, 1992; Young, 1992), tools from network analysis (Costa et al., 2007b)
have been applied to study these questions in neural systems.
What are the benefits of using network analysis in neuroimaging research? First, networks provide an
abstraction that can reduce the complexity when dealing with neural networks. Human brains show a
large variability in size and surface shape (Van Essen and Drury, 1997). Network analysis, by hiding
these features, can help to identify similarities and differences in the organization of neural networks.
Second, the overall organization of brain networks has been proven reliable in that features such as
small-worldness and modularity, present but varying to some degree, could be found in all human
brain networks (and other species, too). Third, using the same frame of reference, given by the
identity of network nodes as representing brain regions, both comparisons between subjects as well as
comparisons of different kinds of networks (e.g. structural versus functional) are feasible (Rubinov
and Sporns, 2010).
The analysis of networks originated from the mathematical field of graph theory (Diestel, 1997) later
leading to percolation theory (Stauffer and Aharony, 2003) or social network analysis (Wasserman
and Faust, 1994). In 1736, Leonhard Euler worked on the problem of crossing all bridges over the
river Pregel in Königsberg (now Kaliningrad) exactly once and returning to the origin, a path now
called an Euler tour. These and other problems can be studied by using graph representations. Graphs
are sets of nodes and edges. Edges can either be undirected going in both directions or directed (arcs
or arrows) in that one can go from one node to the other but not in the reverse direction. A path is a
walk through the graph where each node is only visited once. A cycle is a closed walk meaning a path
that returns back to the first node. A graph could also contain loops that are edges that connect a node
to itself; however, for analysis purposes we only observe simple graphs without loops. In engineering,
graphs are called networks if there is a source and sink of flow in the system and a capacity for flow
through each edge (e.g. flow of water or electricity). However, following conventions in the field of
network science, we will denote all brain connectivity graphs as networks.
For brain networks, nodes could be neurons or cortical areas and edges could be axons or fibre tracts.
Thus, edges could refer to the structural connectivity of a neural network. Alternatively, edges could
signify correlations between the activity patterns of nodes forming functional connectivity. Finally, a
directed edge between two nodes could exist if activity in one node modulates activity in the other
node forming effective connectivity!(Sporns et al., 2004). Network representations are an abstract way
to look at neural systems. Among the factors missing from network models of nodes, say brain areas,
are the location, the size, and functional properties of the nodes. In contrast, geographical or spatial
networks also give information about the spatial location of a node. Two- or three-dimensional
Euclidean coordinates in a metric space indicate the location of neurons or areas. However, location
can also be non-metric where the distance between two nodes has ordinal values (e.g. location of
proteins given by a reaction compartment within a cell).
The application of network analysis identified several changes during aging and disease that can form
biomarkers for clinical applications. During aging, for example, functional connectivity to neighbors
at the local level and other brain areas at the global level is reduced particularly affecting frontal and
!
#!
temporal cortical and subcortical regions (Achard and Bullmore, 2007). In schizophrenia, such small-
world features were also altered regarding functional connectivity in EEG and fMRI (Micheloyannis
et al., 2006; Skudlarski et al., 2010) and structural connectivity using diffusion tensor imaging (van
den Heuvel et al., 2010). Alzheimer’s disease patients show a link between highly-connected nodes in
functional networks and high amyloid-! deposition (Buckner et al., 2009) and abnormal small-world
functional connectivity both in EEG (Stam et al., 2007) and in fMRI (He et al., 2008). Functional
connectivity for epilepsy patients is enhanced in EEG (Bettus et al., 2008) and shows altered modular
organization in MEG (Chavez et al., 2010) whereas DTI structural connectivity showed reduced
fractional anisotropy both adjacent to and further apart from cortical lesions in patients with partial
intractable epilepsy!(Dumas de la Roque et al., 2004). For healthy subjects, the number of steps to go
from one node in the fMRI functional network to another was linked to the IQ of that subject (van den
Heuvel et al., 2009).
Network analysis techniques can be applied to the analysis of brain connectivity and we will discuss
structural connectivity as an example. Using neuroanatomical or neuroimaging techniques, it can be
tested which nodes of a network are connected, i.e. whether projections in one or both directions exist
between a pair of nodes (Figure 1A). How can information about brain connectivity be represented? If
a projection between two nodes is found, the value ’1’ is entered in the adjacency matrix; the value ’0’
defines absent connections or cases where the existence of connections was not tested (Figure 1B).
The memory demands for storing the matrix can be prohibitive for large networks as N2 elements are
stored for a network of N nodes. As most neuronal networks are sparse, storing only information
about existing edges can save storage space. Using a list of edges, the adjacency list (Figure 1C),
stores each edge in one row listing the source node, the target node, and—for networks with variable
connection weight—the strength of a connection.
! "
"
"
"
"
""""# "
"
"""$"
Fig. 1. Representations of networks. (A) Directed graph with two directed edges or arcs (A!C and
D!C), and one undirected edge being equivalent to a pair of directed edges in both directions
(B"C). (B) The same graph can be represented in a computer using an adjacency matrix where a
value of 1 denotes the existence of an edge and 0 the absence of an edge. In this example, rows show
outgoing connections of a node and columns show incoming connections. (C) Sparse matrices (few
edges) can also be represented as adjacency lists to save memory. Each edge is represented by the
source node, the target node, and the weight of the edge (here: uniform value of 1).
In this tutorial, we will neither focus on listing tools for network analysis (Costa et al., 2007b) such as
the Brain Connectivity Toolbox nor review connectome analysis results (Bullmore and Sporns, 2009;
Sporns et al., 2004). Instead, we will introduce concepts of network analysis to new researchers in the
field of brain connectivity. A previous article (Rubinov and Sporns, 2010) provided a classification
and discussion of topological network measures relevant to neuroscience. The tutorial here will cover
features of the topological but also the spatial organization of neural systems. The spatial location of
neurons or regions in three dimensions and the delays for transmitting information over a distance are
properties unlike those found in many network models. However, spatial networks with delays for
propagation are frequent in real-world artificial and social networks, for example, for epidemic
!
$!
spreading (Hufnagel et al., 2004; Marcelino and Kaiser, 2009; May and Lloyd, 2001), transportation
networks (Guimerà and Amaral, 2004; Kaiser and Hilgetag, 2004b), or the Internet (Vázquez et al.,
2002; Waxman, 1988; Yook et al., 2002). We discuss the benefits and limits of individual analysis
methods. In addition, we will show how to interpret measurements and outline the common pitfalls
and misconceptions in the field. This review, for the first time, also introduces the concepts of single
node motifs extending notions of hub nodes and provides a classification of cluster detection
algorithms. We divide this review into six parts: (remainder of this section) yielding brain networks
and the role of node and edge definitions, (section 2) features of individual nodes, (section 3) features
of whole networks, (section 4) groups within networks, (section 5) types of networks, and (section 6)
spatial network features. Terms, which are also defined in the glossary at the end, are set in italics.
=>=!@&'<8*&;!8&'!A')(+!.&++-.%(B(%0!)+)*01(1!
How can one get information about brain connectivity between regions, the macroscopic connectome
(Akil et al., 2011)? The classical way to find out about structural connectivity is to inject dyes into a
brain region. The dye is then taken up by dendrites and cell bodies and travels within a neuron either
in an anterograde (from soma to synapse) or a retrograde (from synapse to soma) direction. Typical
dyes are Horseradish peroxidase (HRP), fluorescent microspheres, Phaseolus vulgaris-leucoagglutinin
(PHA-L) method, Fluoro-Gold, Cholera B-toxin, DiI, and tritiated amino acids. Allowing some time
for the tracers to travel, which could be several weeks for the large human brain, the neural tissue can
be sliced up and dyes can indicate the origin and target of cortical fiber tracts. Whereas this approach
yields high-resolution information about structural connectivity it is an invasive technique usually
unsuitable for human subjects (however, there are some post-mortem studies). In the following, we
will therefore present non-invasive neuroimaging solutions to yield structural and functional
connectivity.
The workflow for yielding human connectivity data starts with anatomical magnetic resonance
imaging (MRI) scans with high resolution (Figure 2). These scans are later used to register the
location of brain regions. For establishing functional connectivity, a time series of brain activity in
different voxels or regions can be derived. The correlation between the time series of different voxels
or, using aggregated measures, brain regions can be detected and represented as a correlation matrix
(value ranging from -1 to 1). This matrix can either directly be interpreted as a weighted network or it
can be transformed into a binary matrix in that only values above a threshold lead to a network
connection.
For establishing structural connectivity, diffusion tensor imaging (DTI) or diffusion spectrum imaging
(DSI) can be applied. Using deterministic tracking, for example, the number of streamlines between
brain regions can be represented in a matrix. For probabilistic tracking, matrix elements would
represent the probability to reach a target node starting from a source node. In both cases, the
weighted matrix can either be analyzed directly or it can be thresholded so that connections are only
formed if a minimum number of streamlines or a minimum probability has been reached.
!
%!
2
4
7
1
6
11
10
9
8
7
6
5
4
3
5
7
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
Fig. 2. Workflow for structural and functional connectivity analysis. High-resolution anatomical MRI
scans of each subject are used as references for further measurements (1). For establishing functional
connectivity, a time series of brain activity in different voxels or regions can be derived (3). The
correlation between the time series of different voxels or, using aggregated measures, brain regions
can be detected and represented as a correlation matrix (5). This matrix can either directly be
interpreted as a weighted network (6) or it can be binarized in that only values above a threshold lead
to a network connection (7). For establishing structural connectivity, diffusion tensor imaging or
diffusion spectrum imaging can be applied (2). Using deterministic tracking, for example, the number
of streamlines between brain regions can be represented in a matrix (4). This weighted matrix can
either be analyzed directly (6) or be tresholded so that connections are only formed if a minimum
number of streamlines has been reached (7).
1.2 Role of node and edge definitions
The choice of nodes and edges can be influenced by the anatomical parcellation schemes and
measures for determining connectivity (Rubinov and Sporns, 2010). This choice must be carefully
considered as different choices might not only change the topology by removing or adding a few
nodes or connections but might alter the local and global network features that will be discussed in the
following sections.
Parcellations of the brain into different nodes should be (a) non-overlapping in that each brain
location only belongs to one region, (b) should assign tissue to one node that has similar connections
to other parts of the brain, and (c) should only be compared with other networks that use the same
parcellation scheme; this is also crucial for comparing structural and functional connectivity in the
same subject (Honey et al., 2009).
!
&!
Connections of a network can be binarized or weighted. Binary connections only report the absence or
presence of a connection. Weighted links can also show the strength of a connection. For structural
connectivity, weights can indicate the number of fibers between brain regions (e.g. the streamline
count of deterministic tracking), the degree of myelination, the probability that a node can be reached
from another node (e.g. probabilistic tracking), or the amount of dye travelling from one node to
another (traditional tract-tracing studies). For functional connectivity, weights can indicate the
correlation in the time course of signals of different nodes. For effective connectivity, weighted links
represent causal relationships between nodes. Weighted networks can be converted to binarized
networks using a threshold in that connections are only established if the weight is above a threshold.
Using binarized networks simplifies the calculation and interpretation of many network measures (we
focus on binarized networks for this tutorial). On the other hand, the choice of a threshold can be
problematic: Using the same threshold can lead to a different number of edges in different networks;
on the other hand, comparing binarized networks with the same number of edge means that a different
threshold may be used for each network (van Wijk et al., 2010). For binarized networks, it therefore
has to be noted for which range of thresholds a phenomenon can be observed. In all cases, loops
(connections of a node to itself) and negative weights must be removed prior to applying network
measures.
C>!D&.)*!1.)*-!E!1(+4*-!+&5-!8-)%$'-1!
Networks can be characterized at different levels ranging from properties characterizing a whole
network at the global scale to properties of network components at the local scale. Starting from the
local scale, components of a network are its nodes and edges. Edges can be weighted, taking
continuous (metric) or discrete (ordinal) values indicating the strength of a connection. Alternatively,
they could just have binary values with zero for absent and one for existing connections.
Thinking about neural systems, there could also be multiple edges between two nodes, e.g. a fiber
bundle connecting two brain regions. However, such multi-graph networks are usually simplified in
that the number of fibers is either neglected (binary values) or included in the strength of a connection.
In addition to fiber count, one might also think of other properties of connections such as delays for
signal propagation or degree of myelination. Whereas such properties likely have significant impact
on network function, they are currently not part of the analysis of network topology.
The other component at the local scale is a network node. A node could be a single neuron but, as for
edges, could also be an aggregate unit of neurons such as a population or a brain area. The degree of a
node is the sum of its incoming (afferent) and outgoing (efferent) connections. The number of afferent
and efferent connections is also called the in-degree and out-degree, respectively. When ki denotes the
degree of the node i of a network with N nodes, the series (k1,…, kN) with increasing degrees (ki " ki+1)
is called the degree sequence of the network. Nodes with a high number of connections, i.e. a large
degree, are called network hubs. For structural and effective connectivity, the ratio between the in-
and out-degree of a node can give information about its function: nodes with predominantly incoming
connections can be seen as integrators (convergence) whereas nodes with mainly outgoing
connections can be seen as distributors (divergence) or broadcasters of information. These distinctions
can be useful when nodes are otherwise similar, e.g. distinguishing different types of network hubs
(Sporns et al., 2007).
For undirected networks, every connection between nodes is bi-directional (e.g. for functional
networks measuring correlation). For such networks, if a node A is connected to a node B with a bi-
directional link, this link is counted as one connection when calculating the degree of the node.
Likewise, it does not make sense to distinguish in-degree and out-degree as they will have the same
values.
Local measures can also refer to the neighborhood of a node. All nodes that directly project to a node
or directly receive projections from that node are called neighbors of that node. The connectivity
between neighbors is used to assess local clustering. The ratio of the number of existing edges
between neighbors and the number of potential connections between neighbors forms the local
!
'!
i
=
clustering coefficient (Figure 3A); a measure of neighborhood connectivity.
The local clustering coefficient for an individual node i with ki neighbors (node degree ki) and #i
edges between its neighbors is
"
C
i
kk
( !
)1
i
i
However, this formula is basically not defined if the number of neighbors ki becomes zero or one as
the denominator becomes zero (Costa et al., 2007b). These cases are usually treated as Ci = 0 although
some authors also set these values to one (Brandes and Erlebach, 2005). Whereas having nodes with
no connections (isolated nodes) or only one connection (leaf nodes) is unlikely for structural
connectivity, such nodes might occur for functional and effective connectivity. We will discuss some
ways to deal with such nodes when describing aggregate measures.
Another local measure of a single node is how many shortest paths (Figure 3B) are containing that
node. The shortest path between two nodes is the length of the path with the lowest-possible number
of connections. For counting how many paths pass through a node, the shortest paths between all pairs
of nodes are calculated. It is then counted how many shortest paths include a certain node. Note that
this measure is listed as a local measure as it is an attribute of a single node even though information
of the whole network is used to calculate this measure. For comparing this measure for two nodes of
the same network, it does not matter whether the absolute number of shortest paths containing one
node (stress centrality) or the relative frequency of how often that node is part of shortest paths
(betweenness centrality) is taken. This measure of how frequently a node is part of shortest paths is
called node betweenness. In a similar way, frequency of shortest paths containing a certain edge is the
edge betweenness of that edge. There are numerous kinds of centrality (Chapter 3, Brandes and
Erlebach, 2005) in addition to measures of how many shortest paths run through network components.
Other node centrality measures include, for example, closeness centrality, which is the reciprocal of
the total distance of a node to any other node of the network, and degree centrality, which is simply
the degree of a node.
! "
"
F
"
B
A
"""""""""""""""""""""""#!!
B
C
A
C
E
D
D
E
Fig. 3. Local and global measures. (A) Local clustering coefficient: All nodes that are connected with
node A are neighbors of that node. The local clustering coefficient of node A is the number of
connections between neighbors (green edges) divided by the number of all potential connections
between its neighbors. In this case, the local clustering coefficient is CA=4/10=0.4 meaning that 40%
of connections between neighbors exist. (B) Shortest paths: The shortest path is the path between two
nodes with the lowest-possible number of connections in the path. The path length of that path is the
number of connections that needs to be crossed to go from one node to another. In this example, the
length of the shortest path between nodes A and C (A!B!C) is 2 and the length of the shortest path
between nodes A and E (A!E) is 1.
!
(!
F>!G*&A)*!1.)*-!E!)44'-4)%-!/-)1$'-1!
We are now zooming out of a network and observe properties that characterize the network as a whole,
leaving out the intermediate regional scale for a moment. Whereas local measures look at properties
of individual components, say the primary visual area V1, global measures at the macroscale look at
the whole network. This is useful when comparing a given neural network with artificially generated
networks called benchmark networks. It is also useful when comparing neural networks from different
species or the same species at different levels of organization (area, column, layer). In these cases, the
number of nodes and edges as well as their identity (e.g. comparing networks that contain V1 with
networks that do not) might differ; however, aggregate measures can still be used to detect changes at
the macroscale.!
The edge density, sometimes called connectivity, of a network is the proportion of connections that
exists relative to the number of potential connections of a network. For a directed network with N
nodes, each node can connect to at most N-1 other nodes. Therefore, the edge density of a network
with E edges and N nodes is d = E / (N (N-1)). For an undirected network, the edge density becomes
d = E / (2 N (N-1)); note the factor 2 in the denominator so that any potential undirected edge between
two nodes is only counted once and not twice. An edge density of 1, corresponding to a percentage of
100%, would mean that all potential edges exist. In biological networks, however, only a small
fraction of potential connections occurs. For the cortico-cortical fiber tract connectivity of the
mammalian brain, for example, the edge density ranges between 10 and 30%. For the connectivity
between neurons in the nematode C. elegans, the edge density is 3.85%.
The edge density gives a first indication how well-connected a network is. However, the time it takes
to go from one node to another might still vary considerably depending on the topology of the
network. A measure of travelling through a network is the number of connections one has to cross, on
average, to go from one node to another. Formally, this average shortest path (ASP) of a network
with N nodes is the average number of edges that has to be crossed on the shortest path from any one
node to another:
1
ASP
!
NN
)1
(
#
ji
,
where d(i, j) is the length of the shortest path between nodes i and j having as few connections
between nodes i and j as possible. Note that the definition for the characteristic path length L is
slightly different (Watts, 1999) in that for each node the average shortest path length to any other
node is calculated and the median, instead of the mean, value over all nodes is returned as L.
In practice, in particular for networks with directed edges, there might be several pairs of nodes for
which no path exists. In such cases, graph theory would demand setting the distance d(i,j) between the
two nodes to infinity. However, having only one such pair in a network would give an ASP of infinity
as well. In practice, such infinite values are excluded that means the average shortest path only takes
the average of existing shortest paths between pairs of nodes. Alternatively, a measure called global
efficiency can be used (Achard and Bullmore, 2007; Latora and Marchiori, 2001). Global efficiency
uses a sum of the inverse of the distance L so that non-existing paths, leading to infinite distance,
contribute a zero value to the sum:
1
1
E
!
NN
L
(
j
i
ji
,
"
j
id
),(
with
global
)1
#
=
=
i
"
j
where Li,j is the length of the shortest path between nodes i and j and N is the number of nodes.
Another aggregate measure based on the local features of individual nodes is average neighborhood
connectivity – the (global) clustering coefficient. The clustering coefficient is just the average of the
local clustering coefficient Ci of all nodes:
!
)!
.
.
9
2
2
density range of 0.6–1.5%. For shifting from C1 to C \x{FFFF} , classification only changed up to 1% of
the cases; a fluctuation which might be due to the small sample size. In addition, a shift also
Contents
Contents
occurred between the classical measures C1 and C2 : whereas changing from C1 to C2 affected
few cases, a shift from C2 to C1 affected up to 3% of the cases for standard rewiring and up
2
1. Introduction
1. Introduction
2
to 14% for inverse rewiring. Therefore, changes in classification are possible for all clustering
3
2. Materials and methods
3
2. Materials and methods
measures; in particular, when using inverse rewiring.
. . . . . .
. . . . . . . . . .
2.1. Networks
.
. . .
.
3
.
.
.
.
.
.
.
.
.
.
.
. . . . .
.
.
. . .
. .
. .
.
.
.
.
.
.
.
2.1. Networks
. .
.
. .
3
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4
. .
. . . . . . .
. .
2.2. Adjusted clustering coefficient definition . . . .
. . . . . .
2.2. Adjusted clustering coefficient definition .
4
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4. Discussion
3. Results
4
3. Results
4
. . .
.
.
. . . . .
. .
. .
. . . . . . .
. .
3.1. Network comparison .
5
.
.
.
.
.
.
.
.
.
.
.
.
. .
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
3.1. Network comparison .
5
We have shown that current definitions underestimate neighborhood clustering in sparse
.
.
. . . . .
. . . . . . . . .
3.2. Changes of small-world features .
. .
. .
. .
6
.
.
3.2. Changes of small-world features .
6
.
.
.
.
.
.
.
.
. .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
networks with many isolated or leaf nodes. In addition, the outcome of comparisons of the
. . .
3.3. Changes of small-world classification .
. .
. . . . . . .
. .
. . . . . . .
8
3.3. Changes of small-world classification .
8
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
extent of small-world features between different networks critically depended on the applied
4. Discussion
9
definition of the clustering coefficient. Furthermore, networks formerly classified as random
4. Discussion
9
10
5. Conclusion
can be classified as small-world when isolated or leaf nodes are excluded from the calculation
5. Conclusion
10
of the average clustering coefficient. This can also happen when switching from C2 to C1 .
10
Acknowledgment
Acknowledgment
10
Could the clustering coefficient definitions impact the analysis of small-world networks?
References
10
References
10
There are three consequences of this study. Firstly, small-world networks regarding previous
measures C1 and C2 will still be detected as small-world using C \x{FFFF} as this value will be higher
1. Introduction
than the previous values. Consequently, the small-worldness s—the ratio of the clustering
1. Introduction
coefficient in the original and random benchmark networks divided by the unchanged ratio of the
Many real-world networks show properties of small-world networks as their neighborhood
Many real-world networks show properties of small-world networks as their neighborhood
characteristic path lengths in original and random networks—will be higher. Secondly, networks
connectivity, generally denoted by the clustering coefficient, is higher than in comparable
connectivity, generally denoted by the clustering coefficient, is higher than in comparable
which are currently not classified as small-world networks may be regarded as small-world due
random networks [1]. The local clustering coefficient for an individual node i with degi
random networks [1]. The local clustering coefficient for an individual node i with degi
to the increase in clustering coefficient. This case will occur when the path length is comparable
neighbors and \x{FFFF}i edges between its neighbors is
neighbors and \x{FFFF}i edges between its neighbors is
to that of random networks but the clustering coefficient, concerning previous definitions C1 and
C2 , is not significantly higher than that of random networks. Thirdly, comparison of networks
\x{FFFF}i
\x{FFFF}i
(1)
C i =
.
(1)
C i =
could lead to opposite conclusions using the new measure. In conclusion, the novel measure C \x{FFFF}
.
degi (degi − 1)
degi (degi − 1)
gives a clearer view of neighborhood connectivity and is more independent of the sparseness of
This formula is basically not defined if the number of neighbors degi becomes zero or one as the
This formula is basically not defined if the number of neighbors degi becomes zero or one as the
edge density.
denominator becomes zero [2]. These cases are usually treated as C i = 0 although some authors
denominator becomes zero [2]. These cases are usually treated as C i = 0 although some authors
A problem of the proposed measure C \x{FFFF} is that the percentage θ of nodes that are excluded
also set these values to one [3]. In the current scheme, these values would be part of the global
also set these values to one [3]. In the current scheme, these values would be part of the global
from analysis could be considerably high (table 1). The percentage of excluded nodes could
calculation
calculation
be as high as 14% for metabolic networks and as high as 59% for man-made networks
N \x{FFFF} C i .
1
(power grid). Note that the value for the protein interaction network in yeast is also high at
N \x{FFFF} C i .
1
(2)
C1 =
(2)
C1 =
56% as edge density is low and isolated nodes are not part of the largest connected cluster
observed here. In general, however, exclusion from the average affected less than 10% for most
In addition, we tested an alternative and more widely used definition of the clustering
In addition, we tested an alternative and more widely used definition of the clustering
Alternatively, a more widely used definition of the clustering coefficient (Newman et al., 2001) is
of the networks. In addition, using a subset of defined nodes is comparable to the procedure
coefficient [4] in which
coefficient [4] in which
for calculating shortest paths or the characteristic path length where unreachable paths with
\x{FFFF} \x{FFFF}i
\x{FFFF} \x{FFFF}i
otherwise infinite distance are not included in calculating the average path length.
(3)
C2 =
\x{FFFF} degi (degi − 1)
.
(3)
C2 =
\x{FFFF} degi (degi − 1)
.
An alternative solution would be to describe the clustering coefficient using inverse
neighborhood clustering. For the shortest paths, for example,
the inverse measure of
Note that the first definition C1 runs into problems if one node’s local clustering coefficient is
This might lead to biased assessments of neighborhood clustering in the sense that values that
This might lead to biased assessments of neighborhood clustering in the sense that values that
efficiency [25] where unreachable paths contribute 1/∞ = 0 to the local efficiency circumvents
undefined (see previous section). A third definition is to describe the clustering coefficient using
are not defined (division by zero) should not be included in the averaging. Thus, instead of using
are not defined (division by zero) should not be included in the averaging. Thus, instead of using
the need for excluding unreachable paths. Similarly, the (neighborhood) disconnectedness D
inverse neighborhood clustering analogous to using inverse shortest path lengths for defining
N as the number of evaluated nodes for the global C1 , a new number N \x{FFFF} indicating all nodes
N as the number of evaluated nodes for the global C1 , a new number N \x{FFFF} indicating all nodes
efficiency. Such a measure of (neighborhood) disconnectedness D could be defined as (Kaiser, 2008):
could be defined as:
with defined local clustering should be used for a global measure C \x{FFFF} . We show that using such
with defined local clustering should be used for a global measure C \x{FFFF} . We show that using such
N \x{FFFF} Di with Di = 1/C i =
1
degi (degi − 1)
an adjusted measure for the clustering coefficient has several implications for network analysis
an adjusted measure for the clustering coefficient has several implications for network analysis
D =
and can help to identify the contribution of leafs and isolated nodes to average clustering.
\x{FFFF}i
and can help to identify the contribution of leafs and isolated nodes to average clustering.
On a conceptual level, the adjusted value C \x{FFFF} is more intuitive as the clustering coefficient
and Di = 0
for \x{FFFF}i = 0.
(5)
On a conceptual level, the adjusted value C \x{FFFF} is more intuitive as the clustering coefficient
is commonly called a measure of neighborhood connectivity: if 30% of the local coefficients
is commonly called a measure of neighborhood connectivity: if 30% of the local coefficients
In relation to the characteristic path length as global efficiency (how well are any two nodes of a
are zeros from cases where no neighbors exist, how can the classical definitions still give
are zeros from cases where no neighbors exist, how can the classical definitions still give
network connected), the clustering coefficient can also be called local efficiency (how well are
New Journal of Physics 10 (2008) 083042 (http://www.njp.org/)
neighbours of a node connected) (Achard and Bullmore, 2007; Latora and Marchiori, 2001). The
New Journal of Physics 10 (2008) 083042 (http://www.njp.org/)
clustering coefficent (C1 or C2) will increase whenever the edge density increases as a higher
New Journal of Physics 10 (2008) 083042 (http://www.njp.org/)
probability that any two nodes are connected also means that connections between neighbours are
more likely. Therefore, comparing clustering coefficents for networks with different edge densities
should either be avoided or they should use a normalized coefficient. Such a normalized coefficient
would be the clustering coefficient relative to the clustering coefficient of a random network (the
clustering coefficient of a random network is the same as the edge density of the original network).
However, normalization does not work if one of the two networks has a much higher edge density:
whereas the clustering coefficient of a network with low edge density can be 2-3 times as high as the
edge density of that network, the clustering coefficient of a network with, say 60% edge density, can
never be 2-3 times as high as the edge density.
H>!I-4(&+)*!1.)*-!E!G'&$31!&8!+-%;&'<!+&5-1!!
Often, we are interested in an intermediate level of organization going beyond single nodes but not
including the whole network. At this scale, we can observe measures for subsets of nodes which share
similar connections, e.g. dealing with visual input. Such measures look at the sets of nodes where the
connectivity within the set is larger than between the set of nodes and the rest of the network. Such
sets of nodes are called clusters, modules, or, following social network analysis, communities.
4.1 Clusters
Clusters or modules are parts of a network with many connections (high edge density) within such a
part and few connections (low edge density) to the remaining nodes of the network. There are many
different algorithms to detect clusters of a network (Girvan and Newman, 2002; Hilgetag et al., 2000b;
Palla et al., 2005). As a general rule, algorithms can be distinguished along three features (many other
classifications exist). First, algorithms could lead to hierarchical or non-hierarchical solutions. Non-
hierarchical solutions just identify different modules that can be displayed in a reordered adjacency
matrix or in a circular graph (Figure 3A). Hierarchical solutions not only identify modules but also
sub-modules within modules, sub-sub-modules within sub-modules, and so on (Clauset et al., 2008).
That means that the modular organization at different hierarchical levels can be observed. Instead of
distinct levels, the parcellations into sub-modules can also be shown in form of a dendrogram – a tree
where nodes in the same cluster are part of the same branch of the tree (Figure 3B). Note that the
hierarchical parcellation of the network critically depends on the threshold that one chooses. Whereas
an early threshold, near the root of the tree, might only show the main clusters – just one hierarchical
level, a late threshold, close to the leafs of the tree, might result in numerous sub-clusters that are so
small that the distinction between clusters becomes weak. Second, algorithms can use a predefined
number of clusters which need to be detected or can determine the number of clusters themselves. If
!
*!
the number of clusters is known, algorithms similar to k-means, where k is the number of clusters, can
be used to detect the clusters. However, in many biological applications, the number of clusters is not
known beforehand, and algorithms that determine the number during the clustering process are
needed. Third, algorithms can lead to overlapping or non-overlapping cluster-classifications of nodes.
For non-overlapping algorithms, a node will belong to one and only one cluster of the network.
However, this assignment to a cluster can often be ambiguous with only a slightly lower preference
for assigning the node to another cluster. For overlapping algorithms, a node can belong to several
clusters with different likelihoods. Say a node could belong to clusters A, B, and C with likelihoods of
20%, 30%, and 50%, respectively. Such overlapping cluster memberships for cortical nodes can point
to nodes that integrate information from several modules, e.g. from the visual and auditory system.
Here, we just show an example of a non-hierarchical, non-overlapping algorithm where the number of
clusters is not known beforehand. To identify clusters, an evolutionary optimization algorithm can be
used (Hilgetag et al., 2000b). This approach is based on the goal that areas should be more frequently
linked to areas in the same cluster than to areas in different clusters. To achieve this goal, Hilgetag et
al. defined a two-component cost function C whose weighted sum was minimized:
C = wattr $ Cattr + wrep $ Crep
The components were Cattr (attraction component), the number of connections existing between
clusters, and Crep (repulsion component), the number of absent connections within clusters, with wattr
and wrep as weights for adjusting the influence of each component, respectively (Hilgetag et al.,
2000a). The first component can be considered as attracting connections to clusters, as it becomes
minimal if no connections between clusters exist. Minimizing the second component, on the other
hand, tends to break up clusters, as it can be reduced to zero by an arrangement that consists of
completely separate areas (Hilgetag et al., 2000a). Only minimizing both components simultaneously
will result in dense connectivity within clusters and few connections between clusters.
As an example, this method can be used for the analysis of structural connectivity using fiber tract
data in the cat and the macaque as tested for the primate visual, global primate cortical, and global cat
cortical network (Hilgetag et al., 2000a). For the cat structural connectivity, based on invasive tract
tracing studies, cortical area groupings largely agreed with functional cortical subdivisions (Hilgetag
et al., 2000b): the four observed clusters consisted predominantly of visual, auditory, somatosensory-
motor, or frontolimbic areas, respectively. In addition, clusters of the primate visual system
corresponded closely to the dorsal and ventral visual streams. In agreement with the idea that
structural clusters correspond to functional subdivisions, cluster analyses of semi-functional
(neuronographic) connections showed functional processing clusters with broadly similar subdivisions
(Stephan et al., 2000).
!
"+!
IP!
FUS!
CUN!
BSTS!
SF!
RMF!
RAC!
PCUN!
PREC!
SMAR!
TT!
CAC!
CMF!
LOF!
MOF!
PARC!
POPE!
PORB!
ST!
SP!
MT!
IP!
LING!
LOCC!
SMAR!
PC!
FUS!
IT!
BSTS!
SF!
PREC!
PCUN!
PSTC!
POPE!
PTRI!
CMF!
PCAL!
CUN!
MOF!
PARC!
ISTC!
RMF!
CAC!
TT!
PORB!
RAC!
LOF!
ISTC! LOCC!LING!
MT!
A
PCAL!
SP!
IT!
B
ST!
PSTC!
PTRI!
PC!
6!
5!
4!
3!
2!
1!
Fig. 4. Clusters. (A) Cluster structure of human cortico-cortical connectivity, based on (Hagmann et
al., 2008). Cortical areas were arranged around a circle by evolutionary optimization, so that highly
inter-linked areas were placed close to each other. Note that nodes in the same cluster, having a high
structural similarity, also have a similar function. (B) Dendrogram of the same network using
hierarchical clustering. A dendrogram running from the root to the leafs (here: from left to the right)
consists of branches connecting objects in the tree. The distance of the branching point on the x-axis
is the rescaled distance when clusters are combined.
Figure 4 shows an example of the modular organization of human corticocortical connectivity, based
on diffusion spectrum imaging (Hagmann et al., 2008). Cortical areas were arranged around a circle
so that highly inter-linked areas were placed close to each other (Figure 4A). Note that nodes in the
same cluster, having a high structural similarity, also have a similar function. The cluster architecture
of the same network can also be represented by a dendrogram using hierarchical clustering (Figure
4B). A dendrogram running from the left to the right consists of branches connecting objects in the
tree. The distance of the branching point on the x-axis is the rescaled distance when clusters are
combined.
This is only a brief overview of clustering but there are numerous approaches for detecting network
modules. For example, clusters can not only be defined by grouping nodes but also by grouping edges
into link communities (Ahn et al., 2010).
4.2 Modularity
A measure that has received a lot of attention for topological clusters in recent times is modularity.
Modularity (Q) is a reflection of the natural segregation within a network (Newman, 2004) and can be
a valuable tool in identifying the functional blocks within. Similar to the measure C discussed in the
previous sub-section, Q can be used to assess how well a parcellation into non-overlapping modules
represents the modular architecture of a network. Given two parcellations into distinct modules for the
same network, the parcellation with the higher value of Q would be preferred. So how can the
modularity Q be computed? Given a parcellation that assigns to each node i a label ci identifying to
which module the node belongs to, the modularity Q is the difference between the number of edges
that lie within a community in the actual network and a random network of the same degree sequence.
A high level of topological clustering is reflected in a high value of modularity. The modularity for a
directed network is given by!(Newman, 2006):
!
""!
Q
=
1
m
"
ij
(
&
&
'
a
ij
)
out
j
kk
in
i
m
%
!
#
i cc
,
#
$
j
!
where m: total number of edges in the network (note that bidirectional links are counted twice); aij :
in : in-degree of node i; kj
out : out-degree of node j; (cid:552)cicj (cid:27)Kronecker
element of adjacency matrix; ki
delta (only one if nodes i and j are in the same module and zero otherwise); cn – label of module to
which node n belongs to.
This measure can be used as a cost function in cluster algorithms where the aim is to maximize the
modularity function Q. As for other optimization problems, a range of methods can be used such as
genetic algorithms (see previous section), simulated annealing, etc. Note that the modularity measure
shows how well a given separation into modules performs. It does not include information about how
many modules exist or about their size or overlap (see previous section about clusters on these
problems). Using this modularity measure Q also has disadvantages such as limited resolution and a
bias towards certain cluster sizes.
Note that this measure differs from the cost function C where the weights wattr and wrep need to be
adjusted, in addition to the testing of different cluster membership configurations, to yield a good
solution for a particular network. Using the notation of Newman (2006), the cost function C can be re-
wC
a
w
a
1(
)
1(
)
!
!
written as:
=
#
#
"
+
"
cc
cc
ij
rep
ij
attr
,
,
i
i
ij
ij
j
j
4.3 Network Motifs
Modules are relatively large structures comprised of tens or hundreds of nodes. Modules are often
linked to function in that nodes of the same module tend to have a similar function. However, there
could also be smaller subgraphs with only a few nodes that could have a specific function for a
network. For a subgraph with only two nodes A and B, there are three ways of how directed edges
could exist between them: a connection in one direction, A#B; a connection in the reverse direction,
A$B; and a bidirectional connection, A"B. For sub-graph counting, the case that no connections
between the nodes exist is not taken into account. Also, the identity of the nodes is not retained;
therefore, A#B and A$B are treated as one pattern. For three nodes, there are already 13 different
patterns how directed edges could be distributed (Figure 5A). For a real-world network it is then
possible to count how often each potential 2-node or 3-node pattern occurs. If a pattern occurs
significantly more often than in a randomly organized network with the same degree distribution, it is
called a network motif. Dubbed the “building blocks” of complex networks, network motifs mimic the
concept of sequence motifs as used in genomics. In a gene sequence, a motif is a recurring
subsequence, a pattern that is conjectured to have some functional significance. In a network, a motif
is a recurring sub-network conjectured to have some significance.
So when does a pattern occur significantly more often than would be expected for a random
organization? To decide this, a set of benchmark networks is generated where the number of nodes
and edges is identical but, starting from the original network, edges are rewired while each node
maintains its original in-degree and out-degree. Thus, the degree distribution of the network remains
unchanged. This means that each node still has the same degree after the rewiring procedure but that
additional information, e.g. the cluster architecture, is lost. In the next step, for each benchmark
network the number of occurrences of a pattern is determined. Then, the pattern count of the real-
world network can be compared with the average pattern count of the benchmark networks; patterns
that occur significantly more often in the real-world network than in the benchmark networks are
called network motifs.
!
"#!
A B
!"#$%&'()*+",&
Human Cortex
-.&
34&
Macaque Visual Cortex 7&
/(#0&
'56&
839&
/#%1+2&
'56&
121.55 (21.03) z = 13.79 &
Macaque Cortex
&
;#)&;+")(<&
!"#$%$&'()#
7&
7&
8&
?&
3:44&
223.66 (34.99) z = 46.22 &
3=3>&
=777&
483@&
472.33 (52.85) z = 14.16 &
1067.03 (121.52) z = 15.98 &
1164.31 (134.71) z = 16.79 &
Fig. 5. Network motifs. (A) Overview of all 13 possible ways to connect three nodes (three nodes
without connections are not considered). (B) Three-node patterns that occur significantly more often
in the human (Iturria-Medina et al., 2008), macaque, cat, and C. elegans structural connectivity than
in rewired networks and are thus network motifs (adapted from (Milo et al., 2002); ID’s refer to the
numbers in (A)).
The seminal paper by Milo et al. (Milo et al., 2002) gave origin to a multitude of definitions and
studies. Network motifs have since been used in the most varied areas. The concept has been applied
to networks in domains like protein-protein interactions (PPI) (Alon, 2003; Wuchty et al., 2003), gene
transcriptional regulation, food webs, and neural systems (Milo et al., 2002; Sporns and Kötter, 2004).
Implementations of motif discovery tools include the original mfinder routine
(http://www.weizmann.ac.il/mcb/UriAlon/groupNetworkMotifSW.html ), the faster FANMOD
routine (http://theinf1.informatik.uni-jena.de/~wernicke/motifs/index.html (Wernicke and Rasche,
2006)), and a Matlab implementation (http://www.brain-connectivity-toolbox.net/ ). Figure 5A shows
all possible patterns for three nodes and Figure 5B shows characteristic network motifs of size 3 for
different neural networks (Milo et al., 2002).
Finding these motifs is a computationally hard problem, because fundamentally we match graph
patterns with the desired patterns, which leads to the well-known problem of graph isomorphism, with
no polynomial time algorithm known (see Box: Runtime complexity). As the size of the motifs gets
bigger, the time needed to calculate grows exponentially. Hence, an exhaustive computation of all
motifs of a network is typically reduced to very small sizes in order to obtain results in a reasonable
amount of time (see (Ribeiro et al., 2009) for a survey of motif detection strategies).
In addition to computational challenges, there are also conceptual problems with motif analysis. The
benchmark for counting the number of motifs is a rewired network with the same degree distribution.
Even if the degree distribution remains identical, random rewiring removes the topological cluster
architecture of the real network that often relates to the underlying spatial clustering of nodes. Taking
into account the modular organization and using a rewiring that maintains both the degree distribution
and the topological cluster architecture leads to a lower number of network motifs. This occurs as
many motifs, such as highly-connected three- or four-node motifs, are frequent when densely
connected modules exist. If the rewired network contains such densely connected modules as well,
those patterns do not occur significantly more often in the original network. Therefore, only few
motifs for the networks considered in (Sporns and Kötter, 2004) remain (Kaiser, unpublished). The
same is true for a network with regular connectivity, but without multiple modules: its neighborhood
connectivity measured through the clustering coefficient is higher than for the rewired network and
thus patterns with dense connectivity arise as network motifs. This is just one example how different
features of a network can be strongly correlated. Other examples are the positive correlations between
a node degree and its node betweenness, the edge density and the clustering coefficient of a network,
and the edge density and the characteristic path length. As mentioned earlier, when comparing
!
"$!
networks it is crucial that differences between networks in one network, say the clustering coefficient,
are not just caused by differences in another feature, say edge density.
Another problem in applying motif analysis to brain connectivity is that many structural and
functional networks yielded by diffusion imaging or time series correlations, respectively, are
undirected. For such networks, the number of patterns and therefore potential motifs is significantly
reduced: observing 3-node patterns yields 13 potential motifs for directed networks (Figure 5A) but
only 3 for undirected networks. This renders motifs less meaningful for undirected networks; however,
measurements yielding directed networks might become available in the future.
Network motifs can give information about characteristic patterns of multiple nodes but what about
individual nodes that are special for a network? Although a measure for a single node would normally
be part of the local scale, we discuss this concept here as it is closely related to multiple-node motifs.
Certain singular node-motifs, such as highly connected nodes or hubs, affect spreading phenomena,
which makes them important components of the network. More complex compound singular node
motifs, which are characterized by multiple features in combination, specify nodes more
comprehensively. With this more precise description new kinds of motifs can be formulated (Costa et
al., 2009). The algorithm first identifies outlier nodes with features that are significantly different
from other nodes in the network. Next, those outlier nodes can be classified into different classes
based on their individual features. Finally, the number of nodes that were found for each class gives a
fingerprint that is characteristic for each network. As the method does not rely on adjustable
parameters, it can be automatically applied to a large number of networks (Echtermeyer et al., 2011).
Implementations of this method are available and provide both a graphical user interface and a
command line version for batch processing (http://www.biological-networks.org/ ).!!
J>!#03-1!&8!+-%;&'<1!
We already mentioned that network measures can be used to compare networks with each other.
Often, we are not only interested in how network measures differ but whether the type of network
differs. Although each neural network has a unique topological and spatial organization, such types or
classes of networks can be used for classification and comparison (Figure 6). Such classes are based
on global features of the degree distribution and the community organization. The following section
shows different types and their characteristic properties. Note that real-world networks, however,
might show a combination of different classes, e.g. being modular and small-world.
5.1 Random networks
Whereas many networks are generated by a random process, the term random network normally refers
to the type of Erdös-Rényi random networks (Erdös and Rényi, 1960). Random networks are
generated by establishing each potential connection between nodes with a probability p. This
probability, for a sufficiently large network, is then equivalent to the edge density of the network; i.e.
the connection density. The process of establishing connections resembles flipping a coin where an
edge is established with probability p and not established with probability q = 1 - p. Therefore, the
distribution of node-degrees follows a binomial probability distribution. For large numbers of nodes,
the probability P(k) that a node has k connections can be approximated by a Poisson distribution, and
hence the term ‘exponential degree distribution’ is also used (Bollobas, 1985). The distribution can be
shown as a histogram where the counts for the different bins are plotted as data points. For the
“exponential” degree distribution, P(k) ~ e-k , of a random network, points are arranged on a line for a
logarithmic plot of log(P(k)).!
!!
!
!
"%!
!!"!!"#$%&'()*+,!#-*$./!!#!!01-23'4#33!!!!!!!!!!!!!!!$!!(3562-#!!!!!!!!!!!!!!!!!!!!!!!!%!!0/-22'7.#2$!!!!!!!!!!!!!&!!8.$62-#!!!!!!!!!!!!!!!!!!!!!'!!9,3#-#1:,1-2!
Fig. 6. Types of networks. Networks contain 24 nodes and 142 edges. The top panel shows individual
networks where nodes are located on a circle. The bottom panel shows the average probability over
100 networks to display a connection in the adjacency matrix: white denotes edges that are always
absent whereas black shows edges that are always present. (A) Erdös-Rényi random network. (B)
Scale-free network with dark edges (top) indicating highly connected nodes or hubs. (C) Regular or
lattice network with high connectivity between neighbors. (D) Small-world network with blue edges
(top) representing short-cuts of the network. Note that for the average probability plot (bottom), short-
cuts are invisible due the averaging over 100 networks. (E) Modular network with two modules. (F)
Hierarchical network with two modules consisting of two sub-modules each. Thus, there are two
hierarchical levels of organization.
!
5.2 Scale-free networks
Scale-free networks are characterized by their specific distribution of node degrees. The degree
distribution follows a power law where the probability that a node with degree k exists, or the
frequency of occurrences for real-world networks, is given by P(k) ~ k -
%. This is different from the
random networks discussed above where the degree distribution follows an exponential distribution,
P(k) ~ e-k. The exponent % of a power-law degree distribution can vary depending on the network
which is studied. For example, for functional connectivity between voxels in human MRI, % was
found to be 2.0 (Eguiluz et al., 2005) whereas % was 1.3 for functional connectivity of neurons in the
hippocampus determined through calcium imaging (Bonifazi et al., 2009).
A B C
8000
)
k
(
N
s
e
c
n
e
r
r
u
c
c
O
6000
4000
2000
0
10(cid:239)1
10(cid:239)2
10(cid:239)3
)
k
(
P
e
v
i
t
a
l
u
m
u
C
10(cid:239)4
100
0
50
100
150
101
102
k
k
Fig. 7. Scale-free networks. (A) Scale-free networks contain highly-connected nodes or hubs (shown
in red). (B) Degree distribution of a scale-free network with 10 000 nodes and 20 145 connections. In
contrast to random networks with one characteristic scale where all node degrees k are close to the
average, scale-free networks can contain nodes with degrees that are several standard deviations away
from the average. In this example, there are 13 nodes with degrees that are nine standard deviations
away from the average degree of 4 (arrows); the maximum degree is 504 (beyond the figure axes). (C)
The cumulative frequency P(k) that a node with degree k occurs in the network follows a power-law
leading to a straight line in a bi-logarithmic plot.
!
"&!
Data points for a power-law degree distribution lie on a straight line for a log-log plot of log(P(k))
against log(k). To test this power-law relationship, the cumulative distribution Pc(k) = P(X > k) = 1 –
F(k) with F(k) = P(X " k) where X is the number of connections of a node is plotted (Figure 7C). As
the histogram uses the same bin widths, the bins for high-degrees have fewer entries than the bins for
nodes with low degree values. Therefore, data points in the histogram will fluctuate more strongly for
the tail of the distribution. In addition to the visual inspection of the log-log plot, a statistical analysis
is needed to test for a power-law behavior (Clauset et al., 2009). Such an analysis will determine the
goodness-of-fit with a power-law distribution and will also compare this result with alternative
hypotheses (e.g. exponential or Yule distributions).
There are two potential problems with determining whether a network shows a power-law degree
distribution and is thus scale-free: One problem arises when only part of the network is known. For
example, one might like to test scale-free properties of neuronal networks but only has connectivity
within one column but not with other parts of the brain. In this case, connectivity is only known for a
subset or sample of the nodes of the whole network. Using such incomplete sampling, is it still
possible to test whether the whole network is scale-free or not? Unfortunately, the amount of
unknown or not included connections or nodes might change the shape of the degree distribution
(Stumpf et al., 2005); in particular missing nodes could mean that the rarely occurring hubs are not
part of the sample resulting in classifying a scale-free network as an Erdös-Rényi random network.
Another problem is networks with a low number of nodes, where the degree distribution only consists
of one or two orders of magnitude. Unfortunately, this is the case for regional brain connectivity
where networks consist of 100 or less nodes: Structural networks in the macaque, cat, and human
within one hemisphere usually consist of around 30-100 nodes. Such a low number of nodes results in
power-law fits that are not robust. For such networks, individual outlier nodes with very high
connectivity might directly alter the tail of the degree distribution, whereas for larger networks the
degrees of several nodes will be binned so that outliers are less influential. However, we can test for
scale-free behavior by using indirect measures whose outcome is not altered significantly by one
individual node (Kaiser et al., 2007b).
5.2.1 Case-Study: Are neural networks scale-free?
Previous studies have shown that functional networks of the human brain, looking at signal
correlations between voxels in fMRI, are scale-free (Eguiluz et al., 2005). However, at the gross level
of signal correlations between brain regions, it was argued that these functional networks are not
scale-free (Achard et al., 2006). We have compared the structural network between cat and macaque
brain regions with different types of benchmark networks, including random, scale-free and small-
world networks, and found strong indications that the brain connectivity networks share some of their
structural properties with scale-free networks. In particular, we compared the effect that the removal
of nodes and connections had on the ASP found in the brain connectivity networks and their
benchmark counterparts (Kaiser et al., 2007b). So even though the degree distribution cannot be tested,
the robustness after simulated lesions is most similar to that of a scale-free network. However, this
does not necessarily mean that cortical networks show a power-law degree distribution. Indeed, the
structural connectivity is unlikely to be scale-free over more than one order of magnitude. Still,
cortical networks contain highly-connected nodes that are unlikely to occur in random networks with
the same number of nodes and edges (Kaiser et al., 2007b). These hubs might be the underlying
reason for the lesion robustness that was comparable to scale-free networks as described above.
If neural networks show more highly-connected nodes than for random benchmark networks, how
could network features such as hubs arise? There are several potential developmental mechanisms that
yield brain networks with highly connected nodes. Work in brain evolution suggests that when new
functional structures are formed by specialization of phylogenetically older parts, the new structures
largely inherit the connectivity pattern of the parent structure (Ebbesson, 1980). This means that the
patterns are repeated and small modifications are added during the evolutionary steps that can arise by
duplication of existing areas (Krubitzer and Kahn, 2003). Such inheritance of connectivity by copying
modules can lead to scale-free metabolic systems (Ravasz et al., 2002). A developmental mechanism
!
"'!
for varying the node degree of regions could be the width of the developmental time window for
synaptogenesis at different regions (Kaiser and Hilgetag, 2007; Nisbach and Kaiser, 2007). Indeed, C.
elegans neurons that are generated early during development tend to accumulate more connections
and tend to be hubs of the adult network (Varier and Kaiser, 2011).
5.3 Small-World networks
Many networks exhibit properties of small-world networks (Watts and Strogatz, 1998). The term
small-world refers to experiments in social networks by Stanley Milgram where a person could reach
any other person through a relatively short chain of acquaintances, the “six degrees of separation”
(Milgram, 1967). However, relatively short does not mean that the average number of connections to
cross from one node to another, the characteristic path length, is minimal. Indeed, the path length is
usually higher than for Erdös–Rényi random networks with the same number of nodes and edges.
However, connectivity between node neighbors, the clustering coefficient, is much higher than for
random networks.
So when can a network be considered a small-world network? Unfortunately, there is no clear
criterion. In general, to classify a network as small-world, its clustering coefficient should be much
higher than the clustering coefficient of Erdös-Rényi random networks. For Erdös-Rényi random
networks, the clustering coefficient has the same value as the edge density (connections between
neighbors are as likely as any other connections of the network) so edge density might be used for the
comparison. In addition, the characteristic path length of the network should be comparable to that of
a random network that means slightly but not excessively higher than that value.
A measure to summarize to what extent a network shows features of a small-world network is small-
worldness S = (C / Crand) / (L / Lrand) where C is the clustering coefficient and L is the characteristic
path length of an observed network and a random network (Humphries and Gurney, 2008). Note that
this measure is useful for comparing small-world networks but not sufficient for determining whether
a network is a small-world network or not: a high value of S might also occur for networks with
extremely high characteristic path length as long as the clustering coefficient is much higher than for
random networks.
5.3.1 Case study: It’s a small brain—small-world properties in neural networks
Small-world properties were found on different organizational levels of neural networks: from the tiny
nematode C. elegans with about 300 neurons (Watts and Strogatz, 1998) over cortical structural
connectivity of the cat and the macaque (Hilgetag et al., 2000a; Hilgetag and Kaiser, 2004; Sporns et
al., 2000) to human structural (Hagmann et al., 2008) and functional (Achard et al., 2006)
connectivity. Whereas the clustering coefficient for the macaque structural connectivity is 49% (16%
in random networks), the characteristic path length is comparatively low with 2.2 (2.0 in random
networks). Similarly, human structural connectivity between brain regions shows small-world
properties with a small-worldness S of 10.6 (Text S2, Hagmann et al., 2008). For human functional
connectivity between brain regions, the clustering coefficient is 53% (22% in random networks) and
the path length 2.5 (2.3 in random networks) (Achard et al., 2006).
That is, on average only one or two intermediate areas are on the shortest path between two areas. An
anatomical basis for small-world features in neural networks is the preference for local short-distance
connections with only few long-distance connections (Kaiser et al., 2009). In that way, most
neighbors of a node are nearby and therefore have a higher probability to be connected (Kaiser and
Hilgetag, 2004a; Kaiser and Hilgetag, 2004b; Nisbach and Kaiser, 2007).
How can small-world networks arise during network development? First, small-world networks, in a
method described in the original article (Watts and Strogatz, 1998), could start from a regular (also
called lattice) network where neighbors are well connected and rewire connections of nodes by
randomly varying the node to which an edge connects to. In this way, edges that do not connect
neighbors but distant nodes in a network can be established; these edges are thus called ‘short cuts’. If
the probability that any edge will be rewired becomes too high, the network turns into a random
!
"(!
network. Second, networks with small-world properties can be generated in the reverse way starting
with random networks and slowly establishing higher neighborhood clustering (Kaiser, 2008). Such
networks can contain isolated nodes and leaf nodes (nodes with only one connection); patterns which
are unlikely to arise following the previous approach of starting with regular networks. Third,
networks could grow in two- or three-dimensional space with a preference for new nodes to connect
to spatially nearby nodes. Such a spatial growth can, in certain parameter regimes, lead to small-world
networks (Kaiser and Hilgetag, 2004a; Kaiser and Hilgetag, 2004b; Nisbach and Kaiser, 2007).
Note that a high clustering coefficient does not necessarily mean that a network contains multiple
clusters! Indeed, the standard model for generating small-world networks by rewiring regular
networks (Watts and Strogatz, 1998) does not lead to multiple clusters. In addition, small-world and
scale-free properties are compatible, but not equivalent; a network might be small-world but not scale-
free and vice versa.
5.4 Modular and hierarchical networks
Two central topological features of brain networks, in particular of the cerebral cortex, are their
modular and hierarchical organization. Modular networks consist of multiple clusters (cf. section 4.1
on clusters). If these clusters occur at different levels, a cluster consisting of multiple sub-clusters,
sub-clusters consisting of several sub-sub-clusters, and so on, the network can be called a hierarchical
modular network. Note that for only one level, the network would be modular but not hierarchical. On
the other hand, networks that are hierarchical but not modular seem impossible.
A modular hierarchical organization of cortical architecture and connections is apparent across many
scales, from cellular microcircuits in cortical columns (Binzegger et al., 2004; Mountcastle, 1997) at
the lowest level, via cortical areas at the intermediate scale, to clusters of highly connected brain
regions at the global systems level (Breakspear and Stam, 2005; Hilgetag et al., 2000a; Kaiser et al.,
2007a). The precise organization of these features at each level is still unknown, and there exists
controversy about the exact organization or existence of modules even at the level of cortical columns
(Horton and Adams, 2005; Rakic, 2008). Nonetheless, current data and concepts suggest that at each
level of neural organization clusters arise, with denser connectivity within than between the modules.
This means that neurons within a column, area or cluster of areas are more frequently linked with
each other than with neurons in the rest of the network.
Several potential biological mechanisms for generating hierarchical modular networks have been
described. One way is to start with an existing network and generate copies of duplicates of such a
network where the copies retain the same internal connectivity as the original network but also
establish connections directly to the original networks. Variations of this method can be used to
generate hierarchical scale-free networks (Ravasz and Barabási, 2003; Ravasz et al., 2002) and were
also thought to lead to cortical connectivity-like networks (Ebbesson, 1980; Krubitzer and Kahn,
2003). For modular networks, time windows during development can lead to multiple clusters where
the cluster number, cluster size, and inter-cluster-connectivity is determined by the number, width and
overlap of developmental time windows for synaptogenesis (Kaiser and Hilgetag, 2007; Nisbach and
Kaiser, 2007).
K>!63).-!E!%L-!8(+)*!8'&+%(-'!
The previous sections have looked at topological properties of neural networks but brain networks
also have spatial properties in that each node and edge has a three-dimensional location and extension
may it be volume for nodes or diameter and trajectory for edges. Given the spatial extent of network
components, space is often a limiting factor for the structural organization of neural systems. For
example, all-to-all connectivity between all neurons of the brain is impossible given the limited
volume available for white matter fiber tracts within the skull. In addition to the feasibility of network
topologies, the actual wiring between nodes can also inform us about functional constraints. For
example, long-distance connections are costly in that their establishment (material) and maintenance
!
")!
(action potential propagation) uses energy. On the other hand, long-distance connections can form
short-cuts that lead to faster information integration and, consequently, accelerated reaction time.
6.1 Connection lengths
Each node and edge in neural networks, at least after potential migration during development, has a
constant spatial position. Such a spatial layout is far from random but to what extent self-organization
or genetic predisposition determines location is still an open question. One first step in observing the
spatial organization of a neural network is to look at the lengths of connections. If two nodes are
connected, the Euclidean distance between the positions of both nodes can be a lower bound of the
length of the connection. Note, that even for cortical fiber tracts this gives a reasonable approximation:
for the prefrontal cortex in the macaque only 15% of the connections are strongly curved and dense
fibers, in particular, tend to be completely straight (Hilgetag and Barbas, 2006). It is often interesting
to observe how many connections go to nearby targets and how many extend over a long distance,
potentially linking different components of the neural network. This can be readily observed using a
histogram of the connection lengths of a network. These histograms for anatomical connection lengths,
ranging from C. elegans and rat neuronal to macaque and human cortical connectivity, all show a
decay of the frequency over distance: short-distance connections are more frequent than long-distance
connections (Figure 8). For these systems, the distribution can best be approximated through a
Gamma distribution (see (Kaiser et al., 2009) for details).
#
=
,
*
+
<
;
+
:
9
/
+
8
.
-
7
0
+
6
!'$
!'#&
!'#
!'"&
!'"
!'!&
!
!
!
!'#&
=
,
*
+
<
;
+
:
9
/
+
8
.
-
7
0
+
6
!'#
!'"&
!'"
!'!&
Human DTI
$
!'#&
=
,
*
+
<
;
+
:
9
/
+
8
.
-
7
0
+
6
!'#
!'"&
!'"
!'!&
Human
rsMRI
#!
%!
?!
>!
()**+,-.)*/0+*1-2/3445
"!!
"#!
#!
%!
?!
>!
()**+,-.)*/0+*1-2/3445
"!!
Macaque
cortical
"
=
,
*
+
<
;
+
:
9
/
+
8
.
-
7
0
+
6
!'$
!'#&
!'#
!'"&
!'"
!'!&
Rat
neuronal
!
!
"!
!
!
!'#
"
&!
"'#
!'%
!'?
!'>
#!
%!
$!
()**+,-.)*/0+*1-2/3445
()**+,-.)*/0+*1-2/3445
Fig. 8. Connection lengths in cortical and neuronal networks. Connection length distributions where
the relative counts of a histogram are plotted as data points (x) fitted with a Gamma function (solid
line, see Appendix for fits and coefficients). For cortical networks, only connectivity within a
hemisphere was considered. Despite different species, levels of organization, and types of
connectivity, all distributions show an early peak and a later distance-dependent decay in the
frequency of connection. (A) Human diffusion tensor imaging network between 55 brain regions
where each existing fiber tract is represented by one (unweighted) network connection. (B) Human
resting-state fMRI network between 55 brain regions where the top 20% of correlation are represented
by a network connection (that means, the threshold was set so that the edge density was 20%). (C)
Macaque cortical fiber-tract network of 95 brain regions. (D) Rat supragranular pyramidal cell
neuronal network of layers II and III of the extrastriate visual cortex. Subplots C and D adapted from
(Kaiser et al., 2009).
!
"*!
Establishing connections involves metabolic structural costs for building connections (especially for
myelinated axons) as well as dynamic costs for transmitting action potentials. It is therefore natural to
assume that these energy costs should be as low as possible!(Cherniak, 1992; Chklovskii et al., 2002;
Wen and Chklovskii, 2008). One possibility for reducing costs is to have a lower number of long-
distance connections. The frequency of such long-distance connections, relative to short-distance
connections, can be seen in the connection length histogram. It can also be tested how far away the
combined length of all connections together, the total wiring length, is from the shortest-possible total
wiring length. Such an optimal solution can be found in two ways: rearranging the connections of
each node whereas the position of a node remains the same or rearranging the position of each node,
swapping around the position of nodes, while retaining the connectivity of each node projecting to the
same target nodes (not target positions).
Reducing the total length by reordering connections could lead to minimal wiring of a system. One
possibility would be to establish connections ranked by their length; connecting nodes that are closest
to each other first. However, this could result in a fragmented network where parts of the network are
unreachable from many starting nodes. To secure reachability, start with a minimum spanning tree (cf.
glossary) that connects N nodes with N-1 edges and a minimal wiring length (Cormen et al., 2009),
then add remaining connections again using short-distance connections as described before.
Alternatively, wiring length reductions in neural systems can be achieved by suitable spatial
arrangement of the components. Under these circumstances, the connectivity patterns of neurons or
regions remain unchanged maintaining their structural and functional connectivity, but the layout of
components is perfected such that it leads to the most economical wiring. In the sense of
this ’component placement optimization’ (CPO;!(Cherniak, 1994)), any rearrangement of the position
of neural components, while keeping their connections unchanged, would lead to an increase of total
wiring length in the network. However, studies on neuronal networks in C. elegans and on cortical
networks in the macaque have shown that a reduction by 30-50% is possible!(Kaiser and Hilgetag,
2006). For a small number of nodes, all possible arrangements of their positions can be tested. For
larger networks, however, such an approach is not feasible: the number of possible layouts for N
nodes is N! (e.g. 10148 possibilities for 95 nodes). In those cases, optimal solutions can only be
approximated through numerical routines such as simulated annealing (Metropolis et al., 1953) or
others.
6.2 Missing links: using spatial and topological features for network reconstruction
As for other biological systems, incomplete data sets are a problem for brain connectivity studies. Are
there ways to predict whether a connection between two nodes exist? One possibility, tested for the
macaque fiber-tract and the C. elegans neuronal network, is to use local features of a pair of nodes to
predict whether they are connected or not (Costa et al., 2007a). Topological features were node degree,
clustering coefficient, characteristic path length, and Jaccard coefficient. Spatial or geometrical
features included local density of nodes, coefficient of variation of the nearest distances, Cartesian
coordinates of the nodes’ center of mass, as well as the area size for nodes in the cortical network.
Such an approach gave good estimates for reconstructing connections of the macaque visual cortex
(Costa et al., 2007a). The prediction performance could be further improved through varying the
contribution of each feature (training weights) (Nepusz et al., 2008).
M>!,&+.*$1(&+!
This tutorial is a first introduction to connectome analysis looking at topological and spatial features
of neural systems. There are several aspects of the connectome that are not covered here such as
complexity (Tononi et al., 1994, 1996), divergence and convergence of information (Tononi and
Sporns, 2003), and the comparison of types of connectivity (e.g. the link between structural and
functional connectivity (Honey et al., 2009)). For the hierarchical organization, we only looked at
topology but hierarchy also relates to the dynamics and spatial organization of neural systems!(Hutt
and Lesne, 2009; Jarvis et al., 2010; Kaiser and Hilgetag, 2010; Kiebel et al., 2009; Krumnack et al.,
2010; Meunier et al., 2009; Rodrigues and Costa, 2009; Zamora-Lopez et al., 2010).!We also only
!
#+!
observed a snapshot of the connectome; however, neural systems change during individual
development (Kaiser and Hilgetag, 2004b; Kaiser et al., 2009; Nisbach and Kaiser, 2007; van Ooyen,
2003; Van Ooyen, 2005), brain evolution (Ebbesson, 1980; Krubitzer and Kahn, 2003; Striedter,
2004), and throughout life through structural and functional plasticity (Butz et al., 2009; Friston et al.,
1994; Hubel et al., 1977; Sur and Leamey, 2001). In addition, we can observe the dynamics in neural
systems and simulate dynamics in network representations of the brain (Qubbaj and Jirsa, 2007). !
In this work, we have outlined how routines from network analysis can be applied to neural systems.
However, neural systems can also inform future work on the analysis of complex biological networks.
Neural systems present several challenges for the field of network analysis. Neural systems differ
from Erdös–Rényi and traditional small-world networks in that they are modular and hierarchical!
(Costa et al., 2011; Meunier et al., 2010; Sporns, 2011b). Network properties can be studied at
different levels ranging from connectivity between brain areas, connectivity within areas, connectivity
within columns (Binzegger et al., 2004), or connectivity of groups and ensembles. Another difference
with respect to standard network models is that nodes, although treated as uniform at the global level
of analysis, differ at the neuronal level in their response modality (excitatory or inhibitory), their
functional pattern due to the morphology of the dendritic tree and properties of individual synapses,
and their current threshold due to the history of previous excitation or inhibition. Such heterogeneous
node properties can also be expected at the global level in terms of the size and layer architecture of
cortical and subcortical regions. Theories where the properties and behavior of individual nodes differ,
beyond their pattern of connectivity, are still rare. Another theoretical challenge is the comparison of
network topologies and dynamics, e.g., between experiments and in silico studies!(Izhikevich and
Edelman, 2008; Kaiser et al., 2007a; Zhou et al., 2006).
In addition to theoretical challenges, network analysis also poses computational problems. The
analysis of experimental network data, such as of correlation network between voxels for MRI
recordings, can take a considerable amount of time. Whereas detecting all network motifs (Ribeiro et
al., 2009) in a single 100-node correlation network is computationally feasible, long recordings could
generate dozens of such correlation networks, which come with enormous computational demands.
Similar demands arise from larger networks such as region of interest (ROI) networks with around
1,000 nodes. These problems also occur for electrophysiological recordings: Multi-electrode units
with 4000 or more electrodes are now available. For the electrophysiology field, the CARMEN
Neuroinformatics project (http://www.carmen.org.uk ) addresses the storing, comparison, and analysis
of large data sets; similar initiatives for neuroimaging are clearly needed and some projects along
these directions already started. High-performance computing is also needed for large-scale
simulations of neural circuits, such as the Blue Brain project for simulating activity within a single
cortical column (Silberberg et al., 2005) or simulations of the whole brain (Izhikevich and Edelman,
2008).
In conclusion, connectome analysis provides several benefits to the field of neuroimaging. First, it
gives a network representation of a human brain in that the size, shape, and position of brain regions
are abstracted away. In this way, networks reduce the complexity of information yielded by
neuroimaging recordings!(Sporns, 2011a). Second, networks can be compared between humans. In
particular, network analysis can identify differences between the brains of patients and control
subjects. These changes can either be used for diagnosis of brain disorders or for evaluating treatment
strategies. Third, connectomes, together with properties of individual nodes and edges as well as input
patterns, form the structural correlate of brain function. Therefore, connectomes now increasingly
form the basis of simulations of brain dynamics. These benefits of the emerging field of connectome
analysis (DeFelipe, 2010) are now within reach due the availability of datasets, e.g. through the
Human Connectome Project or the 1000 Functional Connectome Project!(Biswal et al., 2010), the
range of network analysis tools, and the computational feasibility of network analysis and simulation.
!
#"!
!
N>!".<+&;*-54-/-+%1!
Part of this manuscript is based on lectures taught at Seoul National University and Newcastle
University and I thank students and members of my labs for helpful comments. M.K. was supported
by WCU program through the National Research Foundation of Korea funded by the Ministry of
Education, Science and Technology (R32-10142), the Royal Society (RG/2006/R2), the CARMEN e-
science project (www.carmen.org.uk) funded by EPSRC (EP/E002331/1), and (EP/G03950X/1).
Appendix A
Table A1
Runtime complexity: Assessing the speed of network analysis algorithms.
Network features differ in the amount of time and computer memory it takes to calculate them. The
runtime depends on the network feature, the implemented algorithm, the number of nodes N and the
number of edges E of the network, and the speed of the processor. The memory size depends on the
network representation and, in some cases, the organization of the algorithm. Using an adjacency
matrix needs N2 units of memory. Using adjacency lists saves memory using on the order of E units
but increases the runtime for several network measures when the edge density is very high.
Rather than using the actual calculation time, which strongly depends on processor speed, algorithms
are categorized by the runtime complexity depending on the network size. This is called asymptotic
analysis that looks at the growth of the running time instead of the absolute running time. For this, the
asymptotic O-notation (pronounced: Big-Oh), meaning ‘order-of’, is used for getting information
about the worst-case scenario of running time. The worst-case running time guarantees that an
algorithm will not go above this upper limit. The O-notation shows the order of calculation steps that
are performed by an algorithm. This means that only the largest terms that determine the runtime are
kept:
O(c N) = O(N)
constants are neglected
O(N2 + N + c) = O(N2)
only the largest term of a polynomial is used
Using this notation, algorithms belong to complexity classes P, NP-hard, or NP-complete!(Cormen et
al., 2009).
Rather than looking at computational complexity theory, I will provide some examples for how long
different network analysis algorithms take. Calculating the degree of a node takes O(N) steps when an
adjacency matrix is used and increases linearly with the number of nodes N. Note that for a directed
network, it takes 2N steps for counting all incoming (column) and outgoing (row) connections, but the
constant 2 is neglected in the O-Notation. Calculating the degree of all N nodes takes O(N2) steps and
calculating the shortest path between all pairs of nodes (characteristic path length) takes O(N3) steps;
both being examples for polynomial increase with the number of nodes. Calculating the characteristic
path is already a problem for large networks, but algorithms with non-polynomial runtime complexity
are even worse: testing whether two graphs are identical, a graph similarity problem that occurs in
motif analysis, takes N factorial, O(N!), steps whereas the travelling salesman problem of finding the
shortest metric path to visit N cities takes O(NN) steps.
For large networks, some measures will take too long to calculate leaving three options: using a
different network measure where calculations are computationally feasible, using parallel computing
which works for some network measures, or applying the measure to a subset (sample) of the
network. An example for sampling is the estimation of the characteristic path length of the world-
wide-web with 8$108 nodes based on a sample of a few thousand nodes (Albert et al., 1999). Note,
however, that sampling can become inaccurate especially when testing for power-law degree
distributions!(Stumpf et al., 2005).
!
##!
Table A2
Glossary of network analysis terms.
Adjacency (connection) matrix: The adjacency matrix of a graph is an n % n matrix with entries aij =
1 if node j connects to node i, and aij = 0 if there is no connection from node i to node j. Sometimes,
non-zero entries indicate the strength of a connection using ordinal scales for fiber strength or metric
scales [-1; 1] for correlation networks.
Average Shortest Path: The average shortest path ASP is the global mean of the entries of the
distance matrix. Normally, infinite values (non-existing paths) and values across the diagonal (loops)
are not taken into account for the calculation of the mean of the distance matrix.
Adjacency list: List where each line represents one edge with information about the source node, the
target node, and (optionally) the weight of the edge connecting both nodes.
Betweenness centrality: what proportion of all shortest paths are going through a node or edge of the
network. These values are then called node betweenness or edge betweenness, respectively. See
(Brandes and Erlebach, 2005) for other measures of centrality or influence.
Characteristic path length: The characteristic path length L (also called “path length”) is the median
of the mean shortest path length of all individual nodes (Watts, 1999).
Clustering coefficient: The clustering coefficient Ci of node i is the number of existing connections
between the node’s neighbors divided by all their possible connections. The clustering coefficient
ranges between 0 and 1 and is typically averaged over all nodes of a graph to yield the graph’s
clustering coefficient C.
Cycle: A path that links a node to itself.
Degree: The degree k of a node is the sum of its incoming (afferent) and outgoing (efferent)
connections. The number of afferent and efferent connections is also called the in-degree and out-
degree, respectively.
Degree distribution: Probability distribution of the degrees of all nodes of the network.
Degree sequence: The N-tuple (k1, …, kN) with ki as degree of node i and ki " ki+1 is called degree
sequence.
Distance: The distance between a source node i and a target node j is equal to the path length of the
shortest path. To talk about spatial distance instead, use the term metric distance.
Distance matrix: The entries dij of the distance matrix correspond to the distance between node i and
j. If no path exists, dij = &.
Graph: Graphs are a set of N nodes (vertices, points, units) and E edges (connections, arcs). Graphs
may be undirected (all connections are symmetrical) or directed. Because of the polarized nature of
most neural connections, we focus on directed graphs, also called digraphs.
Hub: Node with a degree that is much higher than for other nodes of the network. In sparse networks,
even the degree of a hub might be relatively low.
Kronecker symbol ! ,!!!
i ,! = 1 for i = j and
i ,! = 0 otherwise.
j
j
Loop: A connection of a node onto itself (in other words: a cycle of length 1).
!
#$!
Minimum spanning tree: A minimum spanning tree is a tree that connects all nodes of a network
with a weight less than or equal to the weight of every other spanning tree. In this context, the weight
is usually the total metric wiring length of the tree when the network is a spatial graph.
Modular graph: A network with multiple modules.
Motif: Sub-graph with a certain number of nodes (usually 3, 4, or 5) that occurs significantly more
often in a given network than in rewired networks with the same degree distribution.
Path: A path is an ordered sequence of distinct connections and nodes, linking a source node i to a
target node j. No connection or node is visited twice in a given path.
Path length: The length of a path is equal to the number of distinct connections.
Random graph: An Erdös-Rényi graph with uniform connection probabilities and a binomial degree
distribution. All node degrees are close to the average degree (“single-scale”).
Scale-free graph: Graph with a power-law degree distribution. “Scale-free” means that degrees are
not grouped around one characteristic average degree (scale), but can spread over a very wide range
of values, often spanning several orders of magnitude.
Small-world graph: A graph in which the clustering coefficient is much higher than in a comparable
random network, but the characteristic path length remains about the same. The term “small-world”
arose from the observation that any two persons can be linked over few intermediate acquaintances.
Spatial graph (or spatial network): A network where each node has a spatial position. Usually, nodes
in neural networks can have two- or three-dimensional positions in metric Euclidean space.
I-8-'-+.-1!
Achacoso, T.B., Yamamoto, W.S., 1992. AY’s Neuroanatomy of C. elegans for Computation. CRC
Press, Boca Raton, FL.
Achard, S., Bullmore, E., 2007. Efficiency and cost of economical brain functional networks. PLoS
Comput Biol 3, e17.
Achard, S., Salvador, R., Whitcher, B., Suckling, J., Bullmore, E., 2006. A resilient, low-frequency,
small-world human brain functional network with highly connected association cortical hubs. J
Neurosci 26, 63-72.
Ahn, Y.-Y., Bagrow, J.P., Lehmann, S., 2010. Link communities reveal multiscale complexity in
networks. Nature 466, 761-764.
Akil, H., Martone, M.E., Van Essen, D.C., 2011. Challenges and opportunities in mining neuroscience
data. Science 331, 708-712.
Albert, R., Jeong, H., Barabási, A.-L., 1999. Diameter of the World-Wide Web. Nature 401, 130-131.
Alon, U., 2003. Biological Networks: The Tinkerer as an Engineer. Science 301, 1866-1867.
Bettus, G., Wendling, F., Guye, M., Valton, L., Régis, J., Chauvel, P., Bartolomei, F., 2008. Enhanced
EEG functional connectivity in mesial temporal lobe epilepsy. Epilepsy Research 81, 58-68.
Binzegger, T., Douglas, R.J., Martin, K.A.C., 2004. A Quantitative Map of the Circuit of Cat Primary
Visual Cortex. J. Neurosci. 24, 8441-8453.
Biswal, B.B., Mennes, M., Zuo, X.N., Gohel, S., Kelly, C., Smith, S.M., Beckmann, C.F., Adelstein,
J.S., Buckner, R.L., Colcombe, S., Dogonowski, A.M., Ernst, M., Fair, D., Hampson, M., Hoptman,
M.J., Hyde, J.S., Kiviniemi, V.J., Kotter, R., Li, S.J., Lin, C.P., Lowe, M.J., Mackay, C., Madden,
D.J., Madsen, K.H., Margulies, D.S., Mayberg, H.S., McMahon, K., Monk, C.S., Mostofsky, S.H.,
Nagel, B.J., Pekar, J.J., Peltier, S.J., Petersen, S.E., Riedl, V., Rombouts, S.A., Rypma, B., Schlaggar,
B.L., Schmidt, S., Seidler, R.D., Siegle, G.J., Sorg, C., Teng, G.J., Veijola, J., Villringer, A., Walter,
M., Wang, L., Weng, X.C., Whitfield-Gabrieli, S., Williamson, P., Windischberger, C., Zang, Y.F.,
Zhang, H.Y., Castellanos, F.X., Milham, M.P., 2010. Toward discovery science of human brain
function. Proc Natl Acad Sci U S A 107, 4734-4739.
!
#%!
Bollobas, B., 1985. Random Graphs. Cambridge University Press.
Bonifazi, P., Goldin, M., Picardo, M.A., Jorquera, I., Cattani, A., Bianconi, G., Represa, A., Ben-Ari,
Y., Cossart, R., 2009. GABAergic Hub Neurons Orchestrate Synchrony in Developing Hippocampal
Networks. Science 326, 1419-1424.
Brandes, U., Erlebach, T., 2005. Network Analysis. Springer, Heidelberg.
Breakspear, M., Stam, C.J., 2005. Dynamics of a neural system with a multiscale architecture. Philos
Trans R Soc Lond B Biol Sci 360, 1051-1074.
Buckner, R.L., Sepulcre, J., Talukdar, T., Krienen, F.M., Liu, H., Hedden, T., Andrews-Hanna, J.R.,
Sperling, R.A., Johnson, K.A., 2009. Cortical Hubs Revealed by Intrinsic Functional Connectivity:
Mapping, Assessment of Stability, and Relation to Alzheimer's Disease. Journal of Neuroscience 29,
1860-1873.
Bullmore, E., Sporns, O., 2009. Complex brain networks: graph theoretical analysis of structural and
functional systems. Nat Rev Neurosci 10, 186-198.
Burns, G.A.P.C., Young, M.P., 2000. Analysis of the connectional organization of neural systems
associated with the hippocampus in rats. Phil. Trans. R. Soc. 355, 55-70.
Butz, M., Wörgötter, F., Van Ooyen, A., 2009. Activity-dependent structural plasticity. Brain
Research Reviews 60, 287-305.
Chavez, M., Valencia, M., Navarro, V., Latora, V., Martinerie, J., 2010. Functional Modularity of
Background Activities in Normal and Epileptic Brain Networks. Physical Review Letters 104.
Cherniak, C., 1992. Local optimization of neuron arbors. Biol Cybern 66, 503-510.
Cherniak, C., 1994. Component Placement Optimization in the Brain. J. Neurosci. 14, 2418-2427.
Chklovskii, D.B., Schikorski, T., Stevens, C.F., 2002. Wiring Optimization in Cortical Circuits.
Neuron 34, 341-347.
Clauset, A., Moore, C., Newman, M.E.J., 2008. Hierarchical structure and the prediction of missing
links in networks. Nature 453, 98-101.
Clauset, A., Shalizi, C.R., Newman, M.E.J., 2009. Power-Law Distributions in Empirical Data. SIAM
Review 51, 661-703.
Cormen, T.H., Leiserson, C.E., Rivest, R.L., Stein, C., 2009. Introduction to Algorithms, 3rd ed. MIT
Press, Cambridge.
Costa, L.D., Kaiser, M., Hilgetag, C.C., 2007a. Predicting the connectivity of primate cortical
networks from topological and spatial node properties. BMC Systems Biology 1.
Costa, L.d.F., Batista, J.L., Ascoli, G.A., 2011. Communication structure of cortical networks.
Frontiers in Computational Neuroscience 5, 6.
Costa, L.d.F., Rodrigues, F.A., Hilgetag, C.C., Kaiser, M., 2009. Beyond the average: Detecting
global singular nodes from local features in complex networks. Europhys. Lett. 87, 18008.
Costa, L.d.F., Rodrigues, F.A., Travieso, G., Villas Boas, P.R., 2007b. Characterization of complex
networks: A survey of measurements. Advances in Physics 56, 167-242.
DeFelipe, J., 2010. From the Connectome to the Synaptome: An Epic Love Story. Science 330, 1198-
1201.
Denk, W., Horstmann, H., 2004. Serial block-face scanning electron microscopy to reconstruct three-
dimensional tissue nanostructure. PLoS Biol 2, e329.
Diestel, R., 1997. Graph Theory. Springer, New York.
Dumas de la Roque, A., Oppenheim, C., Chassoux, F., Rodrigo, S., Beuvon, F., Daumas-Duport, C.,
Devaux, B., Meder, J.-F., 2004. Diffusion tensor imaging of partial intractable epilepsy. European
Radiology 15, 279-285.
Ebbesson, S.O.E., 1980. The Parcellation Theory and its Relation to Interspecific Variability in
Brain Organization, Evolutionary and Ontogenetic Development, and Neuronal Plasticity. Cell
Tissue Res. 213, 179-212.
Echtermeyer, C., Costa, L.d.F., Rodrigues, F.A., Kaiser, M., 2011. Automatic Network Fingerprinting
Through Singular Node Motifs. PLoS ONE 6, e15765.
Eguiluz, V.M., Chialvo, D.R., Cecchi, G.A., Baliki, M., Apkarian, A.V., 2005. Scale-free brain
functional networks. Phys Rev Lett 94, 018102.
Erdös, P., Rényi, A., 1960. On the evolution of random graphs. Publ. Math. Inst. Hung. Acad. Sci. 5,
17-61.
!
#&!
Felleman, D.J., van Essen, D.C., 1991. Distributed hierarchical processing in the primate cerebral
cortex. Cereb. Cortex 1, 1-47.
Friston, K.J., Tononi, G., Reeke, G.N., Jr., Sporns, O., Edelman, G.M., 1994. Value-dependent
selection in the brain: simulation in a synthetic neural model. Neuroscience 59, 229-243.
Girvan, M., Newman, M.E.J., 2002. Community Structure in Social and Biological Networks. Proc.
Natl. Acad. Sci. 99, 7821-7826.
Guimerà, R., Amaral, L.A.N., 2004. Modeling the world-wide airport network. The European
Physical Journal B - Condensed Matter 38, 381-385.
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C.J., Wedeen, V.J., Sporns, O., 2008.
Mapping the structural core of human cerebral cortex. PLoS Biol 6, e159.
He, Y., Chen, Z., Evans, A., 2008. Structural Insights into Aberrant Topological Patterns of Large-
Scale Cortical Networks in Alzheimer's Disease. Journal of Neuroscience 28, 4756-4766.
Hilgetag, C.C., Barbas, H., 2006. Role of Mechanical Factors in the Morphology of the Primate
Cerebral Cortex. PLoS Computational Biology 2, e22.
Hilgetag, C.C., Burns, G.A.P.C., O'Neill, M.A., Scannell, J.W., Young, M.P., 2000a. Anatomical
Connectivity Defines the Organization of Clusters of Cortical Areas in the Macaque Monkey and the
Cat. Phil. Trans. R. Soc. Lond. B 355, 91-110.
Hilgetag, C.C., Kaiser, M., 2004. Clustered organization of cortical connectivity. Neuroinformatics 2,
353-360.
Hilgetag, C.C., O'Neill, M.A., Young, M.P., 2000b. Hierarchical Organization of Macaque and Cat
Cortical Sensory Systems Explored with a Novel Network Processor. Philos. Trans. R. Soc. Lond.
Ser. B 355, 71-89.
His, W., 1888. Zur Geschichte des Gehirns sowie der Centralen und Peripherischen Nervenbahnen
beim menschlichen Embryo. Abhandlungen der mathematisch-physikalischen Classe der Königlichl.
Sachsichen Gesellschaft der Wissenschaften 14.
Honey, C.J., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J.P., Meuli, R., Hagmann, P., 2009.
Predicting human resting-state functional connectivity from structural connectivity. Proc Natl Acad
Sci U S A 106, 2035-2040.
Horton, J.C., Adams, D.L., 2005. The cortical column: a structure without a function. Philosophical
Transactions of the Royal Society B-Biological Sciences 360, 837-862.
Hubel, D.H., Wiesel, T.N., LeVay, S., 1977. Plasticity of Ocular Dominance Columns in Monkey
Striate Cortex. Philos. Trans. R. Soc. Lond. Ser. B 278, 377-409.
Hufnagel, L., Brockmann, D., Geisel, T., 2004. Forecast and Control of Epidemics in a Globalized
World. Proc. Natl. Acad. Sci. USA 101, 15124-15129.
Humphries, M.D., Gurney, K., 2008. Network 'small-world-ness': a quantitative method for
determining canonical network equivalence. PLoS ONE 3, e0002051.
Hutt, M.T., Lesne, A., 2009. Interplay between Topology and Dynamics in Excitation Patterns on
Hierarchical Graphs. Front Neuroinformatics 3, 28.
Iturria-Medina, Y., Sotero, R.C., Canales-Rodríguez, E.J., Alemán-Gómez, Y., Melie-García, L.,
2008. Studying the human brain anatomical network via diffusion-weighted MRI and Graph Theory.
Neuroimage 40, 1064-1076.
Izhikevich, E., Edelman, G., 2008. Large-scale model of mammalian thalamocortical systems. Proc
Natl Acad Sci U S A 105, 3593-3598.
Jarvis, S., Rotter, S., Egert, U., 2010. Extending stability through hierarchical clusters in echo state
networks. Front Neuroinformatics 4.
Kaiser, M., 2007. Brain architecture: a design for natural computation. Philos Transact A Math Phys
Eng Sci 365, 3033-3045.
Kaiser, M., 2008. Mean clustering coefficients: The role of isolated nodes and leafs on clustering
measures for small-world networks. New Journal of Physics 10, 083042.
Kaiser, M., Görner, M., Hilgetag, C.C., 2007a. Functional Criticality in Clustered Networks without
inhibition. New Journal of Physics 9, 110.
Kaiser, M., Hilgetag, C.C., 2004a. Modelling the Development of Cortical Networks.
Neurocomputing 58-60, 297-302.
Kaiser, M., Hilgetag, C.C., 2004b. Spatial growth of real-world networks. Phys Rev E Stat Nonlin
Soft Matter Phys 69, 036103.
!
#'!
Kaiser, M., Hilgetag, C.C., 2006. Nonoptimal Component Placement, but Short Processing Paths, due
to Long-Distance Projections in Neural Systems. PLoS Computational Biology 2, e95.
Kaiser, M., Hilgetag, C.C., 2007. Development of multi-cluster cortical networks by time windows
for spatial growth. Neurocomputing 70, 1829-1832.
Kaiser, M., Hilgetag, C.C., 2010. Optimal hierarchical modular topologies for producing limited
sustained activation of neural networks. Front Neuroinformatics 4, 8.
Kaiser, M., Hilgetag, C.C., van Ooyen, A., 2009. A simple rule for axon outgrowth and synaptic
competition generates realistic connection lengths and filling fractions. Cereb Cortex 19, 3001-3010.
Kaiser, M., Martin, R., Andras, P., Young, M.P., 2007b. Simulation of robustness against lesions of
cortical networks. Eur J Neurosci 25, 3185-3192.
Kiebel, S.J., Daunizeau, J., Friston, K.J., 2009. Perception and hierarchical dynamics. Front
Neuroinformatics 3, 20.
Krubitzer, L., Kahn, D.M., 2003. Nature versus Nurture Revisited: An Old Idea with a New Twist.
Prog. Neurobiol. 70, 33-52.
Krumnack, A., Reid, A.T., Wanke, E., Bezgin, G., Kötter, R., 2010. Criteria for optimizing cortical
hierarchies with continuous ranges. Front Neuroinformatics 4, 7.
Latora, V., Marchiori, M., 2001. Efficient Behavior of Small-World Networks. Phys. Rev. Lett. 87,
198701.
Lichtman, J.W., Livet, J., Sanes, J.R., 2008. A technicolour approach to the connectome. Nat Rev
Neurosci 9, 417-422.
Marcelino, J., Kaiser, M., 2009. Reducing influenza spreading over the airline network. PLoS Curr,
RRN1005.
May, R., Lloyd, A., 2001. Infection dynamics on scale-free networks. Physical Review E 64.
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E., 1953. Equation of State
Calculations by Fast Computing Machines. Journal of Chemical Physics 21, 1087-1092.
Meunier, D., Lambiotte, R., Bullmore, E.T., 2010. Modular and hierarchically modular organization
of brain networks. Front Neurosci 4, 200.
Meunier, D., Lambiotte, R., Fornito, A., Ersche, K.D., Bullmore, E.T., 2009. Hierarchical modularity
in human brain functional networks. Front Neuroinformatics 3, 37.
Micheloyannis, S., Pachou, E., Stam, C.J., Breakspear, M., Bitsios, P., Vourkas, M., Erimaki, S.,
Zervakis, M., 2006. Small-world networks and disturbed functional connectivity in schizophrenia.
Schizophrenia Research 87, 60-66.
Milgram, S., 1967. The Small-World Problem. Psychology Today 1, 60-67.
Milo, R., Shen-Orr, S., Itzkovitz, S., Kashtan, N., Chklovskii, D., Alon, U., 2002. Network Motifs:
Simple Building Blocks of Complex Networks. Science 298, 824-827.
Mountcastle, V.B., 1997. The columnar organization of the neocortex. Brain 120 ( Pt 4), 701-722.
Nepusz, T., Négyessy, L., Tusnády, G., Bazsó, F., 2008. Reconstructing cortical networks: case of
directed graphs with high level of reciprocity. Handbook of Large-Scale Random Networks, 325-368.
Newman, M.E.J., 2004. Fast algorithm for detecting community structure in networks. Phys Rev E
Stat Nonlin Soft Matter Phys 69, 066133.
Newman, M.E.J., 2006. Modularity and community structure in networks. Proc. Natl. Acad. Sci. U. S.
A. 130, 8577-8582.
Newman, M.E.J., Strogatz, S.H., Watts, D.J., 2001. Random Graphs with Arbitrary Degree
Distributions and their Applications. Phys. Rev. E 64, 026118.
Nisbach, F., Kaiser, M., 2007. Developmental time windows for spatial growth generate multiple-
cluster small-world networks. European Physical Journal B 58, 185-191.
Palla, G., Derenyi, I., Farkas, I., Vicsek, T., 2005. Uncovering the Overlapping Community Structure
of Complex Networks in Nature and Society. Nature 435, 814-818.
Qubbaj, Jirsa, 2007. Neural field dynamics with heterogeneous connection topology. Phys Rev Lett
98, 238102.
Rakic, P., 2008. Confusing cortical columns. Proc Natl Acad Sci U S A 105, 12099-12100.
Ramón y Cajal, S., 1892. The Structure of the Retina. Thomas, Springfield, Illinois, 1972.
Ravasz, E., Barabási, A.-L., 2003. Hierarchical Organization in Complex Networks. Phys. Rev. E 67,
026112.
!
#(!
Ravasz, E., Somera, A.L., Mongru, D.A., Oltvai, Z.N., Barabási, A.-L., 2002. Hierarchical
Organization of Modularity in Metabolic Networks. Science 297, 1551-1555.
Ribeiro, P., Silva, F., Kaiser, M., 2009. Strategies for Network Motifs Discovery. IEEE Conference
on e-Science, Oxford, pp. 80-87.
Rodrigues, F.A., Costa, L.d.F., 2009. Signal propagation in cortical networks: a digital signal
processing approach. Front Neuroinformatics 3, 24.
Rubinov, M., Sporns, O., 2010. Complex network measures of brain connectivity: Uses and
interpretations. Neuroimage 52, 1059-1069.
Scannell, J.W., Blakemore, C., Young, M.P., 1995. Analysis of Connectivity in the Cat Cerebral
Cortex. J. Neurosci. 15, 1463-1483.
Seung, H.S., 2009. Reading the book of memory: sparse sampling versus dense mapping of
connectomes. Neuron 62, 17-29.
Silberberg, G., Grillner, S., LeBeau, F.E.N., Maex, R., Markram, H., 2005. Synaptic pathways in
neural microcircuits. Trends Neurosci 28, 541-551.
Skudlarski, P., Jagannathan, K., Anderson, K., Stevens, M.C., Calhoun, V.D., Skudlarska, B.A.,
Pearlson, G., 2010. Brain Connectivity Is Not Only Lower but Different in Schizophrenia: A
Combined Anatomical and Functional Approach. Biological Psychiatry 68, 61-69.
Sporns, O., 2011a. The human connectome: a complex network. Annals of the New York Academy of
Sciences 1224, 109–125.
Sporns, O., 2011b. The Non-Random Brain: Efficiency, Economy, and Complex Dynamics. Frontiers
in Computational Neuroscience 5, 5.
Sporns, O., Chialvo, D.R., Kaiser, M., Hilgetag, C.C., 2004. Organization, development and function
of complex brain networks. Trends Cogn Sci 8, 418-425.
Sporns, O., Honey, C.J., Kötter, R., 2007. Identification and classification of hubs in brain networks.
PLoS ONE 2, e1049.
Sporns, O., Kötter, R., 2004. Motifs in brain networks. PLoS Biol 2, e369.
Sporns, O., Tononi, G., Edelman, G.M., 2000. Theoretical Neuroanatomy: Relating Anatomical and
Functional Connectivity in Graphs and Cortical Connection Matrices. Cereb. Cortex 10, 127-141.
Sporns, O., Tononi, G., Kötter, R., 2005. The human connectome: A structural description of the
human brain. PLoS Comput Biol 1, e42.
Stam, C.J., Jones, B.F., Nolte, G., Breakspear, M., Scheltens, P., 2007. Small-world networks and
functional connectivity in Alzheimer's disease. Cereb Cortex 17, 92-99.
Stauffer, D., Aharony, A., 2003. Introduction to percolation theory, Rev. 2. ed. Routledge, London.
Stephan, K.E., Hilgetag, C.C., Burns, G.A.P.C., O'Neill, M.A., Young, M.P., Kötter, R., 2000.
Computational analysis of functional connectivity between areas of primate cerebral cortex. Phil.
Trans. R. Soc. 355, 111-126.
Striedter, G.F., 2004. Principles of Brain Evolution. Sinauer.
Stumpf, M.P.H., Wiuf, C., May, R.M., 2005. Subnets of Scale-Free Networks are Not Scale-Free:
Sampling Properties of Networks. Proc. Natl. Acad. Sci. USA 102, 4221-4224.
Sur, M., Leamey, C.A., 2001. Development and Plasticity of Cortical Areas and Networks. Nature
Rev. Neurosci. 2, 251-262.
Tononi, G., Sporns, O., 2003. Measuring information integration. BMC Neurosci 4, 31.
Tononi, G., Sporns, O., Edelman, G.M., 1994. A measure for brain complexity: relating functional
segregation and integration in the nervous system. Proc Natl Acad Sci U S A 91, 5033-5037.
Tononi, G., Sporns, O., Edelman, G.M., 1996. A complexity measure for selective matching of
signals by the brain. Proc Natl Acad Sci U S A 93, 3422-3427.
Tuch, D.S., Reese, T.G., Wiegell, M.R., Wedeen, V.J., 2003. Diffusion MRI of Complex Neural
Architecture. Neuron 40, 885-895.
van den Heuvel, M.P., Mandl, R.C.W., Stam, C.J., Kahn, R.S., Hulshoff Pol, H.E., 2010. Aberrant
Frontal and Temporal Complex Network Structure in Schizophrenia: A Graph Theoretical Analysis.
Journal of Neuroscience 30, 15915-15926.
van den Heuvel, M.P., Stam, C.J., Kahn, R.S., Hulshoff Pol, H.E., 2009. Efficiency of Functional
Brain Networks and Intellectual Performance. Journal of Neuroscience 29, 7619-7624.
Van Essen, D.C., Drury, H.A., 1997. Structural and Functional Analyses of Human Cerebral Cortex
Using a Surface-Based Atlas. J. Neurosci. 17, 7079-7102.
!
#)!
van Ooyen, A., 2003. Modeling Neural Development. MIT Press.
Van Ooyen, A., 2005. Competition in neurite outgrowth and the development of nerve connections.
Prog Brain Res 147, 81-99.
van Wijk, B.C.M., Stam, C.J., Daffertshofer, A., 2010. Comparing Brain Networks of Different Size
and Connectivity Density Using Graph Theory. PLoS ONE 5, e13701.
Varier, S., Kaiser, M., 2011. Neural development features: Spatio-Temporal development of the C.
elegans neuronal network. PLoS Comput Biol 7, e1001044.
Vázquez, A., Pastor-Satorras, R., Vespignani, A., 2002. Large-scale topological and dynamical
properties of the Internet. Physical Review E 65.
Wasserman, S., Faust, K., 1994. Social network analysis : methods and applications. Cambridge Univ.
Press, Cambridge.
Watts, D.J., 1999. Small Worlds. Princeton University Press.
Watts, D.J., Strogatz, S.H., 1998. Collective Dynamics of 'small-World' Networks. Nature 393, 440-
442.
Waxman, B.M., 1988. Routing of Multipoint Connections. IEEE J. Sel. Areas Commun. 6, 1617-
1622.
Wen, Q., Chklovskii, D.B., 2008. A cost-benefit analysis of neuronal morphology. Journal of
Neurophysiology 99, 2320-2328.
Wernicke, S., Rasche, F., 2006. FANMOD: a tool for fast network motif detection. Bioinformatics 22,
1152-1153.
White, J.G., Southgate, E., Thomson, J.N., Brenner, S., 1986. The structure of the nervous system of
the nematode C. elegans. Phil. Trans. R. Soc. of London B 314, 1.
Wuchty, S., Oltvai, Z.N., Barabási, A.-L., 2003. Evolutionary Conservation of Motif Constituents in
the Yeast Protein Interaction Network. Nature genetics 35, 176-179.
Yook, S.-H., Jeong, H., Barabási, A.-L., 2002. Modeling the Internet's Large-Scale Topology. Proc.
Natl. Acad. Sci. 99, 13382-13386.
Young, M.P., 1992. Objective Analysis of the Topological Organization of the Primate Cortical
Visual System. Nature 358, 152-155.
Zamora-Lopez, G., Zhou, C., Kurths, J., 2010. Cortical hubs form a module for multisensory
integration on top of the hierarchy of cortical networks. Front Neuroinformatics 4, 1.
Zhou, C., Zemanova, L., Zamora, G., Hilgetag, C.C., Kurths, J., 2006. Hierarchical organization
unveiled by functional connectivity in complex brain networks. Phys Rev Lett 97, 238103.
!
#*!
|
1705.10478 | 3 | 1705 | 2017-08-07T18:24:45 | Adaptive Estimation of the Neural Activation Extent in Computational Volume Conductor Models of Deep Brain Stimulation | [
"q-bio.NC",
"cs.CE"
] | Objective: The aim of this study is to propose an adaptive scheme embedded into an open-source environment for the estimation of the neural activation extent during deep brain stimulation and to investigate the feasibility of approximating the neural activation extent by thresholds of the field solution. Methods: Open-source solutions for solving the field equation in volume conductor models of deep brain stimulation and computing the neural activation are embedded into a Python package to estimate the neural activation dependent on the dielectric tissue properties and axon parameters by employing a spatially adaptive scheme. Feasibility of the approximation of the neural activation extent by field thresholds is investigated to further reduce the computational expense. Results: The varying extents of neural activation for different patient-specific dielectric properties were estimated with the adaptive scheme. The results revealed the strong influence of the dielectric properties of the encapsulation layer in the acute and chronic phase after surgery. The computational time required to determine the neural activation extent in each studied model case was substantially reduced. Conclusion: The neural activation extent is altered by patient-specific parameters. Threshold values of the electric potential and electric field norm facilitate a computationally efficient method to estimate the neural activation extent. Significance: The presented adaptive scheme is able to robustly determine neural activation extents and field threshold estimates for varying dielectric tissue properties and axon diameters while reducing substantially the computational expense. | q-bio.NC | q-bio | Adaptive Estimation of the Neural Activation Extent
in Computational Volume Conductor Models of
Deep Brain Stimulation
Christian Schmidt∗, Ursula van Rienen
7
1
0
2
g
u
A
7
]
.
C
N
o
i
b
-
q
[
3
v
8
7
4
0
1
.
5
0
7
1
:
v
i
X
r
a
Abstract-Objective: The aim of this study is to propose an
adaptive scheme embedded into an open-source environment for
the estimation of the neural activation extent during deep brain
stimulation and to investigate the feasibility of approximating
the neural activation extent by thresholds of the field solution.
Methods: Open-source solutions for solving the field equation
in volume conductor models of deep brain stimulation and
computing the neural activation are embedded into a Python
package to estimate the neural activation dependent on the
dielectric tissue properties and axon parameters by employing
a spatially adaptive scheme. Feasibility of the approximation of
the neural activation extent by field thresholds is investigated to
further reduce the computational expense. Results: The varying
extents of neural activation for different patient-specific dielectric
properties were estimated with the adaptive scheme. The results
revealed the strong influence of the dielectric properties of the
encapsulation layer in the acute and chronic phase after surgery.
The computational time required to determine the neural activa-
tion extent in each studied model case was substantially reduced.
Conclusion: The neural activation extent is altered by patient-
specific parameters. Threshold values of the electric potential and
electric field norm facilitate a computationally efficient method to
estimate the neural activation extent. Significance: The presented
adaptive scheme is able to robustly determine neural activation
extents and field threshold estimates for varying dielectric tissue
properties and axon diameters while reducing substantially the
computational expense.
Index Terms-Deep brain stimulation (DBS), Finite element
methods, Neural activation, Open-source
D EEP BRAIN STIMULATION (DBS) is a widely em-
ployed effective procedure to treat symptoms of motor
disorders such as Parkinson's disease (PD), essential tremor
and dystonia [1], [2] and consists in the implantation of an
electrode lead into deep brain target areas. A common target
in DBS is the subthalamic nucleus (STN), which constitutes
a preferred target for the treatment of PD. The STN consists
of different functional zones, which are classified into limbic,
associative, and sensorimotor zones [3], from which electrical
stimulation of the sensorimotor zone is mostly associated with
the relief of motor symptoms of PD [4]. Due to patient-specific
parameters, such as brain structure anatomy, dielectric tissue
properties, electrode location, and severity of symptoms, the
adjustment and optimization of stimulation parameters during
and after surgery can be rather time consuming and connected
Manuscript submitted to IEEE TBME on April, 26 2017.
Asterisk indicates corresponding author.
∗C. Schmidt
Institute of General Electrical Engi-
neering, University of Rostock, 18059 Rostock, Germany (e-mail:
[email protected])
is with the
U. van Rienen is with the Institute of General Electrical Engineering,
University of Rostock, 18059 Rostock, Germany
to additional costs. Computational models provide a possibility
to estimate the stimulation impact by determining activated
areas in the deep brain based on the given patient-specific
parameters. The extent of neural activation, or the volume of
tissue activated (VTA), is a common computational modeling
approach to estimate the size of the activated tissue during
DBS and has been applied in various computational studies in
this area including homogeneous [5], rotationally symmetric
[6], heterogeneous [7], [8], and anisotropic volume conductor
models [9] of the human brain and deep brain target areas.
In general, the approach is based on positioning a number
of models of mammalian nerve fibers (axon models) in a
grid located in a plane perpendicular to the electrode lead.
For each axon model the computational goal of finding the
minimum stimulation amplitude required to activate the axon
is solved. From the resulting threshold values at the grid
points, a threshold isoline for a given stimulation amplitude
is determined. This procedure is repeated for multiple planes
rotated around the electrode lead. In case of a rotationally
symmetric field model, it is sufficient to compute the threshold
isoline in one plane and revolve the solution around the elec-
trode lead. The resulting threshold isolines then provide the
measure for computing the VTA. The drawback of the method
with respect to computational ressources and adaptivity is
that the location of the axon nodes to include a range of
desired stimulation amplitudes has to be available prior to the
simulation, which involves several pre-simulation runs, which
often are carried out manually. The field solution in the target
area is commonly computed by creating a volume conductor
model of the DBS electrode and the surrounding tissue.
For solving the governing equations, which are typically the
stationary current field or electro-quasistatic equation for DBS
applications [7], [10], often commercial software solutions,
such as COMSOL Multiphysics R(cid:13) (http://www.comsol.com)
are used [5], [7], [8], [11], [12], [13], while, to our knowledge,
no studies employing open-source solutions on the field model
have been published in scientific journals yet. Regarding the
coupling of the neuronal activation in axon models and the
extracellular field distribution, a Python package with the
purpose to compute the local field potentials for a given axon
distribution of defined activation was presented in [14]. To
date, no open-source solution to model the field distribution
during DBS and to estimate the resulting neural activation
extent exists.
The computation of the VTA is a computationally de-
manding task. In a previous computational study investigating
the relationship between the neural activation and the field
solution, an approximation of the extent of neural activation
by threshold values of the field solution and their derivates
have been investigated [13]. The results suggested that electric
field norm thresholds are a good estimator for the extent of
neural activation. In such a case only a small neural activation
extent has to be computed to get the initial field norm threshold
estimate, from which the neural activation extents for varying
stimulation amplitudes can be derived. Especially for large
diameter axon fibers, the electric field norm constituted a good
estimate. The relationship between electric field norm and
neural activation was determined by positioning axons normal
to the electrode lead along a line originating at the active
electrode contact center in a homogeneous volume conductor
model for DBS. The threshold value of the electric field norm
was equivalent to its value at the maximum neural activation
distance for a given stimulation amplitude. The assumption of
this approach is that the shape of the neural activation extent is
spherical, which is true for a homogeneous volume conductor
model and a spherical or point source. For heterogeneous
volume conductor models and DBS electrode geometries, the
shape deviates from the spherical shape.
The goal of this study is to investigate the feasibility of
approximating the neural activation extent by the field solution
of a heterogeneous volume conductor model for DBS incorpo-
rating a DBS electrode, encapsulation layer, brain tissue, and
the STN as target area. To automate the task of determining the
neural activation extent and further reduce the computational
demand, an algorithm is proposed which determines adaptively
the location of axons being activated within a defined stim-
ulation amplitude range or distance range. Besides dropping
the need for manually determining the number of axons in
the target area for a given stimulation amplitude range, the
approach reduces the computational expense by omitting the
computation of activation for axons which are located outside
the activation volume. The model pipeline for the computation
of the field solution and the neural activation is implemented
in a Python package and embedding open-source tools for the
model generation, meshing, and solving. The field solution
and the neural activation are validated using analytical models
as well as reference data published in literature. The Python
package is designed modular, which allows to interchange field
as well as neuron models and to adjust model parameters
accordingly. The Python package, as well as the code to
replicate the data and figures of this study are made available
open-source1.
A. Model geometry
I. METHODS
Following the approach in [15], the model geometry con-
sists of a DBS electrode model located in a bounding box
comprising the different tissue compartments. The geometry
of the DBS electrode represents a Medtronic lead model
3387 (Medtronic, Inc., Minneapolis, MN). To account for the
inflammatory response of the body tissue to the electrode im-
plant, an encapsulation layer with a thickness of 0.2 mm was
1https://bitbucket.org/ChrSchmidt83/fanpy/get/fanpy-1.2.zip
2
Fig. 1. The model geometry consists of a DBS electrode, an encapsulation
layer around the DBS electrode, and a STN model. The model compartments
are surrounded by a bounding box.
Fig. 2. Manually refined mesh of the computational domain in the xz-plane
for y = 0. The different computational subdomains including the bounding
box, electrode lead, electrode contacts, encapsulation layer, and STN, are
shown with different colors.
incorporated around the electrode body. The bounding box size
was determined by an edge length of 100 mm. The geometry
model of the STN based on a functional zones atlas [16] was
generated by creating a surface model out of the right STN
threshold maps of the atlas using the open-source software
platform 3D Slicer (https://www.slicer.org/, version 4.2.1).
The whole geometry model was generated with the open-
source software SALOME (http://www.salome-platform.org/,
version 7.8.0). After creating and merging of the different
compartments, the STN surface model was converted to a
solid and positioned in the geometry with the second electrode
contact of the DBS electrode located in the sensorimotor zone
(Fig. 1).
B. Manual mesh refinement and subdomain generation
The meshing of the computational domain is carried out
with the open-source mesh generator Gmsh (http://gmsh.info,
version 2.10.1). Therefore, the model geometry is exported
from SALOME in the brep format and loaded into gmsh.
Since the geometry contains entities of varying scales (Fig. 1),
the mesh was manually refined at the surfaces of the entities
by specifying a characteristic length of the finite elements.
Based on values for the manual mesh refinement for DBS
volume conductor models of the human brain [9], which
provided a sufficient refinement of the computational domain,
the characteristic length was set to 0.1 mm for the electrode
lead, electrode contacts, and encapsulation layer, to 0.2 mm
for the STN, and 5.0 mm for the bounding box. Additionally,
a refined cubical mesh region with an edge length of 20 mm
100nmm100nmm1.5nmm1.5nmm1.27nmmGeometrynModelDBSnElectrodenModelSubthalamicNucleusxyzElectrodecontact432123 mm14.5 mmxyz 4(cid:88)
i=1
and a characteristic length of 0.5 mm was defined in the target
area. The final mesh contained approximately 240,000 vertices
and 1.5 million cells (Fig. 2). In order to assign material
properties and boundary conditions to the model compartments
and surfaces, the subdomains of the geometry model were
assigned and grouped to physical volumes and surfaces. The
information on the manual mesh refinement and the defined
physical volumes and surfaces is stored in a geo file, from
which automatically the mesh can be generated.
C. Incorporating the model into FEniCS
∇ · [σ(r)∇(ϕ(r))] = 0, r ∈ Ω
The mesh generated with Gmsh is converted into a for-
mat readable by the open-source simulation software FEniCS
(https://fenicsproject.org, version 2016.2.0) by using the com-
mand line tool dolfin-convert. To faciliate the inter-
change of different model geometries and meshes, information
on the model's subdomains, boundaries, and default boundary
conditions and material properties are defined in an xml
file. The FieldModel module of the designed Python pack-
age loads the definitions in the model xml file, checks the
information for consistency, generates the mesh and applies
the material properties and boundary conditions. While the
conductivity of the electrode lead (insulation) and the electrode
contacts (platinum-iridium) are kept constant with a value
of 1 · 10−7 Sm−1 for the electrode lead, and 1 · 107 Sm−1
for the electrode contacts,
the conductivity values for the
encapsulation layer, brain tissue, and STN varied for the
different study cases. The field equation is determined by a
stationary current field problem, which is commonly applied
in various computational modeling studies for DBS [13], [15]:
(1)
with the conductivity σ,
the electric potential ϕ and the
computational domain Ω. If the capacitive and dispersive
properties of human tissue are taken into account, the field
equation (1) has to be reformulated for complex materials with
(2)
with the imaginary unit j, the angular frequency ω, the electric
field constant 0, and the relative electric permittivity r, which
placed in equation (1) resembles a quasistatic field problem.
The field equations are solved with the finite element method
by formulating the variational problem within FEniCS. Dirich-
let boundary conditions are applied to the surface of the second
electrode contact, located within the STN (Fig. 1), with a unit
value of 1.0 V, and the exterior boundary of the bounding
box with a value of 0.0 V, serving as ground. The resulting
linear system of equations was solved for quadratic nodal basis
functions using the generalized minimal residual method with
a relative tolerance of 1 · 10−6 and an absolute tolerance of
1 · 10−7, employing an algebraic multigrid preconditioner in
case of the stationary current field problem. It was ensured
that the deviation in the electric potential and electric field
norm in the prescribed activation distances (see section I-H)
as well as the impedance of the model was below 1 % if cubic
ansatz functions were employed. Plots of the field distribution
were visualized with the open-source visualization application
Paraview (http://www.paraview.org/, version=5.0.1).
σc(ω, r) := σ(ω, r) + jω0r(ω, r)
D. Electrical properties of human tissue
3
The conductivity and relative permittivity of biological
tissue show a frequency dependence which can be described
by different dispersion regions and parametrized by assembled
Cole-Cole equations [17] representing a complex conductivity
σc(ω) = ∞ +
σion
jω0
+
∆i
1 + (jωτi)1−αi
(3)
with the static ionic conductivity σion, the relaxation time
the dispersion constant αi ∈ [0, 1], and the
constants τi,
relative permittivity at high frequency ∞ as well as the
difference of the low and high frequency relative permittivity
∆i. The tissue model parameters were taken for white matter,
grey matter, and cerebrospinal fluid from [17]. The electrical
tissue properties of the encapsulation layer vary over time
from an acute phase immediately after surgery to a chronic
phase after some weeks due to cell growth in the layer [18].
The acute phase was modeled by the dielectric properties of
cerebrospinal fluid [7], while the chronic phase was modeled
by dividing the values for white matter by a factor of 2 [15].
E. Voltage-controlled and current-controlled stimulation
The time-dependent electrical potential in the target area
for a given stimulation signal was determined by forming the
outer product of the field solution for a unit voltage set to the
active second electrode contact (Fig. 1) and the voltage- or
current-controlled stimulation signal. The stimulation signals
commonly applied in DBS therapy in humans consist of a
monophasic square-wave signal with pulse durations in the
range of 60 µs - 100 µs and a repetition frequency in the
range of 130 Hz - 150 Hz [15], [13], [7], [19]. To avoid
charge accumulation in the tissue, the monophasic stimulation
pulse is often followed by a reversed charge-balancing pulse
of substantially smaller amplitude compared to the active
stimulation pulse. Considering that the activation of a neuron
is mainly influenced by the amplitude of the stimulation pulse
[7], the reversed charge-balancing pulse is not considered in
this study. While the time-dependent electrical potential in the
target area for voltage-controlled stimulation is provided by
the outer product of the field solution for a unit voltage and
the voltage-controlled stimulation signal, the time-dependent
electrical potential for current-controlled stimulation requires
an additional scaling of the field solution by the electrode
impedance, which is computed by dividing the square of the
unit voltage by the electric power P of the field model.
(cid:90)
P =
(cid:104)σ(r)∇ϕ(r),∇ϕ(r)(cid:105) dx
(4)
Ω
with the inner product (cid:104),(cid:105), corresponding to a scaling of the
unit voltage at the active electrode contact boundary condition
by a factor equal to a unit current flowing through its surface.
F. Neuronal activation model
The neural activation model is based on a myelinated axon
cable model, which includes 21 nodes of Ranvier, paranodal
and internodal segmenets as well as the myelin sheath [20],
following the assupmtion that activation occurs along the axon
[12]. Based on the model parameters given in [20], the model
is parametrized with respect to the fiber diameter comprising
nine distinct diameters between 5.7 µm and 16.0 µm, which
were extented by the parameters for 2.0 µm and 3.0 µm
fiber diameter taken from [21]. The model is implemented
in the open-source simulation environment NEURON (https:
//www.neuron.yale.edu/neuron/, version=7.4)2 with the time-
dependent electrical potential for an applied stimulation signal
at the location of each axon compartment applied to its extra-
cellular mechanism node neglecting any axon contribution to
the extracellular field distribution [15]. The axon activation
is determined by solving the linear system of differential
equations resulting from the membrane dynamics with the
backward Euler method for a time step of 5 µs [15]. An axon
was considered to be activated when the inner potential at the
exterior nodes of Ranvier of the model obtained a threshold
value of larger than 0 mV, representing a generated spike as
result of the stimulation pulse, in a 1-to-1 ratio for 10 delivered
stimulation pulses.
G. Adaptive estimation of the neural activation extent
The proposed algorithm determines adaptively the required
location of axons in the target area for a given range of
stimulation amplitudes, which ensures that the extent of neural
activation for any stimulation amplitude within the given range
can be computed from the determined axon locations. The
required axon locations are determined in cutting planes,
which are located around the electrode lead. First, a seed
point, located at a distance of 0.85 mm to the active elec-
trode contact's center, which corresponds to the extent of the
electrode with the encapsulation thickness, is placed in such
a plane. Next, the algorithm determines whether the axon,
which is positioned perpendicular and centered to the electrode
lead at the seed point location, is activated for the minimal
given stimulation amplitude. If an activation was recorded,
the algorithm continues by placing further axons around the
activated axon, with their center node location positioned
radial ( + ∆s) and parallel (z + ∆s, z − ∆s) to the electrode
lead with a step size ∆s, and determining their activation.
If the axon was not activated, the algorithm stops for this
location. The procedure is continued until an inactivated hull
of axons is determined, which is achieved if in each line radial
to the electrode the axon furthest away to the electrode is
not activated by the stimulation. This inactivated hull is then
used as seed points for determining the axon locations for
the maximum stimulation amplitude within the given range
by applying the same algorithm. Finally, the interior points,
which are located inside the activated hull of axons for the
minimum given stimulation amplitude are removed from the
set of locations, resulting in a shell of axon locations for the
given stimulation amplitude range. Determining the minimally
required stimulation amplitude to elicit an action potential in
the axons at these locations represents a root-finding problem,
which is solved using the bisection method (binary search
method) with a tolerance of 1· 10−6 (Fig. 3). The algorithm is
2https://senselab.med.yale.edu/modeldb/showModel.cshtml?model=3810
4
Fig. 3.
Illustration of the adaptive algorithm for the estimation of the neural
activation extent in the range of given stimulation amplitudes. A: Seed points
are placed in front of the active stimulation electrode (here one seed point for
monopolar stimulation). If the axon is activated by the stimulation new seed
points are placed around the activated axon. If the axon is not activated, the
algorithm stops to place new seed points around this axon. The procedure is
continued until a closed hull of inactivated axons is found. B: For a given
stimulation amplitude range, the algorithm determines closed activation hulls
for the minimum and maximum stimulation amplitude and removes unneeded
interior points. For a given tolerance, the minimally required stimulation
amplitude is computed at each point resulting in a threshold map. From this
map, the activation isolines can be computed for any stimulation amplitude
in the given range. The procedure is repeated for several planes around the
stimulation electrode lead, from which the extent of neural activation is finally
estimated. The threshold map computed for the non-adaptive approach is
shown as comparison.
carried out for nα = (cid:100)360/∆α(cid:101) planes around the electrode
with the rotational degree step size ∆α. The subsequent
computation of the neural activation for each axon model
introduced by the algorithm as well as finding the minimally
required stimulation amplitude to activate the axon model are
carried out by the model pipeline in parallel with worker
threads adding and withdrawing axon points to a pool (Python
Queue) of axons. In order to prevent the placement of new
axon points at the same location to the pool, a new axon point
is only added if an axon point with the same location was not
added before by another worker thread.
H. Approximating the neural activation by field thresholds
The proposed adaptive algorithm is used to assess the
feasibility of approximating the neural activation extent for
various stimulation amplitudes by threshold values of the
electrical potential and the electric field norm. A current-
controlled stimulation is applied, which is gaining increased
interest in clinical application due to the reduced side ef-
fects and sensitivity to inter-individual variabilities [22], [23].
Different values for the dielectric properties of the volume
[V]ActivatedVolumes12312AdaptivefApproachftofDeterminefInactivationfHullfinfEachfPlaneSeedfpointrepresentingthefcenterfnodeoffanfaxonPlacementfofnewfaxonsaroundftheactivatedfaxonPlacementfofnewfaxonsstopsfforinactivefaxonsAlgorithmstopsfiffclosedinactivationhullfisffoundSeedfpointActivatedfAxonInactivatedfAxonABActivationmapThresholdmapNon-adaptiveapproachVV CONDUCTIVITY OF THE VOLUME CONDUCTOR COMPARTMENTS FOR THE
DIFFERENT STUDY CASES. THE VALUES WERE DETERMINED BY USING
THE TISSUE PARAMETERS FROM [17].
TABLE I
Study Case
Conductivity [Sm−1]
Model 1 (homogeneous)
Model 2 (acute phase)
Model 3 (chronic phase)
Model 4 (with STN)
encapsulation
layer
0.064
2.000
0.032
0.032
brain
tissue
0.064
0.064
0.064
0.064
subthalamic
nucleus (STN)
0.064
0.064
0.064
0.103
conductor model compartments are employed to investigate
the approximation quality for different cases of tissue het-
erogeneity in the target area. The dielectric properties are
obtained from equation (3) using the parameters for white
matter, grey matter, and cerebrospinal fluid from [17] at a
frequency of 2 kHz, which constitutes a good approximation
of the dispersive nature of the tissue properties for common
DBS signals [8]. The corresponding conductivity values are
approximately 0.064 Sm−1 for white matter, 0.103 Sm−1 for
grey matter, and 2.000 Sm−1 for cerebrospinal fluid. The
cases comprise a homogeneous model (Model 1) with the
conductivity of the encapsulation layer, the brain tissue, and
STN set to the value for white matter, a model with a high
conductive encapsulation layer (Model 2) set to the value of
cerebrospinal fluid [7] representing the acute phase, as well as
with a low conductive encapsulation layer (Model 3), set to
half the value of white matter [12] representing the chronic
phase, and a heterogeneous model (Model 4) with a low
conductive encapsulation layer, brain tissue set to the value of
white matter, and the STN set to the value of grey matter. Since
the varying dielectric tissue properties result in a variation of
the conduction in the volume conductor model, the extent of
neural activation is dependent on the tissue properties [11].
Therefore, DBS with the same stimulation amplitude results in
different sizes of the VTA depending on the tissue properties in
the models. In order to ensure a stimulation of the target region
between a distance of 2.0 mm (activation of sensorimotor
functional zone) and 4.0 mm (activation of larger parts of
the STN) from the electrode center, the required stimulation
amplitude to activate a homogeneous volume within this mini-
mum and maximum distance was computed for each model in
advance. The determinted stimulation amplitude range is then
prescribed to the adaptive algorithm to constrain the extent of
the estimated neural activation. Except for the latter model,
the neural activation extent is computed exploiting rotational
symmetry by computing the extent in a reference plane and
revolving the solution around the electrode. A spatial step size
of ∆s = 0.5 mm and a rotational step size of ∆α = 10 ◦. The
volumes of the neural activation extent and the corresponding
extents determined by the threshold values of the electric
potential and the electric field norm were computed by using
the Qhull library (http://www.qhull.org/) implemented in SciPy
(https://www.scipy.org/, version 0.17.0).
A previous computational study investigated the feasibility
of approximating the neural activation extent determined by
5
the coupling of the field distribution with axon models by
using constant field threshold values, depending on the stimu-
lation protocol and the axon diameter [13]. The used axon
models were based on a multi-compartment mathematical
model employing the cable equation, while in this study a
double cable axon model is applied [20]. The results showed
that iso-volumes for threshold values determined using the
electric field norm allowed for a close approximation of the
neural activation extent determined from the coupled field-
axon models. The deviation between these iso-volumes and
the neural activation extents further decreased with increasing
fiber diameter of the axons. To determine these threshold val-
ues, axon models were used to compute the maximum distance
along a line radial to the active electrode center, for which an
axon model placed at this distance is still activated by the
stimulation. The electric potential or the electric field norm at
this distance provided the field threshold value for computing
the neural activation extent. This procedure was repeated for
a set of stimulation amplitudes in a given range, resulting
in averaged field threshold values over the given stimulation
amplitude range. The neural activation volumes resulting from
this approach and using the electric field norm to determine
the field threshold values are denoted as VTAE,const in this
study. For evaluating the feasibility of this approach in [13],
the threshold values for a given stimulation amplitude range
were normalized with respect to the threshold value for the
minimum stimulation amplitude.
In addition to this methodology, this study incorporates an
approach, which does not average the field threshold values
in the given stimulation amplitude range, but computes the
threshold-distance relationship along the line radial
to the
active electrode center, from which the maximum distance
to activate an axon for a given stimulation amplitude can
be computed in post-processing. In addition, the proposed
methodology accounts also for rotationally asymmetric field
distributions by computing the threshold-distance relationship
in each plane around the stimulation electrode lead and av-
eraging the resulting field threshold values. The correspond-
ing iso-volumes are denoted as VTAϕ if electric potential
threshold values and VTAE if electric field norm threshold
values were applied. To investigate the dependence of the
resulting approximation on the fiber diameter as reported in
[13], the proposed study cases are carried out for axon fibers
with diameters of 2.0 µm, 3.0 µm, 5.7 µm, 7.3 µm, 8.7 µm, an
10.0 µm. Following the approach in [13], the thresholds are
determined by the mean value of the electric potential and
the electric field norm at the activation distance radial to the
center of the active electrode contact for a given stimulation
amplitude within the prescribed range in each plane.
A. Validation of the field solution
II. RESULTS
In order to validate the field solution obtained by the
implemented pipeline, a simplified multi-compartment volume
conductor model for which the analytical solution is known
is used. The analytical example model represents the prob-
lem of determining the electric potential distribution in a
6
Fig. 4. Geometry of the analytical validation example consisting out of a
layered sphere located in an homogeneous electric field E0.
Fig. 6. A: Comparison of the voltage-distance relationship computed for a
5.7 µm axon with the reference data provided by [12, Fig. 2]. B: Threshold-
distance relationships illustrated by quadratic polynomials fitted to the data
for axon models with varying fiber diameter fd in µm.
analytically using a separation approach, which is explained in
detail in the appendix V. The relative deviation of the solution,
determined as the norm of the respective field quantity between
the computational model and the analytical model along a
cut line through the layered sphere was below 7.0 · 10−3 and
2.9 · 10−2 for the electric potential and electric field norm of
the stationary field equation and below 1.1·10−2 and 2.1·10−2
for the norm of the complex-valued electric potential and the
electric field norm of the quasistatic field equation (Fig. 5).
B. Validation of the neural activation solution
To assess the validity of the implementation of the sim-
ulation pipeline to determine the neural activation with the
myelinated axon cable model [20],
the volume conductor
model for DBS used in [15] is adapted in order to compare
the distance-threshold relation for a fiber diameter of 5.7 µm.
The volume conductor model used in the mentioned study
comprises the same DBS electrode as in this study as well as
an encapsulation layer. A voltage-controlled stimulation signal
with a frequency of 150 Hz and a pulse duration of 100 µs is
applied. To match the dielectric tissue properties, the conduc-
tivity of brain tissue and of the STN was set to 0.3 Sm−1 and
the conductivity of the encapsulation layer to 0.15 Sm−1. The
relative deviation between the threshold-distance relationship
was determined to a value below 5.5 % by computing the
norm between the thresholds determined with the implemented
simulation pipeline and the data extracted from [15, Fig. 2],
which presents a good agreement of the computed threshold-
distance relationship with the data from the original model.
Varying the fiber diameter in the considered range described
in section I-H resulted in the expected characteristic decrease
in the threshold-distance relationships with increasing fiber
diameter (Fig. 6).
Fig. 5. Comparison of the electric potential and electric field norm obtained
by the computational and analytical field solution for the field validation
model (Fig. 4). A: The potential distribution for the stationary field and
quasistatic equation. The real and imaginary part of the electric potential
for the quasistatic equation are shown in the top and bottom of the image,
respectively. Potential isolines between −200 mV and 200 mV are shown. B,
C: Comparison of the electric potential and electric field norm along a cut line
through the layered sphere for the computational and the analytical solution
for the stationary field and quasistatic field equation. For the quasistatic field
equation, the norm of the electric potential multiplied by the sign of its real
part is shown.
homogeneous electric field, which is locally disturbed by a
conducting layered sphere (Fig. 4). Following the modeling
and simulation pipeline described in the methods section, a
3D volume conductor model was composed with a radius for
the inner sphere of 0.5 cm and 1.5 cm for the outer sphere.
The bounding box around the layered sphere is determined
by an edge length of 10 cm. A homogeneous electric field
of 10 Vm−1 was generated by applying Dirichlet boundary
conditions of 0.5 V and −0.5 V on the opposing faces of the
bounding box in the yz−plane. The conductivity was set to
2.0 Sm−1 for the inner sphere, 0.1 Sm−1 for the outer sphere,
and 1.0 Sm−1 for the bounding box. For the quasistatic field
problem, relative permittivities of 120 for the inner sphere,
2 · 106 for the outer sphere, and 80 for the bounding box, for
a frequency of 35 kHz were applied. The mesh was manually
refined at the spheres and bounding box surfaces, resulting in
a total number of approximately 1.1 million cells. The analytic
solution is determined by solving the field equations (1)
zxyE0InnersphereOutersphereRiƟRer100 mm100 mm100 mm0.020.06-0.02-0.06-0.2-0.050.050.2ImaginaryRealPotentialDistributionB[V]0.50.25-0.25-0.50.00.090.045-0.045-0.090.0-0.2-0.050.050.20.50.25-0.25-0.50.0ABCAB 7
Fig. 7. Approximation of the neural activation extent by threshold values of the electric potential ϕ and the electric field norm E. A: Normalized activation
thresholds for the different models. B: Neural activation extent (shown in black) and iso-volumes of the activation threshold for the electric potential (green)
and electric field norm (blue) for the minimum and maximum stimulation amplitude required to activate neural tissue in a distance between 2 mm and 4 mm
radial to the active electrode contact center. For Model 4, the STN is illustrated in red.
Fig. 8. Electrode impedance determined for the homogeneous model (Model
1), high encapsulation conductivity model (Model 2),
low encapsulation
conductivity model (Model 3), and the model including the STN as target
area (Model 4).
C. Approximating the neural activation by field thresholds
The adaptive scheme proposed in I-H was used to determine
the neural activation extent for four model setups with varying
electrical properties: A homogeneous model (Model 1), a high
encapsulation conductivity model (Model 2), a low encapsu-
lation conductivity model (Model 3), and a model including
the STN as target area (Model 4). The resulting stimulation
amplitude ranges, required to activate a region between 2 mm
and 4 mm distance to the active electrode contact, and the
corresponding neural activation extents showed a variation
for the different models, which results from their varying
electrical
tissue properties (Fig. 7). This influence is also
noticeable in the determined activation thresholds for the
electric potential with a threshold value of −0.25 V for the
minimum stimulation extent with a normalized increase of
2.64 for the homogeneous model and a threshold value of
−0.32 V with a normalized increase of 1.97 for the model
including the STN. The electric field norm threshold value for
the minimum stimulation extent varied between −129 Vm−1
and −141 Vm−1, but showed a similar normalized increase
between 1.29 and 1.35 for the models. The varying electrical
tissue properties of the models and the resulting varying
neural activation extents correspond to changes in the electrode
impedance with the high encapsulation conductivity model
Fig. 9. Volume of the neural activation extent (VTA) and the iso-volume
extent for corresponding threshold values of the electric potential (VTAϕ in
green) and the electric field norm (VTAE in blue) as well as the approach
using the constant threshold value of the electric field norm (VTAE,const in
red) for the different models.
(Model 2) showing a substantially smaller impedance and also
a substantially larger neural activation extent than the other
models (Fig. 8).
In order to assess the quality of approximating the neural
activation extent by threshold values of the electric potential
and the electric field norm, the volumes of the neural activation
extent and the extents resulting from the determined threshold
values were compared (Fig. 9). While the extents for the neural
activation and for the threshold values were in good agreement
for the homogeneous model (Model 1), the low encapsulation
conductivity model (Model 3), and the model including the
STN as target area (Model 4) with average volume deviations
of 4.0 %± 1.8 %, the extents for the threshold values understi-
mated the neural activation extent for the high encapsulation
conductivity model (Model 2) with average volume deviations
of 14.1 %± 2.8 %, as also noticeable in Figure 7B for the
minimum and maximum stimulation amplitude. Besides the
approach using the threshold-distance relationship to deter-
-0.9 mA-3.4 mA-0.9 mA-3.4 mAElectric potentialElectric field norm-0.4 mA-2.2 mA-0.4 mA-2.2 mAElectric potentialElectric field norm-0.26 V-0.65 V-141 V/m-179 V/m-0.6 mA-2.0 mA-0.6 mA-2.0 mAElectric potentialElectric field norm-0.32 V-0.62 V-134 V/m-171 V/m-0.36 V-0.80 V-129 V/m-171 V/m-0.4 mA-2.2 mA-0.4 mA-2.2 mAElectric potentialElectric field norm-0.25 V-0.65 V-138 V/m-178 V/mAB, STIMULATION AMPLITUDES AND NORMALIZED ACTIVATION THRESHOLDS
FOR APPROXIMATING THE NEURAL ACTIVATION EXTENT IN A DISTANCE
OF 2 mm TO 4 mm RADIAL TO THE ACTIVE ELECTRODE CONTACT.
TABLE II
Study Case
Model 1
Model 2
Model 3
Model 4
Axon
diameter
in µm
2.0
3.0
5.7
7.3
8.7
10.0
2.0
3.0
5.7
7.3
8.7
10.0
2.0
3.0
5.7
7.3
8.7
10.0
2.0
3.0
5.7
7.3
8.7
10.0
Amplitude
Maximum
range
in -mA
[1.2, 8.3]
[0.7, 4.1]
[0.4, 2.2]
[0.3, 1.3]
[0.2, 0.9]
[0.2, 0.8]
[2.6, 13.0]
[1.4, 6.3]
[0.9, 3.4]
[0.6, 2.0]
[0,4, 1.4]
[0.4, 1.1]
[1.2, 8.1]
[0.6, 4.0]
[0.4, 2.2]
[0.3, 1.3]
[0.2, 0.9]
[0.2, 0.8]
[1.8, 6.0]
[0.9, 3.3]
[0.6, 2.0]
[0.4, 1.2]
[0.3, 0.9]
[0.2, 0.7]
normalized threshold
for E2
for ϕ
3.37
2.91
2.64
2.18
2.13
1.96
1.70
1.49
1.30
1.11
1.02
0.96
2.98
2.69
2.29
2.02
2.01
1.72
3.30
3.14
2.63
2.18
2.12
1.95
2.11
2.18
1.97
1.79
1.73
1.83
1.74
1.57
1.35
1.19
1.11
1.04
1.67
1.54
1.29
1.11
1.01
0.96
1.47
1.67
1.29
1.18
1.11
1.00
mine the field threshold values, the approach using the mean
value of the determined normalized field threshold values as
carried out in [13] was used in this study to approximate the
neural activation extent by threshold values of the electric field
norm. Comparing both approaches for all models by comput-
ing the relative deviation between the approximated activation
volumes and the neural activation extents, the approach using
the threshold-distance relationship (VTAE) showed smaller
average volume deviations of 8.3 ± 3.9% (Min: 3.2 %, Max:
15.6 %) compared to 8.8 ± 6.5% (Min: 1.0 %, Max: 22.2 %)
for the approach using the average value of the normalized
field thresholds (VTAE,const).
Changing the fiber diameter from 5.7 µm to smaller and
larger diameters resulted in a variation of the determined
stimulation amplitude ranges. The required amplitudes to
activate the prescribed distance of 2 mm to 4 mm decreased
for increasing fiber diameter, which is also noticeable in
the threshold-distance relationship in Figure 6B. The largest
increases in the normalized activation thresholds were found
for small fiber diameters with values up to 3.37 for the electric
potential and 1.74 for the electric field norm. With increasing
fiber diameter a smaller increase of the normalized activation
thresholds for the electric potential and the electric field norm
8
is noticeable, which corresponds to the results reported in
[13]. The slope for the normalized activation thresholds for
the electric field norm tends to a value of 1, which allows for
approximation of the neural activation extent by using only
one threshold value of the electric field norm at an arbitrary
stimulation amplitude.
The proposed scheme for the adaptive estimation of the
neural activation extent during DBS required approximately
31 % to 43 % less axons than a non-adaptive approach, where
a prescribed number of axons is positioned in a rectangular
grid in each rotational plane (Table III). While the estimation
whether an axon is activated or not activated by the field for
a given stimulation amplitude was carried out by one run of
the neuron model, the estimation of the minimum stimulation
amplitude for one axon required approximately 23.2 runs for a
tolerance of 1·10−6. Considering the number of required opti-
mization runs, a speed up between 38 % and 66 % is obtained
by the adaptive scheme, which reduced the computation time
for the rotationally symmetric neural activation computations
to 24 - 45 minutes saving between 10 - 27 minutes on a
12×2.4 GHz, 48 GB workstation. The longest computation
time was required for the high encapsulation conductivity
model (Model 2), since it employed a wide spread of the
neural activation extent along the electrode lead due to the
highly conductive encapsulation layer. For the heterogeneous
case, the adaptive algorithm reduced the computation time
from 19 hours 41 minutes to 11 hours 36 minutes. When the
neural activation extent is estimated by the threshold values of
the electric potential and the electric field norm, the number
of required axons is substantially reduced from determining
the activation thresholds at several locations in each plane
around the electrode to determining the threshold-distance
relationship along a line radial to the active electrode contact.
This threshold-distance relationship computation required for
a maximum distance of 4 mm radial to the active electrode
contact only the computation of the activation threshold of
eight axons, requiring approximately 186 runs of the neuron
model and a computation time of below four minutes, which
is less than the computation time for the field solution, which
required approximately five minutes for all models. In case
of the heterogeneous model (Model 4), where no rotational
symmetry could be exploited, 288 axons and a computation
time of approximately one hour seven minutes was required.
III. DISCUSSION
The present paper proposes an adaptive scheme to estimate
the neural activation extent during DBS, which is embedded
into a Python package using the open-source solutions FEniCS
and NEURON. The field solution of the volume conductor
model computed with FEniCS as well as the threshold-
distance relationship of the axon model computed with NEU-
RON were compared with analytical solutions as well as
reference data from literature and show a good agreement
with deviations below 2.9 % for the field solution and 5.5 %
for the threshold-distance relationship, respectively (Fig. 5
and Fig. 6). The field model is able to compute stationary
current fields (purely resistive material properties) as well as
electro-quasistatic fields (complex material properties includ-
ing conductivity and relative permittivity) for heterogeneous
and rotationally asymmetric tissue distributions. The support
for incorporating complex material properties further allows
for computing the time-dependent field solution dependent on
the dispersive electrical properties of biological tissue for any
applied voltage- or current-controlled DBS signal using the
Fourier Finite element method [8]. The implementation of the
field and neuron parts in one Python package made it possible
to adaptively estimate the neural activation extent based on
the computed field solution and stimulation signal. Instead of
solving the field problem and exporting the time-dependent
electric potential at the nodes of several axons located at
prescribed positions around the electrode lead to determine the
minimum stimulation amplitude to elicit an action potential for
each axon [7], [11], the adaptive scheme positioned axons only
in those regions, where a neural stimulation by the given field
solution and stimulation amplitude range would occur. With
that, the adaptive scheme requires no pre-knowledge on the
neural activation extent for the given volume conductor model
and model parameters and, in addition, requires substantially
less computational resources and time.
The adaptive scheme was applied to estimate the neural
activation extent for model cases with varying tissue properties
and axon diameters. The tissue properties were chosen to
represent different post-operative stages during DBS as well as
a homogeneous and rotationally asymmetric case, where the
STN was explicitly modeled as target area. Compared to a non-
adaptive approach under the assumption that the minimally
required grid size for the axon positions is already known, a
speed up of up to 66 % was achieved by applying the adaptive
scheme.
Nevertheless, even with the adaptive scheme the total
computation time for determining the neural activation extent
can still be substantially larger than for determining the
field solution. To investigate possibilities to further reduce
the computational expense, a field threshold approach using
the relationship between the field solution and the neural
activation suggested in [13] was applied by computing the
neural activation along a line radial to the center of the active
electrode contact and using the resulting threshold-distance
relationship to determine threshold values of the electric
potential and electric field norm. In [13], the iso-volumes for
these threshold values constituted a good estimate of the neural
activation in homogeneous and rotationally symmetric volume
conductor models for DBS. In [13], a single-cable axon model
was used compared to a double-cable axon model from [20]
used in this study. They show a different threshold-distance
relationship for the same fiber diameters, with a 3.0 µm single-
cable axon model correlating with a 5.7 µm double-cable
axon model [13, Fig. 6]. The different axon models and
different applied stimulation amplitude ranges impede a direct
comparison of the normalized field threshold values in this
study with the data from [13]. Nevertheless, larger normalized
field thresholds were observered in both studies, when using
the electric potential compared to the electric field norm for
the field threshold computation (Tab. II and [13, Table I]).
In addition, in both studies, the maximum normalized field
9
thresholds decreased for increasing fiber diameters tending
to a value close to 1 when using the electric field norm
(1.06 for a 7.5 µm axon in [13] and 0.96 for a 10.0 µm
for Model 1, see Table II). Furthermore, the results suggest
that also for heterogeneous and rotationally asymmetric field
distributions with substantially varying electrode impedances
(Fig. 8), threshold values and corresponding iso-volumes of
the electric potential and the electric field norm generally
constitute a good approximation of the neural activation extent
(Fig. 9). However, the results revealed as well that in case
of a highly conductive encapsulation layer, as in the phase
directly after the surgery, the deviation between the extent
approximated by field thresholds and the neural activation
extent becomes larger (Fig. 7 and Fig. 9). This increased extent
is a result of the increased conductivity of the encapsulation
layer, which is spatially connected to the active electrode
contact, leading to a higher electric field strength and, with
that, an increased activity along the electrode. For this case,
the neural activation outside the target area is underestimated,
which could have possible implications for determining the
impact of unwanted side effects. Nevertheless, the activation
in the target area is still approximated with a good quality by
the determined field thresholds (Fig. 9). The underestimation
of the neural activation for larger distances away from the
active electrode by the field threshold approach may be due
in part to an approximation artifact. While the response of the
axon to a stimulus depends on the electric potential distribution
along the whole axon, especially on the second derivative of
it [24], the field threshold approximation approach determines
the field value for a given stimulation amplitude only at the
center node of the axon. For a cathodic stimulation pulse
and an axon positioned centered to the active stimulation
electrode in an isotropic homogeneous medium, the maximum
depolarization occurs at its center node, which means that the
determined field threshold is connected to the value of the
second derivative of the electric potential along the axon. For
larger distances away from the electrode, the electric potential
and its second derivative along the axon attenuates. Since the
field thresholds were determined for distances between 2 mm
and 4 mm from the electrode center, this might present an
explanation for the strong correlation between iso-volumes of
field threshold values and neural activation extents in the target
area, but not for larger distances away from the electrode.
Regarding the given model parameters, the required stim-
ulation distance or the prescribed stimulation amplitude range
and whether rotational symmetry can be exploited, the field
threshold can be determined by computing the solution of a
few hundred to thousand runs of the axon model, which is
substantially less than using only spatially distributed axon
models in the non-adaptive and adaptive-approach, which
requires 104 to 105 runs of the axon model. The number
of required axon model runs directly depends on the neural
activation extent. For instance, a larger stimulation amplitude
results in a larger neural activation extent, requiring more axon
models run to determine its outer shape. Besides the reduction
of the computational time achieved by the proposed adaptive
and field threshold approximation approach, the computation
time could also be decreased by employing a larger time step
NUMBER OF REQUIRED AXON MODELS, SPEED UP AND COMPUTATION TIME FOR THE DIFFERENT STUDY CASES (MODELS). THE NUMBER OF AXONS
AND THE COMPUTATION TIMES ARE DETERMINED FOR THE ADAPTIVE SCHEME, THE NON-ADAPTIVE SCHEME, AND FOR APPROXIMATING THE
NEURONAL ACTIVATION EXTENT (VTA) BY FIELD THRESHOLD VALUES FOR AXONS WITH 5.7 µm FIBER DIAMETER. THE COMPUTATION TIME IS
MEASURED ON A 12×2.4 GHz, 48 GB WORKSTATION. † ROTATIONAL SYMMETRY WAS USED FOR COMPUTING THE NEURAL ACTIVATION EXTENT. ∗
VALUES DETERMINED BY MAPPING NUMBER OF AXONS AND COMPUTATION TIME TO RECTANGULAR GRIDS OF SAME EXTENT.
TABLE III
10
VTA
non-adaptive∗
VTA
VTA
adaptive
field threshold
adaptive
Speed up
VTA
Study Case
Number of Axons
Model 1†
Model 2†
Model 3†
Model 4
152
288
136
4,896
94
173
94
2,813
8
8
8
288
54 %
59 %
38 %
66 %
field
solution
4.9 min
4.7 min
4.8 min
5.0 min
Computation time
VTA
VTA
non-adaptive∗
36.7 min
71.3 min
35.8 min
1,155.8 min
adaptive
23.8 min
44.8 min
25.9 min
696.3 min
VTA
field threshold
3.5 min
3.4 min
3.4 min
67.0 min
and smaller number of stimulation pulses. In this study, a time-
step of 5 µs based on [20] and a number of 10 pulses were used
in order to achieve the 1-to-1 ratio in the firing of the axon
models with the DBS pulse train. In any case, the accuracy
and convergence of the results have to be carefully checked,
when larger time steps or shorter pulse trains are used.
The proposed approach reduces the computation time for
a model with a rotationally asymmetric field distribution
from more than 11 hours to about one hour on a common
workstation. Therefore, we belief that this approach has the
potential to take the computation of the neural activation extent
closer to a real-world application in clinical practice, where
(computation) time is an important constraint.
In [13], the neural activation extent is approximated by a
constant field threshold for a stimulation amplitude, which is
computed from the normalized activation thresholds for the
corresponding field quantity, such as the electric potential and
the electric field norm, by determining a mean value from
the linear fit of the field thresholds for increasing stimulation
amplitudes. Field threshold values determined with this ap-
proach are used in several computational studies to estimate
the neural activation extent during DBS [25], [26], [27].
The determined normalized activation thresholds determined
in this study with growth factors between 29 % to 35 %
for a 5.7 µm and 47 % to 74 % for a 2.0 µm axon suggest
that this approach leads to an over- and underestimation of
the neural activation extent with volume deviations of up to
24 % by using a constant electric field norm threshold for all
stimulation amplitudes (Table II). Similar to the results in [13],
the deviations decreased for increasing axon diameter (Table
II). Considering that generally smaller axon diameters, such as
5.7 µm and below, are used to estimate the neural activation
extent during DBS [5], [7], [11], [21], this approach might
lead to substantial deviations in the estimated extents when
a constant field threshold is used for its approximation. The
proposed approach to determine for each model the threshold-
distance relationship along a line radial to the active electrode
contact to determine the corresponding field threshold for the
given stimulation protocol was able to estimate the neural acti-
vation extents with deviations below 7.6 % using electric field
norm threshold values and below 3.2 % using electric potential
threshold values for the corresponding stimulation amplitudes.
The model case representing the acute post-operative phase
by a highly conductive encapsulation layer constituted the
exception to the approximation quality with deviations of
11.9 % and 17.5 %, respectively, which can be accounted to the
spread of the neural activation extent along the electrode lead
(Fig. 7). Nevertheless, in contrast to the constant field thresh-
old approach, the suggested threshold-distance field threshold
approach ensures that the activation distance in the target
region and with that the neural activation extent in the target
region equals the extent computed with solely using the axon
models. Therefore, the results suggest that this approach is
feasible to estimate the neural activation extent dependent on
varying dielectric tissue properties, especially for chronic post-
operative phases with a low conductive encapsulation layer,
and of varying axon diameters while reducing substantially
the computational expense.
The axon models to determine the neural activation extent
are distributed along normal trajectories in planes around the
stimulation electrode, which is a common positioning used
for the estimation of the neural activation [5], [7], [11]. This
positioning scheme was used in this study in order to compare
the simulation results with data from [13]. Considering the
anatomy of the target nuclei for DBS, additional knowledge
on the orientation and topology of the axons in the target area
could provide a more target-specific and realisitic estimation
of the neural activation. For instance, the used positioning of
the axon models around the electrode is a major simplification
compared to the more non-uniform orientation of axonal fibers
in and around the STN [28]. The consideration of mutually
varying axon fiber orientations and geometries would require
a modification of the scheme to estimate the neural activation
extent to determine an activation ratio or percentage in the
target area rather than a closed volume.
The described process of finding the inactivated hull in
each plane around the electrode lead accounts for a monopolar
electrode configuration, where one axon seed point is placed
in front of the active stimulation electrode in each plane. In
case of a multipolar electrode setup, axon seed points have
to be placed in front of each active stimulation electrode in
order to ensure that a closed inactivated hull is determined
for the given electrode configuration. The advantage of a
multipolar electrode setup is that the stimulation amplitude for
each active stimulation electrode can be adjusted to achieve
the optimal neural activation extent. The adaptive algorithm to
determine the inactivated hull for all possible combinations of
stimulation amplitudes of a multipolar electrode setup would
have to be adjusted to include the minimum and maximum
stimulation amplitude scenario. Even for multipolar electrode
configurations, this would allow to profit from the speed up
of the proposed adaptive algorithm in comparison to the non-
adaptive algorithm. Regarding the approximation of the neural
activation extent by field threshold values, this is not generally
given: For a monopolar electrode configuration with one active
stimulation electrode, the field threshold approach determines
from one threshold-distance relationship one distance value
and one corresponding field threshold value for a given stim-
ulation amplitude. This is carried out in each plane around
the stimulation electrode lead. In case of a multiplolar elec-
trode configuration, the field distribution varies for the given
stimulation amplitude at each active stimulation electrode,
which results in interaction effects modifying the threshold-
distance relationship for each active stimulation electrode.
Therefore, for each given stimulation amplitude configuration
of a multipolar electrode configuration, the threshold-distance
relationships have to be recomputed for each active stimu-
lation electrode, which reduces the computational speed up.
Nevertheless, such an approach would be still substantially
faster than recomputing the whole neural activation extent
using axon models by the non-adaptive approach. Besides
these required modifications on the adaptive neural activation
algorithm, the implementation of multipolar electrode con-
figuration support requires also the consideration of varying
stimulation amplitudes for the active electrode contacts, which
will be the focus of future releases of the FanPy Python
package.
the results of previous studies point out
The volume conductor model used in this study is em-
bedded into a Python package, which accounts for stationary
as well as electro-quasistatic field problems and, therefore, is
able to compute the time-dependent field solution employing
for resistive as well as dispersive dielectric properties of
biological tissue for voltage-controlled and current-controlled
stimulation signals. To date, current-controlled stimulation is
the favoured stimulation protocol due to the reduced sensitivity
of the stimulation impact regarding the dielectric tissue prop-
erties and effects of the electrode-tissue interface compared
to voltage-controlled sitmulation [22]. For the estimation of
the time-dependent field solution during voltage-controlled
stimulation,
the
necessity to incorporate the dielectric effects at the electrode-
tissue interface, which show a dispersive as well as non-linear
behaviour with respect to the intensity of the stimulation signal
[29], [30], [31]. Voltage-controlled stimulation was used in
this study to validate the threshold-distance relationship of
the implemented axon model [20] neglecting the dielectric
effects of the electrode-tissue-interface in order to adapt the
model used to generate the reference data [15]. The cur-
rently embedded volume conductor model is already able to
account for isotropic heterogeneous and dispersive dielectric
tissue properties. Besides tissue heterogeneity, the anisotropic
dielectric properties of brain tissue can have a substantial
influence on the estimation of the neural activation extent and
the prediction of side effects during DBS [32]. Therefore, it is
planned to incorporate the support for anisotropic conductivity
tensors into the field model for future studies.
11
IV. CONCLUSION
In this study, an adaptive scheme to estimate the neural
activation extent during DBS is presented. The computation of
the field solution as well as the coupling to axon models and
the adaptive computation of their response to the stimulation
signal
is embedded into an open-source Python package,
which was used to estimate the neural activation for vary-
ing axon diameters and electrical tissue properties rendering
different post-operative stages and target area properties. The
determined neural activation extents were used to assess the
feasibility of their approximation by field threshold values.
By using the threshold-distance relationship for determining
the field thresholds and corresponding iso-volumes dependent
on the stimulation amplitude, a close approximation of the
determined neural activation extents could be achieved, while
substantially reducing the computational expense.
V. APPENDIX
The field equation (1) for a conducting layered sphere with
radius Ri of the inner sphere and Re of the outer sphere in an
external homogeneous electric field as illustrated in Figure 4
can be formulated using spherical coordinates and a rotational
symmetry with respect to the azimuthal angle φ. Within an
homogeneous isotropic medium, the field equation for the
given problem has the form
(cid:18)
(cid:19)
(cid:18)
(cid:19)
1
r2
∂
∂r
r2 ∂ϕ
∂r
+
1
r2 sin θ
∂
∂θ
sin θ
∂ϕ
∂θ
= 0
(5)
with the radial distance r and the polar angle θ. Since r and
θ can be varied independently, a separation of the potential
ϕ = R(r)Θ(θ) results in the angular and radial equations
1
sin θ
d
dθ
(cid:18)
d
dr
sin θ
(cid:18)
dΘ(θ)
dθ
r2 dR(r)
r
(cid:19)
(cid:19)
+ λΘ(θ) = 0
− λR(r) = 0
(6)
(7)
(cid:0)(x2 − 1)n(cid:1) (8)
with the separation constant λ, where the solution to the
angular equation is given by
Θn(θ) = DnPn(cos θ), Pn(x) =
1
2nn!
dn
dxn
using the substitution x = cos(θ) and λ = n(n + 1)
with a constant Dn and the Legendre polynomial Pn(x) for
n = 0, 1, 2, . . . [33]. The radial equation represents an Euler-
Cauchy differential equation, which solution is given by
Rn(r) = C1,nrn + C2,nr−(n+1)
(9)
with the constants C1,n and C2,n. Applying the separation
equation, the general solution of the potential ϕ := ϕ(r, θ) is
then given by
∞(cid:80)
∞(cid:80)
n=0
ϕ =
∞(cid:80)
n=0
−E0rx +
AnrnPn(x)
(cid:0)Bnrn + Cnr−(n+1)(cid:1) Pn(x)
, 0 ≤ r ≤ Ri
, Ri < r ≤ Re
Dnr−(n+1)Pn(x)
, Re < r
n=0
(10)
12
[14] H. Lind´en et al., "Analysis of the quasi-static approximation for calcu-
lating potentials generated by neural stimulation," J Neural Eng, vol. 5,
pp. 44–53, Mar 2008.
[15] C. C. McIntyre et al., "Electric field and stimulating influence generated
by deep brain stimulation of the subthalamic nucleus," Clin Neurophys-
iol, vol. 115, pp. 589–595, Mar 2004.
[16] E. A. Accolla et al., "Brain tissue properties differentiate between motor
and limbic basal ganglia circuits," Hum Brain Mapp, vol. 35, pp. 5083–
5092, Oct 2014.
[17] S. Gabriel, R. W. Lau, and C. Gabriel, "The dielectric properties of
biological tissues: III Parametric models for the dielectric spectrum of
tissues," Phys Med Biol, vol. 41, pp. 2271–2293, Nov 1996.
[18] W. M. Grill and J. T. Mortimer, "Electrical properties of implant
encapsulation tissue," Ann Biomed Eng, vol. 22, pp. 23–33, Jan 1994.
[19] J. Gimsa et al., "Choosing electrodes for deep brain stimulation experi-
ments - electrochemical considerations," J Neurosci Meth, vol. 142, pp.
251–265, Mar 2005.
[20] C. C. McIntyre, A. G. Richardson, and W. M. Grill, "Modeling the
excitability of mammalian nerve fibers: influence of afterpotentials on
the recovery cycle." J Neurophysiol, vol. 87, pp. 995–1006, Feb 2002.
[21] S. N. Sotiropoulos and P. N. Steinmetz, "Assessing the direct effects
of deep brain stimulation using embedded axon models," J Neural Eng,
vol. 4, pp. 107–119, Jun 2007.
[22] S. F. Lempka et al., "Current-controlled deep brain stimulation reduces
in vivo voltage fluctuations observed during voltage-controlled stimula-
tion," Clin Neurophysiol, vol. 12, pp. 2128–2133, Dec 2010.
[23] C. Lettieri et al., "Clinical outcome of deep brain stimulation for dysto-
nia: constant-current or constant-voltage stimulation? A non-randomized
study," Eur J Neurol, vol. 22, pp. 919–926, Jun 2015.
[24] F. Rattay, "Analysis of models for external stimulation of axons," IEEE
Trans Biomed Eng, vol. 33, pp. 974–977, Oct 1986.
[25] F. Alonso et al., "Investigation into Deep Brain Stimulation Lead
Designs: A Patient-Specific Simulation Study," Brain Sci, vol. 6, p. E39
(16pp), Sep 2016.
[26] S. Hemm et al., "Patient-Specific Electric Field Simulations and Acceler-
ation Measurements for Objective Analysis of Intraoperative Stimulation
Tests in the Thalamus," Front Hum Neurosci, vol. 10, p. 577 (14pp), Sep
2016.
[27] T. A. Dembek et al., "Probabilistic mapping of deep brain stimulation
effects in essential tremor," Neuroimage Clin, vol. 13, pp. 164–173, Nov
2017.
[28] M. Axer et al., "A novel approach to the human connectome: Ultra-high
resolution mapping of fiber tracts in the brain," NeuroImage, vol. 54, pp.
1091–1101, Sep 2011.
[29] A. Richardot and E. T. McAdams, "Harmonic analysis of low-frequency
bioelectrode behavior," IEEE Trans Med Imaging, vol. 21, pp. 604–612,
Jun 2002.
[30] D. R. Cantrell et al., "Incorporation of the electrode-electrolyte interface
into finite-element models of metal microelectrodes," J Neural Eng,
vol. 5, pp. 54–67, Mar 2008.
[31] B. Howell, L. E. Medina, and W. M. Grill, "Effects of frequency-
dependent membrance capacitance on neural excitability," J Neural Eng,
vol. 12, p. 056015 (32pp), Oct 2015.
[32] B. Howell and C. C. McIntyre, "Analyzing the tradeoff between elec-
trical complexity and accuracy in patient-specific computational models
of deep brain stimulation," J Neural Eng, vol. 13, p. 036023, Jun 2016.
[33] V. L. Sukhorukov et al., "A single-shell model for biological cells ex-
tended to account for the dielectric anisotropy of the plasma membrane,"
Journal of Electrostatics, vol. 50, pp. 191–204, Feb 2001.
using the
substitution x = cos(θ) with constants
An, Bn, Cn, Dn for an external homogeneous electric field in
z-direction −E0z = −E0r cos(θ) and a vanishing influence
of the conducting layered sphere for r → ∞. Applying the
continuity conditions
ϕ1 − ϕ2 = 0
n(J 1 − J 2) = 0
(11)
(12)
for the electrical potential ϕ and the current density J at the
boundaries of the layered sphere and exploting the orthnormal-
ity of the Legendre polynomials results in a system of linear
equations which has only for n = 1 a non-trivial solution and
non-zero right hand side
with σi,i = σiR3
The solution of (13) is then given by
i , σs,i = σsR3
i , σs,e = σsR3
R3
i
i −R3
−1
σi,i −σs,i
2σs
R3
0
1
σs,e −2σs
e
0
B1
C1
D1
0
0
−1
2σe
A1
·
A1r cos θ
(cid:0)B1r + C1r−2(cid:1) cos θ
(cid:0)D1r−2 − E0r(cid:1) cos θ
ϕ(r, θ) =
=
0
(13)
0−E0,e
−σeE0,e
e, E0,e = E0R3
e.
, 0 ≤ r ≤ Ri
, Ri < r ≤ Re
, Re < r .
(14)
REFERENCES
[1] A. L. Benabid et al., "Deep brain stimulation of the subthalamic nucleus
for the treatment of Parkinson's disease," Lancet Neurol, vol. 8, pp. 67–
81, Jan 2009.
[2] M. D. Johnson et al., "Neuromodulation for brain disorders: challenges
and opportunities," IEEE Trans Biomed Eng, vol. 60, pp. 610–624, Mar
2013.
[3] C. Guigoni et al., "Involvement of Sensorimotor, Limbic, and Associa-
tive Basal Ganglia Domains in L-3,4-Dihydroxyphenylalanine-Induced
Dyskensia," J Neurosci, vol. 25, pp. 2912–2107, 2005.
[4] K. A. Follett et al., "Pallidal versus Subthalamic Deep-Brain Stimulation
for Parkinson's Disease," New Engl J Med, vol. 362, pp. 2077–2091,
2010.
[5] C. R. Butson and C. C. McIntyre, "Tissue and electrode capacitance
reduce neural activation volumes during deep brain stimulation," Clin
Neurophysiol, vol. 116, pp. 2490–2500, Oct 2005.
[6] C. R. Butson, C. B. Maks, and C. C. McIntyre, "Sources and effects of
electrode impedance during deep brain stimulation," Clin Neurophysiol,
vol. 117, pp. 447–454, Feb 2006.
[7] P. F. Grant and M. M. Lowery, "Effect of Dispersive Conductivity and
Permittivity in Volume Conductor Models of Deep Brain Stimulation,"
IEEE Trans Biomed Eng, vol. 57, pp. 2386–2393, Oct 2010.
[8] C. Schmidt, T. Flisgen, and U. van Rienen, "Efficient Computation of
the Neural Activation during Deep Brain Stimulation for Dispersive
Electrical Properties of Brain Tissue," IEEE Trans Magn, vol. 52, p.
7203004 (4 pages), Mar 2016.
[9] C. Schmidt and U. van Rienen, "Modeling the field distribution in deep
brain stimulation: the influence of anisotropy of brain tissue," IEEE
Trans Biomed Eng, pp. 1583–1592, Jun 2012.
[10] C. A. Bossetti, M. J. Birdno, and W. M. Grill, "Analysis of the
quasi-static approximation for calculating potentials generated by neural
stimulation," J Neural Eng, vol. 5, pp. 44–53, Mar 2008.
[11] C. Schmidt et al., "Influence of Uncertainties in the Material Properties
of Brain Tissue on the Probabilistic Volume of Tissue Activated," IEEE
Trans Biomed Eng, vol. 60, pp. 1378–1387, May 2013.
[12] C. C. McIntyre et al., "Cellular effects of deep brain stimulation: model-
based analysis of activation and inhibition," J Neurophysiol, vol. 91, pp.
1457–1469, Apr 2004.
[13] M. Astrom et al., "Relationship between Neural Activation and Electric
Field Distribution during Deep Brain Stimulation," IEEE Trans Biomed
Eng, vol. 62, pp. 664–672, Feb 2015.
|
1812.01727 | 2 | 1812 | 2018-12-21T05:39:45 | Optimal models of decision-making in dynamic environments | [
"q-bio.NC"
] | Nature is in constant flux, so animals must account for changes in their environment when making decisions. How animals learn the timescale of such changes and adapt their decision strategies accordingly is not well understood. Recent psychophysical experiments have shown humans and other animals can achieve near-optimal performance at two alternative forced choice (2AFC) tasks in dynamically changing environments. Characterization of performance requires the derivation and analysis of computational models of optimal decision-making policies on such tasks. We review recent theoretical work in this area, and discuss how models compare with subjects' behavior in tasks where the correct choice or evidence quality changes in dynamic, but predictable, ways. | q-bio.NC | q-bio |
Optimal models of decision-making in dynamic
environments
Zachary P. Kilpatrick1, William R. Holmes2, Tahra L. Eissa1, and Kresimir Josi´c3,4
1Department of Applied Mathematics, University of Colorado, Boulder, Colorado, USA
2Department of Physics and Astronomy, Department of Mathematics, Quantitative Systems Biology Center,
Vanderbilt University, Nashville, Tennessee, USA
3Department of Mathematics, Department of Biology and Biochemistry, University of Houston, Houston, Texas, USA
4Department of BioSciences, Rice University, Houston, Texas, USA
*[email protected], [email protected]
ABSTRACT
Nature is in constant flux, so animals must account for changes in their environment when making
decisions. How animals learn the timescale of such changes and adapt their decision strategies
accordingly is not well understood. Recent psychophysical experiments have shown humans and
other animals can achieve near-optimal performance at two alternative forced choice (2AFC) tasks
in dynamically changing environments. Characterization of performance requires the derivation and
analysis of computational models of optimal decision-making policies on such tasks. We review recent
theoretical work in this area, and discuss how models compare with subjects' behavior in tasks where
the correct choice or evidence quality changes in dynamic, but predictable, ways.
Introduction
To translate stimuli into decisions, animals interpret sequences of observations based on their prior
experiences1. However, the world is fluid: The context in which a decision is made, the quality of the
evidence, and even the best choice can change before a judgment is formed, or an action taken. A source
of water can dry up, or a nesting site can become compromised. But even when not fully predictable,
changes often have statistical structure: Some changes are rare, others are frequent, and some are more
likely to occur at specific times. How have animals adapted their decision strategies to a world that is
structured, but in flux?
Classic computational, behavioral, and neurophysiological studies of decision-making mostly involved
tasks with fixed or statistically stable evidence1 -- 3. To characterize the neural computations underlying
decision strategies in changing environments, we must understand the dynamics of evidence accumulation4.
This requires novel theoretical approaches. While normative models are a touchstone for theoretical
studies5, 6, even for simple dynamic tasks the computations required to optimally translate evidence
into decisions can become prohibitive7. Nonetheless, quantifying how behavior differs from normative
predictions helps elucidate the assumptions animals use to make decisions8, 9.
We review normative models and compare them with experimental data from two alternative forced
choice (2AFC) tasks in dynamic environments. Our focus is on tasks where subjects passively observe
streams of evidence, and the evidence quality or correct choice can vary within or across trials. Humans
and animals adapt their decision strategies to account for such volatile environments, often resulting in
performance that is nearly optimal on average. However, neither the computations they use to do so nor
their neural implementations are well understood.
Figure 1. Two alternative forced choice (2AFC) tasks in dynamic environments. (A) Possible
timescales of environmental dynamics: 1) The state (S+ or S−), or the quality of the evidence (e.g.,
coherence of random dot motion stimulus) may switch within a trial5, 6, 10, or 2) across trials11 -- 13; 3) The
hazard rate, h, can change across blocks of trials6, 9. (B) In a dynamic 2AFC task, a two-state Markov
chain with hazard rate h determines the state. (Bi) The current state (correct hypothesis) is either S+ (red)
or S− (yellow). (Bii) Conditional densities of the observations, f±(ξ ) = P(ξS±), shown as Gaussians
with means ±µ and standard deviation σ. (C) Evidence discounting is shaped by the environmental
timescale: (Top) In slow environments, posterior probabilities over the states, P(S±ξ1:4), are more
strongly influenced by past observations, ξ1:3, (darker shades of the observations, ξi, indicate higher
weight) and thus points to S+. (Bottom) If changes are fast, beliefs depend more strongly on the current
observation, ξ4, which outweighs older evidence and points to S−.
Optimal evidence accumulation in changing environments
Normative models of decision-making typically assume subjects are Bayesian agents14, 15 that probabilisti-
cally compute their belief of the state of the world by combining fresh evidence with previous knowledge.
Beyond normative models, notions of optimality require a defined objective. For instance, an observer may
need to report the location of a sound16, or the direction of a moving cloud of dots5, and is rewarded if the
report is correct. Combined with a framework to translate probabilities or beliefs into actions, normative
models provide a rational way to maximize the net rewards dictated by the environment and task. Thus an
optimal model combines normative computations with a policy that translates a belief into the optimal
action.
How are normative models and optimal policies in dynamic environments characterized? Older
observations have less relevance in rapidly changing environments than in slowly changing ones. Ideal
observers account for environmental changes by adjusting the rate at which they discount prior information
when making inferences and decisions17. In Box 1 we show how, in a normative model, past evidence is
nonlinearly discounted at a rate dependent on environmental volatility5, 17. When this volatility8 or the
underlying evidence quality13, 18 are unknown, they must also be inferred.
In 2AFC tasks, subjects accumulate evidence until they decide on one of two choices either freely or
when interrogated. In these tasks, fluctuations can act on different timescales (Fig. 1a): 1) on each trial
2/11
Discounting in Dynamic Environments BiiiTwo-Alternative Forced Choice ModelsADynamicTimescalesHazard Rate ChangeTimescale-µ𝒇−(𝝃)=𝑷(𝝃𝑺−)𝒇+(𝝃)=𝑷(𝝃𝑺+)ξChS-S+1-h1-hh+µ2σ2σSlowFastS+S-S+S-Across TrialsWithin Trialsξ2ξ3ξ1ξ2ξ3ObservationsProbability of S-Probability of S+ξnξ1Across TrialsStateChangesQualityChangesAcross TrialsStateChangesQualityChangesP(siξ1:4)ξ4ξ4P(siξ1:4) Box 1 -- Normative evidence accumulation in dynamic environments.
Discrete time. At times t1:n an observer receives a sequence of noisy observations, ξ1:n, of the state S1:n,
governed by a two-state Markov process (Fig. 1b). Observation likelihoods, f±(ξ ) = P(ξS±), determine
the belief (log-likelihood ratio: LLR), yn = log P(Sn=S+ξ1:n)
P(Sn=S−ξ1:n), after observation n. If the observations are
conditionally independent, the LLR can be updated recursively5, 17:
yn = log
(cid:124)
f+(ξn)
f−(ξn)
current evidence
(cid:123)(cid:122)
(cid:125)
+log (1− h)exp(yn−1) + h
(cid:125)
(cid:124)
hexp(yn−1) + (1− h)
(cid:123)(cid:122)
discounted prior belief
,
(1)
where h is the hazard rate (probability the state switches between times tn−1 and tn). The belief prior to
the observation at time tn, yn−1, is discounted according to the environment's volatility h. When h = 0,
Eq. (1) reduces to the classic drift-diffusion model (DDM), and evidence is accumulated perfectly over
time. When h = 1/2, only the latest observation, ξn, is informative. For 0 < h < 1/2, prior beliefs are
discounted, so past evidence contributes less to the current belief, yn, corresponding to leaky integration.
When 1/2 < h < 1, the environment alternates.
Continuous time. When tn −tn−1 = ∆t (cid:28) 1, and the hazard rate is defined ∆t · h, LLR evolution can be
approximated by the stochastic differential equation5, 17:
,
(2)
(cid:124)(cid:123)(cid:122)(cid:125)
dy = g(t)dt
drift
(cid:124)(cid:123)(cid:122)(cid:125)
+ dWt
noise
(cid:124)
(cid:123)(cid:122)
−2hsinh(y)dt
nonlinear filter
(cid:125)
where g(t) jumps between +g and −g at a rate h, Wt is a zero mean Wiener process with variance ρ2, and
the nonlinear filter −2hsinh(y) optimally discounts prior evidence. In contrast to the classic continuum
DDM, the belief, y(t), does not increase indefinitely, but saturates due to evidence-discounting.
(Fig. 1b,c)5, 6, 2) unpredictably within only some trials19, 20, 3) between trials in a sequence11, 16, or 4)
gradually across long blocks of trials21. We review findings in the first three cases and compare them to
predictions of normative model.
Within trial changes promote leaky evidence accumulation
Normative models of dynamic 2AFC tasks (Fig. 1b,c and 2a, Box 1) exhibit adaptive, nonlinear discounting
of prior beliefs at a rate adapted to expectations of the environment's volatility (Fig. 1c), and saturation of
certainty about each hypothesis, regardless of how much evidence is accumulated (Fig. 2a). In contrast,
ideal observers in static environments weigh all past observations equally, and their certainty grows
without bound until a decision1, 3. Also, in dynamic environments, the performance of ideal observers at
change points -- times when the correct choice switches -- depends sensitively on environmental volatility
(Fig. 2aiii). In slowly changing environments, optimal observers assume that changes are rare, and thus
adapt slowly after one has occured. In contrast, in rapidly changing environments, observers quickly
update their belief after a change point.
The responses of humans and other animals on tasks in which the correct choice changes stochastically
during a trial share features with normative models: In a random dot-motion discrimination (RDMD) task,
where the motion direction switches at unsignaled changepoints, humans adapt their decision-making
process to the switching (hazard) rate (Fig. 2ai)5. However, on average, they overestimate the change rates
3/11
Figure 2. Dynamic State Changes. (A) State changes within trials in a (Ai) random dot motion
discrimination (RDMD) task, in which drift direction switches throughout the trial5, and (Aii) dynamic
auditory clicks task, in which the side of the higher rate stream alternates during the trial6. (Aiii) An ideal
observer's LLR (See Eq. (2) in Box 1) when the hazard rate is low (top panels: h=0.1Hz) and high
(bottom panels: h=1Hz). Immediately after state changes, the belief typically does not match the state. (B)
State changes across trials. (Bi) In the triangles task5, samples (star) are drawn from one of two Gaussian
distributions (yellow and red clouds) whose centers are represented by triangles. The observers must
choose the current center (triangle). (Bii) In an RDMD task, dots on each trial move in one of two
directions (colored arrows) chosen according to a two-state Markov process. Depending on the switching
rate, trial sequences may include excessive repetitions (Top), or alternations (Bottom). (Biii) (Top)
Responses can be biased by decisions from previous trials. (Bottom) Probabilistic feedback ('O': correct;
'X': incorrect) affects initial bias (e.g., trials 3, 4, and 5), even when not completely reliable.
of rapidly switching environments and underestimate the change rates of slowly switching environments.
In a related experiment (Fig 2aii), rats were trained to identify which of two Poisson auditory click streams
arrived at a higher rate22. When the identity of the higher-frequency stream switched unpredictably during
a trial, trained rats discounted past clicks near-optimally on average, suggesting they learned to account
for latent environmental dynamics6.
However, behavioral data are not uniquely explained by normative models. Linear approximations of
normative models perform nearly identically17, and, under certain conditions, fit behavioral data well5, 6, 23.
Do subjects implement normative decision policies or simpler strategies that approximate them? Subjects'
decision strategies can depend strongly on task design and vary across individuals5, 9, suggesting a need
for sophisticated model selection techniques. Recent research suggests normative models can be robustly
distinguished from coarser approximations when task difficulty and volatility are carefully tuned24.
4/11
AiWithin Trial State Changesiiiiiℎ=1𝐻𝑧StateBeliefTime (within a trial)Time (s)ℎ=0.1𝐻𝑧StateBeliefTime (s)mismatchLLRLLRTime high rate in right earhigh rate in left earclicksh4-4002468102-200246810iiiAcross Trial State Changes*Time (trial-by-trial) AlternatingRepetitiveiiiTrial1234567StateResponseBiasTrial1234567StateResponseBiasFeedbackOOXXXOONo FeedbackFeedbackB Subjects account for correlations between trials by biasing initial beliefs
Natural environments can change over timescales that encompass multiple decisions. However, in many
experimental studies, task parameters are fixed or generated independently across trials, so evidence from
previous trials is irrelevant. Even so, subjects often use decisions and information from earlier trials to
(serially) bias future choices25 -- 27, reflecting ingrained assumptions about cross-trial dependencies21, 28.
To understand how subjects adapt to constancy and flux across trials, classic 2AFC experiments have
been extended to include correlated cross-trial choices (Fig. 2b) where both evidence accumulated during
a trial, and probabilistic reward provide information that can be used to guide subsequent decisions16, 29.
When a Markov process30 (Fig. 1b) is used to generate correct choices, human observers adapt to these
trial-to-trial correlations and their response times are accurately modeled by drift diffusion11 or ballistic
models16 with biased initial conditions.
Feedback or decisions across correlated trials impact different aspects of normative models31 including
accumulation speed (drift)32 -- 34, decision bounds11, or the initial belief on subsequent trials12, 35, 36. Given
a sequence of dependent but statistically identical trials, optimal observers should adjust their initial belief
and decision threshold16, 28, but not their accumulation speed in cases where difficulty is fixed across
trials18. Thus, optimal models predict that observers should, on average, respond more quickly, but not
more accurately28. Empirically, humans12, 35, 36 and other animals29 do indeed often respond faster on
repeat trials, which can often be modeled by per trial adjustments in initial belief. Furthermore, this bias
can result from explicit feedback or subjective estimates, as demonstrated in studies where no feedback is
provided (Fig. 2biii)16, 36.
The mechanism by which human subjects carry information across trials remains unclear. Different
models fit to human subject data have represented intertrial dependencies using initial bias, changes in
drift rate, and updated decision thresholds11, 16, 34. Humans also tend to have strong preexisting repetition
biases, even when such biases are suboptimal25 -- 27. Can this inherent bias be overcome through training?
The answer may be attainable by extending the training periods of humans or nonhuman primates5, 9, or
using novel auditory decision tasks developed for rodents6, 29. Ultimately, high throughput experiments
may be needed to probe how ecologically adaptive evidence accumulation strategies change with training.
Time-varying thresholds account for heterogeneities in task difficulty
Optimal decision policies can also be shaped by unpredictable changes in decision difficulty. For instance,
task difficulty can be titrated by varying the signal-to-noise ratio of the stimulus, so more observations
are required to obtain the same level of certainty. Theoretical studies have shown that it is optimal to
change one's decision criterion within a trial when the difficulty of a decision varies across trials13, 18, 37.
The threshold that determines how much evidence is needed to make a decision should vary during the
trial (Fig. 3a) to incorporate up-to-date estimates of trial difficulty18. There is evidence that subjects use
time-varying decision boundaries to balance speed and accuracy on such tasks38, 39.
Dynamic programming can be used to derive optimal decision policies when trial-to-trial difficulties
or reward sizes change. For instance, when task difficulty changes across trials in a RDMD task, opti-
mal decisions are modeled by a DDM with a time-varying boundary, in agreement with reaction time
distributions of humans and monkeys18, 38. Both dynamic programming18 and parameterized function38, 40
based models suggest decreasing bounds maximize reward rates (Fig. 3a,b). This dynamic criterion helps
participants avoid noise-triggered early decisions or extended deliberations18. An exception to this trend
was identified in trial sequences without trials of extreme difficulty13, in which case the optimal strategy
used a threshold that increased over time.
Time-varying decision criteria also arise when subjects perform tasks where information quality
5/11
Figure 3. Dynamic Evidence Quality. (A) Trial-to-trial two-state Markovian evidence quality
switching: (Ai) Evidence quality switches between easy (Qeasy) and hard (Qhard) with probability Pswitch.
(Aii) Optimal decision policies require time-varying decision thresholds. An observer who knows the
evidence quality (easy or hard) uses a fixed threshold (grey traces, dashed lines) to maximize reward rate,
but thresholds must vary when evidence quality is initially unknown (black trace, green gradient). (B)
Different triangle task difficulties (from Fig. 2Ai): Triangles are spaced further apart in easy trials
compared to hard trials. (C) Changes in quality within trials: (Ci) An RDMD task in which the drift
coherence increases mid-trial, providing stronger evidence later in the trial. (Cii) The corresponding LLR
increases slowly early in the trial, and more rapidly once evidence becomes stronger.
changes within trials (Fig. 3c)40, especially when initially weak evidence is followed by stronger evidence
later in the trial. However, most studies use heuristic models to explain psychophysical data19, 20, suggest-
ing a need for normative model development in these contexts. Decision threshold switches have also been
observed in humans performing changepoint detection tasks, whose difficulty changes from trial-to-trial41,
and in a model of value-based decisions, where the reward amounts change between trials42. Overall,
optimal performance on tasks in which reward structure or decision difficulty changes across trials require
time-varying decision criteria, and subject behavior approximates these normative assumptions.
One caveat is that extensive training or obvious across-trial changes are needed for subjects to learn
optimal solutions. A meta-analysis of multiple studies showed that fixed threshold DDMs fit human
behavior well when difficulty changes between trials were hard to perceive43. A similar conclusion
holds when changes occur within trials44. However, when nonhuman primates are trained extensively
on tasks where difficulty variations were likely difficult to perceive, they appear to learn a time-varying
criterion strategy45. Humans also exhibit time-varying criteria in reward-free trial sequences where
6/11
ADynamic Evidence Quality ModelBAcross Trial Quality ChangesCWithin Trial Quality ChangesEasy HardiiiLLRTime (s)06481216LLRTime (s)3421001Hard EasyQEasyQHardPsameiiiPsamePswitchPswitch interrogations are interspersed with free responses46. Thus, when task design makes it difficult to perceive
task heterogeneity or learn the optimal strategy, subjects seem to use fixed threshold criteria43, 44. In
contrast, with sufficient training45, or when changes are easy to perceive46, subjects can learn adaptive
threshold strategies.
Questions remain about how well normative models describe subject performance when difficulty
changes across or within trials. How distinct do task difficulty extremes need to be for subjects to use
optimal models? No systematic study has quantified performance advantages of time-varying decision
thresholds. If they do not confer a significant advantage, the added complexity of dynamic thresholds may
discourage their use.
When and how are normative computations learned and achieved?
Except in simple situations, or with overtrained animals, subjects can at best approximate computations
of an ideal observer14. Yet, the studies we reviewed suggest that subjects often learn to do so effectively.
Humans appear to use a process resembling reinforcement learning to learn the structure and parameters of
decision task environments47. Such learning tracks a gradient in reward space, and subjects adapt rapidly
when the task structure changes48. Subjects also switch between different near-optimal models when mak-
ing inferences, which may reflect continuous task structure learning9. However, these learning strategies
appear to rely on reward and could be noisier when feedback is probabilistic or absent. Alternatively,
subjects may ignore feedback and learn from evidence accumulated within or across trials28, 46.
Strategy learning can be facilitated by using simplified models. For example, humans appear to use
sampling strategies that approximate, but are simpler than, optimal inference9, 49. Humans also behave in
ways that limit performance by, for instance, not changing their mind when faced with new evidence50.
This confirmation bias may reflect interactions between decision and attention related systems that are
difficult to train away51. Cognitive biases may also arise due to suboptimal applications of normative
models52. For instance, recency bias can reflect an incorrect assumption of trial dependencies53. Subjects
seem to continuously update latent parameters (e.g., hazard rate), perhaps assuming that these parameters
are always changing21, 29.
The adaptive processes we have discussed occur on disparate timescales, and thus likely involve neural
mechanisms that interact across scales. Task structure learning occurs over many sessions (days), while
the volatility of the environment and other latent parameters can be learned over many trials (hours)6, 49.
Trial-to-trial dependencies likely require memory processes that span minutes, while within trial changes
require much faster adaptation (milliseconds to seconds).
This leaves us with a number of questions: How does the brain bridge timescales to learn and
implement adaptive evidence integration? This likely requires coordinating fast neural activity changes
with slower changes in network architecture8. Studies of decision tasks in static environments suggest that
a subject's belief and ultimate choice is reflected in evolving neural activity1 -- 3, 54. It is unclear whether
similar processes represent adaptive evidence accumulation, and, if so, how they are modulated.
Conclusions
As the range of possible descriptive models grows with task complexity8, 49, optimal observer models
provide a framework for interpreting behavioral data5, 6, 34. However, understanding the computations
subjects use on dynamic tasks, and when they depart from optimality, requires both careful comparison of
models to data and comparisons between model classes55.
While we mainly considered optimality defined by performance, model complexity may be just as
important in determining the computations used by experimental subjects56. Complex models, while more
7/11
accurate, may be difficult to learn, hard to implement, and offer little advantage over simpler ones8, 9.
Moreover, predictions of more complex models typically have higher variance, compared to the higher
bias of more parsimonious models, resulting in a trade-off between the two9.
Invasive approaches for probing adaptive evidence accumulation are a work in progress57, 58. However,
pupillometry has been shown to reflect arousal changes linked to a mismatch between expectations and
observations in dynamic environments10, 27, 59. Large pupil sizes reflect high arousal after a perceived
change, resulting in adaptive changes in evidence weighting. Thus, pupillometry may provide additional
information for identifying computations underlying adaptive evidence accumulation.
Understanding how animals make decisions in volatile environments requires careful task design.
Learning and implementing an adaptive evidence accumulation strategy needs to be both rewarding and
sufficiently simple so subjects do not resign themselves to simpler computations43, 44. A range of studies
have now shown that mammals can learn to use adaptive decision-making strategies in dynamic 2AFC
tasks5, 6. Building on these approaches, and using them to guide invasive studies with mammals offers
promising new ways of understanding the neural computations that underlie our everyday decisions.
Acknowledgements
We are grateful to Joshua Gold, Alex Piet, and Nicholas Barendregt for helpful feedback. This work was
supported by an NSF/NIH CRCNS grant (R01MH115557) and an NSF grant (DMS-1517629). ZPK was
also supported by an NSF grant (DMS-1615737). KJ was also supported by NSF grant DBI-1707400.
WRH was supported by NSF grant SES-1556325.
References
1. Gold, J. I. & Shadlen, M. N. The neural basis of decision making. Annu. review neuroscience 30
(2007).
2. Britten, K. H., Shadlen, M. N., Newsome, W. T. & Movshon, J. A. The analysis of visual motion: a
comparison of neuronal and psychophysical performance. J. Neurosci. 12, 4745 -- 4765 (1992).
3. Bogacz, R., Brown, E., Moehlis, J., Holmes, P. & Cohen, J. D. The physics of optimal decision
making: a formal analysis of models of performance in two-alternative forced-choice tasks. Psychol.
review 113, 700 (2006).
4. Gao, P. et al. A theory of multineuronal dimensionality, dynamics and measurement. bioRxiv 214262
(2017).
5. Glaze, C. M., Kable, J. W. & Gold, J. I. Normative evidence accumulation in unpredictable environ-
ments. Elife 4, e08825 (2015).
6. **Piet, A. T., El Hady, A. & Brody, C. D. Rats adopt the optimal timescale for evidence integration in
a dynamic environment. Nat. Commun. 9, 4265 (2018).
Rats can learn to optimally discount evidence when deciding between two dynamically switching auditory click
streams, and they adapted to shifts in environmental change rates.
7. Adams, R. P. & MacKay, D. J. Bayesian online changepoint detection. arXiv preprint arXiv:0710.3742
(2007).
8. Radillo, A. E., Veliz-Cuba, A., Josi´c, K. & Kilpatrick, Z. P. Evidence accumulation and change rate
inference in dynamic environments. Neural computation 29, 1561 -- 1610 (2017).
8/11
9. **Glaze, C. M., Filipowicz, A. L., Kable, J. W., Balasubramanian, V. & Gold, J. I. A bias -- variance
trade-off governs individual differences in on-line learning in an unpredictable environment. Nat.
Hum. Behav. 2, 213 (2018).
Human performing a dynamic triangles task use decision strategies that obey a trade-off in which history-
dependent adaptive strategies lead to higher choice variability. A sampling strategy best accounted for subject
data.
10. Krishnamurthy, K., Nassar, M. R., Sarode, S. & Gold, J. I. Arousal-related adjustments of perceptual
biases optimize perception in dynamic environments. Nat. human behaviour 1, 0107 (2017).
11. Goldfarb, S., Wong-Lin, K., Schwemmer, M., Leonard, N. E. & Holmes, P. Can post-error dynamics
explain sequential reaction time patterns? Front. Psychol. 3, 213 (2012).
12. Purcell, B. A. & Kiani, R. Hierarchical decision processes that operate over distinct timescales
underlie choice and changes in strategy. Proc. Natl. Acad. Sci. 113, E4531 -- E4540 (2016).
13. **Malhotra, G., Leslie, D. S., Ludwig, C. J. & Bogacz, R. Overcoming indecision by changing the
decision boundary. J. Exp. Psychol. Gen. 146, 776 (2017).
Humans' decision strategies in tasks where difficulty varies trial-to-trial are well approximated by a drift-
diffusion model with time-varying decision boundaries. Subject deviations from this normative model did little
to impact the reward rate.
14. Geisler, W. S. Ideal observer analysis. The visual neurosciences 10, 12 -- 12 (2003).
15. Knill, D. C. & Pouget, A. The Bayesian brain: the role of uncertainty in neural coding and computation.
Trends Neurosci. 27, 712 -- 719 (2004).
16. Kim, T. D., Kabir, M. & Gold, J. I. Coupled decision processes update and maintain saccadic priors
in a dynamic environment. J. Neurosci. 3078 -- 16 (2017).
17. *Veliz-Cuba, A., Kilpatrick, Z. P. & Josic, K. Stochastic models of evidence accumulation in changing
environments. SIAM Rev. 58, 264 -- 289 (2016).
Derivation and analysis of nonlinear stochastic models of evidence accumulation in dynamic environments for
decisions between two and more alternatives. Optimal evidence discounting can be implemented by a mutual
excitatory neural population model.
18. Drugowitsch, J., Moreno-Bote, R., Churchland, A. K., Shadlen, M. N. & Pouget, A. The cost of
accumulating evidence in perceptual decision making. J. Neurosci. 32, 3612 -- 3628 (2012).
19. *Holmes, W. R., Trueblood, J. S. & Heathcote, A. A new framework for modeling decisions about
changing information: The piecewise linear ballistic accumulator model. Cogn. psychology 85, 1 -- 29
(2016).
In this study, humans performed a RDMD task in which the direction of dots sometimes switched midtrial.
Hierarchal Bayes' fits of a piecewise linear accumulator model demonstrate subjects react slowly to new
evidence and that the percieved strength of post-switch evidence is influenced by pre-switch evidence strength.
20. Holmes, W. R. & Trueblood, J. S. Bayesian analysis of the piecewise diffusion decision model. Behav.
research methods 50, 730 -- 743 (2018).
21. Yu, A. J. & Cohen, J. D. Sequential effects: Superstition or rational behavior? Adv. Neural Inf.
Process. Syst. 21, 1873 -- 1880 (2008).
22. Brunton, B. W., Botvinick, M. M. & Brody, C. D. Rats and humans can optimally accumulate
evidence for decision-making. Sci. 340, 95 -- 98 (2013).
9/11
23. Ossmy, O. et al. The timescale of perceptual evidence integration can be adapted to the environment.
Curr. biology : CB 23, 981 -- 986 (2013).
24. Tavoni, G., Balasubramanian, V. & Gold, J. I. On the complexity of predictive strategies in noisy and
changing environments. In Computational and Systems Neuroscience (CoSyNe) (Denver CO, Mar
1-4, 2018).
25. Fernberger, S. W. Interdependence of judgments within the series for the method of constant stimuli.
J. Exp. Psychol. 3, 126 (1920).
26. Frund, I., Wichmann, F. A. & Macke, J. H. Quantifying the effect of intertrial dependence on
perceptual decisions. J. vision 14, 9 -- 9 (2014).
27. *Urai, A. E., Braun, A. & Donner, T. H. Pupil-linked arousal is driven by decision uncertainty and
alters serial choice bias. Nat. Commun. 8, 14637 (2017).
Increases in pupil diameter can be used to predict choice alternations in serial decisions, providing a promising
non-invasive avenue for validating adaptive decision making behavior.
28. Nguyen, K. P., Josi´c, K. & Kilpatrick, Z. P. Optimizing sequential decisions in the drift -- diffusion
model. J. Math. Psychol. 88, 32 -- 47 (2019).
29. Hermoso-Mendizabal, A. et al. Response outcomes gate the impact of expectations on perceptual
decisions. bioRxiv 433409 (2018).
30. Anderson, N. Effect of first-order conditional probability in two-choice learning situation. J. Exp.
Psychol. 59, 73 -- 93 (1960).
31. White, C. N. & Poldrack, R. A. Decomposing bias in different types of simple decisions. J. Exp.
Psychol. Learn. Mem. Cogn. 40, 385 (2014).
32. Ratcliff, R. Theoretical interpretations of the speed and accuracy of positive and negative responses.
Psychol. review 92, 212 (1985).
33. Diederich, A. & Busemeyer, J. R. Modeling the effects of payoff on response bias in a perceptual
discrimination task: Bound-change, drift-rate-change, or two-stage-processing hypothesis. Percept. &
Psychophys. 68, 194 -- 207 (2006).
34. Urai, A. E., de Gee, J. W. & Donner, T. H. Choice history biases subsequent evidence accumulation.
bioRxiv 251595 (2018).
35. Olianezhad, F., Tohidi-Moghaddam, M., Zabbah, S. & Ebrahimpour, R. Residual Information of
Previous Decision Affects Evidence Accumulation in Current Decision. arXiv.org (2016). 1611.
03965v2.
36. Braun, A., Urai, A. E. & Donner, T. H. Adaptive history biases result from confidence-weighted
accumulation of past choices. J. Neurosci. 2189 -- 17 (2018).
37. Deneve, S. Making decisions with unknown sensory reliability. Front. neuroscience 6, 75 (2012).
38. Zhang, S., Lee, M. D., Vandekerckhove, J., Maris, G. & Wagenmakers, E.-J. Time-varying boundaries
for diffusion models of decision making and response time. Front. Psychol. 5, 1364 (2014).
39. Purcell, B. A. & Kiani, R. Neural mechanisms of post-error adjustments of decision policy in parietal
cortex. Neuron 89, 658 -- 671 (2016).
40. Thura, D., Beauregard-Racine, J., Fradet, C.-W. & Cisek, P. Decision making by urgency gating:
theory and experimental support. J. neurophysiology 108, 2912 -- 2930 (2012).
10/11
41. Johnson, B., Verma, R., Sun, M. & Hanks, T. D. Characterization of decision commitment rule
alterations during an auditory change detection task. J. neurophysiology 118, 2526 -- 2536 (2017).
42. Tajima, S., Drugowitsch, J. & Pouget, A. Optimal policy for value-based decision-making. Nat.
communications 7, 12400 (2016).
43. Hawkins, G. E., Forstmann, B. U., Wagenmakers, E.-J., Ratcliff, R. & Brown, S. D. Revisiting the
evidence for collapsing boundaries and urgency signals in perceptual decision-making. J. Neurosci.
35, 2476 -- 2484 (2015).
44. Evans, N. J., Hawkins, G. E., Boehm, U., Wagenmakers, E.-J. & Brown, S. D. The computations that
support simple decision-making: A comparison between the diffusion and urgency-gating models.
Sci. reports 7, 16433 (2017).
45. Hawkins, G., Wagenmakers, E., Ratcliff, R. & Brown, S. Discriminating evidence accumulation from
urgency signals in speeded decision making. J. neurophysiology 114, 40 -- 47 (2015).
46. Palestro, J. J., Weichart, E., Sederberg, P. B. & Turner, B. M. Some task demands induce collapsing
bounds: Evidence from a behavioral analysis. Psychon. bulletin & review 1 -- 24.
47. Khodadadi, A., Fakhari, P. & Busemeyer, J. R. Learning to allocate limited time to decisions with
different expected outcomes. Cogn. psychology 95, 17 -- 49 (2017).
48. Drugowitsch, J., DeAngelis, G. C., Angelaki, D. E. & Pouget, A. Tuning the speed-accuracy trade-off
to maximize reward rate in multisensory decision-making. Elife 4, e06678 (2015).
49. Wilson, R. C., Nassar, M. R. & Gold, J. I. Bayesian online learning of the hazard rate in change-point
problems. Neural Comput. 22, 2452 -- 2476 (2010).
50. Bronfman, Z. Z. et al. Decisions reduce sensitivity to subsequent information. Proc. Royal Soc. B:
Biol. Sci. 282 (2015).
51. Talluri, B. C., Urai, A. E., Tsetsos, K., Usher, M. & Donner, T. H. Confirmation bias through selective
overweighting of choice-consistent evidence. Curr. Biol. 28, 3128 -- 3135 (2018).
52. Beck, J. M., Ma, W. J., Pitkow, X., Latham, P. E. & Pouget, A. Not noisy, just wrong: the role of
suboptimal inference in behavioral variability. Neuron 74, 30 -- 39 (2012).
53. Feldman, J. & Hanna, J. F. The structure of responses to a sequence of binary events. J. Math. Psychol.
3, 371 -- 387 (1966).
54. Hanks, T., Kiani, R. & Shadlen, M. N. A neural mechanism of speed-accuracy tradeoff in macaque
area lip. Elife 3, e02260 (2014).
55. Wu, Z., Schrater, P. & Pitkow, X. Inverse POMDP: Inferring what you think from what you do. arXiv
preprint arXiv:1805.09864 (2018).
56. Bialek, W., Nemenman, I. & Tishby, N. Predictability, complexity, and learning. Neural computation
13, 2409 -- 2463 (2001).
57. Thura, D. & Cisek, P. The basal ganglia do not select reach targets but control the urgency of
commitment. Neuron 95, 1160 -- 1170 (2017).
58. Akrami, A., Kopec, C. D., Diamond, M. E. & Brody, C. D. Posterior parietal cortex represents sensory
history and mediates its effects on behaviour. Nat. 554, 368 (2018).
59. Nassar, M. R. et al. Rational regulation of learning dynamics by pupil-linked arousal systems. Nat.
neuroscience 15, 1040 (2012).
11/11
|
1902.07509 | 1 | 1902 | 2019-02-20T11:12:11 | Auditory information loss in real-world listening environments | [
"q-bio.NC"
] | Whether animal or speech communication, environmental sounds, or music -- all sounds carry some information. Sound sources are embedded in acoustic environments that contain any number of additional sources that emit sounds that reach the listener's ears concurrently. It is up to the listener to decode the acoustic informational mix, determine which sources are of interest, decide whether extra resources should be allocated to extracting more information from them, or act upon them. While decision making is a high-level process that is accomplished by the listener's cognition, selection and elimination of acoustic information is manifest along the entire auditory system, from periphery to cortex. This review examines latent informational paradigms in hearing research and demonstrates how several hearing mechanisms conspire to gradually eliminate information from the auditory sensory channel. It is motivated through the computational need of the brain to decomplexify unpredictable real-world signals in real time. Decomplexification through information loss is suggested to constitute a unifying principle of the mammalian hearing system, which is specifically demonstrated in human hearing. This perspective can be readily generalised to other sensory modalities. | q-bio.NC | q-bio | Auditory information loss in real-world listening environments
9
1
0
2
b
e
F
0
2
]
.
C
N
o
i
b
-
q
[
1
v
9
0
5
7
0
.
2
0
9
1
:
v
i
X
r
a
Adam Weisser
Macquarie University, Sydney, Australia
February 21, 2019
Abstract
Whether animal or speech communication, environmental sounds, or music -- all sounds carry
some information. Sound sources are embedded in acoustic environments that contain any number of
additional sources that emit sounds that reach the listener's ears concurrently. It is up to the listener
to decode the acoustic informational mix, determine which sources are of interest, decide whether extra
resources should be allocated to extracting more information from them, or act upon them. While
decision making is a high-level process that is accomplished by the listener's cognition, selection and
elimination of acoustic information is manifest along the entire auditory system, from periphery to
cortex. This review examines latent informational paradigms in hearing research and demonstrates
how several hearing mechanisms conspire to gradually eliminate information from the auditory sensory
channel. It is motivated through the computational need of the brain to decomplexify unpredictable
real-world signals in real time. Decomplexification through information loss is suggested to constitute
a unifying principle of the mammalian hearing system, which is specifically demonstrated in human
hearing. This perspective can be readily generalised to other sensory modalities.
1 Background
Various types of acoustic information are regularly invoked in hearing research -- e.g., spectral, tem-
poral, spatial, envelope, intensity, speech -- yet information theory itself (Shannon 1948) is generally
not introduced as a milestone of the field. Nevertheless, the explicit influence it has had on the re-
lated fields of cognitive psychology and neuroscience, and the more implicit roles in psychoacoustics
and signal processing, cannot be overstated. The split between formally introducing and implicitly
harnessing information in the various auditory domains may be understood given the conceptually
strict way in which information is defined and used, and the historical backlash against applying it
loosely (Shannon 1956). Additionally, projecting information theoretical concepts on the operation of
the human brain is not universally accepted, even if it dominates neuroscience (Searle 1990). Still, it is
possible to reinterpret a very broad range of auditory phenomena as motivated by information process-
ing economy, without entering any of these formal debates, by considering hearing as a communication
system operating in real-world conditions.
A brief review of some concepts from information theory that are pertinent to hearing is presented
in Section 2, followed by a breakdown of auditory research paradigms compared to real-world acoustic
environments in informational terms in Sections 3 and 4. The role of the auditory channel is then
contextualised within the general information processing of the brain in Section 5. It is then argued in
Section 6 that auditory information loss is a necessary goal and not a side-effect of the system. Lossy
compression of speech is given as a concrete example in Section 7. The lossy perceptual framework is
suggested to be applicable in other modalities in Section 8.
1
2 General concepts of information
Below is a qualitative introduction to information theory, relevant to hearing research. Comprehensive
treatments of the subjects are found, for example, in (Shannon 1948, Pierce 1980, Cover & Thomas
2006).
A generic communication system comprises a source that transmits a message to a receiver, through
a physical channel (Shannon 1948). The amount of information that a given message carries corre-
sponds to the level of uncertainty or unpredictability that it has, relative to the ensemble of all
possible messages. Thus, the degree of uncertainty (quantified using Shannon's entropy and measured
in bits) is computed solely based on the probability distribution of the messaging system, which is
independent of the meaning of the messages. The messaging convention, or its code, is unique for the
transmitter-receiver pair and the context of their communication.
Messages are transmitted over time in sequences of symbols over the information channel. The
channel is generally susceptible to noise that can lead to ambiguous reception and results in errors.
However, it is possible to combine individual symbols in sequences (codewords) so that the reception
error caused by the noise is made arbitrarily small, at the cost of a lower rate of information transmis-
sion on the channel. This is achieved by increasing the amount of redundancy in the code -- repeated
information that can be used by the receiver to disambiguate the reception in noisy channels. Alter-
natively, the code can be made more efficient by removing redundancies in a process called lossless
compression, which can significantly decrease the size of the message, but make it more susceptible
to decoding errors and ambiguity. A more aggressive process is called lossy compression, whereby
non-redundant information is removed and cannot be recovered, even when the message retains its
intelligibility. Because of its physical nature, the channel has a finite capacity to carry information
per unit time, which influences what code should be used, how efficient it should be, and what kind
of compression is entailed by it. As information can flow between systems and channels, codes and
physical constraints change throughout, but the amount of information in a channel can never exceed
the amount in the preceding channel.
The concept of redundancy can be taken a step further than in Shannon's information theory,
where patterns are compressed based on their relative probabilities in the ensemble. Algorithmic
complexity (also referred to as Kolmogorov complexity, Kolmogorov 1965) is defined as the shortest
computer program that can be written on a universal Turing machine (a computer programmed with a
generic language; Turing 1937), which reproduces the pattern. Algorithmic complexity asymptotically
matches the entropy of the sequence when it cannot be compressed further (i.e., when it is random).
Importantly, algorithmic complexity is not universally computable, i.e., there is no general way to
find the shortest computer program that can reproduce an arbitrary sequence, except for trying all
programs possible (see discussion in Cover & Thomas 2006, p. 482-484).
While messages are conveniently treated as made of discrete symbols and sequences, all communi-
cation has to be realised in continuous physical means. Physical systems, however, have unbounded
bandwidth and dynamic range, which means they produce infinite amount of information. Fortu-
nately, it is always possible to transform between the continuous and the discrete representations by
bandlimiting the communicated signal, and quantizing its instantaneous levels to fixed steps through
sampling. This discretisation may be done at any level of precision (or reproduction error) to rep-
resent the original signal using the proper choice of bandlimiting and quantization. Without loss of
generality, the bandlimited continuous signal can then be realised by modulating a fast carrier wave
with slow changing information in its complex amplitude (e.g., Couch II 2007, 237-241).
2
3 Hearing research paradigms in informational terms
The concepts of information theory -- transmitter, receiver, and channel -- have not been formally
defined in hearing research, except for sporadic applications. Yet, the informational framework is
necessary in order to account for information flow and loss produced by auditory mechanisms, con-
trast them with real-world conditions, and relate them to brain theories. Therefore, several standard
experimental concepts, methods, and tools in hearing research are functionally analysed below in in-
formational terms. This analysis does not pertain to any specific experiment, but is abstracted from
standard practices of the field (e.g., Moore 2013, Gelfand 2017).
3.1 The sound source
The message-containing source, target signal, or stimulus, is typically produced electroacoustically
using loudspeakers, or headphones, and seldom using live sources (e.g., talkers, musical instruments).
Unlike played-back recorded material of real sources, artificially generated stimuli offer superior experi-
mental control. The building blocks of hearing-test stimuli have traditionally been the pure tone, noise
burst, and impulse, and many variants thereof obtained by filtering, modulation, and superposition.
These sounds carry only the information required to elucidate specific aspects of the system operation.
For example, a (realistic, finite) pure tone is generated using five pieces of information: frequency, am-
plitude, onset and offset times, and initial phase. Therefore, it very low algorithmic complexity, given
an algorithm that can generate sinusoids. Such synthetic sounds are considerably less information-
ally rich and complex than naturally produced ecologically-relevant vocalisations (Theunissen & Elie
2014).
In any case, the choice of source constrains the kind of information that the listener must
auditorily process.
3.2 The sound receiver
The listener's auditory system, or any of its subsystems, is the recipient of the messages from the
acoustic source. Its role may be completely passive, as when acoustic properties of the ear are measured
(e.g., ear canal transfer function, otoacoustic emissions). It can also demand active listening when
decision-making about the sound is required (e.g., threshold measurement, speech intelligibility). The
signals are often designed in an attempt to tap only part of the auditory system, so that the information
they carry targets only that part and interacts with other parts in a predictable manner. For example,
if a stimulus is identical but is out-of-phase on the two ears, then whatever effect this interaural phase
difference has must be the result of binaural processing. Since the auditory system in all mammals is
very similar -- albeit with different tuning -- most in-vivo physiological data (and some psychoacoustic
data as well) about it have been obtained using animal models, and the knowledge is assumed valid for
humans as well in many cases (Long 1994). The source-receiver pair forms the basic communication
system in hearing that is mediated via the channel.
3.3 The information channel
The information channel depends on the input and output points of the measurement, and it is
usually a cascade of several distinct media. In the case of loudspeaker presentation, the room is the
immediate medium. Acoustic reflections and reverberation from the room boundaries distort the signal
and are preferably removed by situating the system in an anechoic chamber or a soundproof booth,
but may also be desirable in some cases. Either way, the experimenter has a choice in demarcating
the stimulus and channel. In headphone presentation, sound travels almost straight into the auditory
periphery, where in a series of transductions, information travels up the auditory pathways. Acoustic
inputs are then bandpass-filtered to narrowband channels, or auditory filters, which retain their order
3
tonotopically along the auditory pathways (e.g., Merzenich & Reid 1974). Typically, only information
involving neuron firing is explicitly referred to as 'coding', where the temporal, level, speech, and other
auditory codes are still being studied (Eggermont 2001, Sayles & Heinz 2017).
Regardless of the auditory channel segment that is involved in the test, all internal and external
noise and distortion sources in the system should be accounted for, to eliminate measurement biases
of the receiver's response. Other communication features should be controlled, as the availability of
information running on a parallel channel (e.g., visual Sumby & Pollack 1954), or the existence of
feedback or memory, complicates one-directional information flow assumption (Marko 1973).
3.4 The task
The experimental task ties together the stimulus, channel, and listener, to bring out only certain
aspects of hearing.
In active tasks, decision-making is required, which implies that an underlying
mechanism must process the input and extract certain information from it. This is true regardless of
the operation -- detection, recognition, comparison, identification, estimation, etc. -- all require some
computation. It may include hard-coded operations such as filtering, or neural recoding, but can also
be much more elaborate, such as recognizing speech in noise and reverberation. Finally, behavioural
studies are administered in trials separated by pauses, which allow listeners to momentarily rest and
sometimes respond not in real-time.
3.5 Functional roles
The functional roles that controlled stimuli and tasks fill for the receiver are determined by design.
Ideally, the message from the information source and its role for the receiver are completely known
by the 'omniscient' experimenter (Eggermont 2001). Stimuli are designated by the experimenter: the
roles of target, noise, masker, distractor, figure, background, etc. In and of themselves, the stimuli are
almost always meaningless for the subject, whose response space is confined to one or few task-relevant
dimensions with well-defined range. Usual tasks are designed to test the capability of the auditory
system to extract relevant information despite poor signal-to-noise ratios, competing signals, or all too
subtle cues. The performance in the task can also be presented as success or failure amid challenging
conditions, where in some cases, confusion, distraction, fatigue, mind-wandering, or other high level
extraneous factors affect the listener.
4 Real-world environments and listening
Everyday listening entails fundamentally different situations to those set up in laboratory-based experi-
ments (Theunissen & Elie 2014), as are prescribed by the experimental paradigms above (Buracas & Albright
1999), and may entail situations that could potentially overwhelm the listener in comparison. Imagine
being in a small and busy street: every minute dozens of cars of different size move in different direc-
tions and speeds. People are walking on sidewalks, coming in and out of stores, chatting, shouting, or
eating in public. Occasional sounds of music and machinery come from within the buildings. Some
passers-by carry devices that emit all sorts of electronic sounds. Birds are calling from the treetops.
Occasionally, an aeroplane flies above, or a siren goes off. All sounds reverberate between the build-
ing facades, and happen more or less concurrently. While each aspect of this acoustical display is
completely controlled in a laboratory setting, it is not the case in the actual street.
Several differences to laboratory-based situations are evident. It might be tempting to refer to such
a scene simply as noise -- a collection of unwanted sounds (cf., Schafer 1994, p. 273) -- but this sound
contains rich information about the street and the life within it. Moreover, the number of acoustic
sources in such an environment are generally unknown and unbounded, so the listener cannot have
4
complete knowledge of the identity of all of them. Also, the listener has no absolute control over what
kind of sounds they will generate, which means that auditory inputs are inherently unpredictable and
cannot be forced into a closed categorical response space. As a corollary to the continuous nature of
real sources (Section 2), the natural acoustic environment contains an infinite amount of information.
Importantly, listeners have no predefined 'task', but rather continuously negotiate what source(s), if
any, should count as an instantaneous 'target', while some undesired sound(s) as 'noise'. This happens
automatically (unconsciously) and dynamically, while other external and internal stimuli that are not
necessarily acoustic compete for the listener's attention. Therefore, the finite-resourced listener must
select what information, if any, to process, whereas unused information is filtered, suppressed, ignored,
deselected, or turned into noise.
5 Brain theories
The auditory system itself constitutes only one channel of perceptual information that feeds into the
brain's decision making processes. As the brain negotiates signals from different internal and external
inputs, infinite information rates from hearing may be unsuitable for efficient real-time operation of the
finite brain without some preprocessing of the inputs by the auditory system itself. Rather disturbing
evidence for the possibility of computationally overloading the system is evident from a review by
(Lipowski 1975) of studies that subjected listeners to stimuli of chaotic displays of loud noises and
bright lights. The resultant sensory stimulation was so overwhelming for some individuals, that they
experienced heightened arousal, anxiety, sadness, aggression, paranoia, and even involuntary sleeping
and hallucinations1. As extreme as these conditions are, they are most definitely something to be
avoided for healthy survival.
Fortunately, auditory information is finite. Hearing is bandlimited and has a finite dynamic range
-- set between the system internal noise (spontaneous neural firing) and the threshold of pain, where
the cochlear hair cells get mechanically damaged. These physical constraints yield an upper bound
on the auditory input information rate. Furthermore, neuronal recordings of modulation frequency
along the auditory pathways (a proxy for information rate) suggest that the maximal rate gradually
decreases on the way to cortex (Joris et al. 2004, Figure 9), so information is eliminated or compressed
the higher up signals are neurally recoded in the ascending auditory pathways. How is this auditory
information reduction is achieved?
According to the influential Efficient Coding Hypothesis (Attneave 1954, Barlow 1961), sensory
information from the environment naturally contains redundancies that are gradually removed as
the neuronal signal gets to cortex and eventually to consciousness. The hypothesis was reiterated
by (Barlow 2001), stressing that redundancies are critical, but not because they are removed, but
rather because they are used to generate statistics -- an internal model about the world that provides
the priors necessary for prediction and error correction of future sensory information2. Redundant
patterns are then represented in multiple neurons rather than eliminated by lossless compression (see
also Chater & Vit´anyi 2003).
However, it is not possible to universally apply lossless compression, for arbitrary degree of com-
plexity of realistic stimuli, since no single algorithm exists that can identify hidden redundancies.
Thus, it is argued here, there is no reason to assume that the brain is always able to detect those
redundancies, if they require more intricate computational processes to uncover than mere statisti-
cal pattern detection.
In other words, because finding the algorithmic complexity of a sequence is
an intractable problem, the brain should be not be capable of reducing just any input signal to its
1These tests undoubtedly placed significant emotional stress on the subjects -- something which could have interacted with
the mere perceptual overloading effects.
2In fact, redundancies are thought to not be compressed in vision (Barlow 2001), but they may be in audition (Chechik et al.
2006, Smith & Lewicki 2006).
5
most efficient form and always simplify its representation losslessly.
Instead, as will be illustrated
in Section 6, it uses a lossy compressive toolkit that aggressively reduces the information rate and
complexity of stimuli from the environment, while still allowing survival. It explains why the realis-
tic street scene (Section 4) does not acoustically overload the normal functioning individual, unlike
the completely chaotic stimuli that are potentially irreducible and incompressible (Lipowski 1975).
This is a demonstration of a rate-distortion theoretical perspective3, which showed that there exists
a trade-off for organisms between the perceptual precision and its energetic cost due to processing
(Marzen & DeDeo 2017). According to this view, perceiving the environment in high fidelity is useful
only inasmuch as it promotes survival, where additional, costly information can be safely discarded.
In another relevant perspective, predictive coding theories emphasise how the brain tries to min-
imise the surprise from perceptual inputs, which can be done using adequate statistical modelling of
the world, combined with action (Clark 2013). While certain variants of this theory are framed in
terms of precision-complexity trade-off (Friston 2010), in many realistic cases it would be incoherent
to talk about active predictive modelling of the world based on information that never reaches the
predicting system, due to its finite input aperture. Instead, by restricting information along the per-
ceptual pathways the system is heuristically geared either to not bother about missing information, or
to accept it as ambient noise. Only when the stimulus results in error signals that cannot be ignored,
does the brain have to respond by acquiring better perceptual information, or forming more precise
models about these otherwise neglected aspects of the world.
Considering the brain function along with real-world information content, it is argued here that
information reduction is a requirement, rather than a side-effect of perception, in order to avert
informational overload on the brain. This is in line with what many auditory mechanisms achieve,
as can be appreciated if their operation is reframed as follows: Instead of focusing on the failure to
extract information from synthetic stimuli, the successful elimination of information in complex scenes
is emphasised, because it relieves the system from avoidable and costly information processing. This
is demonstrated in the next section.
6 Auditory mechanisms
Interpreting how auditory processing of real-world signals operates based on known mechanisms from
controlled laboratory observations is not trivial for several reasons. First, applying these mechanisms
quantitatively requires complete knowledge of all sound sources in the environments, their absolute and
relative levels, and their time-dependent relative positions -- knowledge that is generally unavailable
and, as was argued above, unknowable. Second, knowledge about auditory mechanisms was obtained
using synthetic stimuli that are rarely -- if ever -- found outside of the lab. Inferring instantaneous
responses to realistic sounds based on responses to idealised pure tones and noise bursts (for example)
is unintuitive and complex. Arguably, this classical approach is of limited value for dynamic signals
such as speech (e.g., Sharma et al. 2017). Third, many mechanisms have counter-mechanisms, e.g.,
masking phenomena have masking release, masked sounds can be glimpsed at or restored, grouping
and segregation in scene analysis, attention can switch or remain focused both due to competing
processes -- top-bottom processes driven by context, or bottom-up driven by the salience of different
events. Observations about when these counter-effects are expected are usually available for particular
conditions only (e.g., spatial release from masking for certain targets that are of equal levels and exactly
0◦ or 90◦ apart, but not for arbitrary targets, levels, and locations).
Figure 1 shows a conceptual model of acoustic information flowing into the auditory periphery,
through its ascending pathways, all the way to cognition and perception. Gradually, with every
processing level, information is eliminated from the original input (cf. Buracas & Albright 1999).
3Rate-distortion theory is the branch of information theory the is concerned with optimizing the received errors in finite-
capacity channels (Shannon 1948, Cover & Thomas 2006).
6
Action
Perception / Cognition
Motivation
Attention
Auditory scene analysis
Other Modalities
Periphery
Source 1
Source 2
Source 3
Source 4
Source 5
Source 6
Figure 1: A conceptual model of auditory information loss with multiple acoustic sources. The wavy
arrows designate flow of acoustic information of arbitrary nature (e.g., spectral, temporal, spatial)
transmitted from the acoustic sources. The number of arrows per source qualitatively indicates
the amount of information it transmits -- not all of it is picked up by the ears. Received sounds
are coded to auditory information (straight lines) and compressed (less arrows per source / colour)
as it travels up the auditory pathways. Source 4 exhibits salience (arbitrarily), so its respective
black arrows dominate perception. On a higher level, perception / cognition receives the input from
attention and inform action, which may also be modulated by motivation. Action, in turn, can affect
what information reaches the periphery (e.g., by head turns). Other general feedback mechanisms
are illustrated between perception / cognition and attention (e.g., context) and attention and scene
analysis (e.g., switched attention), and back to the periphery.
Feedback mechanisms exist between all levels that can adapt the system response to the instantaneous
needs of the listener, also considering non-auditory inputs. Each processing level encompasses several
mechanisms that can be observed using particular experimental paradigms. However, some mecha-
nisms may overlap in function and their physiological realization is distributed over different auditory
circuits, in a way that is not always fully understood. A subset of these mechanisms is described below
with emphasis on their informational lossiness, but a more comprehensive set is provided in (Weisser
2018, p. 143-162). None of the phenomena is exclusive to humans, unless otherwise stated.
6.1 Low-level mechanisms
The first group of mechanisms contains phenomena that are largely associated with the periphery (the
outer, middle and inner ears, and the auditory nerve), but are not confined to it, as some mechanisms
have correlates in neural auditory circuits. Limited bandwidth and dynamic range, which have already
been mentioned above, are the two fundamental properties that turn hearing to finite. They are related
mainly to the ear mechanics and geometry and are species-specific (Heffner 2004).
The two ears are sometimes treated as two independent channels of acoustic information (e.g.,
Cherry 1953). In reality, natural signals tend to be highly correlated between the two ears, and sound
events are experienced to be unitary -- auditory information is used to identify and to localise acoustic
events in space (Rauschecker 2017). Hence, acoustic sources in space can be efficiently expressed using
a monaural signal along with additional interaural parameters (time, phase, or level differences) or
functions (head-related transfer function, interaural coherence). Importantly, the binaural parameters
are also updated relatively sluggishly (Grantham & Wightman 1978). This suggests that correlated
left- and right-ear time signals can be compressed to a single signal plus a low-rate spatial transfor-
7
mation (e.g., Johnston & Ferreira 1992).
More complex loss mechanisms involve multiple sources. 'Energetic masking' is generally associated
with the cochlea and auditory nerve, and may be classified as simultaneous, forward or backward
(Moore 2013, Chapter 3). All types are characterised by elevation of threshold in specific auditory
bands as a result of a masking sound. In informational terms, because masking limits the channel-
specific dynamic range, its channel capacity is reduced. With rare exceptions (Ben-Shalom et al. 2003),
masking has not been portrayed as a desirable property of the hearing system, because it results in
loss of information. Somewhat confusingly, 'informational masking' occurs whenever energetic masking
cannot explain a change of threshold, as in speech-on-speech tests (Brungart et al. 2006)4.
Applying masking to real-world situations is particularly confusing, because the loss from masking
is not inevitable, as some acoustic cues can 'unmask', or release the signal from energetic or informa-
tional masking (Culling & Stone 2017, Kidd Jr & Colburn 2017, Kidd Jr et al. 2008). For example,
depending on the shape of the animal's outer ears and head (Holt & Schusterman 2007), spatial re-
lease from masking can happen when the sources are not exactly co-located (unlike how most masking
experiments are set up). Realistic sources are almost never co-located, as it is a physical impossibility,
unless they all come from a single loudspeaker. Similarly for other release cues, realistic sources are
generally uncorrelated, so their modulations and fundamental frequencies rarely match exactly. Still,
it is likely that energetic masking is so dominant in some cases that no cues could release it, espe-
cially at large sound level differences between target and masker, or for large incoherent sources that
occupy a wide spatial angle. Therefore, information is lost from masking when possible unmasking
fails. In this case, sound glimpsing and restoration may be sometimes applied to reconstruct partially
corrupted messages (Miller & Licklider 1950, Cooke 2006, Warren 2008, Chapter 6).
6.2 Auditory scene analysis
After the acoustic information is spectrally analyzed in the cochlea and gets neurally recoded, bits
of auditory information are (re-)grouped to produce contingent auditory objects. This preattentive,
complex, and lossy process is called auditory scene analysis (Bregman 1990). Spectro-temporal pat-
terns that share the same acoustical properties such as being harmonically related, timbrally identical,
having common onset time, or coming from the same direction, are often processed as a single au-
ditory object or event, effectively transforming multi-event / channel input to a grouped perceptual
stream (Bregman 1990, Bee & Micheyl 2008). Grouping is a powerful method to cognitively han-
dle high information load (Miller 1956), but it is often lossy, as it is accompanied by the removal
of fine-grained details of the time series, which are not needed to maintain the sound event identity
over time (McAdams 1993, p. 182-183). The supremacy of grouped or categorised auditory informa-
tion can be seen in the categorization of acoustically ambiguous phonemes (Ganong 1980), in timbral
constancy despite acoustic manipulations (Charbonneau 1981), and in the consistent identification of
glass breaking sounds that were heavily tampered with acoustically (Warren & Verbrugge 1984).
If the auditory system creates objects and streams and only selects them later (Sussman et al.
2007, but see Shinn-Cunningham et al. 2017), then it requires the computational resources to be able to
process multiple streams in parallel. However, the number of simultaneous streams that can be handled
is typically smaller than four (in humans), depending on the precise stimuli and measurement setup
(Kawashima & Sato 2015, Zhong & Yost 2017, Weller et al. 2016), so complex scenes may transmit
more acoustic information than can be tracked in the scene analysis stage.
4Informational masking has not been studied yet in non-human animals (Reichmuth 2012).
8
6.3 Attention
Attention is the mental stage in which competing sensory inputs are singled out to be further pro-
cessed by cognition, which can drive subsequent decision making and action. William James famously
asserted that it is impossible to attend to more than one thing at once, unless it implies very simple
or highly automated processes (William 1890, p. 402-458). This idea metamorphosed to modern
attention theories, which combined it implicitly with the concept of channel capacity, by emphasizing
the finite resources that attention has to deal with multiple sensory inputs (e.g., Broadbent 1958,
Kahneman 1973)5. Thus, attentional suppression of auditory streams may be the only mechanism
that is openly acknowledged in literature to have a desirable role in eliminating auditory informa-
tion (Golumbic et al. 2013, Ding & Simon 2012). Attention appears to gradually evolve or emerge
in the ascending auditory pathways (Shinn-Cunningham et al. 2017), while unattended streams may
have to be suppressed to not interfere with the attended signal (Fritz et al. 2007), as was found in
several auditory evoked potential studies in humans (e.g., Horton et al. 2013, Kong et al. 2014). Un-
der certain conditions, attention modulates peripheral responses via the medial olivocochlear efferents
(Hern´andez-P´erez 2018), and is likely modulated in itself by higher-level factors, such as motivation,
as was shown in vision (Balcetis & Dunning 2006) (see Figure 1).
In real environments, not all acoustic sources are equal and some are more salient (e.g., louder,
having impulsive onset), which can serve as cue for capturing attention through bottom-up processes
(Koch & Ullman 1987, Kayser et al. 2005). This comes at the expense of other streams due to a
'Winner-Take-All' architecture that may underlie this type of processing (De Coensel & Botteldooren
2010, Koch & Ullman 1987). Similarly, during attentive listening, certain sound events can be percep-
tually placed in the foreground and dominate most or all of the listener's focused attention, making
other surrounding sounds effectively inaudible. This 'inattentional deafness' takes place when (human)
listeners attend to specific elements of a sound scene, but they can be completely oblivious to other
sounds that are otherwise clearly audible (Koreimann et al. 2014, Dalton & Fraenkel 2012).
6.4 Cognitive and other high-level mechanisms
Information loss may occur despite successful auditory stream analysis and focused attention. This can
happen because the sounds cannot be categorised by listeners (see Pylyshyn 2001), the sounds are not
represented in long-term memory in any accessible form (McAdams 1993), or the sound event context
is unclear and not helpful for determining their identity with certainty (Ballas 1993). Unfamiliarity
also seems to require more resources, as coding of unknown sounds may be less efficient compared to
known sounds (Attias & Schreiner 1997, Fernald et al. 2006).
Sound processing and perception depend on different forms of auditory memory. Such are the
real-time generation of streams from sequences of sounds (short-term auditory memory), promotion
of sound identification (long-term auditory memory), or even the facilitation of forward masking and
temporal integration (Demany & Semal 2008). Additionally, basic computational processes such as
comparison, discrimination, and adaptation also require the function of memory. Unsurprisingly,
there are multiple possibilities for information loss because of memory-related phenomena, or rather,
forgetting. This can be part of the normal operation of the auditory system (McKeown & Mercer
2012), or due to interference of one stimulus with another similar, ongoing one, so that the information
about the former is lost (Deutsch 1970, Starr & Pitt 1997). Informational bottlenecks can stem from
limited working memory, or other executive processing measures, where processing of information over
longer-time frames is required (e.g. Baddeley 2012, Cowan 2010).
5Neither Broadbent nor Kahneman cited Shannon's work explicitly, but both mentioned information theory. However,
all modern attention theories were inspired by the seminal cocktail party problem (Cherry 1953), where Shannon's work
on written English redundancy was explicitly cited (Shannon 1951). Information channels are still central concepts in the
attention literature (e.g., Pashler 1998, Wickens & McCarley 2008).
9
Talker / Source
Listener / Receiver
Neural Codes
Thought
Language
Motor
Speech
Signal
Hearing
Language
Thought
Acoustics
Other Modalities
Linguistics
Psychology
Figure 2: Different levels of information flow of human speech. The lowest level (innermost in the
figure) is the exact acoustic signal that the source transmits to the receiver (the talker says to the
listener). This signal contains the highest information rate. One level up (outward), the motor system
is responsible for modulating the airflow that results in speech production, which is maybe mirrored
in speech perception as well (Liberman et al. 1967, Poeppel et al. 2007), among other parallel non-
acoustic channels (e.g., Sumby & Pollack 1954, McGurk & MacDonald 1976). The linguistic level is
one level higher, which already has much lower and more efficient informational content than the
acoustic signal (Chomsky 1956). Finally, on an abstract psychological level, there is thought (goals,
emotion, motivation, etc.) that generates / parses the messages to / from the linguistic modality. All
information transfers that occurs within the brain are coded in intermediate neural signals, which
may also include feedback and feedforward loops. Noise sources for all channels are not shown.
If exacerbated, low motivation (Balcetis & Dunning 2006), physical or mental fatigue, or competing
sensory information (e.g., Stein 1998, Lavie 2010, Parmentier 2014), can affect performance and lead to
switching of the listener's attention away from the auditory stream, in favour of an altogether different
sensory input, or mind-wandering by task-irrelevant thoughts (Lavie 2010). Similarly, listeners may
physically react to what they hear (in decision-making) or engage in another behaviour that competes
for attentional resources, or reorients perception, to optimise their situation through action (Friston
2010).
7 Auditory information rate examples
Because of the multitude of auditory subsystems and special classes of sound, it is possible to tap
the information flow at different places along the auditory pathways, or pick only a subset of acoustic
situations and model them using a code that explains most of the effect. An example of the multiple
processing levels, codes, and rates of information involved is given in Figure 2, which presents a speech-
centric adaptation of Shannon's original communication system (Shannon 1948). In the figure, the
information communicated -- human language in this case -- may be analysed using codes in different
levels of abstraction: acoustic, sensory-motor (or visual), neuronal, linguistic, and psychological. Some
of these codes are much more readily operationalised than others (cf. Wiener 1950).
The lowest level of information input to the hearing system is through sound waves that impinge
on the outer ears. Each ear receives an approximately one-dimensional pressure wave reaching its
eardrum. The acoustic signal can be sampled at more than twice its bandwidth, so typically at 44.1
kHz and bit depth of 16 bits/samples, for humans, the continuous external acoustic world is brought
to a finite level of 705,600 bits/s per ear. A 4-11 times reduction of this rate is achieved with negligible
loss of fidelity (depending on the sampled signal) by eliminating parts of sound, which would anyway be
10
inaudible due to masking, among other methods of data compression (Musmann 2006). However, the
acoustic fidelity can be sacrificed to reduce the rate further, with negligible loss in speech intelligibility.
In modern telephony standards, the signal bit rate can be compressed at 100-200 times the size of the
uncompressed signal (4.75-23.85 kbits/s) (Gibson 2016).
Continuing to focus on speech communication as a particularly well-studied case -- if the acoustic
packaging of language production is completely abstracted, then the linguistic message is much smaller
than its full carrier signal (Flanagan 1965, p. 3-6). It has been empirically estimated using information
processing response time task that the human cognition processes a maximum of 52 bits/s (Mart´ın
2011), and about 35 bits/s for speech using acoustic methods (Leijon 2001). Even if these bounds
are underestimated, they are still 3-4 orders of magnitude lower than the full acoustic signal, which
contains much unnecessary information insofar as the high-level message is concerned, as well as useful
information that may not be recoverable such as intonation, emotion, and voice-specific cues. The
lossy aspect of language compression is most evident in written language, which is very low in bit rate
compared to sound6. English text, for example, can be compressed by 50-75% based on redundancy
alone (Shannon 1951), so that a speaking rate of three five-letter words per second can be textually
transmitted at about 18 bit/s.
8 Beyond hearing
The unifying principle of the above auditory mechanisms -- being informationally lossy -- can be readily
generalised to other modalities (cf., Buracas & Albright 1999). The laboratory-based methods and
relevance of the communication-theoretical framework can be translated to other senses, without loss of
generality. Many of the mechanisms reviewed above, such as dynamic range, masking, masking release,
scene analysis, selective attention, and memory, have parallels in other modalities. Additionally,
natural sensory inputs are often multimodal, which can result in complex interactions, when cross-
modal information is complementary, or in no interactions when they are completely independent
(Partan & Marler 1999). Whenever the modalities interact or are informationally redundant, some
form of compression is inevitable.
9 Conclusion
It was demonstrated mainly through human hearing -- that largely generalises to mammalian hearing
-- how the brain handles the arbitrary sensory inputs of everyday environments, by being regularly
engaged in lossy information processing. Cumulative and compressive information loss in hearing
constitutes a unifying principle of many multilevel auditory effects, which have been usually studied
in isolation in the hearing research literature.
It is maintained that acoustic stimuli of unknown
complexity are turned to processable, low-capacity output that the brain can use to enable survival,
inform world models, and allow decision making and action, by avoiding overloading the system. This
approach may be seen as a step toward parsimony in accounting for otherwise disjointed auditory
phenomena, as well as toward real-world understanding thereof. These ideas can be readily generalised
to other sensory modalities, as many of the underlying mechanisms have parallels across senses.
Acknowledgments
The author would like to thank Jorg Buchholz and Gitte Keidser for supervising the lead-up to this
work, and for help with formulating the conceptual model of Figure 1. Special thanks to Timothy
6Defining a relevant code for language is much more complicated than sound, though, since there are many speech elements
that are not embodied in written language (e.g., Beechey 2019, p. 1-10).
11
Beechey, Kelly Miles, and Liviu Sigler for their useful comments. This work was supported by the
Oticon Foundation and Macquarie University via an iMQRES PhD scholarship.
References
Attias, H. & Schreiner, C. E. (1997), Temporal low-order statistics of natural sounds, in M. C. Mozer,
M. Jordan, M. Kearns & S. Solla, eds, 'Advances in neural information processing systems', MIT
Press, Cambridge, MA, pp. 27 -- 33.
Attneave, F. (1954),
'Some informational aspects of visual perception.', Psychological Review
61(3), 183.
Baddeley, A. (2012), 'Working memory: theories, models, and controversies', Annual review of psy-
chology 63, 1 -- 29.
Balcetis, E. & Dunning, D. (2006), 'See what you want to see: motivational influences on visual
perception.', Journal of Personality and Social Psychology 91(4), 612.
Ballas, J. A. (1993), 'Common factors in the identification of an assortment of brief everyday sounds',
Journal of Experimental Psychology: Human Perception and Performance 19(2), 250.
Barlow, H. (2001),
'Redundancy reduction revisited', Network: Computation in Neural Systems
12(3), 241 -- 253.
Barlow, H. B. (1961), Possible principles underlying the transformations of sensory messages, in W. A.
Rosenblith, ed., 'Sensory Communication. Contributions to the symposium on principles of sensory
communication, July 19 -- August 1, 1959, Endicott House, M.I.T.', M.I.T. Press, pp. 217 -- 234.
Bee, M. A. & Micheyl, C. (2008), 'The cocktail party problem: what is it? how can it be solved? and
why should animal behaviorists study it?', Journal of Comparative Psychology 122(3), 235.
Beechey, T. (2019), Communication difficulty and effort in conversation, PhD thesis, Macquarie Uni-
versity.
Ben-Shalom, A., Werman, M. & Dubnov, S. (2003), Improved low bit-rate audio compression using
reduced rank ica instead of psychoacoustic modeling, in 'Acoustics, Speech, and Signal Processing,
2003. Proceedings.(ICASSP'03). 2003 IEEE International Conference on', Vol. 5, IEEE, pp. V -- 461 --
464.
Bregman, A. S. (1990), Auditory scene analysis: The perceptual organization of sound, MIT press.
Broadbent, D. E. (1958), Perception and communication, 2nd reprint edn, Pergamon Press Ltd.
Brungart, D. S., Chang, P. S., Simpson, B. D. & Wang, D. (2006), 'Isolating the energetic component
of speech-on-speech masking with ideal time-frequency segregation', The Journal of the Acoustical
Society of America 120(6), 4007 -- 4018.
Buracas, G. T. & Albright, T. D. (1999), 'Gauging sensory representations in the brain', Trends in
neurosciences 22(7), 303 -- 309.
Charbonneau, G. R. (1981), 'Timbre and the perceptual effects of three types of data reduction',
Computer Music Journal 5(2), 10 -- 19.
12
Chater, N. & Vit´anyi, P. (2003), 'Simplicity: A unifying principle in cognitive science?', Trends in
Cognitive Sciences 7(1), 19 -- 22.
Chechik, G., Anderson, M. J., Bar-Yosef, O., Young, E. D., Tishby, N. & Nelken, I. (2006), 'Reduction
of information redundancy in the ascending auditory pathway', Neuron 51(3), 359 -- 368.
Cherry, E. C. (1953), 'Some experiments on the recognition of speech, with one and with two ears',
The Journal of the Acoustical Society of America 25(5), 975 -- 979.
Chomsky, N. (1956), 'Three models for the description of language', IRE Transactions on Information
Theory 2(3), 113 -- 124.
Clark, A. (2013), 'Whatever next? predictive brains, situated agents, and the future of cognitive
science', Behavioral and Brain Sciences 36(3), 181 -- 204.
Cooke, M. (2006), 'A glimpsing model of speech perception in noise', The Journal of the Acoustical
Society of America 119(3), 1562 -- 1573.
Couch II, L. W. (2007), Digital and analog communication systems, 8th edn, Pearson Education Inc.,
Upeer Saddle River, NJ.
Cover, T. M. & Thomas, J. A. (2006), Elements of information theory, John Wiley & Sons, Hoboken,
NJ.
Cowan, N. (2010), 'The magical mystery four: How is working memory capacity limited, and why?',
Current directions in psychological science 19(1), 51 -- 57.
Culling, J. F. & Stone, M. A. (2017), Energetic masking and masking release, in J. C. Middlebrooks,
J. Z. Simon, A. N. Popper & R. R. Fay, eds, 'The Auditory System at the Cocktail Party', ASA
Press / Springer International Publishing AG, pp. 41 -- 73.
Dalton, P. & Fraenkel, N. (2012), 'Gorillas we have missed: Sustained inattentional deafness for
dynamic events', Cognition 124(3), 367 -- 372.
De Coensel, B. & Botteldooren, D. (2010), A model of saliency-based auditory attention to environ-
mental sound, in '20th International Congress on Acoustics (ICA-2010)', pp. 1 -- 8.
Demany, L. & Semal, C. (2008), The role of memory in auditory perception, in W. A. Yost, R. R. Fay
& A. N. Popper, eds, 'Auditory perception of sound sources', Springer Science & Business Media,
New York, NY, pp. 77 -- 113.
Deutsch, D. (1970), 'Tones and numbers: Specificity of interference in immediate memory', Science
168(3939), 1604 -- 1605.
Ding, N. & Simon, J. Z. (2012), 'Emergence of neural encoding of auditory objects while listening to
competing speakers', Proceedings of the National Academy of Sciences 109(29), 11854 -- 11859.
Eggermont, J. J. (2001), 'Between sound and perception: reviewing the search for a neural code',
Hearing Research 157(1-2), 1 -- 42.
Fernald, A., Perfors, A. & Marchman, V. A. (2006), 'Picking up speed in understanding: Speech pro-
cessing efficiency and vocabulary growth across the 2nd year.', Developmental Psychology 42(1), 98.
Flanagan, J. L. (1965), Speech Analysis Synthesis and Perception, 1st edn, Springer-Verlag, Berling,
Heidelberg.
13
Friston, K. (2010), 'The free-energy principle: a unified brain theory?', Nature Reviews Neuroscience
11(2), 127.
Fritz, J. B., Elhilali, M., David, S. V. & Shamma, S. A. (2007), 'Auditory attention -- focusing the
searchlight on sound', Current Opinion in Neurobiology 17(4), 437 -- 455.
Ganong, W. F. (1980), 'Phonetic categorization in auditory word perception.', Journal of experimental
psychology: Human perception and performance 6(1), 110.
Gelfand, S. A. (2017), Hearing: An introduction to psychological and physiological acoustics, 6th edn,
CRC Press, Taylor and Francis Group, Boca Raton, FL.
Gibson, J. (2016), 'Speech compression', Information 7(2), 32.
Golumbic, E. M. Z., Ding, N., Bickel, S., Lakatos, P., Schevon, C. A., McKhann, G. M., Goodman,
R. R., Emerson, R., Mehta, A. D., Simon, J. Z. et al. (2013), 'Mechanisms underlying selective
neuronal tracking of attended speech at a "cocktail party"', Neuron 77(5), 980 -- 991.
Grantham, D. W. & Wightman, F. L. (1978), 'Detectability of varying interaural temporal differ-
encesa', The Journal of the Acoustical Society of America 63(2), 511 -- 523.
Heffner, R. S. (2004), 'Primate hearing from a mammalian perspective', The Anatomical Record Part
A: Discoveries in Molecular, Cellular, and Evolutionary Biology: An Official Publication of the
American Association of Anatomists 281(1), 1111 -- 1122.
Hern´andez-P´erez, H. (2018), Disentangling the influence of attention in the auditory efferent system
during speech processing, PhD thesis, Macquarie University.
Holt, M. M. & Schusterman, R. J. (2007), 'Spatial release from masking of aerial tones in pinnipeds',
The Journal of the Acoustical Society of America 121(2), 1219 -- 1225.
Horton, C., D'Zmura, M. & Srinivasan, R. (2013), 'Suppression of competing speech through entrain-
ment of cortical oscillations', Journal of Neurophysiology 109(12), 3082 -- 3093.
Johnston, J. & Ferreira, A. J. (1992), Sum-difference stereo transform coding, in 'Acoustics, Speech,
and Signal Processing, 1992. ICASSP-92., 1992 IEEE International Conference on', Vol. 2, IEEE,
pp. 569 -- 572.
Joris, P., Schreiner, C. & Rees, A. (2004), 'Neural processing of amplitude-modulated sounds', Physi-
ological Reviews 84(2), 541 -- 577.
Kahneman, D. (1973), Attention and effort, Prentice-Hall Englewood Cliffs, NJ.
Kawashima, T. & Sato, T. (2015), 'Perceptual limits in a simulated "cocktail party"', Attention,
Perception, & Psychophysics 77(6), 2108 -- 2120.
Kayser, C., Petkov, C. I., Lippert, M. & Logothetis, N. K. (2005), 'Mechanisms for allocating auditory
attention: an auditory saliency map', Current Biology 15(21), 1943 -- 1947.
Kidd Jr, G. & Colburn, H. S. (2017), Informational masking in speech recognition, in J. C. Middle-
brooks, J. Z. Simon, A. N. Popper & R. R. Fay, eds, 'The Auditory System at the Cocktail Party',
ASA Press / Springer International Publishing AG, pp. 75 -- 109.
14
Kidd Jr, G., Mason, C. R., Richards, V. M., Gallun, F. J. & Durlach, N. I. (2008), Informational
masking, in W. A. Yost, R. R. Fay & A. N. Popper, eds, 'Auditory perception of sound sources',
Springer Science & Business Media, New York, NY, pp. 143 -- 189.
Koch, C. & Ullman, S. (1987), Shifts in selective visual attention: towards the underlying neural
circuitry, in 'Matters of intelligence', Springer, pp. 115 -- 141.
Kolmogorov, A. N. (1965), 'Three approaches to the quantitative definition of information', Problems
of Information Transmission 1(1), 1 -- 7.
Kong, Y.-Y., Mullangi, A. & Ding, N. (2014), 'Differential modulation of auditory responses to at-
tended and unattended speech in different listening conditions', Hearing Research 316, 73 -- 81.
Koreimann, S., Gula, B. & Vitouch, O. (2014), 'Inattentional deafness in music', Psychological Research
78(3), 304 -- 312.
Lavie, N. (2010), 'Attention, distraction, and cognitive control under load', Current Directions in
Psychological Science 19(3), 143 -- 148.
Leijon, A. (2001), 'Estimation of sensory informaion transmission using a hidden markov model of
speech stimuli', Acta Acustica united with Acustica 88, 423 -- 432.
Liberman, A. M., Cooper, F. S., Shankweiler, D. P. & Studdert-Kennedy, M. (1967), 'Perception of
the speech code.', Psychological Review 74(6), 431.
Lipowski, Z. (1975), 'Sensory and information inputs overload: Behavioral effects', Comprehensive
Psychiatry 16(3), 199 -- 221.
Long, G. R. (1994), Psychoacoustics, in R. R. Fay & A. N. Popper, eds, 'Comparative hearing:
mammals', Springer-Verlag, New York, Inc., pp. 18 -- 56.
Marko, H. (1973), 'The bidirectional communication theory -- a generalization of information theory',
IEEE Transactions on Communications 21(12), 1345 -- 1351.
Mart´ın, F. M. d. P. (2011), 'Macroscopic thermodynamics of reaction times', Journal of Mathematical
Psychology 55(4), 302 -- 319.
Marzen, S. E. & DeDeo, S. (2017), 'The evolution of lossy compression', Journal of The Royal Society
Interface 14(130), 20170166.
McAdams, S. (1993), Recognition of sound sources and events, in S. McAdams & E. Bigand, eds,
'Think in Sound. The Cognitive Psychology of Human Audition', Clarendon Press / Oxford Uni-
versity Press, pp. 146 -- 198.
McGurk, H. & MacDonald, J. (1976), 'Hearing lips and seeing voices', Nature 264(5588), 746.
McKeown, D. & Mercer, T. (2012), 'Short-term forgetting without interference.', Journal of Experi-
mental Psychology: Learning, Memory, and Cognition 38(4), 1057.
Merzenich, M. M. & Reid, M. D. (1974), 'Representation of the cochlea within the inferior colliculus
of the cat', Brain Research 77(3), 397 -- 415.
Miller, G. A. (1956), 'The magical number seven, plus or minus two: Some limits on our capacity for
processing information', Psychological Review 63(2), 81.
15
Miller, G. A. & Licklider, J. C. (1950), 'The intelligibility of interrupted speech', The Journal of the
Acoustical Society of America 22(2), 167 -- 173.
Moore, B. C. (2013), An introduction to the psychology of hearing, 6th edn, Brill, Leiden, Boston.
Musmann, H. G. (2006), 'Genesis of the mp3 audio coding standard', IEEE Transactions on Consumer
Electronics 52(3), 1043 -- 1049.
Parmentier, F. B. (2014), 'The cognitive determinants of behavioral distraction by deviant auditory
stimuli: A review', Psychological Research 78(3), 321 -- 338.
Partan, S. & Marler, P. (1999), 'Communication goes multimodal', Science 283(5406), 1272 -- 1273.
Pashler, H. E. (1998), The Psychology of Attention, MIT Press, Cambridge, MA.
Pierce, J. R. (1980), An introduction to information theory: symbols, signals and noise, 2nd edn, Dover
Publications, Inc., Mineola, NY.
Poeppel, D., Idsardi, W. J. & Van Wassenhove, V. (2007), 'Speech perception at the interface of
neurobiology and linguistics', Philosophical Transactions of the Royal Society B: Biological Sciences
363(1493), 1071 -- 1086.
Pylyshyn, Z. W. (2001), 'Visual indexes, preconceptual objects, and situated vision', Cognition 80(1-
2), 127 -- 158.
Rauschecker, J. P. (2017), 'Where, when, and how: Are they all sensorimotor? towards a unified view
of the dorsal pathway in vision and audition', Cortex .
Reichmuth, C. (2012), Psychophysical studies of auditory masking in marine mammals: key concepts
and new directions, in A. Popper & A. Hawkins, eds, 'The Effects of Noise on Aquatic Life', Springer
Science+Business Media, pp. 23 -- 27.
Sayles, M. & Heinz, M. G. (2017), Afferent coding and efferent control in the normal and impaired
cochlea, in G. A. Manley, A. W. Gummer, A. N. Popper & R. R. Fay, eds, 'Understanding the
Cochlea', ASA Press / Springer International Publishing AG, Cham, Switzerland, pp. 215 -- 252.
Schafer, R. M. (1994), The soundscape: Our sonic environment and the tuning of the world, Destiny
Books, Rochester, VT.
Searle, J. R. (1990), Is the brain a digital computer?, in 'Proceedings and Addresses of the American
Philosophical Association', Vol. 64, pp. 21 -- 37.
Shannon, C. E. (1948), 'A mathematical theory of communication', The Bell System Technical Journal
27(3), 379 -- 423, 623 -- 656.
Shannon, C. E. (1951), 'Prediction and entropy of printed english', Bell System Technical Journal
30(1), 50 -- 64.
Shannon, C. E. (1956), 'The bandwagon', IRE Transactions on Information Theory 2(1), 3.
Sharma, R., Vignolo, L., Schlotthauer, G., Colominas, M. A., Rufiner, H. L. & Prasanna, S. (2017),
'Empirical mode decomposition for adaptive am-fm analysis of speech: a review', Speech Commu-
nication 88, 39 -- 64.
16
Shinn-Cunningham, B., Best, V. & Lee, A. K. (2017), Auditory object formation and selection, in
J. C. Middlebrooks, J. Z. Simon, A. N. Popper & R. R. Fay, eds, 'The Auditory System at the
Cocktail Party', ASA Press / Springer International Publishing AG, pp. 7 -- 40.
Smith, E. C. & Lewicki, M. S. (2006), 'Efficient auditory coding', Nature 439(7079), 978.
Starr, G. E. & Pitt, M. A. (1997), 'Interference effects in short-term memory for timbre', The Journal
of the Acoustical Society of America 102(1), 486 -- 494.
Stein, B. E. (1998), 'Neural mechanisms for synthesizing sensory information and producing adaptive
behaviors', Experimental Brain Research 123(1-2), 124 -- 135.
Sumby, W. H. & Pollack, I. (1954), 'Visual contribution to speech intelligibility in noise', The Journal
of the Acoustical Society of America 26(2), 212 -- 215.
Sussman, E. S., Horv´ath, J., Winkler, I. & Orr, M. (2007), 'The role of attention in the formation of
auditory streams', Perception & Psychophysics 69(1), 136 -- 152.
Theunissen, F. E. & Elie, J. E. (2014), 'Neural processing of natural sounds', Nature Reviews Neuro-
science 15(6), 355.
Turing, A. M. (1937), 'On computable numbers, with an application to the entscheidungsproblem',
Proceedings of the London mathematical society 2(1), 230 -- 265.
Warren, R. M. (2008), Auditory perception: An Analysis and Synthesis, 3rd edn, Cambridge University
Press, Cambridge, UK.
Warren, W. H. & Verbrugge, R. R. (1984), 'Auditory perception of breaking and bouncing events:
a case study in ecological acoustics.', Journal of Experimental Psychology: Human perception and
performance 10(5), 704.
Weisser, A. (2018), Complex Acoustic Environments: Concepts, Methods, and Auditory Perception,
PhD thesis, Macquarie University, Sydney, Australia.
Weller, T., Best, V., Buchholz, J. M. & Young, T. (2016), 'A method for assessing auditory spatial
analysis in reverberant multitalker environments', Journal of the American Academy of Audiology
27(7), 601 -- 611.
Wickens, C. D. & McCarley, J. S. (2008), Applied attention theory, CRC press, Boca Raton, FL.
Wiener, N. (1950), 'Speech, language, and learning', The Journal of the Acoustical Society of America
22(6), 696 -- 697.
William, J. (1890), The Principles Of Psychology in Two Volumes. Vol. I, Henry Holt and Company,
New York.
Zhong, X. & Yost, W. A. (2017), 'How many images are in an auditory scene?', The Journal of the
Acoustical Society of America 141(4), 2882 -- 2892.
17
|
1301.4247 | 1 | 1301 | 2013-01-16T01:17:44 | Functional Magnetic Resonance Imaging: a study of malnourished rats | [
"q-bio.NC"
] | Malnutrition is a main public health problem in developing countries. Incidence is increasing and the mortality rate is still high. Malnutrition can leads mayor problems that can be irreversible if it is present before brain development is completed. We used BOLD (blood oxygen level-dependent effect) Functional Magnetic Resonance Imaging to investigate the regions of brain activity in malnourished rats. The food competition method was applied to a rat model to provoke malnutrition during lactation. The weight increase is delayed even if there is plenty milk available. To localize those regions of activity resulting from the trigeminal nerve stimulation, the vibrissae-barrel axis was employed due to the functional and morphological correlation between the vibrissae and the barrels. BOLD response changes caused by the trigeminal nerve stimulation on brain activity of malnourished and control rats were obtained at 7T. Results showed a major neuronal activity in malnourished rats on regions like cerebellum, somatosensorial cortex, hippocampus, and hypothalamus. This is the first study in malnourished rats and illustrates BOLD activation in various brain structures. | q-bio.NC | q-bio |
Functional Magnetic Resonance Imaging: a study of
malnourished rats
R. Martin∗, R. Godinez, A. O. Rodriguez
November 15, 2018
Department of Electrical Engineering, Universidad Autonoma Metropolitana Iztapalapa, Av. San Rafael
Atlixco 186, Mexico, DF, 09340, Mexico
Abstract
Malnutrition is a main public health problem in developing countries. Incidence is increasing and
the mortality rate is still high. Malnutrition can leads mayor problems that can be irreversible if
it is present before brain development is completed. We used BOLD Functional Magnetic Resonance
Imaging to investigate the regions of brain activity in malnourished rats. The food competition method
was applied to a rat model to provoke malnutrition during lactation. The weight increase is delayed
even if there is plenty milk available. To localize those regions of activity resulting from the trigeminal
nerve stimulation, the vibrissae-barrel axis was employed due to the functional and morphological
correlation between the vibrissae and the barrels. BOLD response changes caused by the trigeminal
nerve stimulation on brain activity of malnourished and control rats were obtained at 7T. Results
showed a major neuronal activity in malnourished rats on regions like cerebellum, somatosensorial
cortex, hippocampus, and hypothalamus. This is the first study in malnourished rats and illustrates
BOLD activation in various brain structures.
1
Introduction
The discovery of the cerebral processes of cerebral maturation in mammals, have held the opportunity
to investigate that exists a certain vulnerability in the cerebral development when there is a malnour-
ished problem, causing cerebral damage [1]. Among the important is that we can see damage at both
morphological and neurotransmitter levels [1].
Different studies in malnourished animal models, such as post-mortem, histological and imaging have
helped to discover some morphological changes including a decrease in the prenatal brain size [2], the
postnatal brain size [3], the hippocampal neural cells [4], the granular volume at the cerebellar cortex [5],
abnormalities in the dendritic spines at the cerebellar cortex [6], reduction in the myelin content [7], and
the protein synthesis of myelin [8], among others.
Other specific consequences in brain function are principally alterations in neurotransmitters. Re-
duction of serotonin levels [9] [10], increase of GABA concentrations at the hippocampus [11], reduction
in the cholinergic cells [12], increase of the AChE at the hippocampus [13], decrease in the segregation
of dopamine [14] [15], have been found such as hypothalamus, hippocampus, and cerebral cortex, areas
related to memory and learning.
Almost all of the functional in vivo records have been made with electroencephalography methods,
using both, conventional and invasive [1]. Electroencephalographic methods can give us many noisy
∗*Corresponding author: M. Sc. Rodrigo A. Martin, email: [email protected]
1
signals that can lead into a malinterpretation of results if there is not an expertise. The quality of the
electroencephalographic signals depends on the type and number of electrodes employed, that can carry
out problems at the moment of placing them. We have to consider that at the moment of inserting the
electrodes we can damage some brain tissues. Magnetic Resonance Imaging (MRI) is a powerful imaging
tool which has many advantages in comparison to other imaging techniques. MRI is a non-ionizing
technique, able to produce anatomical and functional information. Image contrast depends on the tissue
intrinsic characteristics making it ideal to study brain activity. In this work, we used functional Magnetic
Resonance Imaging (fMRI), which provides a method for mapping brain functional activity based on the
blood oxygen level-dependent effect (BOLD). The BOLD MRI technique, allows indirectly to study the
whole brain activation, in an animal model under a malnourished program.
As far as we know, this is the first fMRI study on malnutrition. To provoke malnutrition, the food
competition method was applied [16]. This method consists in inducing malnutrition during lactation
through food competition. A large number of pups cannot be sufficiently fed by one nursing mother. Then,
a delay in the weight increase is observed even if there is plenty milk available. We chose the stimulation of
the vibrissae-barrel axis because it is suitable for studying structure, function, development and plasticity
within the somatosensory cortex, due to the functional and morphological correlation between the vibrissae
and the barrels [17].
2 Materials and methods
2.1 Animal preparation
All animal procedures were performed according to the federal guidelines of the Animal Care and approved
by the local authority. Twelve male Wistar rats (provided by the closed colony of breeding of the Division
of Biological and Health Sciences of the Autonomous Metropolitan University, Mexico) were used. The
animals were kept in an environment at a temperature of 22 - 25 0C, with 45% relative humidity. Animals
were kept under 12 hours controlled light-darkness cycles. Animals were divided into two groups according
to nutritional manipulation: control and experimental groups. Both groups consisted of 6 rats each (aged
18 to 21 days). To provoke malnutrition, the food competition method was applied [17]. Control rats
weighted 41.94±4.9g and malnourished rats weighted 29.09±3.3g. The weights of the experimental group
correspond to a second degree malnourished level, according to [16]. This setup can be seen in Figure 1.
Figure 1. Animal Setup employed for all BOLD experiments. Stimulating electrodes were placed on the whiskers (anode) and in the
masticatory muscles (cathode). The anesthesia was delivered through the tube attached to the stereotactic frame-like. To minimise
motion artefacts, the rat's head and ears were fixed to an Agilent stereotactic frame. This setup was also used to administer
anesthesia at 1% of isoflurane with 2 l/min of O2 during all the fMRI experiment. EKG, breathing monitoring, and rat's temperature
control (370C ± 10C) were done with a small animal monitoring and gating system (Model 1025, SA Instruments, NY, USA).
2
2.2 Trigeminal nerve stimulation
The trigeminal nerve (the fifth cranial nerve, also called the fifth nerve, or simply CNV or CN5) contains
both sensory and motor fibers. It is responsible for sensation in the face and certain motor functions such
as biting, chewing, and swallowing. Sensory information from the face and body is processed by parallel
pathways in the central nervous system. In animal models, more specific in rodents, it has been used as
an attractive model the whisker sensory system for studying this nerve. This structure helps us to study
and understand structure, function, development, and plasticity within the somatosensory cortex [17].
fMRI Acquisition and data processing
2.3
All experiments were performed on a 7T/21cm Agilent system (Agilent Technologies, Inc, Palo Alto,
CA) equipped with DirectDrive technology and a transceiver 16-rung birdcage coil (12 cm long and a
6 cm diameter). Rat's brain images were acquired using a standard gradient echo sequence and the
following parameters: TR/TE=107.82/3.8 ms, Flip angle=200, FOV=30x30mm, matrix size= 128x128,
thickness=0.3mm, and NEX=1, thus a spatial resolution of 0.23 x 0.23 x 0.3 mm3 was achieved. Gradient-
echo BOLD fMRI has reasonably high spatiotemporal resolution can be routinely performed on this type
of animal models at high field.
Images for BOLD fMRI were taken for both control and experimental groups. The left trigeminal nerve
was stimulated using percutaneously inserted stainless steel electrodes, whose cathode was positioned in
the whiskers and the anode was inserted in the masticatory muscles. The trigeminal nerve was stimulated
using a stimulator with 10 ms constant current pulses (500 mV and 2 mA) applied every second (Grass
S-48 Stimulator, Grass Technologies, RI, USA) and at 1Hz. 60s OFF was alternated with 60s ON periods
[17].
To determine the regions of interest in the somatosensory cortex where neuronal activity is expected
to happen, all brain images were digitally processed using the toolbox SPMMouse [18]. Data were overlaid
onto anatomical images acquired in the same image plane. For comparison between control and experi-
mental somatosensory cortex we used a p < 0.005, and for the analysis in the experimental group a p <
0.05 was used, for estimating the SPM results.
3 Results
fMRI maps
3.1
To study the possible areas of activation, representative activation maps of stimulus-correlated intensity
changes in the rat brain of both groups were obtained. Fig. 2 shows five different brain regions from
the olfactory bulb to the cerebellum on both groups, so we can appreciate different activation between
groups. Schematics diagrams taken from Paxino's rat brain atlas were added to facilitate localization of
the coronal cuts in the rat's brain [19]. Under normal conditions, the expected responses are similar to
the response of the control group, as shown in Figure 2.
3
Figure 2. Coronal brain maps of BOLD intensity of one rat of each group showing activation from the olfactory bulb to cerebellum.
From left to right, we see areas corresponding from olfactory bulb to the cerebellum. Notice that in the experimental group exist random
pattern activation in comparison with the control group (expected response).
This is an important activation only in somatosensory cortex per se. But in the case of a malnourished
brain, it can be appreciated responses from other brain structures, such as hippocampus, hypothalamus,
and cerebellum besides somatosensoy cortex. In general, all BOLD fMRI maps of malnourished rats show
more activation than those in the control group in the entire brain. However, the brain may be regulated by
a fine mechanism of excitatory and inhibitory entries at the interneuron level [20]. Our results indicate an
abnormal increment in the electrical activity among different regions of the cerebral tissue. This is probably
the result of a decrease on the neural inhibitory discharges of the inhibitory postsynaptic potentials (IPSP)
and/or an increase in the excitatory postsynaptic potentials (EPSP) [20]. Results reported by Segura and
collaborators [21] demonstrated the existence of a demyelinating process in the peripheral nervous system
in malnourished rats and, that a similar effect is happening at the encephalic level. Malnutrition probably
is affecting the inhibitory interneurons, so that their demyelination processes cause a decrease in the IPSP,
and then increasing the neuronal activity in a random way as depicted in Figure 2. From our results, it
is not yet possible to determine if abnormal activation from different brain regions is a consequence of
brain's tissue damage or these activations are the result from adaptive process of neural plasticity. The
changes are so dramatic that suggest the existence of preexistent synaptic pathways that are not normally
expressed, while in a brain of a malnourished subject these pathways are unmasked.
3.2 BOLD response
Control and experimental data at the somatosensory cortex were then fitted to a non-linear regression over
time and shown in Figure 3. Nonlinearity of the BOLD response is observed for both groups. There is great
similarity in the pattern of BOLD response for both groups. These results showed a great concordance
with results already reported by Silva and Korestky [22]. The BOLD responses experiment a similar time
delay for both groups and reach their maximum value at 5 seconds during the stimulation. Experimental
group results suggest larger oxygen consumption when compared with the control group. Fig. 3c) shows
the averaged BOLD signal intensity time course for control and experimental groups were computed with
6 rats: 30 seconds per rat. A sign change can be appreciated between the regions I and II of Fig. 3.c),
roughly after 12 s for both cases.
4
Figure 3. Adjusted BOLD signal from the somatosensory cortex obtained from a) control group, b) experimental group, and c) the
mean response from each group of 6 rats each.
This has been already reported as a negative BOLD response and it is caused by inhibition mecha-
nisms [23]. A t student test was run to investigate statistical independence of the measurements of the
response amplitude. At the 0.05 level, the difference of the population means (meancontrol=1879.39 &
meanexp=1254.34) is statistically different. In order to quantify the BOLD response of the malnourished
rats, the method reported in [24] was used. Then, the response amplitudes were calculated from the peak
of BOLD response intensity. Additionally, the response integral was defined as the product of the ampli-
tude and the full width at half maximum (FWHM) for both groups of rats the response integral values
are 5638.18 and 3763.01 for the control and malnourished rats, respectively. Also a simple trapezoid-
integration under the positive curve was made; the results were 20460.15 and 13708.85 for control and
malnourished rats, respectively. This parameter indicates a clear increment in the neuronal activity for
the malnourished rats.
We also analyzed responses in the experimental group, in Figure 4, we can see BOLD responses for dif-
ferent structures that were activated, and we can see responses from cerebellum, hypothalamus, hippocam-
pus, and somatosensory cortex. We made the same measurements of area under the positive curve of the
mean responses and the results obtained by simple trapezoid-rule integration were: cerebellum≈21625.49,
hypothalamus≈15606.45, hippocampus≈15272.86, and somatosensory cortex≈13708.85.
5
Figure 4. Adjusted BOLD signals from different brain areas from experimental (malnourished) group. Also there is represented the
mean BOLD signal from each brain areas. All figures show same maximum intensity scale so we can notice the differences between
structures. The lower figure show areas from which this signals were taken.
A similar behavior was obtained with the FWHM method, described for the comparison between
in the malnourished analysis the results were cerebellum≈5936.09,
control and experimental groups;
hypothalamus≈4283.9, hippocampus≈4192.33, and somatosensory cortex≈3763.01. A greater area at the
cerebellum region can be appreciated, hypothalamus and hippocampus have a very similar behavior, and
somatosensory cortex area has a lower area in comparison to the cerebellum one, and larger area means
a greater activity in the region. Our results agree with the results obtained with invasive techniques and
demonstrated that cerebellum is one of the structures more affected by malnutrition as demonstrated by
Hillman [5]. Other structures related to learning and memory processes are also affected because of the
malnutrition process, such as hippocampus and hypothalamus [2-14]. Finally a three-dimensional ren-
dering process was computed to observe anatomical differences between both groups. In Fig. 5 we can
clearly observe significant differences in brain structure, mainly at the cerebellum development. Also we
can notice a big difference at the development of both brain hemispheres.
Figure 5. 3D reconstruction from the anatomical images obtained from the a) top row, control and b) bottom row, experimental group.
We can see an important difference between both brains, at the cerebellum (solid line arrow), in which a notorious low development is
present. Also an irregular development from both cerebral hemispheres (dashed line arrow) can be noticed in the experimental brain.
6
4 Discussion
This fMRI study is the first performed in rats and illustrates a similar pattern of BOLD responses between
malnourished and control rats. Negative BOLD responses can be also appreciated for the two groups of
rats caused by inhibition mechanisms. Vascular nonlinearities are the major contribution to the observed
nonlinearities and they are neuronal in origin. Our results make a pathway to continue study this health
problem and its consequences in brain function, because we cannot still say if the resulting changes in the
brain correspond totally to a brain damage or if there is a process of adaptive plasticity.
Results obtained were not the expected ones, although it has been described brain damages by mal-
nutrition in structures like cerebellum, hypothalamus, and hippocampus; according to the experimental
protocol used. We only expected limited alterations at the somatosensoy cortex at first, and no significant
changes in the function of other structures. The brain changes observed with this experiment may imply
that the mechanisms of neurotransmitters like GABA (inhibitory) or Glutamate (excitatory) are affected,
but others neurotransmitters like ACh, Dopamine, or Noradrenaline may be not working as they should.
Also, our results may imply that exist problems with the hypothalamic pathway with other structures.
These pathways are shown in the brain scheme of Fig. 5. These problems may be involved with the
neurotransmitters cycle.
5 Conclusion
We used BOLD fMRI and trigeminal nerve stimulation to study the BOLD changes during sensory stim-
ulation in malnourished rats. This study is the first performed in malnourished rats and illustrates BOLD
activation in various brain areas, specially cerebellum is the structure with more activation. Further in-
vestigation should be carried out to determine whether plasticity or demyelination, or even combinations
of both are responsible for the activation of the brain areas reported in this work. These results may pave
the way into the treatment of this health problem and other possible rehabilitation procedure.
Acknowledgments
We would like to acknowledge financial funding from CONACYT-Mexico under grant no. 166404 and a
Ph. D. scholarship.
References
1. M. Medina, et. al. Consequences of Malnutrition. World Federation of Neurology. Seminars in
Clinical Neurology. Vol 6. 2007. pp 1-36.
2. Zeman FJ, et. al. Teratog Carcinog Mutagen, 6: 339-347. 1986.
3. Fukuda MT, et. al. Behav Brain Res, 133: 271-277. 2002.
4. Lister JP, et. al. Hippocampus, 15: 393-403. 2005.
5. Hillman DE, Chen S. Neuroscience, 6: 1249-1262. 1981
6. Benitez-Bribiesca L, et. al. Pediatrics, 104: e21. 1999
7. Reddy PV, et. al. Brain Res,161: 227-235 . 1979.
8. Montanha-Rojas EA, et. al. Nutr Neurosci, 8: 49-56. 2005.
9. Mazer C, et. al. Brain Res, 760: 68-73. 1997.
7
10. Chen JC, et. al. Int J Dev Neurosci, 15: 257-263. 1997.
11. Chang YM, et. al. Nutr Neurosci, 6: 263-267. 2003.
12. Nakagawasai O. Yakugaku Zasshi, 125: 549-554. 2005.
13. Cermak JM, et. al. Dev Neurosci, 21: 94-104. 1999.
14. Zimmer L, et. al. Neurosci Lett, 284: 25-28. 2000.
15. Chalon S, et. al. Lipids, 36: 937-944. 2001.
16. Ortiz R. et al. Med. Sci. Res. 24:843. 1996
17. Just N. et al. Mag. Res. Imaging. 28:1143-1151. 2010.
18. Sawiak SJ. et. al. 17th ISMRM. 2009:1086
19. Paxinos G, Watson Ch. The Rat Brain in Stereotaxic Coordinates, 4th Ed. Academic Press, 1998
20. Kandel, ER. Principles of Neural Science, 4th Ed. McGraw-Hill, New York, USA. 2000.
21. Segura B. et. al. Neurosc. Lett. 2004;354:181
22. Silva AC, Koretsky AP. PNAS. 2002;99:15182
23. Vandervliet E. et al. ESMRMB. 2006:624.
24. Yesilyrut B. et al. MRI. 2008;26:85.
8
|
1705.00063 | 1 | 1705 | 2017-04-28T20:15:45 | Derivation of a novel efficient supervised learning algorithm from cortical-subcortical loops | [
"q-bio.NC"
] | Although brain circuits presumably carry out useful perceptual algorithms, few instances of derived biological methods have been found to compete favorably against algorithms that have been engineered for specific applications. We forward a novel analysis of function of cortico-striatal loops, which constitute more than 80% of the human brain, thus likely underlying a broad range of cognitive functions. We describe a family of operations performed by the derived method, including a nonstandard method for supervised classification, which may underlie some forms of cortically-dependent associative learning. The novel supervised classifier is compared against widely-used algorithms for classification, including support vector machines (SVM) and k-nearest neighbor methods, achieving corresponding classification rates --- at a fraction of the time and space costs. This represents an instance of a biologically-derived algorithm comparing favorably against widely used machine learning methods on well-studied tasks. | q-bio.NC | q-bio | COMPUTATIONAL NEUROSCIENCE
ORIGINAL RESEARCH ARTICLE
published: 10 January 2012
doi: 10.3389/fncom.2011.00050
Derivation of a novel efficient supervised learning
algorithm from cortical-subcortical loops
Ashok Chandrashekar 1* and Richard Granger 2
1 Department of Computer Science, Dartmouth College, Hanover, NH, USA
2 Psychological and Brain Sciences, Thayer School of Engineering and Computer Science, Dartmouth College, Hanover, NH, USA
Edited by:
Hava T. Siegelmann, Rutgers
University, USA
Reviewed by:
Thomas Boraud, Universite de
Bordeaux, France
Taro Toyoizumi, RIKEN/BSI, Japan
*Correspondence:
Ashok Chandrashekar, Department of
Computer Science, Dartmouth
College, Hinman Box 6211, Hanover,
NH 03775, USA.
e-mail: ashok.chandrashekar@
dartmouth.edu
Although brain circuits presumably carry out powerful perceptual algorithms, few instances
of derived biological methods have been found to compete favorably against algorithms
that have been engineered for specific applications.We forward a novel analysis of a subset
of functions of cortical–subcortical loops, which constitute more than 80% of the human
brain, thus likely underlying a broad range of cognitive functions. We describe a family of
operations performed by the derived method, including a non-standard method for super-
vised classification, which may underlie some forms of cortically dependent associative
learning. The novel supervised classifier is compared against widely used algorithms for
classification, including support vector machines (SVM) and k-nearest neighbor methods,
achieving corresponding classification rates – at a fraction of the time and space costs.
This represents an instance of a biologically derived algorithm comparing favorably against
widely used machine learning methods on well-studied tasks.
Keywords: biological classifier, hierarchical, hybrid model, reinforcement, unsupervised
INTRODUCTION
1.
Distinct brain circuit designs exhibit different functions in human
(and other animal) brains. Particularly notable are studies of the
basal ganglia (striatal complex), which have arrived at closely-
related hypotheses, from independent laboratories, that the system
carries out a form of reinforcement learning (Sutton and Barto,
1990; Schultz et al., 1997; Schultz, 2002; Daw, 2003; O'Doherty
et al., 2003; Daw and Doya, 2006); despite ongoing differences
in the particulars of these approaches, their overall findings are
surprisingly concordant, corresponding to a still-rare instance of
convergent hypotheses of the computations produced by a partic-
ular brain circuit. Models of thalamocortical circuitry have not yet
converged to functional hypotheses that are as widely agreed-on,
but several different approaches nonetheless hypothesize the abil-
ity of thalamocortical circuits to perform unsupervised learning,
discovering structure in data (Lee and Mumford, 2003; Rodriguez
et al., 2004; Granger, 2006; George and Hawkins, 2009). Yet thal-
amocortical and striatal systems do not typically act in isolation;
they are tightly connected in cortico-striatal loops such that virtu-
ally each cortical area interacts with corresponding striatal regions
(Kemp and Powell, 1971; Alexander and DeLong, 1985; McGeorge
and Faull, 1988). The resulting cortico-striatal loops constitute
more than 80% of human brain circuitry (Stephan et al., 1970,
1981; Stephan, 1972), suggesting that their operation provides the
underpinnings of a very broad range of cognitive functions.
We forward a new hypothesis of the interaction between cor-
tical and striatal circuits, carrying out a hybrid of unsupervised
hierarchical
learning and reinforcement, together achieving a
cortico-striatal loop algorithm that performs a number of dis-
tinct operations of computational utility, including supervised
and unsupervised classification, search, object and feature local-
ization, and hierarchical memory organization. For purposes of
the present paper we focus predominantly on the particular task
of supervised learning.
Traditional supervised learning methods typically identify class
boundaries by focusing primarily on the class labels, whereas unsu-
pervised methods discover similarity structure occurring within a
dataset; two distinct tasks with separate goals, typically carried out
by distinct algorithmic approaches.
Widely used supervised classifiers such as support vector
machines (Vapnik, 1995), supervised neural networks (Bishop,
1996), and decision trees (Breiman et al., 1984; Buntine, 1992), are
so-called discriminative models, which learn separators between
categories of sample data without learning the data itself, and with-
out illuminating the similarity structure within the data set being
classified.
The cortico-striatal loop (CSL) algorithm presented here is
"generative,"i.e., it is in the category of algorithms that models data
occurring within each presented class, rather than seeking solely
to identify differences between the classes (as would a "discrimi-
native" method). Generative models are often taken as performing
excessive work in cases where the only point is to distinguish
among labeled classes (Ng and Jordan, 2002). The CSL method
may thus be taken as carrying out more tasks than classification,
which we indeed will see it does. Nonetheless, we observe the
behavior of the algorithm in the task of classification, and com-
pare it against discriminative classifiers such as support vectors,
and find that even in this restricted (though very widely used)
domain of application, the CSL method achieves comparable clas-
sification as discriminative models, and uses far less computational
cost to do so, despite carrying out the additional work entailed in
generative learning.
The approach combines the two distinct tasks of unsupervised
classification and reinforcement, producing a novel method for yet
www.frontiersin.org
January 2012 Volume 5 Article 50 1
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
another task: that of supervised learning. The new method iden-
tifies supervised class boundaries, as a byproduct of uncovering
structure in the input space that is independent of the super-
vised labels. It performs solely unsupervised splits of the data into
similarity-based clusters. The constituents of each subcluster are
checked to see whether or not they all belong to the same intended
supervised category. If not, the algorithm makes another unsu-
pervised split of the cluster into subclusters, iteratively deepening
the class tree. The process repeats until all clusters contain only (or
largely) members of a single supervised class. The result is the con-
struction of a hierarchy of mostly mixed classes, with the leaves of
the tree being "pure" categories, i.e., those whose members contain
only (or mostly) a single shared supervised class label.
Some key characteristics of the method are worth noting.
contain only members that are similar to each other.
• Only unsupervised splits are performed, so clusters always
• In the case of similar-looking data that belong to distinct super-
vised categories (e.g., similar-looking terrains, one leading to
danger and one to safety), these data will constitute a diffi-
cult discrimination; i.e., they will reside near the boundary that
partitions the space into supervised classes.
• In cases of similar data with different class labels, i.e., difficult
discriminations, the method will likely perform a succession
of unsupervised splits before happening on one that splits the
dangerous terrains into a separate category from the safe ones.
In other words, the method will expend more effort in cases of
difficult discriminations. (This characteristic is reminiscent of the
mechanism of support vectors, which identify those vectors near
the intended partition boundary, attempting to place the boundary
so as to maximize the distance from those vectors to the bound-
ary.) Moreover, in contrast to supervised methods that provide
expensive, detailed error feedback at each training step (instruct-
ing the method as to which supervised category the input should
have been placed in), the present method uses feedback that is
comparatively far more inexpensive, consisting of a single bit at
each training step, telling the method whether or not an unsuper-
vised cluster is yet "pure"; if so, the method stops for that node; if
not, the method performs further unsupervised splits.
This deceptively simple mechanism not only produces a super-
vised classifier, but also uncovers the similarity structure embed-
ded in the dataset, which competing supervised methods do not.
Despite the fact that competing algorithms (such as SVM and Knn)
were designed expressly to obtain maximum accuracy at super-
vised classification, we present findings indicating that even on
this task, the CSL algorithm achieves comparable accuracy, while
requiring significantly less computational resource cost.
In sum, the CSL algorithm, derived from the interaction of
cortico-striatal loops, performs an unorthodox method that rivals
the best standard methods in classification efficacy, yet does so in
a fraction of the time and space required by competing methods.
2. CORTICO-STRIATAL LOOPS
The basal ganglia (striatal complex), present in reptiles as well as
in mammals, is thought to carry out some form of reinforcement
learning, a hypothesis shared across a number of laboratories (Sut-
ton and Barto, 1990; Schultz et al., 1997; Schultz, 2002; Daw, 2003;
O'Doherty et al., 2003; Daw and Doya, 2006). The actual neural
mechanisms proposed involve action selection through a maxi-
mization of the corresponding reward estimate for the action on
the task (see Brown et al., 1999; Gurney et al., 2001; Daw and Doya,
2006; Leblois et al., 2006; Houk et al., 2007 for a range of views on
action selection). This reward estimation occurs in most models
of the striatum through the regulation of the output of the neuro-
transmitter dopamine. Therefore, in computational terms we can
characterize the functionality of the striatum as an abstract search
through the space of possible actions, guided by dopaminergic
feedback.
The neocortex and thalamocortical loops are thought to hier-
archically organize complex fact and event information, a hypoth-
esis shared by multiple researchers (Lee and Mumford, 2003;
Rodriguez et al., 2004; Granger, 2006; George and Hawkins, 2009).
For instance, in Rodriguez et al. (2004) the anatomically recog-
nized "core" and "matrix" subcircuits are hypothesized to carry
out forms of unsupervised hierarchical categorization of static
and time-varying signals; and in Lee and Mumford (2003), George
and Hawkins (2009), Riesenhuber and Poggio (1999), and Ullman
(2006) and many others, hypotheses are forwarded of how corti-
cal circuits may construct computational hierarchies; these studies
from different labs propose related hypotheses of thalamocortical
circuits performing hierarchical categorization.
It is widely accepted that these two primary telencephalic struc-
tures, cortex and striatum, do not act in isolation in the brain;
they work in tight coordination with each other (Kemp and Pow-
ell, 1971; Alexander and DeLong, 1985; McGeorge and Faull,
1988). The ubiquity of this repeated architecture (Stephan et al.,
1970, 1981; Stephan, 1972) suggests that cortico-striatal circuitry
underlies a very broad range of cognitive functions. In particular,
it is of interest to determine how semantic cortical informa-
tion could provide top-down constraints on otherwise too-broad
search during (striatal) reinforcement learning (Granger, 2011).
In the present paper we study this interaction in terms of subsets
of the leading extant computational hypotheses of the two com-
ponents: thalamocortical circuits for unsupervised learning and
the basal ganglia/striatal complex for reinforcement of matches
and mismatches. If these bottom-up analyses of cortical and stri-
atal function are taken seriously, it is of interest to study what
mechanisms may emerge from the interaction of the two mech-
anisms when engaged in (anatomically prevalent) cortico-striatal
loops.. We adopt straightforward and tractable simplifications of
these models, to study the operations that arise when the two
are interacting. Figure 1 illustrates a hypothesis of the func-
tional interaction between unsupervised hierarchical clustering
(uhc; cortex) and match-mismatch reinforcement (mm; striatal
complex), constituting the integrated mechanism proposed here.
The interactions in the simplified algorithm are modeled in
part on mechanisms outlined in Granger (2006): a simplified
model of thalamocortical circuits produces unsupervised clus-
ters of the input data; then, in the CSL model, the result of the
clustering, along with the corresponding supervised labels, are
examined by a simplified model of the striatal complex. The full
computational models of the thalamocortical hierarchical clus-
tering and sequencing circuit and striatal reinforcement-learning
circuit yield interactions that are under ongoing study, and will, it
is hoped, lead to further derivation of additional algorithms. For
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 2
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
are readily predictable from their appearance. This corresponds
to an "easy" discrimination task. When this is not the case, i.e., in
instances where similar-looking data belong to different labeled
categories (e.g., similar mushrooms, some edible and some poiso-
nous), the mechanism will be triggered to successively subdivide
clusters into subclusters, as though searching for the characteristics
that effectively separate the members of different labels.
In other words, less work is done for "easy" discriminations;
and only when there are difficult discriminations will the mech-
anism perform additional steps. The tree becomes intrinsically
unbalanced as a function of the lumpiness of the data: branches
of the tree are only deepened in regions of the space where the
discriminations are difficult, i.e., where members of two or more
distinct supervised categories are close to each other in the input
space. This property is reminiscent of support vectors, which iden-
tify boundaries in the region where two categories are closest (and
thus where the most difficult discriminations occur).
A final salient feature of the mechanism is its cost. In con-
trast to supervised methods, which provide detailed, expensive,
error feedback at each training step (telling the system not only
when a misclassification has been made but also exactly which class
should have occurred), the present method uses feedback that by
comparison is extremely inexpensive, consisting of a single bit, cor-
responding to either "pure" or "impure" clusters. For pure clusters,
the method halts; for impure clusters, the mechanism proceeds to
deepen the hierarchical tree.
As mentioned, the method is generative, and arrives at rich
models of the learned input data. It also produces multiclass
partitioning as a natural consequence of its operation, unlike
discriminative supervised methods which are inherently binary,
requiring extra mechanisms to operate on multiple classes.
Overall, this deceptively simple mechanism not only produces
a supervised classifier, but also uncovers the similarity structure
embedded in the dataset, which competing supervised methods do
not. The terminal leaves of the tree provide final class information,
whereas the internal nodes provide further information: they are
mixed categories corresponding to meta labels (e.g., superordinate
categories; these also can provide information about which classes
are likely to become confused with one another during testing.
In the next section we provide an algorithm that retains func-
tional equivalence with the biological model for supervised learn-
ing described above while abstracting out the implementation
details of the thalamocortical and striatal circuitry. Simplifying
the implementation enables investigation of the algorithmic prop-
erties of the model independent of its implementation details
(Marr, 1980). It also, importantly, allows us to test our model on
real-world data and compare directly against standard machine
learning methods. Using actual thalamocortical circuitry to per-
form the unsupervised data clustering and the mechanism for
the basal ganglia to provide reinforcement feedback, would be
an interesting task for the distinct goal of investigating potential
implementation-level predictions; this holds substantial potential
for future research.
We emphasize that our focus is to use existing hypotheses of
telencephalic component function already posited in the literature;
these mechanisms lead us to specifically propose a novel method
by which supervised learning is achieved by the unlikely route of
FIGURE 1 Simplified hypothesis of the computations of
cortico-striatal interaction. (Left) Inputs (e.g., images) arrive together with
associated labels (A, B, etc.). (Top right) Unsupervised hierarchical
clustering (uhc) categorizes the inputs by their similarity, without initial
regard to labels. (Far right) The match–mismatch mechanism (mm) matches
the class membership labels within each of these unsupervised categories.
If the members of a category all have the same (matching) labels, a "+" is
returned; if there are mismatches, a "−" is returned. (Bottom right) In the
latter case, the clustering mechanism iteratively deepens the hierarchical
tree by creating subclusters (e.g., central node in the diagram); these in
turn are checked for label consistency, as before. The process iterates until
the leaf nodes of the unsupervised tree contain only category members of
a single label. (See text).
the present paper, we use just small subsets of the hypothesized
functions of these structures: solely the hypothesized hierarchical
clustering function of the thalamocortical circuit, and a very-
reduced subset of the reinforcement-learning capabilities of the
striatal complex, such that it does nothing more than compare
(match / mismatch) the contents of a proposed category, and
return a single bit corresponding to whether the contents all have
been labeled as "matching" each other (1) or not (0). This very-
reduced RL mechanism can be thought of simply as rewarding
or punishing a category based on its constituents. In particu-
lar the proposed simplified striatal mechanism returns a single
bit (correct/incorrect) denoting whether the members of a given
unsupervised cluster all correspond to the same supervised"label."
If not, the system returns a "no" ("–") to the unsupervised cluster-
ing mechanism, which in turn iterates over the cluster producing
another, still unsupervised, set of subclusters of the "impure"
cluster. The process continues until each unsupervised subcluster
contains members only (or mostly, in a variant of the algorithm)
of a single category label.
In sum, the mechanism uses only unsupervised categorization
operations, together with category membership tests. These two
mechanisms result in the eventual iterative arrival at categories
whose members can be considered in terms of supervised classes.
Since only unsupervised splits are performed, categories (clus-
ters) always contain only members that are similar to each other.
The tree may generate multiple terminal leaves corresponding to
a given class label; in such cases, the distinct leaves correspond
to dissimilar class subcategories, eventually partitioned into dis-
tinct leaf nodes. The mechanism can halt rapidly if all supervised
classes correspond to similarity-based clusters; i.e., if class labels
www.frontiersin.org
January 2012 Volume 5 Article 50 3
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
combining unsupervised learning with reinforcement. This kind
of computational-level abstraction and analysis of biological enti-
ties continues in the tradition of many prior works, including
Suri and Schultz (2001), Schultz (2002), Daw and Doya (2006),
Lee and Mumford (2003), Rodriguez et al. (2004), George and
Hawkins (2009), Marr (1980), Riesenhuber and Poggio (1999),
Ullman (2006), and many others.
3. SIMPLIFIED ALGORITHM
In our simplified algorithm, we refer to a method which we
term PARTITION, corresponding to any of a family of cluster-
ing methods, intended to capture the clustering functionality of
thalamocortical loops as described in the previous sections; and
we refer to a method we term SUBDIVIDE, corresponding to
any of a family of simple reinforcement methods, intended to
capture the reinforcement-learning functionality of the basal gan-
glia/striatal complex as described in the previous sections. These
operate together in an iterative loop corresponding to cortico-
striatial (cluster–reinforcement) interaction: SUBDIVIDE checks
for the "terminating" conditions of the iterative loop by examin-
ing the labels of the constituents of a given cluster and returning
a true or false response. The resulting training method builds a
tree of categories which, as will be seen, has the effect of per-
forming supervised learning of the classes. The leaves of the tree
contain class labels; the intermediate nodes may contain members
of classes with different labels. During testing, the tree is traversed
to obtain the label prediction for the new samples. Each data sam-
ple (belonging to one of K labeled classes) is represented as a vector
x ∈ Rm. During training, each such vector xi has a corresponding
label yi ∈ 1, . . . K. (The subsequent "Experiments" section below
describes the methods used to transform raw data such as nat-
ural images into vector representations in a domain-dependent
fashion.)
TRAINING
3.1.
The input to the training procedure is the training dataset consist-
ing of ⟨xi, yi⟩ pairs where xi is an input vector and yi is its intended
class label, as in all supervised learning methods. The output is a
tree that is built by performing a succession of unsupervised splits
of the data. The data corresponding to any given node in the tree
is a subset of the original training dataset with the full dataset cor-
responding to the root of the tree. The action performed with the
data at a node in the tree is an unsupervised split, thereby generat-
ing similarity-based clusters (subclusters) of the data within that
tree node. The unsupervised split results in expansion (deepening)
of the tree at that node, with the child nodes corresponding to the
newly created unsupervised data clusters. The cluster represen-
tations corresponding to the children are recorded in the current
node. These representations are used to determine the local branch
that will be taken from this node during testing, in order to obtain
a class prediction on a new sample. For each of the new children
nodes, the labels of the samples within the cluster are examined,
and if they are deemed to be sufficiently pure, i.e., a sufficient per-
centage of the data belong to the same class, then the child node
becomes a (terminal) leaf in the tree. If not, the node is added
to a queue which will be subjected to further processing, growing
the tree. This queue is initialized with the root of the tree. The
procedure (sketched in Algorithm 1 below) proceeds until the
queue becomes empty.
To summarize the mechanism, the algorithm attempts to find
clusters based on appearance similarity, and when these clusters
don't match with the intended (supervised) categories, reinforce-
ment simply gives the algorithm the binary command to either
split or not split the errant cluster. The behavior of the algorithm
on sample data is illustrated in Figure 2. The input space of images
is partitioned by successively splitting the corresponding training
samples into subclusters at each step.
3.1.1. Picking the right branch factor
Since the main free parameter in the algorithm is the number
of unsupervised clusters to be spawned from any given node in
the hierarchy, the impact of that parameter on the performance
of the algorithm should be studied. This quantity corresponds
to the branching factor for the class tree. We initially propose a
single parameter as an upper bound for the branch factor: Kmax,
which fixes the largest number of branches that can be spawned
from any node in the tree. Through experimentation (discussed
in the Results section) we have determined that (i) very small val-
ues for this parameter result in slightly lower prediction accuracy;
Algorithm 1 A sketch of the CSL learning algorithm. The method
constructs a meta class tree, which records unsupervised structure within
the data, as well as providing a means to perform class prediction on novel
samples. The function termed PARTITION denotes an unsupervised
clustering algorithm, which can in principle be any of a family of clustering
routines; selections for this algorithm are described later in the text. The
subroutine SUBDIVIDE determines if the data at a tree node qn all belong
to the same class or not. If the data come from multiple classes,
SUBDIVIDE returns true and otherwise, false. See text for further
description of the derivation of the algorithm.
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 4
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
FIGURE 2 A simplified illustration of the iterative learning process with
labeled data. Images are successively split into two partitions in an
unsupervised fashion (i.e., by similarity). The partitioning of data proceeds
iteratively until the clusters formed are pure with respect to their labels. For
each split, on either side of the dividing hyperplane, the means are shown as
an overlay of the images that fall on the corresponding side of the hyperplane.
(ii) for sufficiently large values, the parameter setting has no sig-
nificant impact on the performance efficacy of the classifier; and
(iii) larger values of the parameter modestly increase the memory
requirements of the clustering algorithm and thus the runtime of
the learning stage (see the Results section below for further detail).
(It is worth noting that selection of the best branch factor value
www.frontiersin.org
January 2012 Volume 5 Article 50 5
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
may be obtained by examination of the distribution of the data
to be partitioned in the input space, enabling automatic selection
of the ideal number of unsupervised clusters without reference
to the number of distinct labeled classes that occur in the space.
Future work may entail the study of existing methods for this
approach, Baron and Cover, 1991; Teh et al., 2004, as potential
adjunct improvements to the CSL method.)
TREE PRUNING
3.2.
Categorization algorithms are often subject to overfitting the data.
Aspects of the CSL algorithm can be formally compared to those
of decision trees, which are subject to overfitting.
Unlike decision trees, the classes represented at the leaves of the
CSL tree need not be regarded as conjunctions of attribute values
on the path from the root, and can be treated as fully represented
classes by themselves. (We refer to this as the "leaf independence"
property of the tree; this property will be used when we describe
testing of the algorithm in the next section.) Also, since the splits
are unsupervised and based on multidimensional similarity (also
unlike decision trees), they exhibit robustness w.r.t. variances in
small subsets of features within a class.
Both of these characteristics (leaf independence and unsuper-
vised splitting) theoretically lead to predictions of less overfitting
of the method.
In addition to these formal observations, we studied overfit-
ting in the CSL method empirically. Analogously to decision trees,
we could choose either to stop growing the tree before all leaves
were perfectly pure (and potentially overfit), or to build a full
tree and then somewhat prune it back. Both methods improve
the overfitting problem observed in decision trees. Experiments
with both methods in the CSL algorithm found that neither one
had a significant effect on prediction accuracy. Thus, surprisingly,
both theoretical and empirical studies find that the CSL class trees
generalize well without overfitting; the method is unexpectedly
resistant to overfitting.
TESTING
3.3.
During testing, the algorithm is presented with previously unseen
data samples whose class we wish to predict. The training
phase created an appearance-based class hierarchy. Since the
tree, including the "pure class" leaves, is generative in nature,
there are two alternative procedures for class prediction. One
is that of descending the tree, as is done in decision trees.
However, in addition, the "leaf independence" property of the
CSL tree, as described in the previous section (which does
not hold for decision trees), enables another testing method,
which we refer to as KNN-on-leaves, in which we only attend
to the leaf nodes of the tree, as described in the second sub-
section below. (This property does not hold for decision trees,
and thus this additional testing method cannot be applied to
decision trees). The two test methods have somewhat different
memory and computation costs and slightly different prediction
accuracies.
Tree descent
3.3.1.
This approach starts at the root of the class tree, and descends. At
every node, the test datum is compared to the cluster centroids
stored at the node to determine the branch to take. The branch
taken corresponds to the closest centroid to the test datum; i.e.,
a decision is made locally at the node. This provides us a unique
path from the root of the class hierarchy to a single leaf; the stored
category label at that leaf is used to predict the label of the input.
Due to tree pruning (described above), the leaves may not be com-
pletely pure. As a result, instead of relying on any given class being
present in the leaves, the posterior probabilities for all the cate-
gories represented at the leaf are used to predict the class label for
the sample.
3.3.2. KNN-on-leaves
In this approach, we make a note of all the leaves in the tree,
along with the cluster representation in the parent of the leaf node
corresponding to the branch which leads to the leaf. We then do K-
nearest neighbor matching of the test sample with all these cluster
centroids that correspond to the leaves. The final label predicted
corresponds to the label of the leaf with the closest centroid. This
approach implies that only the leaves of the tree need to be stored,
resulting in a significant reduction in the memory required to
store the learned model. However, a penalty is paid in recognition
time, which in this case is proportional to the number of leaves in
the tree.
The memory required to store the model in the tree descent
approach is higher than that for the KNN-on-leaves approach.
However, tree descent offers a substantial speedup in recognition,
as comparisons need to be performed only along a single path
through the tree from the root to the final leaf. The algorithm is
sketched below in Algorithm 2.
We expect that the KNN-on-leaves variant will yield better pre-
diction accuracy as the decision is made at the end of the tree and
Algorithm 2 A sketch of the tree descent algorithm for classifying a new
data sample. The method starts at the root node and descends, testing the
sample datum against each node encountered to determine the branch to
select for further descent. The result is a unique path from the root to a
single leaf; the stored category at that leaf is the prediction of the label of
the input. In the event of impure leaves, the posterior probabilities for all
categories in the leaf are used to predict the class label of the sample. See
text for further description.
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 6
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
hence the partitioning of the input space is expected to exhibit
better generalization. In the case of tree descent, since decisions
are made locally within the tree, if the dataset has high variance,
then it is possible that a wrong branch will be taken early on in the
tree, leading to inaccurate prediction. This problem is common to
a large family of algorithms, including decision trees. We have per-
formed experiments to compare the two test methods; the results
confirm that the KNN-on-leaves method exhibits marginally bet-
ter prediction than the tree-descent method. The behavior of the
two methods is illustrated in Figure 3.
4. CLUSTERING METHODS
The only remaining design choice is which unsupervised clus-
tering algorithm to employ for successively partitioning the data
during training, and the corresponding similarity measure. The
choice can change depending on the type of data to be classified,
while the overall framework remains the same, yielding a poten-
tial family of closely related variants of the CSL algorithm. This
enables flexibility in selecting a particular unsupervised clustering
algorithm for a given domain and dataset, without modifying any-
thing else in the algorithm. (Using different clustering algorithms
within the same class tree is also feasible as all decisions are made
locally in the tree.)
There are numerous clustering algorithms from the simple and
efficient k-means (Lloyd, 1982), self organizing maps (SOM; Kaski,
1997) and competitive networks (Kosko, 1991), to the more elabo-
rate and expensive probabilistic generative algorithms like mixture
of Gaussians, Probabilistic latent semantic analysis (PLSA; Hoff-
man, 1999) and Latent Dirichlet Allocation (LDA; Blei et al., 2003);
each has merits and costs. Given the biological derivation of the
system, we began by choosing k-means, a simple and inexpen-
sive clustering method that has been discussed previously as a
candidate system for biological clustering (Darken and Moody,
1990); the method could instead use SOM or competitive learn-
ing,two highly related systems. (It remains quite possible that more
robust (and expensive) algorithms such as PLSA and LDA could
provide improved prediction accuracy. Improvements might also
arise by treating the data at every node as a mixture of Gaus-
sians, and estimating the mixture parameters using the expectation
maximization (EM) algorithm.)
4.1. k-MEANS
k-Means is one of the most popular algorithms to cluster n vec-
tors based on distance measure into k partitions, where k < n. It
attempts to find the centers of natural clusters in the data. The
objective that k-means tries to minimize is the total intra cluster
FIGURE 3 Two methods by which the CSL algorithm predicts
category membership at test time. (Left) Class prediction via
hierarchical descent. At each step, a new (unknown) sample will fall
on one or the other side of a classification hyperplane. The decision
provides a path through the class tree at each node. At the leaves, the
class prediction is obtained. The numbering gives the order in which
the hyperplanes are probed. (Right) Class prediction using only leaves
of class tree. All leaves are considered simultaneously; the test
sample is compared to each leaf and the class prediction is obtained
using KNN.
www.frontiersin.org
January 2012 Volume 5 Article 50 7
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
variance, or, the squared error function:
=
K!i=1 !xj∈Ci
(xj − µi )2
where there are K clusters Si, i= 1, 2, . . ., K and µi is the centroid
or mean point of all the points xj ∈ Ci.
When k-means is used for the unsupervised appearance-based
clustering at the nodes of the class tree, the actual means obtained
are stored at each node, and the similarity measure is inversely
proportional to the Euclidean distance.
Initializing clusters
4.1.1.
In general, unsupervised methods are sensitive to initialization.
We initialize the clustering algorithm at every node in the class
tree as follows.
If we are at node l, with samples having one of Kl labels, we first
determine the class averages of the Kl categories. (For every class,
we remove the samples which are at least 2 standard deviations
away from the mean of the class for the initialization These sam-
ples are considered for the subsequent unsupervised clustering.)
If the number of clusters (branches), K∗ = min (Kl, Kmax) turns
out to be equal to Kl, then the averages are used as the seeds for
the clustering algorithm. If however K∗ < Kl, then we use a simple
and efficient method for obtaining the initial clusters by using an
initial run of k-means on the Kl averages in order to obtain the
K∗ initial centroids. The data samples are assigned to the clus-
ters using nearest neighbor mapping, and the averages of these K∗
clusters are used as seeds for a subsequent run of the unsuper-
vised clustering algorithm. (In our empirical experiments we have
used the k-means++ variant of the popular clustering algorithm
to obtain the initial cluster seeds; Arthur and Vassilvitskii, 2007.)
Figure 4 illustrates the initialization method. (While the method
works relatively well, further studies indicate that other meth-
ods, which directly utilize the semantic structure of the labeled
dataset, can result in even better performance. These alternate
approaches are not discussed in this paper in order to keep the
focus on introducing the core algorithm.) It is worth noting that
the initialization method can be thought of in terms of a logically
prior "developmental" period, in which no data is actually stored,
but instead sampling of the environment is used to set parameters
of the method; those parameters, once fixed, are then used in the
subsequent performance of the then-"adult" algorithm (Felch and
Granger, 2008).
EXPERIMENTS
5.
The proposed algorithm performs a number of operations on
its input, including the unsupervised discovery of structure in
the data. However, since the method, despite being composed
only of unsupervised clustering and reinforcement learning, can
nonetheless perform supervised learning, we have run tests that
involve using the CSL method solely as a supervised classifier.
In addition to these tests of supervised learning alone, we then
briefly describe some additional findings illustrating the CSL algo-
rithm's power at tasks beyond the classification task (including the
tasks of identifying structure in data, and localizing objects within
images).
When viewed solely as a supervised classifier, the CSL method
bears resemblances to two well-studied methods in machine learn-
ing and statistics, and we rigorously compare these. We compared
FIGURE 4 Initialization of the unsupervised partitioning for a set of
labeled training examples. The illustration depicts the process when
working with image datasets. The method needs to be applied only when the
desired number of clusters (in this example, 2) is less than the actual number
of labeled classes represented in the dataset (in this example, 4) to be
clustered.
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 8
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
the accuracy, and the time and space costs, of the CSL algo-
rithm as a supervised classifier, against the support vector machine
(SVM) and k-nearest neighbor (KNN) algorithms. Performance
was examined on two well-studied public datasets.
For SVM, we have used the popular LibSVM implementation
that is publicly available (Chang and Lin, 2001). This package
implements the "one vs one" flavor of multiclass classification,
rather than "one vs rest" variant based on the findings reported
in Hsu and Lin (2002). After experimenting with a few kernels,
we chose the linear kernel since it was the most efficient and
especially since it provided the best SVM results for the high-
dimensional datasets we tested. It is known that for the linear
kernel a weight vector can be computed and hence the support
vectors need not be kept in memory, resulting in low memory
requirements and fast recognition time. However, this is not true
for non-linear kernels where support vectors need to be kept
in memory to get the class predictions at run time. Since we
wish to compare the classifiers in the general setting and it is
likely that the kernel trick may need to be employed to separate
non-linear input space, we have retained the implementation of
LibSVM as it is (where the support vectors are retained in mem-
ory and used during testing to get class prediction). We realize
this may not be the fairest comparison for the current set of
experiments, however, we believe that this setting is more reflec-
tive of the typical use case scenario where the algorithms will be
employed.
For KNN we have hand coded the implementation and set
the parameter K= 1 for maximum efficiency. (For the CSL algo-
rithm with KNN-on-leaves, we use K= 1 as well.) The test bed is
a machine running windows XP 64 with 8GB memory. We have
not used hardware acceleration for any of the algorithms to keep
the comparison fair.
We have used two popular datasets from different domains
with very different characteristics (including dimensionality of
the data) to fully explore the strengths and weaknesses of the algo-
rithm. One is a subset of the Caltech-256 image set, and the other is
a very high-dimensional dataset of neuroimaging data from fMRI
experiments, that has been widely studied.
For both experiments, we performed multiple runs, differently
splitting the samples from each class into training and testing sets
(roughly equal in number). The results shown indicate the means
and standard deviations of all runs.
5.1. OBJECT RECOGNITION
Our first experiment tests the algorithm for object recognition
in natural still image datasets. The task is to predict the label
for an image, having learned the various classes of objects in
images through a training phase. We report empirical findings
for prediction accuracy and computational resources required.
5.1.1. Dataset
The dataset used consists of a subset of the Caltech-256 dataset
(Griffin et al., 2007) using 39 categories, each with roughly 100
instances. The categories were specifically chosen to exhibit very
high between-category similarity, intentionally selected as a very
challenging task, with high potential confusion among classes. The
categories are:
leopard, raccoon, zebra
swan, bat, cormorant, butterfly
• Mammals: bear, chimp, dog, elephant, goat, gorilla, kangaroo,
• Winged: duck, goose, hummingbird, ostrich, owl, penguin,
• Crawlers
iguana,
cockroach, grasshopper, housefly, praying mantis, scorpion,
snail, spider, toad
(reptiles/insects/arthropods/amphibians):
• Inanimate objects: backpack, baseball glove, binoculars, bull-
ipod,
dozer, chandeliers, computer monitor, grand piano,
laptop, microwave.
We have chosen an extremely simple (and very standard) method
for representing images in order to maintain focus on the descrip-
tion of the proposed classifier. First a feature vocabulary consisting
of SIFT features (Lowe, 2004) is constructed by running k-means
on a random set of images containing examples from all classes of
interest; each image is then represented as a histogram of these fea-
tures. The positions of the features and their geometry is ignored,
simplifying the process and reducing computational costs. Thus
each image is a vector x ∈ Rm, where m is the size of the acquired
vocabulary. Each dimension of the vector is a count of the number
of times the particular feature occurred in the image. This rep-
resentation, known as the "Bag of Words," has been successfully
applied before in several domains including object recognition in
images (Sivic and Zisserman, 2003).
We ran a total of 8 trials, corresponding to 8 different random
partitionings of the Caltech-256 data into training and testing sets.
In each trial, we ran the test for each of a range of Kmax values, to
test this free parameter of the CSL model.
5.1.2. Prediction accuracy
The graph in top left of Figure 5 compares the classifier pre-
diction accuracy of the proposed algorithm with that of SVMs
on the 39 subsets of Caltech-256 described earlier. As expected,
the simplistic image representation scheme, and the readily con-
fused category members, renders the task extremely difficult. It
will be seen that all classifiers perform at a very modest success
rate with this data, indicating the difficulty of the dataset and
the considerable room for potential improvement in classification
techniques.
The two variants of the CSL algorithm are competitive with
SVM: SVM has an average accuracy of 23.9%; CSL with tree
descent has an average accuracy of 19.4%; and CSL with KNN-
on-leaves has an average prediction accuracy of 21.3%. The KNN
algorithm alone performs relatively poorly, with an average predic-
tion accuracy of 13.6%. Chance probability of correctly predicting
a class is 1 out of 39 (2.56%).
It can be seen that the branch factor does not have a signif-
icant impact on error rates. This is possibly because the class
tree grows until the leaves are pure, and the resulting internal
structure, though different across choices of Kmax, does not signif-
icantly impact the ultimate classifier performance as the hierarchy
adapts its shape. Different internal structure could significantly
affect the performance of the algorithm on tasks that depended
on the similarity structure of the data, but for the sole task of
supervised classification, the tree's internal nodes have little effect
on prediction accuracy.
www.frontiersin.org
January 2012 Volume 5 Article 50 9
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
FIGURE 5 Comparison of CSL and other standard classifiers. (Top left)
Prediction accuracy of the CSL classifier on the Caltech-256 subsets. The
scores in blue are the rates achieved by the CSL classifier. Scores in pink are
from standard multiclass SVM (see text). (Top right) Memory requirements for
CSL with respect to the branch factor parameter. The figure shows that the
parameter does not significantly impact the size of the tree. Also, we can see
a clear difference between the memory usage of CSL and the other
supervised classifiers after training. (Bottom left) Run times for the training
stage. CSL (red) requires roughly an order of magnitude less training run time
than SVM (blue). (Bottom right) Average time to recognize a new image after
training for the different algorithms. The y axis (logarithmic scale) shows CSL
outperforming SVM and KNN by an order of magnitude.
5.1.3. Memory usage
The graph in top right of Figure 5 shows the relationship between
the overall number of nodes in the tree to be retained (and hence
vectors of dimensionality M ) and the branch factor for CSL clas-
sifier. CSL with tree descent had to store an average of 1036.25
vectors, while the knn-on-leaves variant had to store 902.21 vec-
tors. SVM required 2286 vectors while the vanilla KNN method
(with k= 1) requires storage of the entire training corpus of 2322
vectors. Thus, the number of vectors retained in memory by the
CSL variants is roughly half the number retained by the SVM and
KNN algorithms. Further, the memory needed to store the trained
model when we predict using the KNN-on-leaves approach is
smaller than when we use tree descent, as we expected and dis-
cussed earlier. As can be seen, there is not much variation in CSL
performance across different branch factor values. This suggests
that after a few initial splits, most of the sub trees have very few
categories represented within them and hence the upper bound
on the branch factor does not play a significant role in ongoing
performance.
5.1.4. Classifier run times
The runtime costs of the algorithms paint an even more startling
picture. The graph in bottom left of Figure 5 shows the plots
comparing the training times of the CSL and SVM algorithms.
The two variants of CSL have the same training procedure and
hence require the same time to train. (KNN has no explicit train-
ing stage.) As can be seen, the training time of the new algorithm
(average of 2.42 s) is roughly an order of magnitude smaller than
that of the SVM (average of 18.54 s). It should be clearly noted
that comparisons between implementations of algorithms will not
necessarily reflect underlying computational costs inherent to the
algorithms, for which further analysis and formal treatment will be
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 10
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
required. Nonetheless, in the present experiments, the empirical
costs were radically different despite efforts to show the SVM in
its best light.
As indicated earlier, the choice of branch factor does not have a
large impact on the training time needed. We also found that the
working memory requirements of our algorithm were very small
compared to that of the SVM. In the extreme, when large represen-
tations were used for images, the memory requirements for SVMs
rendered the task entirely impracticable. In such circumstances,
the CSL method still performed effectively. The working amount
of memory we need is proportional to the largest clustering job
that needs to be performed. By choosing low values of Kmax, we
empirically find that we can keep this requirement low without
loss of classifier performance.
The bottom right plot of Figure 5 shows how the average
time for recognizing a new image varies with branch factor.
The times are shown in logarithmic scale. The CSL variants are
an order of magnitude faster than KNN and SVM algorithms
with the tree descent variant being the fastest. This shows the
proposed algorithm in its best light. Once training is complete,
recognition can be extremely rapid by doing hierarchical descent,
making the CSL method unusually well suited for real-time
applications.
5.2. HAXBY fMRI DATASET, 2001
5.2.1. Dataset
Having demonstrated the CSL system on image data, we selected
a very different dataset to test: neuroimaging data collected from
the brain activity of human subjects who were viewing pictures.
As with the Caltech-256 data, we selected a very well-studied set
of fMRI data, from a 2001 study by Haxby et al. (2001).
Six healthy human volunteers entered an fMRI neuroimaging
apparatus and viewed a set of pictures while their brain activity
(blood oxygen-level dependent measures) was recorded. In each
run, the subjects passively viewed gray scale images of eight object
categories, grouped in 24 s blocks separated by rest periods. Each
image was shown for 500 ms and was followed by a 1500-ms inter-
stimulus interval. Each subject carried out twelve of these runs.
The stimuli viewed by the subjects consisted of images from the
following eight classes: Faces, Cats, Chairs, Scissors, Houses, Bot-
tles, Shoes, and random scrambled pictures. Full-brain fMRI data
were recorded with a volume repetition time of 2.5 s, thus, a stim-
ulus block was covered by roughly 9 volumes. For a complete
description of the experimental design and fMRI acquisition para-
meters, see Haxby et al. (2001). (The dataset is publicly available.)
Each fMRI recording corresponding to 1volume in a block for
a given input image can be thought of as a vector with 163840
dimensions. The recordings for all the subjects have the same vec-
tor length. (In the original work, "masks" for individual brain
areas were provided, retaining only those voxels that were hypoth-
esized by the experimenters to play a significant role in object
recognition. Using these masks reduces the data dimensional-
ity by a large factor. However, the masks are of different lengths
for different subjects, thus preventing meaningful aggregation of
recordings across subjects. Thus, we have not used the masks and
instead trained the classifiers in the original high dimensional
space.)
Testing on individual subjects
5.2.2.
For each subject who participated in the experiment, we have neu-
roimaging data collected as that subject viewed images from each
of the eight classes. The task was to see whether, from the brain data
alone, the algorithms could predict what type of picture the subject
was viewing. Top left in Figure 6 shows the prediction accuracy
of the various classifiers we tried. On the whole, all the classifiers
exhibit similar performance with SVM performing slightly better
on a couple of the subjects.
Top right of Figure 6 shows the memory requirements for all
the algorithms. The CSL variants require significantly less mem-
ory to store the model learned during training compared to SVM
and KNN. SVM requires a large number of support vectors to
fully differentiate the data from different classes leading to large
memory consumption, whereas KNN needs to store all the train-
ing data in memory. For CSL, if the testing method is tree descent,
then the entire hierarchy needs to be kept in memory. For the
KNN-on-leaves testing method, only the leaves of the tree are
retained, rendering even a smaller memory requirement for the
stored model.
Bottom left of Figure 6 shows the training time for the CSL
algorithm being an order of magnitude smaller than that of SVM.
KNN does not have any explicit training stage. Finally, bottom
right of Figure 6 compares the recognition time for the differ-
ent algorithms, again on a log scale. The average recognition time
on a new sample for the CSL tree descent variant is a couple of
orders of magnitude smaller than both KNN and SVM. For the
KNN-on-leaves variant of the CSL method, the recognition time
grows larger (while still being significantly smaller than KNN or
SVM). Therefore the fastest approach is performing a tree descent
(paying a penalty in terms of memory requirements for storing
the model).
5.2.3. Aggregating data across subjects
Since the recordings from all the subjects have the same dimen-
sionality, we can merge all the data from the different subjects
into 1 large dataset and partition it into the training and test-
ing datasets. This way we can study the performance trends with
increasing datasets. The SVM system, unfortunately, was unable
to run on pools containing more than two subjects, due to the
SVM system's high memory requirements. Nonetheless, the two
variants of the CSL algorithm, and the KNN algorithm, ran suc-
cessfully on collections containing up to five subjects' aggregated
data.
The subplot on the left of Figure 7 shows that the classification
prediction accuracy of the different classifiers remain competitive
with each other as we increase the pool. The subplot on the right
of Figure 7 shows the trend of memory consumption by the dif-
ferent algorithms as we increase the number of subjects included.
Compared to standard KNN, the increase in memory consump-
tion is much slower (sub linear) for the CSL algorithm, with the
KNN-on-leaves variant of the CSL algorithm growing very slowly.
Finally, in Figure 8, we examine the growth in the average recog-
nition time with increasing pool size. The costs of adding data
cause the recognition time to grow for the KNN algorithm more
than for either variant of the CSL algorithm (either tree descent
or KNN-on-leaves versions).
www.frontiersin.org
January 2012 Volume 5 Article 50 11
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
FIGURE 6 Comparison of classifiers. The task was to determine,
from neuroimaging data alone, what type of image the subject was
viewing. The classifiers tested were SVM, KNN, CSL:KNN-on-leaves
and CSL:Tree descent. Each test was run on the data from an individual
subject in the fMRI experiment (see text). (Top left) Accuracy of
classifier predictions. All classifiers have similar prediction accuracy for
four of the subjects, while SVM performs slightly better on two
subjects (S3, S6). (Top right) Memory usage across different algorithms.
CSL variants exhibit significantly lower memory usage than the other
two methods. (Bottom left) Time required to train classifiers (CSL vs.
SVM; the two CSL variants are tested differently but trained identically;
KNN algorithms are not trained). CSL requires roughly an order of
magnitude less time to train than SVM. (Bottom right) Average time to
recognize a new input after training. The y axis (a logarithmic scale)
shows clearly that CSL with tree descent outperforms the other
classifiers by an order of magnitude.
Between these two CSL algorithm variants, the latter exhibits
some modest time growth as data is added, whereas the former
(tree descent) version of CSL exhibits no significant increase in
recognition time whatsoever as more data is added to the task.
It is notable that the reason for this is that the tree depth has
not increased with increasing size of the dataset; that is, as more
data is added, the learned CSL tree arrives at the ability to suc-
cessfully classify the data early on, and adding new data does
not require the method to add more to the tree. Interestingly,
the trees become better balanced as we increase the number of
subjects, but their sizes do not increase. The results suggest that
the CSL algorithm is better suited to scale to extremely large
data sets than either of the competing standard SVM or KNN
methods.
6. ANALYSES AND EXTENSIONS
6.1. ALGORITHM COMPLEXITY
When k-means is used for clustering, the time complexity for
each partitioning is O(NtK), where N is the number of samples,
K is the number of partitions and t is the number of itera-
tions. If we fix t to be a constant (by putting an upper limit
on it), then each split takes O(NK ). Since we also put a bound
on K (Kmax), we can assume that each split is O(N). Further
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 12
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
FIGURE 7 (Left) Accuracy of classifier predictions on split tests.
The accuracy of the different algorithms remain competitive as we
increase the subject pool. (The memory usage by the LibSVM
implementation of SVM was too large for testing on subject pools
larger than 2). (Right) Tracks the trend of memory consumption with
increasing size of the subject pool for the classifiers. KNN's memory
usage grows linearly, whereas the CSL variants grow at a lower rate,
illustrating their scalability.
can be further treated in parallel. However, in the experiments
reported here, we have as yet made no attempt to parallelize the
code, seeking instead to compare the algorithm directly against
current standard SVM implementations.
6.2. COMPARISON WITH OTHER HIERARCHICAL LEARNING
TECHNIQUES
The structure of the algorithm makes it very similar to CART (and
in particular, decision trees; Buntine, 1992) since both families of
algorithms partition the non-linear input space into discontinuous
regions such that the individual sub regions themselves provide
effective class boundaries. However, there are several significant
differences.
• Perhaps the most substantial difference is that decision trees
use the labels of the data to perform splits, whereas the CSL
algorithm partitions based on unsupervised similarity.
• The CSL algorithm splits in a multivariate fashion, taking into
account all the dimensions of the data samples, as opposed
to decision trees where most often, a single dimension which
results in the largest demixing of the data, is used to make splits.
The path from the root to a leaf in a decision tree is a con-
junction of local decisions on feature values and as a result
is prone to over fitting. As discussed before, the CSL tends to
exhibit little overfitting, and we can understand why this is the
case (see Discussion in the Simplified Algorithm section ear-
lier). The leaves can be treated independently of the rest of
the tree and KNN can be used on them to obtain the class
predictions.
• Decision trees are by nature a 2 class discriminative approach
(multiclass problems can be handled using binary decision trees;
Lee and Oh, 2003) whereas the CSL algorithm is a natural
multiclass generative algorithm.
FIGURE 8 Average recognition time required for a new test sample, as
a function of the amount of data trained. As expected, KNN grows
linearly. CSL with KNN-on-leaves grows more slowly. Interestingly, CSL
with tree descent hardly shows any increase, suggesting its scalability to
extremely large datasets.
analysis is needed on the total number of paths and their con-
tribution to runtime. The maximum amount of memory needed
is for the first unsupervised partitioning. This is proportional
to O(NK ). When we have small K, the amount of memory is
directly proportional to the number of data elements being used
in training.
As mentioned earlier, the algorithm is intrinsically highly par-
allel. After every unsupervised partitioning, each of the partitions
www.frontiersin.org
January 2012 Volume 5 Article 50 13
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
Most importantly, the goals of these systems differ. The primary
goal of the CSL algorithm is to uncover natural structure within the
data. The fact that the label-based impurity of classes is reduced,
resulting in the ability to classify labeled data, falls out as a (very
valuable) side effect of the procedure. The CSL algorithm thus will
carry out a range of additional tasks, beyond supervised classifi-
cation, that use deeper analysis of the underlying structure of the
data, not apparent through supervised labeling alone.
6.3. DISCOVERY OF STRUCTURE
For purposes of this paper we have focused solely on the classifica-
tion abilities of the algorithm, though the algorithm can perform
many other tasks outside the purview of classification. Here we
will briefly cover two illustrative additional abilities: (i) uncover-
ing secondary structure of data, and (ii) localization of objects
within images.
6.3.1. Haxby dataset
Once a model is trained, for each training sample if we do hier-
archical descent and aggregate the posterior probabilities of the
nodes along the path, we get a representation for the sample.
When we do an agglomerative clustering on that representation, we
uncover secondary structure suggesting meta classes occurring in
the dataset. Figure 9 captures the output of such an agglomerative
clustering for the recordings of one subject (S1). Here we can see
extensive structure relations among the responses to various pic-
tures; perhaps most prominent is a clear separation of the data
into animate and inanimate classes. The tree suggests the struc-
ture of information that is present in the neuroimaging data; the
subjects' brain responses distinguish among the different types of
pictures that they viewed. Related results were shown by Hanson
et al. (2004); these were arrived at by analysis of the hidden node
activity of a back propagation network trained on the same data.
In contrast, it is worth noting that the CSL classifier obtains this
structure as a natural byproduct of the tree-building process.
Image localization
6.3.2.
A task quite outside the realm of supervised classification is
that of localizing, i.e., finding an object of interest within an
image. This task is useful to illustrate additional capabilities of the
algorithm beyond just classification, making use of the internal
representations it constructs.
We assume for this example that the clustering component of
the algorithm is carried out by a generative method such as PLSA
(Sivic et al., 2005); we then can assume that the features specific
to the object class will contribute to the way in which an image
becomes clustered, and that those features will contribute more
than will random background features in the image.
FIGURE 9 Agglomerative clustering of tree node activity for all the data from Subject 1. The hierarchy shows extensive internal structure in the
responses, including a clear separation of animate and inanimate stimuli. See text for further discussion.
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 14
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
Figure 10 shows an example of localization of a face within
an image. The initial task was to classify images of faces, cars,
motorcycles, and airplanes (from Caltech 4). PLSA was used for
clustering and for determining the cluster membership of a previ-
ously unseen image x. For every cluster z, we can obtain the pos-
terior probability p(zx, w), for every feature w in the vocabulary.
FIGURE 10 An illustration of object localization on an image from
Caltech-256. (A) The original image. (B–E) Positive, neutral, and negative
features (green, blue, red, respectively) shown at levels 1 through 4 along the
path in the CSL tree. (F) A thresholded map of the aggregate feature scores.
Points in green indicate feature scores above threshold and those in red
indicate below-threshold scores. Note that although green dots occur in
multiple regions, the presence of red dots (negative features) is limited only
to regions outside the face region.
www.frontiersin.org
January 2012 Volume 5 Article 50 15
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
Thus, we can test all features in the image to see which ones max-
imize the posterior, indicating strong influence on the eventual
cluster membership. The location of those features can then be
used to identify the vicinity of the object.
As the path from root to leaf in the CSL hierarchy is traversed
for a particular test image, the posterior at a given node determines
the contribution of the feature to the branch selected. Let y be the
final object label prediction for image x.
Consider feature fi from the vocabulary. At any given node at
height l along the path leading to prediction of y for x, let dl
i be
the branch predicted by feature fi, i.e., among all branches at node
l, the posterior for that branch is highest for that feature. dl
i is
actually a set of labels that can be reached at various leaves using
the branch and finally let the overall branch taken at l be bl.
i ), and finally, negative if 1(dl
i
At level l, fi, can be classified as positive if 1(dl
i == bl ), neutral
= bl ) and 1(y ∈ dl
if 1(dl
= bl ) and
i
1(y /∈ dl
i ). The overall score for fi is a weighted sum Si of all the
scores (negative features getting a negative score) along the path.
Since we know the locations of the features, we can transfer the
scores to actual locations on the images (more than one location
may map to the same feature in the vocabulary). When a simple
threshold is applied, we get the map seen in the final image. The
window most likely to contain the object can then be obtained by
optimization of the scores on the map using branch and bound
techniques.
7. CONCLUSION
We have introduced a novel, biologically derived algorithm that
carries out similarity-based hierarchical clustering combined with
simple matching, thus determining when nodes in the tree are to
be iteratively deepened. The clustering mechanism is a reduced
subset of published hypotheses of thalamocortical function; the
match/mismatch operation is a reduced subset of proposed basal
ganglia operation; both are described in Granger (2006). The
resulting algorithm performs a range of tasks, including identify-
ing natural underlying structure among object in the dataset; these
abilities of the algorithm confer a range of application capabilities
beyond traditional classifiers. In the present paper we described
in detail just one circumscribed behavior of the algorithm: its
ability to use its combination of unsupervised clustering and rein-
forcement to carry out the task of supervised classification. The
experiments reported here suggest the algorithm's performance
is comparable to that of SVMs on this task, yet requires only a
fraction of the resources of SVM or KNN methods.
It is worth briefly noting that the intent of the research
described here has not been to design novel algorithms, but rather
to educe algorithms that may be at play in brain circuitry. The
two brain structures referenced here, neocortex and basal gan-
glia, when studied in isolation, have given rise to hypothesized
operations of hierarchical clustering and of reinforcement learn-
ing, respectively (e.g., Sutton and Barto, 1998; Rodriguez et al.,
REFERENCES
Alexander, G., and DeLong, M. (1985).
Microstimulation of
the primate
neostriatum. I. Physiological prop-
erties of
striatal microexcitable
zones.
J.
1401–1416.
Neurophysiol.
53,
Arthur, D., and Vassilvitskii, S. (2007).
the advantages of
seeding," in SODA '07:
"k-Means++:
careful
2004). These structures are connected in a loop, such that (striatal)
reinforcement learning can be hypothesized to selectively interact
with (thalamocortical) hierarchies being constructed. We conjec-
ture that the result is a novel composite algorithm (CSL), which
can be thought of as iteratively constructing rich representations
of sampled data.
Though the algorithm was conceived and derived from analysis
of cortico-striatal circuitry, the next aim was to responsibly analyze
its efficacy and costs and compare it appropriately against other
competing algorithms in various domains. Thus we intention-
ally produced very general algorithmic statements of the derived
cortico-striatal operations, precisely so that (1) we can retain func-
tional equivalency with the referenced prior literature (Schultz
et al., 1997; Suri and Schultz, 2001; Schultz, 2002; Rodriguez et al.,
2004; Daw and Doya, 2006); and (2) the derived algorithm can
be responsibly compared directly against other algorithms. The
algorithm can be applied to a number of tasks; for purposes of
the present paper we have focused on supervised classification
(though we also briefly demonstrated the utility of the method for
different tasks, including identification of structure in data, and
localization of objects in an image).
It is not yet known what tasks or algorithms are actually being
carried out by brain structures. Brain circuits may represent com-
promises among multiple functions, and thus may not outperform
engineering approaches to particular specialized tasks (such as
classification). In the present instance, individual components are
hypothesized to perform distinct algorithms, hierarchical clus-
tering and reinforcement learning, and the interactions between
those components perform still another composite algorithm, the
CSL method presented here. (And, as mentioned, the studied
operations are very-reduced subsets of the larger hypothesized
operations of these thalamocortical and basal ganglia systems; it
is hoped that ongoing study will yield further algorithms arising
from richer interactions of these cortical and striatal structures,
beyond the reduced simplifications studied in the present paper.)
As might be expected of a method that has been selectively devel-
oped in biological systems over evolutionary time, these compo-
nent operations may represent compromises among differential
selectional pressures for a range of competing tasks, carried out
by combined efforts of multiple distinct engines of the brain. This
represents an instance in which models of a biological system lead
to derivation of tractable algorithms for real-world tasks. Since
the biologically derived method studied here substantially outper-
forms extant engineering methods in terms of efficacy per time
or space cost, we forward the conjecture that brain circuitry may
continue to provide a valuable resource from which to mine novel
algorithms for challenging computational tasks.
ACKNOWLEDGMENTS
This work was supported in part by grants from the Office of Naval
Research and the Defense Advanced Research Projects Agency.
Proceedings of the Eighteenth Annual
ACM-SIAM Symposium on Discrete
Algorithms, New Orleans.
Baron, A., and Cover, T.
Minimum complexity
(1991).
density
estimation. IEEE Trans. Inf. Theory
4, 1034–1054.
Bishop, C. (1996). Neural Networks
for Pattern Recognition. New York:
Oxford University Press.
Frontiers in Computational Neuroscience Iterative computations of cortico-striatal loops
January 2012 Volume 5 Article 50 16
Chandrashekar and Granger
Iterative computations of cortico-striatal loops
Blei, D., Ng, A., and Jordan, M. (2003).
Latent Dirichlet allocation. J. Mach.
Learn. Res. 3, 993–1022.
Breiman, L., Friedman, J., Olshen, R.,
and Stone, C. (1984). Classification
and Regression Trees. Belmont, CA:
Wadsworth.
Brown, J., Bullock, D., and Grossberg,
S. (1999). How the basal ganglia
use parallel excitatory and inhibitory
learning pathways
to selectively
respond to unexpected reward. J.
Neurosci. 19, 10502–10511.
Buntine, W. (1992). Learning classifica-
tion trees. Stat. Comput. 2, 63–73.
Chang, C., and Lin, C. (2001). Lib-
svm: A Library for Support Vector
Machines. Available at: http://www.
csie.ntu.edu.tw/cjlin/libsvm
Darken, C., and Moody, J. (1990). "Fast,
adaptive k-means clustering: some
empirical results," in Proceedings of
the IEEE IJCNN Conference (San
Diego: IEEE Press).
Daw, N. (2003). Reinforcement Learning
Models of the Dopamine System and
Their Behavioral Implications. Ph.D.
Thesis, Carnegie Mellon University,
Pittsburgh, PA.
Daw, N., and Doya, K. (2006). The com-
putational neurobiology of learning
and reward. Curr. Opin. Neurobiol.
16, 199–204.
Felch, A., and Granger, R. (2008).
The hypergeometric connectivity
hypothesis: divergent performance
of brain circuits with different
synaptic connectivity distributions.
Brain Res. 1202, 3–13.
George, D., and Hawkins, J. (2009).
Towards a mathematical theory of
cortical microcircuits. PLoS Comput.
Biol. 5, e1000532. doi:10.1371/jour-
nal.pcbi.1000532
Granger, R. (2006). Engines of the brain:
the computational
instruction set
of human cognition. AI Mag. 27,
15–32.
Granger, R. (2011). How Brains are
Built: Principles of Computational
Neuroscience. Cerebrum; The Dana
Foundation, Available at: http://
dana.org/news/cerebrum/detail.as
px?id=30356.
Griffin, G., Holub, A., and Perona, P.
(2007). Caltech-256 Object Category
Dataset. California Institute of Tech-
nology.
Gurney, K., Prescott, T., and Redgrave,
P. (2001). A computational model of
action selection in the basal ganglia:
I. a new functional anatomy. Biol.
Cybern. 84, 401–410.
Hanson, S., Matsuka, T., and Haxby, J.
(2004). Combinatorial codes in ven-
tral temporal lobe for object recog-
nition: Haxby (2001). Revisited: is
there a face area? Neuroimage 23,
156–166.
Haxby, J., Gobbini, M., Furey, M., Ishai,
A., Schouten, J., and Pietrini, P.
(2001). Distributed and overlapping
representations of faces and objects
in ventral temporal cortex. Science
293, 2425–2430.
Hoffman, T. (1999)."Probabilistic latent
semantic indexing," in SIGIR '99:
Proceedings of
the 22nd Annual
International ACM SIGIR Confer-
ence on Research and Development
in Information Retrieval, Berkeley,
50–57.
Houk, J., Bastianen, C., Fansler, D., Fish-
bach, A., Fraser, D., Reber, P., Roy, S.,
and Simo, L. (2007). Action selection
and refinement in subcortical loops
through basal ganglia and cerebel-
lum. Philos. Trans. R. Soc. Lond. B
Biol. Sci. 362, 1573–1583.
Hsu, C., and Lin, C. (2002). A compari-
son of methods for multi-class sup-
port vector machines. IEEE Trans.
Neural Netw. 13, 415–425.
Kaski, S. (1997). Data exploration using
self-organizing maps. Acta Polytech-
nica Scand. Math. Comput. Manag.
Eng. Ser. 82.
Kemp, J., and Powell, T. (1971). The con-
nexions of the striatum and globus
pallidus: synthesis and speculation.
Philos. Trans. R. Soc. Lond. B Biol.
Sci. 262, 441–445.
Kosko, B. (1991). Stochastic competitive
learning. IEEE Trans. Neural Netw. 2,
522–529.
Leblois, A., Boraud, T., Meissner,
W., Bergman, H., and Hansel,
D. (2006). Competition between
feedback loops underlies normal
and pathological dynamics in the
basal
26,
3567–3583.
J. Neurosci.
ganglia.
Lee, J., and Oh, I. (2003). "Binary classi-
fication trees for multi-class classifi-
cation problems" in ICDAR '03 Pro-
ceedings of the Seventh International
Conference on Document Analysis
and Recognition, Edinburgh.
Lee, T., and Mumford, D. (2003). Hier-
archical Bayesian inference in the
visual cortex. Opt. Soc. Am. 20,
1434–1448.
Lloyd, S. (1982). Least squares quantiza-
tion in pcm. IEEE Trans. Inf. Theory
28, 129–136.
Lowe, D. (2004). Distinctive image fea-
tures from scale-invariant keypoints.
Int. J. Comput. Vis. 60, 91–110.
Marr, D. (1980). Vision. MIT press.
McGeorge, A., and Faull, R. (1988). The
organization of the projection from
the cerebral cortex to the striatum in
the rat. Neuroscience 29, 503–537.
Ng, A., and Jordan, M. (2002). On dis-
criminative vs. generative classifiers.
a comparison of logistic regression
and naive Bayes. Neural Inf. Process.
Syst. 2, 841–848.
O'Doherty,
J., Dayan, P., Friston,
K., Critchley, H., and Dolan, R.
(2003). Temporal difference mod-
els
and reward-related learning
in the human brain. Neuron 38,
329–337.
Riesenhuber, M., and Poggio, T. (1999).
Hierarchical models of object recog-
nition in cortex. Nat. Neurosci. 2,
1019–1025.
J.,
Rodriguez, A., Whitson,
and
Granger, R. (2004). Derivation and
analysis of basic
computational
operations
thalamocortical
circuits.
J. Cogn. Neurosci. 16,
856–877.
of
Schultz, W. (2002). Getting formal with
dopamine and reward. Neuron 36,
241–263.
Schultz, W., Dayan, P., and Montague,
R. (1997). A neural substrate of
prediction and reward. Science 175,
1593–1599.
Sivic, J., Russell, B., Efros, A., Zisserman,
A., and Freeman, W. (2005). Discov-
ering objects and their location in
images. IEEE Int. Conf. Comput. Vis.
1, 370–377.
Sivic, J., and Zisserman, A. (2003).
"Video Google: a text
retrieval
approach to object matching in
videos," in ICCV '03: Proceedings of
the Ninth IEEE International Confer-
ence on Computer Vision, Nice.
Stephan, H. (1972). "Evolution of pri-
mate brains: a comparative anatomi-
cal approach,"in Functional and Evo-
lutionary Biology of Primates, ed. R.
Tuttle (Chicago: Aldine-Atherton),
155–174.
Stephan, H., Bauchot, R., and Andy, O.
(1970). Data on size of the brain and
of various brain parts in insectivores
and primates. Advances in Primatol-
ogy, 289–297.
Stephan, H., Frahm, H., and Baron, G.
(1981). New and revised data on vol-
umes of brain structures in insecti-
vores and primates. Folia Primatol.
35, 1–29.
Suri, R., and Schultz, W. (2001). Tem-
poral difference model reproduces
anticipatory neural activity. Neural
Comput. 13, 841–862.
Sutton, R., and Barto, A.
(1990).
"Time derivative models of Pavlov-
ian reinforcement," in Learning and
Computational Neuroscience: Foun-
dations of Adaptive Networks, eds. M.
Gabriel, and J. Moore (MIT Press),
497–537.
Sutton, R., and Barto, A. (1998). Rein-
forcement Learning: An Introduction.
MIT press.
Teh, Y., Jordan, M., Beal, M., and Blei,
D. (2004). "Sharing clusters among
related groups: hierarchical dirichlet
processes," in Proceedings of Neural
Information Processing Systems, Van-
couver.
Ullman, S. (2006). Object recognition
and segmentation by a fragment-
based hierarchy. Trends Cogn. Sci.
(Regul. Ed.) 11, 58–64.
Vapnik, V. (1995). The Nature of Sta-
tistical Learning Theory. New York:
Springer.
Conflict of Interest Statement: The
authors declare that the research was
conducted in the absence of any
commercial or financial relationships
that could be construed as a potential
conflict of interest.
A
Chandrashekar
Received: 27 May 2011; accepted: 28
October 2011; published online: 10 Jan-
uary 2012.
Citation:
and
Granger R (2012) Derivation of a
novel efficient supervised learning algo-
rithm from cortical-subcortical
loops.
Front. Comput. Neurosci. 5:50. doi:
10.3389/fncom.2011.00050
This article was submitted to Frontiers in
Iterative computations of cortico-striatal
loops, a specialty of Frontiers in Compu-
tational Neuroscience.
Copyright © 2012 Chandrashekar and
Granger. This is an open-access arti-
cle distributed under
terms of
the Creative Commons Attribution Non
Commercial License, which permits non-
commercial use, distribution, and repro-
duction in other forums, provided the
original authors and source are credited.
the
www.frontiersin.org
January 2012 Volume 5 Article 50 17
|
1808.04262 | 1 | 1808 | 2018-08-10T08:54:31 | Connectivity-Driven Brain Parcellation via Consensus Clustering | [
"q-bio.NC",
"cs.LG",
"stat.ML"
] | We present two related methods for deriving connectivity-based brain atlases from individual connectomes. The proposed methods exploit a previously proposed dense connectivity representation, termed continuous connectivity, by first performing graph-based hierarchical clustering of individual brains, and subsequently aggregating the individual parcellations into a consensus parcellation. The search for consensus minimizes the sum of cluster membership distances, effectively estimating a pseudo-Karcher mean of individual parcellations. We assess the quality of our parcellations using (1) Kullback-Liebler and Jensen-Shannon divergence with respect to the dense connectome representation, (2) inter-hemispheric symmetry, and (3) performance of the simplified connectome in a biological sex classification task. We find that the parcellation based-atlas computed using a greedy search at a hierarchical depth 3 outperforms all other parcellation-based atlases as well as the standard Dessikan-Killiany anatomical atlas in all three assessments. | q-bio.NC | q-bio |
Connectivity-Driven Brain Parcellation via
Consensus Clustering
Anvar Kurmukov1,2, Ayagoz Mussabayeva1, Yulia Denisova1, Daniel Moyer4,
and Boris Gutman3,1
1 The Institute for Information Transmission Problems
2 National Research University Higher School of Economics
3 Illinois Institute of Technology
4 University of Southern California
Abstract. We present two related methods for deriving connectivity-
based brain atlases from individual connectomes. The proposed methods
exploit a previously proposed dense connectivity representation, termed
continuous connectivity, by first performing graph-based hierarchical clus-
tering of individual brains, and subsequently aggregating the individual
parcellations into a consensus parcellation. The search for consensus min-
imizes the sum of cluster membership distances, effectively estimating a
pseudo-Karcher mean of individual parcellations. We assess the quality
of our parcellations using (1) Kullback-Liebler and Jensen-Shannon di-
vergence with respect to the dense connectome representation, (2) inter-
hemispheric symmetry, and (3) performance of the simplified connec-
tome in a biological sex classification task. We find that the parcellation
based-atlas computed using a greedy search at a hierarchical depth 3
outperforms all other parcellation-based atlases as well as the standard
Dessikan-Killiany anatomical atlas in all three assessments.
1
Introduction
The ability to quantify how the human brain is interconnected in vivo has opened
the door to a number of possible analyses. In nearly all of these, brain parcel-
lation plays a crucial role. Variations in parcellation significantly impact con-
nectome reproducibility, derived graph-theoretical measures, and the relevance
of connectome measures with respect to biological questions of interest [16]. A
natural approach is then to use individual densely sampled connectomes to drive
the parcellation directly, leading to a more compact, connectivity-aware set of
brain regions and resulting graph, as done in e.g. [10]. A comprehensive review
of parcellation methods and their effects on the derived connectome quality is
given in [17]. Because individual connectivity data is at once very informative
and highly redundant, there is a great flexibility in how parcels can be derived
from dense, highly resolute graphs. It is possible for example to derive (1) a uni-
fied population-based atlas, (2) individual-level parcellations with cross-subject
label mapping, or (3) individual parcellations with no inter-subject label corre-
spondence. While the first approach is appealing for its simplicity and ease of
2
Anvar Kurmukov et al.
interpretation, the second and third may enable the researcher to reveal some
individual aspect of the connectome that is lost in the aggregate atlas.
In this work, we attempt to bridge these three approaches by first construct-
ing maximally flexible hierarchical parcellations, and then finding a unifying set
of labels and parcels to maximize individual agreement. We use the a contin-
uous representation of a brain connectivity [8] as our initial dense connectome
representation. Continuous connectivity is a parcellation-free representation of
tractography-based, or "structural" connectomes that is based on the Poisson
point process. Once individual parcellations are computed, we obtain a group-
wise parcellation using partition ensemble algorithm. We access quality of the re-
sulting parcellations in three ways. (1) We use the continuous connectome frame-
work to compare parcellation-approximate and exact edge distribution functions.
(2) We compare perfomance of the resulting graphs on a gender classification
task. (3) We also show that without any explicit knowledge of brain geome-
try and based solely on graph connectivity we obtain comparatively symmetric
parcellations.
2 Methods
2.1 Continuous Connectome
The continuous connectome model (ConCon) treats each tract as an observation
of an inhomogeneous symmetric Poisson point process with the intensity function
given by
λ : Ω × Ω → R+,
(1)
where Ω denote union of two disjoint toplogically spherical brain hemispheres,
representing cortical white matter boundaries. In practice, ConCon uses cortical
mesh vertices as nodes of connectivity graph. From such a representation, a
"discrete" connectivity graph could be computed from any particular cortical
parcellation P . We follow definitions from [8] and call P = {Ei}N
i=1 a parcellation
of Ω if E1 . . . Ek ⊆ Ω such that ∪iEi = Ω, and N is the number of parcels (ROIs).
Edges between regions Ei and Ej can then be computed by integration of the
intensity function:
C(Ei, Ej) =
(2)
Due to properties of the Poisson Process, C(Ei, Ej) is the expectation of the
number of observed tracts between Ei and Ej. In the context of connectomics,
this is the expected edge strength.
λ(x, y)dxdy,
Ei,Ej
2.2 Graph Clustering
Once we obtain all individual continuous connectomes, we partition each inde-
pendently into a set of disjoint communities. For graph clustering we use the
Louvain modularity algorithm [1], as it has shown good results in multiple neu-
roimaging studies [9], [12], [5], [7]. This algorithm consist of two steps. The first
Connectivity-Driven Brain Parcellation via Consensus Clustering
3
step combines locally connected nodes into communities, while the second step
builds new meta graph. The nodes of the meta-graph are communities from the
previous step, and the edges are defined as the sum of all inter-community con-
nections of the new nodes. The algorithm in [1] cycles over these steps iteratively,
converging when further node clustering leads to no increase in modularity. We
Fig. 1. Adjacency matrix of a sample continuous connectome. Rows and columns are
reordered according to partition of the third hierarchical level. Boxes of different color
represents clusters of different hierarchical levels. P I clusters are obtained first, next
we reapply clustering on each detected P I cluster and obtain P II. This is repeated once
more to obtain P III
follow the hierarchical brain concept [7], repeating the clustering procedure iter-
atively. After the initial parcellation, we further cluster each individual parcel as
an independent graph. In this work, we repeat the process three times. For each
(i'th) continuous connectome this procedure yields a three-level hierarchically
embedded partition: P I
, (see Figure 1).
i
i , P II
i , P III
2.3 Consensus clustering
In order to obtain a unified parcellation for all subjects, we use consensus cluster-
ing. The concept was developed for aggregating multiple partitions of the same
data into a single partition. We define the average partition over all individual
partitions {Pi} as:
(cid:88)
¯P == argminP
d(P, Pi),
(3)
where ¯P is used to denoted desirable average partition, K is a number of averaged
partitions, d(Pi, Pj) is a distance measure between two partitions and we want
i
4
Anvar Kurmukov et al.
to minimize average distance from ¯P to all given partitions Pi. All partitions are
represented by a vector of length M , where M is a number of clustered objects
(vertices of a graph in our case). It contains values from 1 up to N , where N
is a number of clusters (parcels). This task is generally NP complete [14], but
there are many approximate algorithms. We use two approaches: Cluster-based
Similarity Partitioning Algorithm (cspa) [11] and greedy algorithm from [2].
CSPA defines a similarity between data points based on co-occurrence in a
same cluster across different partitions, and then partitions a graph induced by
this similarity. Specifically, given multiple partitions P1, . . . PK of a data points
x1, . . . xM . One can define similarity between points xi, xj as follow:
K(cid:88)
(cid:88)
k=1...N
S(xi, xj) =
δ(Pk(xi), Pk(xj)),
(4)
k=1
Here δ is Kroneker delta. Thus S(xi, xj) is just number of partitions in which
points xi and xj were in the same cluster. Next we build a graph, with nodes
correspond to data points and edge between node xi and xj is equal to S(xi, xj).
We the partition this graph into communities using some clustering algorithm
and the resulting partition is our clustering consensus partition.
Another way to find such average clustering is to optimize loss function given
by Equation 3.
The authors of [2] propose a greedy approach (Hard Ensemble - HE). Given
multiple partitions P1 . . . PK it combines them iteratively, first it finds average
of ¯P1,2 = min ¯P (d( ¯P , P1) + d( ¯P , P2)), next average of P1,2 and P3 and so on.
As a measure of distance the authors take the average square distance between
membership functions:
d(Pi, Pj) =
1
N
pk
i − pk
j2,
(5)
Exclusively for this definition we use another way to encode object's member-
ships: Pi is a matrix of size M × N (number of objects times number of clusters)
(cid:40)
P m,n
i
=
if m'th object belongs to n'th cluster
1
0 otherwise.
(6)
i and pk
In Equation 5 pk
j are kth rows of memberships matrices Pi and Pj respec-
tively. They correspond to membership vector of the kth object. Since we are
looking for disjoint clusters, only a single element of such row vector is equal to 1.
This representation is defined up to any column permutation π of matrix P , thus
the optimization procedure is done subject to all possible column permutations.
2.4 Comparison metrics
Once we find individual partitions and combine them into an average partition,
we want to access their quality. We use two different approaches.
Connectivity-Driven Brain Parcellation via Consensus Clustering
5
First, we compare representation strength of different parcellations by mea-
suring distance between original λ(x, y) and its piece-wise approximation given
by:
1
EiEjC(Ei, Ej),
γ(x, y) =
(7)
where x ∈ Ei and y ∈ Ej. Natural way to compare two statistical distributions
is to measure distance between their probability density functions, we will use
Kullback-Leibler divergence [4]. For two probability distributions with densities
λ(x) and γ(x) the KL divergence is:
KL(λ, γ) =
λ(x) log
−∞
λ(x)
γ(x)
dx,
(8)
It takes values close to 0 if two distributions are equal almost everywhere. Similar
but symmetrized version of KL divergence is Jensen-Shannon divergence [6].
Again for two probability distributions with densities λ(x) and γ(x) it is given
by:
JS(λ, γ) =
(KL(λ, r) + KL(γ, r)),
(9)
where r(x) = 1
2 (λ(x) + γ(x)).
Second, we compare performance of different parcellations on a gender clas-
sification task. We use Logistic Regression model with (small) l1 regularization
on a vectors of edge weights (the upper triangle of adjacency matrix excluding
diagonal). Classification perfomance is measured in terms of ROC AUC score,
which is typical for binary classification tasks.
Finally, in order to quantify goodness of consensus clustering and access
hemisphere symmetry we use Adjusted Mutual Information [15]. It measures
similarity between two partitions, with value 1 corresponds to identical partitions
and values close to zero for partitions that are very different.
Using AMI we access ensemble goodness (how good clustering ensemble al-
gorithm combines multiple partitions) using modified 3:
Ensemble goodness =
AMI( ¯P , Pi),
(10)
We compute parcellation symmetry by comparing hemisphere parcels (labels):
Symmetry = AMI( ¯PLH, ¯PRH).
(11)
i
3 Experiments
3.1 Data description
We use construct continuous connetocmes of 400 subjects from the Human Con-
nectome Project S900 release [13] following [8]. We use an icosahedral spehrical
∞
1
2
K(cid:88)
6
Anvar Kurmukov et al.
sampling, at a resolution of 10242 mesh vertices per hemisphere. We used Dipy's
implementation of constrained spherical deconvolution (CSD) to perform prob-
abilistic tractography. Prior to clustering, we exclude all mesh vertices that were
labeled by FreeSurfer as corpus callosum or cerebellum.
3.2 Experimental pipeline
Our experiments are summarized as follows:
1. For each subject we reconstruct its Continuous Connectome.
2. For each Continuous Connectome we iteratively run Louvain clustering algo-
rithm, as described above. Subgraphs of having less then 1 percent of original
graph vertices were not divided.
3. Next we aggregate individual subject partitions and obtain consensus clus-
tering. Aggregation was done over 400 HCP subjects. Further, after finding
the optimal parcellation, we obtain two parcellations based on two disjoint
sets of 200 HCP subjects in order to compute reproducibility.
4. We aggregate partitions of the same level (I-II-III) using CSPA and HE.
5. We compare obtained partitions between themselves and with FreeSurfer's
Desikan-Killiani parcellation using Kullback-Leibler and Jensen-Shannon di-
vergence. We compute goodness of an ensemble and parcellation symmetry
using AMI.
6. We compare performance of simplified connectomes on a binary classification
task using Logistic Regression with l1 penalty. Classification results are mea-
sured in terms of ROC AUC score, with averaging over 10 cross-validation
folds.
3.3 Results
Table 1 represent all comparison results. First we can see that CSPA algorithm
failed to find good clustering ensemble which result in poor classification per-
formance and high KL and JS divergences. Greedy algorithm performed on P III
on the other hand outperforms standard Desikan atlas accross all comparison
metrics (except number of parcels, 68 versus 83). Surprisingly, greedy ensemble
of second level partition (P II) performs comparatively with Desikan , despite
having twice as lower number of parcels (30 versus 68).
Another interesting property that we get automatically is parcellation sym-
metry. Our clustering algorithm known nothing about brain topology (all in-
formation was contained in graph connectivity), still reconstruct parcellations
which are highly symmetrical. For standard Desikan atlas hemisphere symme-
try is 0.64, and for our best parcellation this value even higher (0.66), and still
remains quite high for second level partition (0.55).
Finally we check if our best ensemble parcellation, which combines 400 indi-
vidual partitions is stable. We split 400 subjects into 2 groups of 200 subjects
and independently combine their partitions. We compare resulting parcellations:
Connectivity-Driven Brain Parcellation via Consensus Clustering
7
cspa P I
cspa P II
cspa P III
HE P I
HE P II
HE P III
DKT
KL
JS
1.22 ± .07 1.18 ± .07
.20 ± .00
.19 ± .00
1.15 ± .07
.19 ± .00
1.16 ± .07 .86 ± .05
.20 ± .00
.17 ± .00
.66 ± .04 .83 ± .05
.14 ± .00 .16 ± .00
Gender
Classification
.63 ± .04
.64 ± .04
.69 ± .03
.64 ± .03
.75 ± .03
.86 ± .02 .81 ± .03
Hemisphere
symmetry
Ensemble
goodness
Number of
ROIs
.15
.24
.32
.26
.55
.66
.47 ± .06
.40 ± .02
.35 ± .00
.53 ± .05
.64 ± .02
.70 ± .01
.64
−
5
7
8
7
30
83
68
Table 1. All results are rounded to 2 significant digits. Where it possible results are
reported with standard deviation. Best result in each row is colored. KL, JS divergences,
lower is better; binary Gender Classification was measured in terms of ROC AUC
score, higher - better; Ensemble goodness and Hemisphere symmetry were measured
using AMI, Ensemble goodness is an average AMI between consensus partition and all
individual partitions, higher - better.
¯P1,200 and ¯P201,400 between themselves and with original ¯P (which is an ensemble
of all 400 subjects) again using Adjusted Mutual Information. Both ¯P1,200 and
¯P201,400 shows AMI value greater than 0.80 (0.83 and 0.82 respectively) when
compare with ¯P , they also highly similar between themselves.
4 Conclusion
We have presented an approach for generating unified connectivity-based human
brain atlases bases on consensus clustering. The method is based on finding a
pseudo average over the set of individual partitions. Our approach outperforms
standard a anatomical parcellation on several important metrics, including agree-
ment with dense connectomes, improved relevance to biological data, and even
improved symmetry. Because our approach is entirely data driven an requires no
agreement between individual parcellation labels, it combines both the flexibility
of individual parcellations and the interpretability of simple unified atlases.
References
1. Blondel, Vincent D., et al. "Fast unfolding of communities in large networks." Jour-
nal of statistical mechanics: theory and experiment 2008.10 (2008): P10008.
8
Anvar Kurmukov et al.
Fig. 2. Left column: Desikan-Killiany parcellation. Right column: HE P III parcellation.
Lateral and Medial views, left hemisphere.
2. Dimitriadou, Evgenia, Andreas Weingessel, and Kurt Hornik. "A combination
scheme for fuzzy clustering." International Journal of Pattern Recognition and Ar-
tificial Intelligence 16.07 (2002): 901-912.
3. Hubert, Lawrence, and Phipps Arabie. "Comparing partitions." Journal of classifi-
cation 2.1 (1985): 193-218.
4. Kullback, Solomon, and Richard A. Leibler. "On information and sufficiency." The
annals of mathematical statistics 22.1 (1951): 79-86.
5. Kurmukov, Anvar, et al. "Classifying Phenotypes Based on the Community Struc-
ture of Human Brain Networks." Graphs in Biomedical Image Analysis, Computa-
tional Anatomy and Imaging Genetics. Springer, Cham, 2017. 3-11.
6. Lin, J. (1991) Divergence measures based on Shannon entropy. IEEE Transactions
on Information Theory, 37, 14, 145 -- 151.
7. Meunier, David, Renaud Lambiotte, and Edward T. Bullmore. "Modular and hi-
erarchically modular organization of brain networks." Frontiers in neuroscience 4
(2010): 200.
8. Moyer, Daniel, et al. "Continuous representations of brain connectivity using spatial
point processes." Medical image analysis 41 (2017): 32-39.
9. Nicolini, Carlo, Ccile Bordier, and Angelo Bifone. "Community detection in
weighted brain connectivity networks beyond the resolution limit." Neuroimage 146
(2017): 28-39.
10. Parisot, Sarah, et al. "GraMPa: Graph-based multi-modal parcellation of the cortex
using fusion moves." International Conference on Medical Image Computing and
Computer-Assisted Intervention. Springer, Cham, 2016.
Connectivity-Driven Brain Parcellation via Consensus Clustering
9
11. Strehl, Alexander, and Joydeep Ghosh. "Cluster ensembles -- a knowledge reuse
framework for combining multiple partitions." Journal of machine learning research
3.Dec (2002): 583-617.
12. Taylor, Peter N., Yujiang Wang, and Marcus Kaiser. "Within brain area trac-
tography suggests local modularity using high resolution connectomics." Scientific
reports 7 (2017): 39859.
13. Van Essen, David C., et al. "The WU-Minn human connectome project: an
overview." Neuroimage 80 (2013): 62-79.
14. Vega-Pons, Sandro, and Jos Ruiz-Shulcloper. "A survey of clustering ensemble
algorithms." International Journal of Pattern Recognition and Artificial Intelligence
25.03 (2011): 337-372.
15. Vinh, Nguyen Xuan, Julien Epps, and James Bailey. "Information theoretic mea-
sures for clusterings comparison: Variants, properties, normalization and correction
for chance." Journal of Machine Learning Research 11.Oct (2010): 2837-2854.
16. Petrov, D., Ivanov, A., and Faskowitz, J., Gutman, B., Moyer, D., Villalon, J., Ja-
hanshad, N., Thompson, P., "Evaluating 35 Methods to Generate Structural Con-
nectomes Using Pairwise Classification", ArXiv e-prints, eprint = 1706.06031, June
(2017)
17. Salim Arslan, Sofia Ira Ktena, Antonios Makropoulos, Emma C. Robinson, Daniel
Rueckert, Sarah Parisot, Human brain mapping: A systematic comparison of par-
cellation methods for the human cerebral cortex, NeuroImage, (170) 5-30, (2018)
|
1606.02344 | 1 | 1606 | 2016-06-07T21:59:31 | Resting state brain networks from EEG: Hidden Markov states vs. classical microstates | [
"q-bio.NC",
"stat.ML"
] | Functional brain networks exhibit dynamics on the sub-second temporal scale and are often assumed to embody the physiological substrate of cognitive processes. Here we analyse the temporal and spatial dynamics of these states, as measured by EEG, with a hidden Markov model and compare this approach to classical EEG microstate analysis. We find dominating state lifetimes of 100--150\,ms for both approaches. The state topographies show obvious similarities. However, they also feature distinct spatial and especially temporal properties. These differences may carry physiological meaningful information originating from patterns in the data that the HMM is able to integrate while the microstate analysis is not. This hypothesis is supported by a consistently high pairwise correlation of the temporal evolution of EEG microstates which is not observed for the HMM states and which seems unlikely to be a good description of the underlying physiology. However, further investigation is required to determine the robustness and the functional and clinical relevance of EEG HMM states in comparison to EEG microstates. | q-bio.NC | q-bio |
Resting state brain networks from EEG:
Hidden Markov states vs. classical microstates
Tammo Rukat1, Adam Baker2, Andrew Quinn2, and Mark Woolrich2
1 Department of Statistics, University of Oxford, Oxford, United Kingdom
2 Oxford Centre for Human Brain Activity, Oxford, United Kingdom
[email protected]
Abstract. Functional brain networks exhibit dynamics on the sub-second tempo-
ral scale and are often assumed to embody the physiological substrate of cognitive
processes. Here we analyse the temporal and spatial dynamics of these states, as mea-
sured by EEG, with a hidden Markov model and compare this approach to classical
EEG microstate analysis. We find dominating state lifetimes of 100–150 ms for both
approaches. The state topographies show obvious similarities. However, they also fea-
ture distinct spatial and especially temporal properties. These differences may carry
physiological meaningful information originating from patterns in the data that the
HMM is able to integrate while the microstate analysis is not. This hypothesis is
supported by a consistently high pairwise correlation of the temporal evolution of
EEG microstates which is not observed for the HMM states and which seems unlikely
to be a good description of the underlying physiology. However, further investigation
is required to determine the robustness and the functional and clinical relevance of
EEG HMM states in comparison to EEG microstates.
1 Introduction
Temporal correlations in the spontaneous oscillatory activity of spatially distinct neuronal as-
semblies are a well established phenomenon described as resting state brain networks (RSNs).
RSNs exhibit functional [1] and clinical [2, 3, 4, 5] significance. They have first been iden-
tified based on blood-oxygen levels measured through functional MRI [6, 7]. While fMRI is
limited in its temporal resolution and captures only slow oscillations with frequencies below
0.1 Hz, it features a high spatial resolution down to 1 mm. In contrast, electroencephalogra-
phy (EEG) and magnetoencephalography (MEG) are techniques that provide a more direct
measure of the electrical activity in the brain [8, 9]. EEG measures the difference in electric
potentials on the scalp and captures high frequency oscillations on the millisecond timescale
that is most relevant for the characterisation of cognitive processes. It therefore is a suitable
tool to characterise the electrophysiological basis of RSNs [10, 11]. Notably, the same resting
state patterns can be observed across the different time scales of fMRI and M/EEG, which is
made conceivable by the dynamics of brain states being scale free across the relevant regime
[12, 13, 14].
Here, we apply a hidden Markov Model (HMM) [15, 16] to the power envelope of the
EEG signal in sensor space, in order to identify quasi-stable networks of correlated activation
that the signal is likely to have emerged from. This model has previously been applied to
resting state MEG power envelopes and identified brain states that appear highly similar
to known RSNs [17], featuring rapid fluctuations with state lifetimes of 100–200 ms. The
aim of this study is to investigate the HMM's potential as an alternative or complementary
method to classical EEG microstate analysis and to compare their spatial and temporal
characteristics. In classical microstate analysis, the EEG signal is thought of as a sequence
of a limited number of quasi-stable EEG topographies, each defining a microstate [18, 19].
2
These are inferred based on the EEG topographies at local maxima of the global field power
(GFP), which is given by the sum squared difference between all electrode potentials Vi and
the mean potential V : GFP(t) = ( 1
n
(cid:80)
i Vi(t) − V (t)) 1
2 .
Topographies at the GFP maxima are of particular interest because they feature the
highest signal-to-noise ratio [20]. The selected topographies are subject to a clustering pro-
cedure that aims to determine a fixed number of states, reflecting typical topographies.
Traditionally this is achieved by an iterative procedure [21], while K-means clustering or
more sophisticated hierarchical clustering methods [8, 22] have become the current stan-
dard. Classical microstate analysis often limits itself to 4 clusters that have repeatedly been
observed to explain most variance in the data. Time courses can then be derived under the
assumption that the switching between mutually exclusive microstates happens only at GFP
peaks. This procedure yields mean state durations of around 100 ms [23]. Microstates have a
variety of clinical applications, e.g. in Schizophrenia [24] and Alzheimer's [25]. where dura-
tion and switching patterns between the four microstates are connected to the disease state.
However, the extent to which they reflect topographies of physiological activation remains
unclear.
2 Data acquisition, preprocessing and classical microstate analysis
For this study we recorded two times ten minutes resting state EEG data in 6 healthy
subjects. Upon identification based on the cardiac signal, the eye blink signal and the signal's
kurtosis and frequency spectrum, the data is manually cleaned from apparent artefacts.
Subsequently it is decomposed into 150 independent components and band-pass filtered
into the 1–40 Hz band. EEG Microstates are inferred, based on global field power (GFP)
time course, smoothed with a Gaussian kernel with a width of 10 time steps and a standard
deviation of 5 time steps. GFP peaks are considered local maxima if all 10 surrounding values
are smaller. Upon identification, the peak topographies are subject to k-means clustering
with a fixed number of clusters, where the objective function is the within sample correlation
for each cluster. The mean distance for each topography from its assigned centroid under
variation of the number of states does not immediately suggesting a certain number of
clusters that is particularly well supported by the data (not shown). This procedure yields
microstates that are shown in Fig. 1 (1) and that appear similar to those found in the
literature [26].
(a)
(b)
(c)
(d)
(a)
(b)
(c)
(d)
(1) Classical EEG microstates
(2) Hidden Markov states
Fig. 1: Quasi-stable spatial EEG topographies – Based on resting state EEG measure-
ment for 2x10 minutes in 6 subjects. The red-blue colour coding shows opposite potentials.
As opposed to the microstates, the range in potentials differs among HMM states. They are
separately normalised to facilitate the comparison between modalities.
3
3 A hidden Markov model for EEG and MEG topographies
In contrast to classical microstate analysis, the hidden Markov model, as proposed here,
is a generative model that describes the observations that emerge from the rapid switching
between quasi-stable topographies with a Gaussian observation model. It promises to be able
to capture temporal and spatial dynamics that are more closely related to the underlying
brain activity than classical microstate analysis and has been successfully applied to the
analysis of MEG RSNs [17]. We now briefly outline the model of which a detailed account
is given elsewhere [27].
3.1 Model derivation
At any given time t the system is in a state k out of fixed number, K, of states, denoted
st. Each of these states is associated with a Gaussian observation model that describes the
mean and covariance for every data point. With yt denoting the vector of observation at
time t we therefore write:
P (ytst = k, µk, Σk) ∼ N (µk, Σk).
(1)
The transition probability between states is Markovian, such that
P (st = kst−1 = k(cid:48), st−2 = k(cid:48)(cid:48), . . .) = P (st = kst−1 = k(cid:48)) = πk,k(cid:48)
(2)
where the transition probabilities from state k to k(cid:48) are described by KxK matrix π. The
initial probability to be in state k is given by π0. The full posterior likelihood is given by:
(cid:89)
P (y, s, π0, π, µk, Σk) =
P (ytst, µk, Σk)P (stst−1, π)P (πt)P (π0)P (µ, Σ)
(3)
t
Choosing conjugate distributions for the priors, P (πt), P (π0), and P (µ, Σ) facilitates the
approximation of the posterior distribution by means of variational Bayes inference [15]. To
this end, the posterior distribution is approximated to factorise, such that
P (y, s, π0, π, µk, Σk) ≈ P (y)P (s)P (π0, π)P (µ, σ) =: Q.
(4)
Q is then determined by minimising the variational free energy [28] between the true pos-
terior and this approximation. Up to an additive constant this free energy equals the KL
divergence between the two distributions.
3.2 Application to EEG envelopes
The number of states that is supported by the EEG data is investigated by plotting the free
energy as a function of the number of states, as shown in Fig. 2. The free energy decreases
steadily for larger numbers of states, which is in agreement with earlier observations by
Baker and colleagues [17] on resting state MEG data. However, we base our choice of the
number of states onto comparability with classical EEG microstate analysis (4 states).
The HMM is applied to the 20 first principal components of the EEG power envelopes.
State topographies are derived with a general linear model with the inferred HMM state
time course as design matrix and the EEG sensor space power as response. The resulting
coefficients are maps of partial correlations, shown in Fig. 1 (2).
4
·106
y
g
r
e
n
e
e
e
r
F
8
7.5
3
4
5
6
7
8
9
10
11
12
Number of hidden states
Fig. 2: Free energy of the variational approximation – Variational inference is repeated
10 times for every fixed number of Markov states.
4 Comparison of EEG microstates and Markov states
EEG microstates and EEG HMM states show some spatial similarities (Fig. 1). The activa-
tion is mostly limited to one specific region of sensor space and both microstate and HMM
state topographies can be very broadly classified as right lateral, left lateral, anterior/central,
and posterior. The lateral microstates expand more into anterior areas, while the correspond-
ing HMM states are laterally confined. Notably, the absolute range of potential differences
differs between HMM states and is virtually identical between microstates.
We further compare the temporal properties of both sets of states. Similarly to the
HMM state analysis, microstate time courses are obtained as partial correlation between
each microstate topography and the EEG power envelope time course. The most probable
state st at every time point is derived by the Viterbi algorithm [15]. It facilitates an estimate
of the overall fractional occupancy of each state, which is similar between states and models.
The difference in the time spent in two corresponding states is 15% or less.
Pairwise correlations of the full time courses are shown in Fig. 3 (a). The corresponding
spatial correlations are shown in Fig. 3 (b). Overall, they exhibit a very similar pattern.
There are however clear difference, as for instance a strong spatial correlation between mi-
crostate (b) and HMM state (b), which is not reflect in the temporal correlation. Conversely
microstate (d) and HMM state (c) feature a moderately positive time course correlation,
while the spatial patterns are negatively correlated. Since the time courses for the two modal-
ities are partial correlations of the topographies with the identical envelope time courses it
seems unsurprising that temporal correlations are in close agreement with the spatial overlap
of the topographies. However, it is clear that the temporal evolution features information
complimentary to the 2D topographies.
To investigate the time scale of the inferred dynamics of state switching as supported by
the envelope data we correlate low pass filtered versions of each state time course with the
envelope fluctuation of a representative EEG sensor. We selected the sensor that had the
highest correlation with the unfiltered state time course and repeated this analysis for every
microstate and every HMM state varying the width of the filter. The results are shown in
Fig. 4 and consistently exhibit maxima in the correlation at window width of 100–150 ms.
This agrees with previous observations for EEG microstates [29]. None of the modalities
shows a steadily higher correlation than the other. Interestingly, we observe a very different
correlations between the microstate time courses as shown in comparison to the HMM
time courses in Fig. 5. The microstate time courses exhibit pairwise correlations that are
consistently larger than 0.6. Conversely the HMM state time courses show correlations down
to -0.5.
5
5 Discussion and conclusion
We identified states of quasi-stable topographies in resting state EEG by means of clas-
sical microstate analysis and proposed an alternative approach based on a hidden Markov
model with a Gaussian observation model that was tractable for approximate inference using
Variational Bayes [16]. The microstate analysis identified topographies that are similar to
known microstates [19]. While the HMM state topographies are less confined, they feature
similar activation patterns and can partly be matched to corresponding microstates. This
matching was shown to be reflected in the temporal evolution of the state time courses.
Notably the absolute values of these correlations reach their maxima at about 0.4, pointing
to a dissimilarity that may correspond to a loss/gain of meaningful information in one of
the methods. The EEG microstate time courses show strong pairwise temporal correlations,
which is not observed for EEG HMM states. While a temporal (and spatial) overlap between
RSNs is entirely possible, networks of different function should also be temporally distinct
[30]. Thus, a high positive correlation between the dynamics of all given states is unlikely
to be a good description of underlying physiology. We find dominating state lifetimes of
100–150 ms, which is consistent with earlier findings [29, 17].
The free energy, a measure for the HMM model fitness, decreases steadily for an increase
in the number of states. A possible explanation for the absence of an optimal number of states
within the investigated range is the following. For a higher number of states, subject specific
activations are introduced in addition to the desirable patterns that are present across
subjects. We frequently observe such subject specific states when increasing the number
of states (not shown here). While this was partly amendable by demeaning and variance
normalising the subject-wise power envelopes, the reason for this behaviour is likely the
actually distinct covariance structure between subjects. To make the HMM analysis robust
and reliable, this issue should be addressed in future work. More generally, a nonparametric
model could automatically infer the optimal number of states.
Further work should include the investigation of the scaling behaviour of HMM states
for both EEG and MEG measurements to ascertain whether they exhibit the long-range
dependencies (LRDs) that were found for EEG microstates [14] and that are hypothesised to
be necessary for the efficient execution of cognitive processes [31, 32, 33]. Recently Gschwind
(a) Correlation structure of microstate and
HMM state time courses
(b) Correlation structure of microstate and
HMM state spatial maps
Fig. 3: Spatial and temporal correlation structure of microstates and HMM state
6
REFERENCES
Fig. 4: Time scale analysis of within-topography fluctuations – The fractional occu-
pancy time window dependency of the correlation between the microstate and HMM state
time course and the envelope of the most representative EEG sensor is shown and exhibits
consistent maxima around 100–150 ms
(a) HMM state time course correlation struc-
ture
(b) Microstate time course correlation
structure
Fig. 5: Correlation structure of the state time courses for the different modalities
– A strong positive correlation between all EEG microstate time courses is visible, while
different HMM state time courses are approximately uncorrelated or negatively correlated.
et. al [34] criticised the lack of an explicit representation of LRDs in the context of different
model [35]. The HMM does not explicitly model LRDs and therefore the same criticism
applies. However, even without explicitly modelling them, we expect the HMM to capture
LRDs, if they are present in the data. Nevertheless, a model that takes explicit account of
LRDs could constitute a useful expansion of the HMM.
Overall our results suggest, that EEG HMM states could serve as a model-based, physio-
logically motivated alternative to classical EEG microstates. However, further work remains
to be done to substantiate this proposition and to better understand the relationship between
between the two approaches.
References
[1] Bel´en Guerra-Carrillo, Allyson P. Mackey, and Silvia A. Bunge. "Resting-state fMRI:
a window into human brain plasticity." eng. In: Neuroscientist 20.5 (2014), pp. 522–
REFERENCES
7
533. doi: 10 . 1177 / 1073858414524442. url: http : / / dx . doi . org / 10 . 1177 /
1073858414524442.
[2] Nicola Filippini et al. "Distinct patterns of brain activity in young carriers of the
APOE-epsilon4 allele." eng. In: Proc Natl Acad Sci U S A 106.17 (2009), pp. 7209–
7214. doi: 10.1073/pnas.0811879106. url: http://dx.doi.org/10.1073/pnas.
0811879106.
[3] Maria Centeno and David W. Carmichael. "Network Connectivity in Epilepsy: Resting
State fMRI and EEG-fMRI Contributions." eng. In: Front Neurol 5 (2014), p. 93. doi:
10.3389/fneur.2014.00093. url: http://dx.doi.org/10.3389/fneur.2014.
00093.
[4] Stefan Lang, Niall Duncan, and Georg Northoff. "Resting-state functional magnetic
resonance imaging: review of neurosurgical applications." eng. In: Neurosurgery 74.5
(2014), 453–64; discussion 464–5. doi: 10.1227/NEU.0000000000000307. url: http:
//dx.doi.org/10.1227/NEU.0000000000000307.
[5] Smadar Ovadia-Caro, Daniel S. Margulies, and Arno Villringer. "The value of resting-
state functional magnetic resonance imaging in stroke." eng. In: Stroke 45.9 (2014),
pp. 2818–2824. doi: 10.1161/STROKEAHA.114.003689. url: http://dx.doi.org/
10.1161/STROKEAHA.114.003689.
[6] Christian F Beckmann et al. "Investigations into resting-state connectivity using inde-
pendent component analysis". In: Philosophical Transactions of the Royal Society of
London B: Biological Sciences 360.1457 (2005), pp. 1001–1013. issn: 0962-8436. doi:
10.1098/rstb.2005.1634.
[7] J. S. Damoiseaux et al. "Consistent resting-state networks across healthy subjects".
In: Proceedings of the National Academy of Sciences 103.37 (2006), pp. 13848–13853.
doi: 10.1073/pnas.0601417103. url: http://www.pnas.org/content/103/37/
13848.abstract.
[8] Michel. Electrical Neuroimaging (Cambridge Medicine). July 2009. url: http : / /
amazon.com/o/ASIN/B002SEKZ4M/.
[9] Malcolm Proudfoot et al. "Magnetoencephalography." eng. In: Pract Neurol 14.5 (Oct.
2014), pp. 336–343. doi: 10.1136/practneurol-2013-000768.
[10] Biyu J. He et al. "Electrophysiological correlates of the brain's intrinsic large-scale
functional architecture". In: Proceedings of the National Academy of Sciences 105.41
(2008), pp. 16039–16044. doi: 10.1073/pnas.0807010105. url: http://www.pnas.
org/content/105/41/16039.abstract.
[11] Zhongming Liu et al. "Large-scale spontaneous fluctuations and correlations in brain
electrical activity observed with magnetoencephalography". In: NeuroImage 51.1 (2010),
pp. 102 –111. issn: 1053-8119. doi: http://dx.doi.org/10.1016/j.neuroimage.
2010 . 01 . 092. url: http : / / www . sciencedirect . com / science / article / pii /
S1053811910001151.
[12] K. Linkenkaer-Hansen et al. "Long-range temporal correlations and scaling behavior
in human brain oscillations." eng. In: J Neurosci 21.4 (2001), pp. 1370–1377.
[13] Manfred G. Kitzbichler et al. "Broadband criticality of human brain network syn-
chronization." eng. In: PLoS Comput Biol 5.3 (Mar. 2009), e1000314. doi: 10.1371/
journal.pcbi.1000314.
[14] dimitri van de ville, juliane britz, and christoph m. michel. "eeg microstate sequences
in healthy humans at rest reveal scale-free dynamics." eng. In: proc natl acad sci u
s a 107.42 (2010), pp. 18179–18184. doi: 10.1073/pnas.1007841107. url: http:
//dx.doi.org/10.1073/pnas.1007841107.
I. Rezek and S. J. Roberts. "Ensemble hidden Markov models for biosignal analysis".
In: International Conference on Digital Signal Processing, 2002. Vol. 1. 2002. doi:
[15]
8
REFERENCES
10.1109/ICDSP.2002.1027907. url: http://ieeexplore.ieee.org/stamp/stamp.
jsp?arnumber=1027907.
[16] diego vidaurre et al. "spectrally resolved fast transient brain states in electrophysio-
logical data". in preparation.
[17] Adam P. Baker et al. "Fast transient networks in spontaneous human brain activity."
In: Elife 3 (2014). doi: 10.7554/eLife.01867.
[18] T. Koenig et al. "A deviant EEG brain microstate in acute, neuroleptic-naive schizophren-
ics at rest". English. In: European Archives of Psychiatry and Clinical Neuroscience
249.4 (1999). issn: 0940-1334. doi: 10.1007/s004060050088. url: http://dx.doi.
org/10.1007/s004060050088.
[19] Arjun Khanna et al. "Microstates in resting-state EEG: current status and future
directions." eng. In: Neurosci Biobehav Rev 49 (2015), pp. 105–113. doi: 10.1016/j.
neubiorev.2014.12.010. url: http://dx.doi.org/10.1016/j.neubiorev.2014.
12.010.
[20] T Koenig et al. "Brain connectivity at different time-scales measured with EEG".
In: Philosophical Transactions of the Royal Society of London B: Biological Sciences
360.1457 (2005), pp. 1015–1024. issn: 0962-8436. doi: 10.1098/rstb.2005.1649.
[21] R. D. Pascual-Marqui, C. M. Michel, and D. Lehmann. "Segmentation of brain electri-
cal activity into microstates: model estimation and validation." eng. In: IEEE Trans
Biomed Eng 42.7 (1995), pp. 658–665. doi: 10.1109/10.391164. url: http://dx.
doi.org/10.1109/10.391164.
[22] Robert Tibshirani and Guenther Walther. "Cluster Validation by Prediction Strength".
In: Journal of Computational and Graphical Statistics 14.3 (2005), pp. 511–528. doi:
10.1198/106186005X59243. url: http://dx.doi.org/10.1198/106186005X59243.
[23] Verena Brodbeck et al. "EEG microstates of wakefulness and NREM sleep." eng. In:
Neuroimage 62.3 (2012), pp. 2129–2139. doi: 10.1016/j.neuroimage.2012.05.060.
url: http://dx.doi.org/10.1016/j.neuroimage.2012.05.060.
[24] Dietrich Lehmann et al. "EEG microstate duration and syntax in acute, medication-
naive, first-episode schizophrenia: a multi-center study." eng. In: Psychiatry Res 138.2
(2005), pp. 141–156. doi: 10 . 1016 / j . pscychresns . 2004 . 05 . 007. url: http :
//dx.doi.org/10.1016/j.pscychresns.2004.05.007.
[25] Keiichiro Nishida et al. "EEG microstates associated with salience and frontoparietal
networks in frontotemporal dementia, schizophrenia and Alzheimer's disease." eng. In:
Clin Neurophysiol 124.6 (2013), pp. 1106–1114. doi: 10.1016/j.clinph.2013.01.
005. url: http://dx.doi.org/10.1016/j.clinph.2013.01.005.
[26] Thomas Koenig et al. "Millisecond by millisecond, year by year: normative EEG mi-
crostates and developmental stages." eng. In: Neuroimage 16.1 (May 2002), pp. 41–48.
doi: 10.1006/nimg.2002.1070.
[27] diego vidaurre et al. "spectrally resolved fast transient brain states in electrophysio-
logical data". In: in preperation ().
[28] Christopher M. Bishop. Pattern Recognition and Machine Learning. 1st Edition. Springer,
2013. isbn: 9788132209065. url: http://amazon.com/o/ASIN/8132209060/.
[29] D. Lehmann, H. Ozaki, and I. Pal. "EEG alpha map series: brain micro-states by
space-oriented adaptive segmentation." eng. In: Electroencephalogr Clin Neurophysiol
67.3 (1987), pp. 271–288.
[30] Stephen M. Smith et al. "Temporally-independent functional modes of spontaneous
brain activity". In: Proceedings of the National Academy of Sciences 109.8 (2012),
pp. 3131–3136. doi: 10.1073/pnas.1121329109.
[31] Dante R. Chialvo. "Critical brain networks". In: Physica A 340 (2008), :756(2004).
REFERENCES
9
[32] Ariel Haimovici et al. "Brain Organization into Resting State Networks Emerges at
Criticality on a Model of the Human Connectome". In: Phys. Rev. Lett. 110 (17 Apr.
2013), p. 178101. doi: 10.1103/PhysRevLett.110.178101.
[33] Enzo Tagliazucchi et al. "Criticality in large-scale brain fMRI dynamics unveiled by a
novel point process analysis." In: Frontiers in Physiology 3.15 (2012). issn: 1664-042X.
doi: 10.3389/fphys.2012.00015.
[34] Markus Gschwind, Christoph M. Michel, and Dimitri Van De Ville. "Long-range de-
pendencies make the difference-Comment on "A stochastic model for EEG microstate
sequence analysis"." eng. In: Neuroimage 117 (2015), pp. 449–455. doi: 10.1016/j.
neuroimage.2015.05.062. url: http://dx.doi.org/ 10.1016/j.neuroimage.
2015.05.062.
[35] Matthias Gartner et al. "A stochastic model for EEG microstate sequence analysis."
eng. In: Neuroimage 104 (2015), pp. 199–208. doi: 10.1016/j.neuroimage.2014.
10.014. url: http://dx.doi.org/10.1016/j.neuroimage.2014.10.014.
|
1702.03409 | 1 | 1702 | 2017-02-11T10:52:36 | Disruptive Behavior Disorder (DBD) Rating Scale for Georgian Population | [
"q-bio.NC",
"stat.AP"
] | In the presented study Parent/Teacher Disruptive Behavior Disorder (DBD) rating scale based on the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV-TR [APA, 2000]) which was developed by Pelham and his colleagues (Pelham et al., 1992) was translated and adopted for assessment of childhood behavioral abnormalities, especially ADHD, ODD and CD in Georgian children and adolescents. The DBD rating scale was translated into Georgian language using back translation technique by English language philologists and checked and corrected by qualified psychologists and psychiatrist of Georgia. Children and adolescents in the age range of 6 to 16 years (N 290; Mean Age 10.50, SD=2.88) including 153 males (Mean Age 10.42, SD= 2.62) and 141 females (Mean Age 10.60, SD=3.14) were recruited from different public schools of Tbilisi and the Neurology Department of the Pediatric Clinic of the Tbilisi State Medical University. Participants objectively were assessed via interviewing parents/teachers and qualified psychologists in three different settings including school, home and clinic. In terms of DBD total scores revealed statistically significant differences between healthy controls (M=27.71, SD=17.26) and children and adolescents with ADHD (M=61.51, SD= 22.79). Statistically significant differences were found for inattentive subtype between control (M=8.68, SD=5.68) and ADHD (M=18.15, SD=6.57) groups. In general it was shown that children and adolescents with ADHD had high score on DBD in comparison to typically developed persons. In the study also was determined gender wise prevalence in children and adolescents with ADHD, ODD and CD. The research revealed prevalence of males in comparison with females in all investigated categories. | q-bio.NC | q-bio | Disruptive Behavior Disorder (DBD) Rating Scale for
Georgian Population
Vera Bzhalava1, Ketevan Inasaridze2
1First Called Georgian University of the Patriarchate of Georgia
2Georgian State University of Physical Education and Sport
Abstract
In the presented study Parent/Teacher Disruptive Behavior Disorder (DBD) rating scale based on
the Diagnostic and Statistical Manual of Mental Disorders (DSM-IV-TR [APA, 2000]) which was
developed by Pelham and his colleagues (Pelham et al., 1992) was translated and adopted for
assessment of childhood behavioral abnormalities, especially ADHD, ODD and CD in Georgian
children and adolescents. The DBD rating scale was translated into Georgian language using back
translation technique by English language philologists and checked and corrected by qualified
psychologists and psychiatrist of Georgia. Children and adolescents in the age range of 6 to 16
years (N 290; Mean Age 10.50, SD=2.88) including 153 males (Mean Age 10.42, SD= 2.62) and
141 females (Mean Age 10.60, SD=3.14) were recruited from different public schools of Tbilisi and
the Neurology Department of the Pediatric Clinic of the Tbilisi State Medical University.
Participants objectively were assessed via interviewing parents/teachers and qualified psychologists
in three different settings including school, home and clinic. In terms of DBD total scores revealed
statistically significant differences between healthy controls (M=27.71, SD=17.26) and children and
adolescents with ADHD (M=61.51, SD= 22.79). Statistically significant differences were found for
inattentive subtype between control (M=8.68, SD=5.68) and ADHD (M=18.15, SD=6.57) groups, for
hyperactive/impulsive subtype ADHD group (M=17.39, SD=6.74) and healthy controls (M=8.12,
SD=6.07), for oppositional defiant disorder in healthy control (M=6.68, SD=4.59) and persons with
ADHD (M=13.7, SD=5.82), for conduct disorder between healthy group (M=2.51, SD=2.99) and group
with ADHD (M=8.31, SD=6.14). In general it was shown that children and adolescents with ADHD had
high score on DBD in comparison to typically developed persons. In the study also was determined
gender wise prevalence in children and adolescents with ADHD, ODD and CD. The research
revealed prevalence of males in comparison with females in all investigated categories.
1
This research was supported by a grant N 021-08 from the Rustaveli National Science Foundation.
Correspondence concerning this article should be addressed to Ketevan Inasaridze
E-mail: [email protected]
The Disruptive Behavior Disorders (DBDs) especially oppositional defiant disorder (ODD) and
conduct disorder (CD) are frequently co-occurring psychiatric disorders in approximately half of
children and adolescents with attention deficit hyperactivity disorder (ADHD) (Waschbusch 2002;
Connor et al. 2010, Anderson and Kiehl, 2013; Masi., 2015). The results of many researches
showed that the development of DBDs are associated with negative interaction of psychosocial and
neurobiological factors (Lavigne et al., 2016; Kerekes et al., 2014; Boden et al., 2010).
ADHD is one of the most common neurodevelopmental disorder which is characterized by
developmentally inappropriate levels of innattention, locomotor hyperactivity and impulsivity
(Faraone et al., 2015). According to the DSM-IV-TR three subtypes of ADHD are defined:
ADHD combined subtype (ADHD-C), ADHD predominantly hyperactive/impulsive subtype
(ADHD-PH) and ADHD predominantly inattentive subtype (ADHD-PI). The validity of
differentiating of these subtypes is rather debatable.
Disruptive Behavior Disorders (DBDs) might best be described along a continuum as the
emergence of ODD may be a precursor to CD (Turgay, 2009; Rowe et al., 2010). According to
DSM-IV children with ODD are characterized by angry/irritable mood, argumentative/defiant
behavior, or vindictiveness which lasts at least six months (APA, 2000). Persons with oppositional
defiant disorder do not reveal aggression toward people and animal, do not destroy property, and do
not show a pattern of theft or deceit (Campbell et al., 2000; Nolen-Hoeksema, 2014).
Conduct disorder (CD) is a behavioral and emotional disorder diagnosed in childhood/adolescence
that presents itself through a repetitive and persistent pattern of behavior in which the basic rights of
others or major age-appropriate norms are violated (Lahey et al., 2003; Baker, 2013). These
behaviors are often referred to as "antisocial behaviors" (Hinshaw SP, Lee SS., 2003). It is often
seen as the precursor to antisocial personality disorder, which is not diagnosed until the individual is
18 years old (APA, 2000).
The frequency of ODD among boys is higher prior to puberty, although the tendency does not
persist after puberty period (Demmer et al., 2017). The CD is also more common in boys in
comparison to girls (Nordström et al., 2013). The manifestation of CD is also different between
males and females. CD onset in girls is generally prior to adolescence (Olino et al., 2010).
2
Long term outcomes of children and adolescents for establishing or excluding comorbidity of
ADHD, CD, ODD and also distinguishing of subtypes of ADHD strong differential diagnosis needs
to be done. One of the common used screening scale in children and adolescents for
abovementioned purposes is parent/teacher DBD rating scale which was developed by Pelham and
his colleagues (Pelham et al., 1992) according to DSM –IV-TR diagnostic criteria (Molina et al.,
1998; Pelham et al., 2005; Evans et al., 2013).The Disruptive Behavior Disorder rating scale
consists of 45 items with response categories ranging from not at all (0) to very much (3) and takes
approximately 10 minutes to administer. This rating scale includes 9 items related to symptoms of
predominantly inattentive subtype of ADHD (ADHD-PI), 9 items for Hyperactive/Impulsive
subtype of ADHD (ADHD-PH), 15 items for Conduct disorder (CD), and 8 items for oppositional
defiant disorder (ODD). The DBD rating scale also has subscale for measuring combined type of
ADHD (ADHD-C) (items, 9,18, 23, 27, 29, 34, 37, 42, 44, 1, 7, 12, 19, 22, 25, 30, 33 and 35) in
children and adolescents. For the combined subtype of ADHD 6 or more items have to be selected
from ADHD-PI subtype and 6 or more items – from ADHD-PH category. Items 10, 14, and 21 were
not included for assessment of children/ adolescents' behavioral disorders since they are from
DSM-III-R and are not included in the scoring for a DSM-IV diagnosis.
Attention-Deficit/Hyperactivity Disorder, Oppositional Defiant Disorder, or Conduct Disorder can
be determined in two ways. The first method involves counting symptoms for each disorder using
the Disruptive Behavior Disorders (DBD) rating scale. The second method includes comparing the
target child's factor scores (e.g., using a 2 SD cutoff) on the DBD Rating Scale to established
norms.
For empirical assessment of ADHD, ODD, CD, and comorbidity of these disorders in Georgian
children and adolescents, Disruptive Behavior Disorder rating scale (Pelham et al., 1992) has been
translated into Georgian language and adapted for Georgian population.
Methods
Objectives
The study was designed for preparation of a reliable and valid scale in Georgian language for
assessment of neurodevelopmental disorders such as an ADHD, ODD and CD and establishment of
norms for Georgian population. Therefore the aims of the research were: (1) Translation of the
Disruptive Behavior Disorder (DBD) rating scale into Georgian language and (2) determination of
psychometric properties of Georgian DBD rating scale.
Based on the aims the study was separated into two steps. The first step was translation of the
Disruptive Behavior Disorder (DBD) rating scale. The second step related to collecting data and
3
establishment of Georgian Disruptive Behavior Disorder (DBD) rating scale norms for Georgian
population.
Step 1. In this step translation of Disruptive Behavior Disorder (DBD) rating scale was done by
English language philologists and assessment of content equivalence between Georgian and English
versions of Disruptive Behavior Disorder (DBD) rating scale was done by qualified two
psychologists and one psychiatrist. For determining the authenticity of Georgian translation the
scale was translated back into English language by two English language philologists who had not
been involved in initial translation of the rating scale.
Step 2. Second step included determination of psychometric parameters of Georgian Disruptive
Behavior Disorder (DBD) rating scale.
The data were analyzed by different descriptive and inferential statistical methods using Statistical
Package for Social Sciences 20.0 (SPSS 20.0).
Participants. Georgian version of Disruptive Behavior Disorder (DBD) rating scale was
administered on a sample of 290 children and adolescents with age range of 6-16 years (mean age
10.50, SD=2.87). The sample included 153 males (mean age 10.42, SD=2.62) and 137 females
(mean age 10.60, SD=3.14). Participants were recruited from different public schools of Tbilisi and
the Neurology Department of the Pediatric Clinic of the Tbilisi State Medical University (Table 1).
Table 1. The distribution of gender, age and ADHD subtypes for controls and children/adolescents with ADHD.
Control type
Gender
N
Mean Age
Standard Deviation
control
Hyperactive/Impulsive
Inattentive
Combined
Total
girl
boy
Total
girl
boy
Total
girl
boy
Total
girl
boy
Total
girl
boy
Total
11.09
10.90
11.00
8.86
9.20
9.09
9.18
9.56
9.41
8.69
9.84
9.45
10.60
10.42
10.50
106
97
203
7
15
22
11
16
27
13
25
38
137
153
290
4
3.203
2.604
2.926
3.237
2.111
2.448
2.401
2.250
2.275
1.437
2.824
2.479
3.138
2.622
2.874
Figure 1. The distribution of gender, age and ADHD subtypes for controls and children/adolescents with ADHD
Procedure. Procedure. The study was conducted in agreement with Helsinki Declaration (1977)
regarding ethical standards for clinical studies in medicine. Recruited participants were assessed by
the qualified neuropsychologists/psychologists in three different settings including school, home
and clinic. During assessing in school environment only those teachers were interviewed who were
teaching children/adolescents participated in the research at least last one year. At home settings
parents/caregivers of participated in the study children and adolescents were interviewed by
psychologists. Raters were required to assess participants based on their behavior during the last six
month in the appropriate settings. In clinic setting qualified neuropsychologists/psychologists based
on their own observation objectively assessed participants.
Results.
In the presented study data were analyzed by different descriptive and inferential statistical methods
using SPSS 20.0.
The mean scores and standart deviations were determined for typically developed and ADHD
children and adolescents (Table 2). In terms of DBD total scores revealed statistically significant
differences between healthy controls (M=27.71, SD=17.26) and children and adolescents with
ADHD (M=61.51, SD= 22.79) (Fig. 2a). For the inattentive subtype significant difference was
found for control (M=8.68, SD=5.68) and ADHD (M=18.15, SD=6.57) group (fig. 2b). For
hyperactive/impulsive subtype statistically significant difference revealed between ADHD group
(M=17.39, SD=6.74) and healthy controls (M=8.12, SD=6.07) (Fig. 2c). For oppositional defiant
disorder in healthy control (M=6.68, SD=4.59) and persons with ADHD (M=13.7, SD=5.82) was
found statistically significant difference (Fig. 2d). For conduct disorder revealed significant
difference between healthy group (M=2.51, SD=2.99) and group with ADHD (M=8.31, SD=6.14)
(Fig. 2e.). In general it was shown that children and adolescents with ADHD had high score on
DBD in comparison to typically developed persons.
5
variable
control
ADHD
variable
control
ADHD
median mode median mode
1.00
.00
1.00
.00
.00
.00
1.00
.00
1.00
.00
.00
1.00
.00
.00
.00
.00
1.00
1.00
1.00
.00
1.00
DBD1
DBD2
DBD3
DBD4
DBD5
DBD6
DBD7
DBD8
DBD9
DBD10
DBD11
DBD12
DBD13
DBD14
DBD15
DBD16
DBD17
DBD18
DBD19
DBD20
DBD21
Table 2. The mean scores of Georgian DBD for Healthy controls and children and adolescents with ADHD
DBD22
DBD23
DBD24
DBD25
DBD26
DBD27
DBD28
DBD29
DBD30
DBD31
DBD32
DBD33
DBD34
DBD35
DBD36
DBD37
DBD38
DBD39
DBD40
DBD41
DBD42
2.00
.00
2.00
1.00
.00
.00
2.00
.00
2.00
1.00
.00
3.00
2.00
1.00
1.00
.00
2.00
2.00
2.00
2.00
2.00
2
0
2
1
0
0
3
0
3
0
0
3
2
0
1
0
2
2
3
2
2
median mode median mode
.00
1.00
1.00
.00
1.00
1.00
1.00
1.00
1.00
.00
.00
1.00
.00
.00
.00
1.00
.00
.00
.00
.00
.00
2.00
2.00
2.00
2.00
2.00
2.00
2.00
2.00
2.00
.00
1.00
2.00
2.00
2.00
.00
2.00
1.00
2.00
.00
.00
2.00
2
3
1
3
2
2
2
3
2
0
0
3
3
2
0
3
1
3
0
0
3
0
1
0
0
0
0
1
0
0
0
0
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
1
0
0
2a
2b
2c
2d 2e
Figure 2. Measures and standard deviations of control and ADHD groups for Georgian DBD topics for 2a –
total, 2b – inattentive, 2c – hyperactive/impulsive, 2d – oppositional defiant disorder and 2e – conduct
disorder categories.
6
Discussions
The study was performed for preparing diagnostic scale for the assessment of children and
adolescents with behavioral disorders such as ADHD, ODD and CD and also with comorbidity of
these disorders in Georgia. For abovementioned purposes parents/teachers DBD rating scale was
translated into Georgian language and psychometric properties of Disruptive Behavioral Disorder
(DBD) rating scale were determined for Georgian population. As was shown in the study Georgian
version of DBD rating scale can be successfully used for assessment of children's and adolescents'
behavioral disorders both in educational and clinical settings. In general, it was shown that children
and adolescents with ADHD had high score on DBD in comparison to typically developed persons. In
the study also was determined gender wise prevalence in children and adolescents with ADHD,
ODD and CD. The research revealed prevalence of males in comparison with females in all
investigated categories. The Georgian version of DBD rating scale proved to be successfully
applicable in the educational and clinical settings for screening and diagnosis for children and
adolescents with behavioral disorders such as ADHD, ODD and CD.
References
1. Anderson, N. E., &Kiehl, K. A. Psychopathy: Developmental Perspectives and their Implications
for Treatment. Restorative Neurology and Neuroscience, 201432(1), 103–117.
2. Baker K. Conduct disorders in children and adolescents. Paediatrics and Child Health, 2013,
23:24–9.
3. Boden JM1, Fergusson DM, Horwood LJ. Risk
factors
for
conduct
disorder
and
oppositional/defiant disorder: evidence from a New Zealand birth cohort. J Am Acad Child
Adolesc Psychiatry, 2010 Nov; 49(11):1125-33.
4. Campbell SB, Shaw DS, Gilliom M. Early externalizing behavior problems: Toddlers and
preschoolers at risk for later maladjustment. Development and Psychopathology, 2000, 12:467–
488.
5. Connor DF1, Steeber J, McBurnett K. A review of attention-deficit/hyperactivity disorder
complicated by symptoms of oppositional defiant disorder or conduct disorder. Journal of
developmental and Behavioral Pediatrics, 2010, Jun;31(5):427-440
6. Demmer DH, Hooley M, Sheen J, McGillivray JA, Lum JA. Sex Differences in the Prevalence
of Oppositional Defiant Disorder During Middle Childhood: a Meta-Analysis. Journal of
Abnormal Child Psychology, 2017, Feb: 45(2): 313-325.
7
7. Diagnostic and Statistical Manual of Mental Disorders 4th Edition, American Psychiatric
Association, 2000.
8. Evans SW, Brady CE, Harrison JR, Bunford N, Kern L, State T, Andrews C. Measuring ADHD
and ODD symptoms and impairment using high school teachers' ratings. Journal of Clinical
Child and Adolescent Psychology, 2013;42(2):197-207.
9. Faraone S., Asherson P., Banaschewski T., Biederman J., Buitelaar J., Ramos‐Quiroga J.,
&Franke B. Attention‐deficit/hyperactivity disorder. Nature Reviews Disease Primers, Advance
online publication, 2015, doi: 10.1038/nrdp.2015.20.
10. Hinshaw SP., Lee SS., Conduct and Oppositional Defiant Disorders. Child psychopathology (E.
J. Mash &e.R. Barkley Ed.) New York: Guilford Press pp. 144-198.
11. Kerekes N, Lundström S, Chang Z, Tajnia A, Jern P, Lichtenstein P, Nilsson T, Anckarsäter H.
Oppositional defiant-and conduct disorder-like problems: neurodevelopmental predictors
and genetic background in boys and girls, in a nationwide twin study.Kyvik K, ed. PeerJ. 2014;2:
e359. doi:10.7717/peerj.359.
12. Lahey BB, Waldman ID. A developmental propensity model of the origins of conduct problems
during childhood and adolescence. In: Lahey BB, Moffitt TE, Caspi A, editors. Causes of
conduct disorder and juvenile delinquency. New York: Guilford Press; 2003. pp. 76–117.
13. Lavigne JV, Gouze KR, Hopkins J, Bryant FB. A multidomain cascade model of early childhood
risk factors associated with oppositional defiant disorder symptoms in a community sample of 6-
year-olds. Development and Psychopathology, 2016 Nov; 28(4pt2):1547-1562.
14. Masi, Laura. "ADHD and Comorbid Disorders in Childhood Psychiatric Problems, Medical
Problems, Learning Disorders
and Developmental Coordination Disorder." Clinical
Psychiatry 2015.
15. Molina BSG, Pelham WE, Blumenthal J, &Galiszewski E. Agreement among Teachers'
Behavior Ratings of Adolescents with a Childhood History of Attention Deficit Hyperactivity
Disorder. Journal of Clinical Child Psychology, 1998, 27(3), 330–339.
16. Nolen-Hoeksema S. Abnormal Psychology. New York, NY; MCGraw Hill. P.323.
17. Nordström T1, Ebeling H, Hurtig T, Rodriguez A, Savolainen J, Moilanen I, Taanila A.
Comorbidity of disruptive behavioral disorders and attention-deficit hyperactivity disorder--
indicator of severity in problematic behavior? Nord J Psychiatry. 2013 Aug; 67(4):240-8. doi:
10.3109/08039488.2012.731431.
18. Olino TM, Seeley JR, Lewinsohn PM. Conduct Disorder and Psychosocial Outcomes at Age 30:
Early Adult Psychopathology as a Potential Mediator. Journal of abnormal child psychology.
2010; 38(8):1139-1149.
8
19. Pelham W, Gnagy EM, Greenslade KE, Milich R. Teacher ratings of DSM-III-R symptoms for
the disruptive behavior disorders. Journal of the American Academy of Child and Adolescent
Psychiatry, 1992, 310, pp. 210-218
20. Pelham WE Jr, Fabiano GA, Massetti GM. Evidence-based assessment of attention deficit
hyperactivity disorder in children and adolescents. Journal of Clinical Child and Adolescent
Psychology. 2005 Sep;34 (3):449-76.
21. Rowe R, Costello EJ, Angold A, Copeland WE, Maughan B. Developmental pathways in
Oppositional Defiant Disorder and Conduct Disorder. Journal of abnormal psychology.
2010;119(4):726-738.
22. Turgay A. Psychopharmacological
treatment of oppositional defiant disorder. CNS
Drugs. 2009, 23(1):1-17.
23. Waschbusch DA. A meta-analytic examination of comorbid hyperactive-impulsive-attention
problems and conduct problems. Psychological Bulletin, 2002 Jan; 128(1): 118-150;
9
|
1701.00390 | 1 | 1701 | 2017-01-02T14:02:24 | Density-based clustering: A 'landscape view' of multi-channel neural data for inference and dynamic complexity analysis | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Simultaneous recordings from N electrodes generate N-dimensional time series that call for efficient representations to expose relevant aspects of the underlying dynamics. Binning the time series defines neural activity vectors that populate the N-dimensional space as a density distribution, especially informative when the neural dynamics performs a noisy path through metastable states (often a case of interest in neuroscience); this makes clustering in the N-dimensional space a natural choice. We apply a variant of the 'mean-shift' algorithm to perform such clustering, and validate it on an Hopfield network in the glassy phase, in which metastable states are uncorrelated from memory attractors. The neural states identified as clusters' centroids are then used to define a parsimonious parametrization of the synaptic matrix, which allows a significant improvement in inferring the synaptic couplings from neural activities. We next consider the more realistic case of a multi-modular spiking network, with spike-frequency adaptation (SFA) inducing history-dependent effects; we develop a procedure, inspired by Boltzmann learning but extending its domain of application, to learn inter-module synaptic couplings so that the spiking network reproduces a prescribed pattern of spatial correlations. After clustering the activity generated by multi-modular spiking networks, we represent their multi-dimensional dynamics as the symbolic sequence of the clusters' centroids, which naturally lends itself to complexity estimates that provide information on memory effects like those induced by SFA. To obtain a relative complexity measure we compare the Lempel-Ziv complexity of the actual centroid sequence to the one of Markov processes sharing the same transition probabilities between centroids; as an illustration, we show that the dependence of such relative complexity on the time scale of SFA. | q-bio.NC | q-bio |
Density-based clustering: A 'landscape view' of
multi-channel neural data for inference and
dynamic complexity analysis
Gabriel Baglietto1, Guido Gigante2 and Paolo Del Giudice3
1INFN-Roma1, Rome, Italy and IFLYSIB, Instituto de F´ısica de L´ıquidos y Sistemas
Biol´ogicos (UNLP-CONICET), La Plata, Argentina
2Italian Institute of Health, Rome, Italy and Mperience srl, Rome, Italy
3Italian Institute of Health and INFN-Roma1, Rome, Italy
November 7, 2018
Abstract
Simultaneous recordings from N electrodes generate N-dimensional
time series that call for efficient representations to expose relevant aspects
of the underlying dynamics.
Binning the time series defines a sequence of neural activity vectors
that populate the N-dimensional space as a density distribution, especially
informative when the neural dynamics proceeds as a noisy path through
metastable states (often a case of interest in neuroscience); this makes
clustering in the N-dimensional space a natural choice.
We apply a variant of the 'mean-shift' algorithm to perform such clus-
tering, and validate it on an Hopfield network in the glassy phase, in which
metastable states are largely uncorrelated from memory attractors.
The neural states identified as clusters' centroids are then used to de-
fine a parsimonious parametrization of the synaptic matrix, which allows
a significant improvement in inferring the synaptic couplings from the
neural activities.
We next consider the more realistic case of a multi-modular spiking
network, with spike-frequency adaptation inducing history-dependent ef-
fects; we develop a procedure, inspired by Boltzmann learning but extend-
ing its domain of application, to learn inter-module synaptic couplings so
that the spiking network reproduces a prescribed pattern of spatial corre-
lations.
After clustering the activity generated by such multi-modular spiking
networks, we cast their multi-dimensional dynamics in the form of the
symbolic sequence of the clusters' centroids; this representation naturally
lends itself to complexity estimates that provide compact information on
memory effects like those induced by spike-frequency adaptation. Specifi-
cally, to obtain a relative complexity measure we compare the Lempel-Ziv
complexity of the actual centroid sequence to the one of Markov processes
sharing the same transition probabilities between centroids; as an illus-
tration, we show that the dependence of such relative complexity on the
characteristic time scale of spike-frequency adaptation.
1
1
Introduction
Technology nowadays allows neuroscientists to simultaneously record brain ac-
tivity from increasingly many channels, at multiple scales; indeed, recent years
witnessed a kind of 'Moore's Law' for neural recordings [1], and this poses new
challenges and opens new opportunities.
One obvious challenge is to devise data representations that easily convey in a
compact form the spatio-temporal structure of the recorded data. Various forms
of dimensional reduction are now commonly used in the analysis of multiple
recordings.
In general terms, if one views recorded experimental data as a
matrix whose columns are the 'feature vectors' (in the case at hand, the set of
recorded activities in a given time bin), and whose rows span the 'sample space'
(in the case at hand, the successive time bins), dimensional reduction across
the column direction provides a reduced representation in terms of few suitably
identified features obtained from the original ones (e.g. principal component
analysis); on the other hand, one can view clustering across the rows direction
as a way to reduce the dimensionality of the data matrix by lumping together,
according to some similarity measure, groups of activity vectors sampled at
different times. This latter viewpoint, which we take here, to our knowledge
has been much less used in neuroscience.
On the other hand, one recently explored opportunity is to take advantage
of multiple recordings to revive old approaches to infer estimates of synaptic
couplings from measured correlations between neural activities. Correlations
measured from single neuron pairs obviously can only provide ambiguous esti-
mates of the direct synaptic couplings (due to confounding causes like common
input to the sampled neurons). However it was noticed in a landmark paper [2]
that when many (order 100 for instance) simultaneous recordings are available,
even though the underlying biological neural network is still dramatically under-
sampled, the global pattern of (individually small) pairwise correlations allows
to extract meaningful information about the synaptic connectivity. This was
achieved by assuming a maximum entropy Ising model, for which an "inverse
Ising" problem was solved to infer the parameters (couplings and external input)
for the given data. Many efforts were subsequently devoted both to extend the
approach to non-equilibrium estimates, and to lighten the computational load of
maximum entropy estimates (Boltzmann learning) through various mean-field
approximations (see e.g. [3][4][5]).
In the present work we propose an approach, based on clustering in the
multi-dimensional state space of simultaneous recordings, that provides both
an advantage for a compact representation of data, also amenable to efficient
estimation of the complexity of the system's dynamics, and besides allows to
improve inference on the network couplings.
After describing the clustering method (which is a slightly modified version
of the 'mean-shift' algorithm [6][7]), we first illustrate its working on time series
generated from the dynamics of a Hopfield network which, since it possesses an
energy function, naturally lends itself to density-based clustering in the state
space; here we choose a 'hard' regime where the Hopfield network is in a disor-
dered phase and spatio-temporal structures related to the energy landscape are
not easily discernible from the time series. At this stage we also formulate a
parametrization of the model's synaptic matrix, based on the identified clusters,
and show that it allows to obtain an inference of the synaptic couplings which
2
is much more insensitive to noise.
We then move on to the more complex and biologically motivated case of a
multi-modular attractor spiking network: its integrate-and-fire neurons are en-
dowed with spike-frequency adaptation (SFA), acting as an activity-dependent
self-inhibition that introduces history-dependent effects making the 'landscape'
dynamic; modules are approximately bistable between high and low activity
states; within-modules connectivity is stronger than between-modules.
Reasons to choose this particular context include recently published evidence
[8][9] that cortical dynamics can occur in the form of abrupt switches, between
an 'Up' and a 'Down' states, of local self-excited modules, such that the whole
dynamics appears as the evolution of a 'binary word', each bit being the 'Up'
or 'Down' state of each cortical module.
In the chosen setting of weakly coupled bistable modules, the network dy-
namics is largely determined by inter-modular couplings; in choosing the latter
we wanted to preserve some a priori knowledge of key features of the state space,
to be later checked against the found cluster structure, and we do this by seeking
inter-modular couplings such that the network possesses a prescribed pattern of
spatial pairwise correlations, derived in turn from a template Hopfield network.
To achieve this we develop a procedure, inspired by Boltzmann learning but
extending its domain of application, which we believe has an interest beyond
the contingent purpose.
We check the good performance of this new learning procedure by direct
comparison between the prescribed spatial correlations and those generated by
the optimal network at the end of learning; by comparing the state-space cluster
structure of the optimal network and the template Hopfield network; by infer-
ring effective inter-module synaptic couplings from simulations of the optimal
network and comparing them to the synaptic efficacies of the template model.
Knowledge of the cluster structure also allows to cast the multi-dimensional
time series generated by the network dynamics in the reduced form of a symbolic
dynamics in the clusters' centroid space. We perform such a reduction to show
that it allows to expose in a compact way time-dependent features like history-
dependent effects; we do this by exploring a large range of characteristic times
of SFA, and showing that the Lempel-Ziv complexity measure, applied to the
centroid sequence, nicely captures the memory effects induced by SFA.
In summary, what we propose here is a way to use knowledge of the spa-
tial correlations to develop informative and compact representations of multi-
dimensional neural data, also allowing for improved inference and useful reduced
representation of the multidimensional dynamics, and to develop a strategy for
data-driven model building with spiking networks.
2 Results
2.1 When the landscape is (partially) known: The Hop-
field Model as a first test ground for state-space clus-
tering
We start by providing an example of the information gained from clustering in
the state space, i.e. finding the local maxima of the density distribution of the
3
configurations generated by a system's dynamics, or - equivalently - the minima
of an effective energy landscape (see Fig. 1).
Though we want to ultimately apply clustering to data from realistic neural
simulations, to illustrate the method we consider here a well known neural
network (the Hopfield model) for which an equilibrium distribution of states is
defined.
Figure 1: Sketch of the steps involved in the density-based clustering of multi-
dimensional time series through the mean shift algorithm. The time-binned N -
dimensional data ('spikes') in panel A define a density profile in a N -dimensional
space, interpreted as sampling a stationary probability distribution (panel B),
from which an effective energy landscape can be defined (panel C). Panel D
illustrates how the mean-shift operates to find the local maxima of the density
distribution. Blue dot: current position of a point to be moved by the algorithm.
Big yellow area: the chosen neighbourhood of the blue point. Green dots:
neighbouring points of the blue point. Red dots: points that are not neighbours
of the blue point. Small yellow dot: the new position of the blue point is given
by the algorithm as the center of mass of its neighbouring points. See Section
4 for details.
For the Hopfield model [10], the equilibrium probability distribution of the
neural states σ = {σ1, σ2, . . . σN } (σi = ±1) is given by:
p(σ) ∝ exp(cid:16) − β H[σ](cid:17),
(1)
where β is interpreted as the inverse of a temperature and H is the energy
function:
N
P
H[σ] = −
1
N
Xi,j=1
Xµ=1
ξµ
i ξµ
j σiσj,
(2)
where the P discrete vectors ξµ are chosen as random uncorrelated configura-
tions of the N spins (ξµ
i = ±1) which, depending on P , N and β can act as
4
'stored patterns', i.e. the dynamics relaxes to the neighborhood of one of the
configurations ξµ, which are minima of the energy (maxima of the probability
distribution) together with their mirror patterns −ξµ [11]. Depending on P , N
and β, besides the patterns ξµ a large number of energy local minima is also
present, notably various linear combinations of the stored patterns themselves
or states uncorrelated with the patterns in the 'spin glass' phase.
In the following we consider the same Hopfield model (same synaptic cou-
plings, therefore same energy minima) in two operating regimes: low noise (i.e.
temperature, where the network is expected to reside most of the time in the
proximity of the deepest energy minima), and high noise (where the network
explores larger regions of the state space, less constrained by the structure of
the underlying energy function). The network comprises N = 50 neurons and
P = 4 stored patterns. Although for such small numbers relying quantitatively
on known theoretical estimates of memory capacity is unwarranted, for the cho-
sen values of N and P the model would be predicted to be well above the
retrieval phase for both temperatures, and for its non-trivial ('glassy') energy
landscape the energy minima are not expected to coincide with the memory
patterns.
The neural configuration time series were generated simulating the Hofpield
model for 20000 Monte Carlo steps (a suitable number of initial thermalization
steps were neglected).
To the N -dimensional resulting time series we applied the (slightly modified)
mean-shift (MS) clustering procedure as described in Section 4: essentially, we
iteratively move the points in the N -dimensional space, each of which represents
one state generated in the Monte Carlo sequence, towards the center of mass of
their neighboring points, thereby identifying at the end the estimated position
of the local maxima of the density distribution in the state space.
The states of the Hopfield model are defined over the N -dimensional hyper-
cube, where N is the number of neurons. In order to perform the mean shift
displacements we take the sign of the mean value of each coordinate, so that
the points remain in the original space. Any time the mean value of a given
coordinate is zero, we don't shift that coordinate.
Finally, after reaching a convergence criterion (see Section 4) of the MS
algorithm, we re-run it over the set of found centroids, with a fixed radius given
by Hamming distance 2 (overlap 0.92) 1; we found this further step to make
clustering more resistant to noise.
In general, we will consider only clusters containing data points above a
minimal fraction ("cutoff") of the whole time series, i.e. minimal mass (1%
unless specified). In this way we avoid to consider as clusters small bumps of
density due to finite sample fluctuations.
Figure 2, left panels, show representative time courses of the neural states for
the low-temperature (β = 1.3, top) and high-temperature (β = 0.83, bottom)
cases. Right panels show, for the two cases, the distribution of the overlaps
between all configurations in the time series.
The subsequent Figure 3 illustrates the result of the clustering procedure for
the low- and high-temperature cases.
1The overlap qαβ between two binary configurations, and weighting each centroid with its
i , and it is
mass in the analogous of Eq. 5. σα and σβ is defined as qαβ = 1
related with the Hamming distance hαβ by qαβ = 1 − 2 hαβ
N .
i σβ
N PN
i=1 σα
5
s
t
i
n
U
s
t
i
n
U
10
20
30
40
50
10
20
30
40
50
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Time (ms)
2
×104
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
Time (MCS)
2
×104
×106
y
c
n
e
u
q
e
r
F
2.5
2
1.5
1
0.5
0
-1
-0.8
-0.6
×106
y
c
n
e
u
q
e
r
F
3
2.5
2
1.5
1
0.5
0
-0.8
-0.6
-0.2
-0.4
0.4
Overlap between data points
0.2
0
-0.4
0
-0.2
0.4
Overlap between data points
0.2
0.6
0.8
0.6
0.8
Figure 2: Left column: The horizontal axis represents the time in MCS, while
the vertical axis identifies the different units. White (black) pixel in the matrix
entry (i, j) means that the i unit had a value -1 (1) at MCS j. Right column:
Histogram of the overlaps (see footnote 1) between the MC configurations.
In the high-temperature case it is very difficult to discern a structure in
the raster plot, and considering also the unimodal overlap distribution, broadly
symmetric around zero, it seems a challenging case for a procedure aimed at
extracting the energy minima.
This temperature dependence is reflected in the distribution of overlaps be-
tween the visited states. We note that, if energy minima would mostly coincide
with the stored (uncorrelated) patterns, the overlap distribution would be tri-
modal, with a peak in zero and two symmetric peaks at high (positive and
negative) overlap, which is not the case even for low temperature (top-right
panel in Figure 2), with the distribution of visited states having average over-
lap about 0.5 in magnitude. For higher temperature (bottom-right panel in
Figure 2) the overlap distribution is approximately Gaussian.
From Fig. 3 it is seen that for low temperature most centroids are indeed
different from the stored patterns; for high temperature more centroids are
identified, including the stored patterns.
While we did not perform a thorough analysis of clusters' centroids in terms
of energy local minima, we checked that clusters are indeed akin to attractor
basins, by measuring the configurations flow at zero temperature: starting the
deterministic (zero-temperature Monte Carlo) dynamics from each one of the
configurations assigned to a given cluster, the fraction of them for which the final
overlap with their assigned centroid is larger than the initial one is 0.90 ± 0.04
(for the clusters found at β = 0.83), and 0.85 ± 0.05 (for the clusters found at
β = 1.3).
6
s
t
i
n
U
s
t
i
n
U
10
20
30
40
50
10
20
30
40
50
2
1
Centroid Index
3
5
7
13 915
1
2
3
4
6
5
7
Centroid Index
8
9 10 11
13 15 18
Figure 3: Clustered configurations for the raster plots in Fig. 2. White is for
units with value -1, while in colors (different for different clusters) are indicated
units with value 1. The clusters whose centroids correspond to stored patterns or
their reflections are marked by red dots. The centroids of the clusters appearing
in the upper panel are a subset of those in the lower panel. Equal numbers refer
to the same centroid. The clusters' masses can change due to finite size sampling,
and clusters in each figure are ordered from biggest to smallest masses.
2.2 Clustering-aided Inference of Synaptic Connectivity
As mentioned in the Section Section 1, several methods have been and are
being developed to infer effective synaptic connectivities from the time series of
simultaneous neural recordings, some of which fall in the category of so-called
"Inverse Ising" problems [2, 4].
Obvious hindrance in the application of such methods is the large number
of parameters to be inferred (effective synaptic efficacies, of order N 2 for N
neurons), which makes them sensitive to noise in the data.
In the following, we show that the proposed clustering method can be used to
formulate the inference problem in a reduced parameter space, using the activity
configurations that are the identified centroids to parametrize a coupling matrix
7
(inspired to the construction of the Hopfield connectivity matrix) with a number
of parameters equal to the number of centroids.
Specifically (see also Section 4), the inference model has the form:
H[σ] = −
1
N
N
Xij=1
σi Jij σj .
(3)
where the effective coupling matrix J is the sum of weighted Hopfield-like terms
cµ
i cµ
j (see Eq. 2)2:
Jij =
1
N
C
Xµ=1
ωµcµ
i cµ
j ,
(4)
cµ being the C centroids identified by the clustering procedure, and the weights
ωµ are to be inferred.
To test this approach, we first use again the Hopfield model data of Fig. 3,
bottom panel. Since 14 out of the 18 identified centroids are pairs of reflected
patterns and give rise to the same Hopfield-like term in Eq. 8, the number of
parameters ω to be estimated is reduced to 11.
In Figure 4, left panel, we show the inferred values of ωµ. Only four of
them (marked with an asterisk) are significantly different from zero, and they
correspond to the centroids that coincide with the 4 stored patterns.
t
i
h
g
e
W
1
0.8
0.6
0.4
0.2
0
-0.2
-0.4
0.1
0.05
0
-0.05
s
e
i
r
t
n
E
x
i
r
t
a
M
d
e
r
r
e
f
n
I
1
2
4
6
9
8
11
Centroid Index
13
14
15
18
-0.1
-0.1
-0.05
0
0.05
0.1
Actual Matrix Entries
Figure 4: Left: Inferred values of the weights ωα when fitting the reduced model
(Eq. 8) to the raster plot shown in the lower column of Fig. 2). The red stars
mark the weights corresponding to the patterns actually stored in the system.
Right: Scatter plot of the inferred (reduced model) synaptic couplings against
the actual ones. The red curve shows the identity as a reference
Notice that, when the ωs are roughly equal, they play a role similar to an
inverse temperature; indeed, the value of β used to generate the time series (0.83)
is very close to the ω values for the centroids corresponding to the patterns.
We also remark that the obtained values for ω would not be trivially expected
from the structure of the state space, in that not only 7 out of 11 centroids (not
counting reflections) do not belong to the stored patterns, but the clusters of
2In the following, depending on the context, σi will be either the activity of a binary neuron
or a suitable binarization of the average activity of a population of spiking neurons.
8
largest mass are not centered on patterns (remember that the Hopfield system
is far from the retrieval phase).
In Figure 4, right panel, we compare the inferred J with the real J (N ×
(N − 1) = 2450 elements, taking only P + 1 = 5 values); the inference is actually
good (the continuous identity line is drawn to guide the eye), and better than
the one obtained inferring directly the full J matrix (i.e. not adopting the
parametrization of the synaptic matrix in Eq. 8), as shown in Fig. 5, where we
compare the relative error ((J Inf erred
− Jij )/hJijiij ) for the two cases.
ij
8
6
4
2
0
y
t
i
s
n
e
D
y
t
i
l
i
b
a
b
o
r
P
d
e
t
a
m
i
t
s
E
-1
-0.5
0
0.5
1
Relative Error
Figure 5: A posteriori probabilities of inference errors for the cases of the reduced
model (red curve) and the full matrix inference (blue curve). Details in the text.
From this simple example, we confirm that the adopted parametrization,
besides the obvious greater simplicity and lesser computational load, is effective
in reducing the effect of noise in the data and makes the inference less prone to
overfitting, while allowing a good match between model and data.
2.3 Towards a more realistic scenario: a multi-modular
network of spiking neurons matching a prescribed spa-
tial correlation structure
In the previous Section we showed that, from the spatial correlation structure
of the configurations sampled by Monte Carlo, MS clustering was effective in
reconstructing the main local energy minima, and allowed for a parsimonious
parametrization of the synaptic matrix that afforded better inference. In that
case, the correspondence between local density maxima in the state space, and
local energy minima, was ensured by the existence of a Gibbs equilibrium prob-
ability for the Hopfield model.
In the perspective of applicability to real electrophysiology data,
in the
present Section we extend the approach to networks of spiking neurons, and
to a situation where the notion of a static energy landscape is no longer strictly
applicable.
To establish a meaningful benchmark, we want to preserve some a priori
knowledge of key features of the state space, to be checked against the found
9
cluster structure, and we do this by setting up a spiking network constructed so
as to possess a prescribed spatial correlation structure, from which we generate
time series to be clustered. To achieve this we develop a method that, we believe,
has a wider interest beyond the case at hand.
The chosen network architecture is composed of strongly self-coupled mod-
ules of spiking (integrate-and-fire) neurons, with much (10−2 − 10−3) weaker
synaptic couplings between modules.
Intra-module synapses are drawn from
a Gaussian distribution, with mean and average chosen (using mean-field pre-
dictions) such that each module in isolation is approximately bistable, between
states of low (DOWN) and high (UP) firing activity.
The interest in this choice for the architecture of the spiking network stems
from accumulating evidence that, not only a modular structure is suggested
by the mesoscopic anatomical organization of the cortex, but it also appears
to be recognizable in the cortical neural dynamics, which can proceed as the
dynamic composition of abrupt jumps between UP and DOWN states ([8, 9]),
as mentioned in the Introduction (see also [12][13][14]).
We remark that synapses are not be constrained to be symmetric, therefore
the clustering procedure will not, strictly speaking, match the minima of an
energy function of the system. Besides, the Integrate-and-fire spiking neuron
model (see Section 4) is endowed with spike-frequency adaptation (SFA), a much
studied (and pervasively observed) self-inhibitory mechanism depending on the
activity history of the neuron; SFA makes the effective landscape, even when
it exists, locally dynamic at the point currently corresponding to the network
state.
Fig. 7 describes the main steps involved in the network construction. The
multi-modular network is sketched (panel A) as a collection of 64 neural mod-
ules, each composed of 32 adapting excitatory and 16 inhibitory neurons (see
see Section 4 for details). The approximate bistability of the single modules,
which is to a large extent preserved in the interacting network, is illustrated
in panel B by the time course of the firing rate of the excitatory neurons from
two sample modules; the resulting bimodal distribution of firing rates allows to
binarize the modules' activity, as shown in panel C 3. Once binarized, the time
course of the whole network activity can be represented as a sequence of binary
vectors (see the 'raster plot' in panel D), to which clustering is applied.
We choose to assign a spatial correlation structure mimicking the one of a
Hopfield attractor network, each module corresponding to one binary neuron of
the Hopfield network: inter-module synapses, initially drawn from a Gaussian
distribution, are subject to an iterative procedure (see below, and Section 4 for
details) to match those of the reference Hopfield model.
Were it not for the effects of SFA, the correlation-matched multi-modular
network would be expected to behave as an attractor, Hopfield-like system,
with the Up and Down states of each module playing the role of the binary
values of the Hopfield neurons. However, as the network enters the basin of
one attractor, the active modules start lowering their activity because of SFA,
thereby destabilizing the state: as noted above, the attractor landscape becomes
'dynamic', in that its depth and curvature around attractor states get lower
when the system visits them, promoting transitions to other basins, which are
3We remark that the binarization step, here instrumental for the construction of the in-
tended benchmark, would not be needed in general, and spate-space clustering can be per-
formed of the raw multi-dimensional time series.
10
Figure 6: Sketch of the main steps involved in the network construction. See
text for details.
biased by the correlations with other attractor states (due to finite-size effects)
4.
The new procedure we developed to determine the inter-module synaptic
couplings is inspired to the Boltzmann learning strategy where, at each iteration,
the change in the couplings is proportional to the difference between the spatial
correlations in the model and in the data (and analogously for the external
fields). For Boltzmann learning, such difference is the gradient of the Kullback-
Leibler distance between the state probabilities in the model and in the data.
In our spiking network the explicit form of such a function is lacking; still,
intuition suggests, and we assume, that a monotonic relation still holds between
the synaptic couplings and the spatial correlations, and this is equivalent to
assume that we know the sign of an unknown gradient. Formulated in this way,
our 'pseudo-Boltzmann' iterative process to find the optimal synaptic couplings
is naturally mapped onto the Rprop algorithm (see Section 4, Eqs. 17, 18).
In summary, the sequence of steps we take is the following: 1) store a set of
patterns a la Hopfield in a network of binary neurons (as in the previous section,
the Hopfield network will be in its glassy phase); 2) from the Hopfield network,
measure spatial correlations and site magnetizations (average activity); 3) set
up a multi-modular spiking network of approximately bistable modules; 4) use
pseudo-Boltzmann learning to find inter-modular couplings and external rates
4The role of SFA in promoting transitions between neural states has been studied in various
contexts, see e.g. [15, 16, 17, 18, 19, 20, 21, 22, 23].
11
to mimic correlations and magnetizations of the Hopfield system; 5) perform
MS clustering on the configurations generated by simulations of the spiking
multi-modular system; 6) check the quality of the result (see below) .
The chosen 'reference' Hopfield model has N = 64, P = 4, β = 1.1, for
which we generate a long Monte Carlo sequence, and measure the spatial cor-
relations and site magnetization that are to be matched by the multi-modular
spiking network with optimal inter-module couplings Jpq and external inputs
νext
p
obtained from the pseudo-Boltzmann procedure explained above.
One may ask whether the simplest choice that intuition would suggest, i.e.
simply taking the inter-module synapses as proportional to the computed Hop-
field ones, would work. We checked this option, with poor results. One reason
is that the low- and high-firing rate states (which are subject to binarization in
the clustering procedure) are not dynamically equivalent (contrary to the binary
neurons of the Hopfield case): one state can be deterministically more stable
than the other (e.g. in the sense of the linear stability of the two corresponding
fixed points in the mean-field approximation); furthermore, noise is higher in the
high-firing rate state (finite-size noise is activity-dependent and acts multiplica-
tively). Finally, there are 'quenched noise' effects: the synaptic connectivity in
each module is a different (small) realization of the same probabilistic model,
and can lead to quite different dynamics between modules.
We checked the success of pseudo-Boltzmann learning in several ways.
First, the correlation between the partial correlations measured from the
optimal spiking network and those from the reference Hopfield model is very
high (R2 = 0.975); the average 'magnetization' for the optimal spiking network
is order 10−3, to be compared with the potential range [−1, 1] and the theoretical
null value for the Hopfield network. The found optimal external firing rates have
large variations between modules (over ±20% with respect to their average),
which is expected, since they contribute to compensate for the heterogeneous
excitability of the different modules.
Second, we checked the similarity between the cluster structures emerging
from the optimal spiking network and that of the reference Hopfield network, by
computing the absolute value q of the overlaps between the centroids found for
the two networks (cS and cH respectively for the spiking and Hopfield networks).
Out of the 12 cH centroids, 9 of them (75%) have q = 1 with at least one
of the cS; among the others, 2 of them have q ∼ 0.97 with at least one of the
cS; 1 of them has q ∼ 0.84 with one of the cS.
Conversely, out of the 13 cS centroids, 8 of them (62%) have q = 1 with at
least one of the cH; among the others, 2 of them have q ∼ 0.97 with at least
one of the cS; 3 of them have q ∼ 0.75 with at least one of the cH .
This shows that the pseudo-Boltzmann iterative procedure, by enforcing
approximately equal mean activities and spatial pair correlations between the
Hopfield model and the modules of the spiking network results in fact in similar
mostly visit regions of the state space for the two systems.
Third, we inferred inter-modular synaptic efficacies (using MPF as in the
previous Section) from the spiking network time series, and found that for all
modules pairs the inferred synapses are close to corresponding synapses of the
reference Hopfield model (the mean absolute value of the relative error, taken
over the J entries, is 11%, likely mostly due to finite-sample noise, as suggested
by comparison with Fig. 5); this confirms that, despite the large differences be-
tween the synapses of the Hopfield network and the optimal inter-modular aver-
12
age synaptic efficacies determined by pseudo-Boltzmann learning (not shown),
the dynamics of the resulting spiking network effectively embodies inter-module
interactions consistent with the reference Hopfield network.
To summarize, the somewhat complex procedure described allowed us to
construct a modular spiking system with some control on desired features of
the state space, not easily enforced by simple ad hoc assignment of the synaptic
structure; while in this case the construction was guided by a reference Hopfield
model, whose neurons were naturally mapped onto the network's modules, more
in general we believe the procedure is interesting per se, as a means to enforce a
prescribed pattern of spatial correlations (and associated state-space structure)
in relatively complex networks of spiking neurons.
2.4 Dynamics in the centroid space
When performing MS clustering in the state space, information about the dy-
namics of the original time series is - by construction - lost. However, once
clustering is done, knowledge of the clusters' centroids allows to go back to the
multi-dimensional dynamics, and cast it in a useful compact form: to each one
of the vectors expressing the states at successive sampling times, we substitute
the label of the cluster that vector was assigned to.
The description of the system's dynamics is reduced to a 'symbolic dynamics'
in the centroids space, which opens up options to expose dynamic features that
may be difficult to uncover directly from the analysis of the spiking activity, as
we illustrate through an example in the present section.
Based on the procedure developed in the previous Section to set up multi-
modular spiking networks, we want to generate a family of networks for which
different degrees of 'complexity' can be expected, instantiated here in differ-
ent dynamic memory span, induced by the SFA component that, as discussed,
affects locally the dynamics in a history-dependent way.
In the previous section SFA was chosen small, just enough to obtain mea-
surable state transition rates in the simulation time. Here we set up a series of
spiking multi-modular networks, each one constructed as in the previous section,
but with different parameters for SFA.
We span a range of values for the timescale of SFA (τSF A), while keeping
the product τSF A gSF A constant (in order to keep the SFA 'strength' constant
and have comparable systems, see Section 4).
We show here that the reduced dynamics in the centroid space lends itself
naturally to methods for symbols-oriented measures of complexity, able to easily
capture memory effects.
It has long been suggested, and reported in several published works, that
the Lempel-Ziv (LZ) complexity measure [24] may be usefully adapted to char-
acterize neural data series [25, 26, 27];
in particular, a suitably normalized
LZ complexity has been successfully employed as a diagnostic measure of the
distance to the conscious state in neurological patients [28].
In essence, the
approach is based on the intuitive idea that the more complex the signal, the
less its compressibility; in other words, more structure in the signal increases
its predictability. Therefore, a memoryless stochastic time series would have
maximal complexity, and any memory embedded in the dynamics generating
the time series would make it decrease. Such entropic measures provide infor-
mation beyond what linear correlation analysis can provide.
13
We therefore measure, for the symbolic sequence reduction of the multi-
dimensional dynamics obtained from networks with different τSF A, LZ com-
plexity and a relative complexity index, as detailed in Section 4.
Expectation is that increasing τSF A generates multidimensional time series
with decreasing complexity. We checked that such expectation is met in our
spiking simulation data, and that differences in LZ complexity for high and low
τSF A are statistically significant for the simulated time span.
In order to quantify the difference, for each value of τSF A and gSF A we also
simulated ten realizations of Markov chains in the centroid space, generated by
the transition probabilities estimated from the simulation, thereby producing
surrogate memoryless sequences with the same transition statistics as the data;
we computed the LZ complexity averaged over the ten Markov chains ('cMarkov'),
and to compare it to the LZ complexity from the simulation data ('csample'), we
computed the ratio R = (cMarkov − csample)/cMarkov (see Section 4), 5.
For a meaningful comparative analysis of the complexity of the dynamics in
the centroid space for widely different values of τSF A (and related gSF A), we
should make sure that the corresponding networks are indeed similar between
themselves, and with the reference Hopfield network, in terms of the state space
structure.
For this purpose we did a preliminary analysis of the clusters found for all
the explored values of τSF A; since, as τSF A increases, an increasing number of
small clusters appears6, with small differences from main ones, we first clustered
the centroids with standard methods (hierarchical clustering using Hamming
distance and complete linkage), to obtain a 'fuzzy' version of the whole set of
main centroids; for those (76 clusters), we found that: the first 8 fuzzy centroids
account for 66.3% of the total configuration mass and are detected, on average,
for 84.0% of the τSF A values; the first 10 fuzzy centroids account for 71.7%
of the total configuration mass and are detected, on average, for 73.0% of the
τSF A values; the centroids of the reference Hopfield network are recovered in
87.2% of cases (by 'recovered' we mean that at least one of the centroids found
for an Hopfield simulation-clustering has overlap > 0.719 with the considered
Hopfield centroid, the threshold value 0.719 being determined such that 95% of
the overlaps between the centroids found for the considered case are below such
threshold.
The results of the analysis are summarized in Figure 7: we see that the
ratio R increases with increasing τSF A or, in other words, that as memory
effects increase the complexity of the actual centroid sequences gets increasingly
larger than that of surrogate Markov sequences, which provides a quantitative
information on the non-Markovian nature7) of dynamics for higher τSF A cases.
In the figure we also report for comparison (red line) the R value obtained for
the centroid sequence extracted from the time series generated by the reference
Hopfield network (which is inherently Markovian); it is seen that, although the
centroid sequence deviates from a Markov process, the relative difference w.r.t.
5including self-transitions would blur the difference and decrease the LZ complexity, since
the corresponding runs of identical centroid labels add to the compressibility of the sequence;
we did not consider self-transitions in the analysis
6This is due in part to the fact that, because of SFA, 'bridge states' appear, with associated
small clusters, where the system transits just after SFA has destabilized one major attractor
state, on its way to make a transition to another major attractor state.
7We use the term here in an informal sense, not distinguishing between higher-order Marko-
vian and strictly non-Markovian processes.
14
the surrogate sequence is very small, much smaller than the one for the spiking
dynamics, where SFA plays a major role.
0.25
0.2
0.15
0.1
0.05
R
0
0
500
1000
1500
2000
!
SFA
(ms)
2500
3000
3500
4000
4500
Figure 7: Dependence of the complexity measure R vs τSF A. Red line: R com-
puted for the Markovian sequence generated by the reference Hopfield system;
blue line: R vs τSF A for the actual sequence.
Of course, deviations from a Markov, memoryless dynamics can take many
forms; a generic expectation is that, for a non-Markov process, sequences of
states of given length are more, or less, likely to occur than predicted based
only on the transition probability matrix.
To gain insight for the case at hand, we considered the centroid sequence
for the spiking network with the highest τSF A = 4s (of length about 1.3 ×
103). For each of the triplets of centroid labels (133 triplets, since 13 centroids
were extracted in this case by MS clustering) the probability of occurrence was
estimated from the actual sample, and computed from the Markov transition
probabilities estimated from the same sample.
Fig. 8, left panel, shows (blue points) a scatter plot of such probabilities
(limited to the triplets occurring more than 10 times). With reference to the
black identity line (close to which the points would obviously cluster if the ac-
tual sequence was Markovian), it is seen that many triplets are over- or under-
represented in the actual sequence (particularly in the region of higher proba-
bilities which matters most), as a reflection of its non-Markovian nature.
In order to assess the significance of the observed differences, we also gener-
ated a truly Markovian sequence with the same transition probabilities and of
the same length as the sample, and from it we estimated the triplets probabil-
ities; the green crosses show the corresponding scatter plot with the computed
Markov probabilities, and it is clearly seen that finite-sample effects are much
smaller than the spread observed for the actual sequence, confirming its genuine
non-Markovian nature.
Finally, given the dependence of R on τSF A shown in Fig. 7, it was natural
to ask how the non-Markovian estimated triplets occurrence probabilities would
depend on τSF A; this is illustrated in Fig. 8, right panel, where we report the
Kullback-Leibler distance between the sampled triplets distributions from the
15
actual sequence and the surrogate Markov sequence, as a function of τSF A;
although in principle for different τSF A we may expect different contributions
to the non-Markov nature of the sequence from sub-sequences of different length,
we observe an approximately monotonic dependence of the KL distance on τSF A.
Sample
Markov surrogates
10-2
l
)
e
p
m
a
s
(
t
l
e
p
i
r
t
P
10-3
10-3
10-2
Ptriplet (Markov estimate)
Figure 8: Left: for all triplets occurring more than 10 times, the blue points
are the estimated probability from the actual sequence vs the computed Markov
probabilities from the estimated transition probability matrix; green crosses are
the estimated probability from the actual sequence vs the estimated probability
from the Markov surrogates. Right: the Kullback-Leibler distance between
the sampled triplets distributions from the actual sequence and the surrogate
Markov sequence, as a function of τSF A.
In summary, reducing the multidimensional time series of spiking data to the
sequence of labels of the clusters identified by the state-space clustering proce-
dure, casts the dynamics in a form easily suited to capture and quantify traces
of memory effects, for instance allowing in principle the comparison between
recordings in different experimental conditions.
3 Discussion
In this work we considered pre-existing strategies and developed new tools that,
taken together, compose a methodology with good potential, we believe, in
analysis and modeling of neuroscience data.
We started from a simple idea (though, to our knowledge, it was not ex-
ploited so far in the analysis of multiple neural recordings): to represent the
multidimensional time series of neural activities as a density distribution in a
corresponding multidimensional space, and perform a density-based clustering
procedure to extract the local density maxima. We strived to show that this type
of dimensional reduction is in fact a versatile instrument; in particular we illus-
trated examples that it can be useful in achieving better inference of synaptic
couplings from neural activities, and also to cast the multi-dimensional neural
dynamics in a compact form amenable to symbol-oriented methods of complex-
ity analysis.
16
The clusters, and the associated centroids, have an obvious interpretation
when the original time series is generated by an attractor dynamics, but retain
an informative value even when (like in the case of SFA in the spiking network)
the picture of a static attractor landscape is no longer appropriate; preliminary
work in progress on multiple in-vivo recordings during motor tasks makes us
confident that the approach can provide compact and usable information even
in strongly non-stationary conditions.
We first validated the method on the activity generated by a simple Hopfield
network, albeit we choose for it a highly nontrivial working regime in which the
energy landscape explored during the network dynamics is 'glassy', with many
local minima uncorrelated from the memory pattern embedded in the Hebbian
synaptic matrix. We showed that our (modified) Mean-Shift clustering is indeed
effective in revealing features of the energy landscape from very noisy time series.
Besides, knowledge of the clusters' centroids allowed an efficient parametrization
of the synaptic matrix, which in turn allowed a much better result when inferring
the synaptic efficacies from the network's sampled activity.
We should remark that the generality and robustness of this latter result is
still to be thoroughly investigated; in the case shown here, the mathematical
form of the inference model is similar (though not equal) to the one of the model
network generating the data used for inference. While this is indeed a limita-
tion shared with most published works dealing with 'inverse Ising' approaches
to synaptic inference, in future work we want to systematically explore more
generic forms of parametrization of the synaptic matrix (still using information
from the clusters extracted from the activity time series).
Furthermore, the proposed approach addresses, though admittedly in a spe-
cific context, a general issue. Whatever the specific method adopted, the extent
to which the inference of single synapses can be trusted can be severely affected
by several factors, like inherent inadequacy of the model used for inference, poor
quality of the data and noise, limited data sample. Still, it is usually (explicitly
or implicitly) assumed that even when the inference procedure fails to match
single synaptic efficacies, if the synaptic matrix has a global structure it should
still be captured in the inferred matrix. This would almost inevitably call for
some kind of dimensional reduction of the inferred synaptic matrix, such that
the informative relevant 'mesoscopic' structure is retained.
In a sense, what
we propose here can be viewed as a way to formulate a 'mesoscopic' inference
problem in the first place; again, we cannot claim full generality here, but our
clear success in the case of attractor networks examined in the present work
makes the approach, we believe, an interesting option to be further explored.
Wanting to move to more realistic network models, we were naturally led to
networks of spiking neurons (Integrate-and-fire, with SFA); in the Introduction
and in the Results we provided motivations for a specific choice of the network
architecture as composed of weakly coupled, individually bistable neural popu-
lations. In order to make contact with the study of clustering in the Hopfield
network and, more importantly, to have some a priori knowledge of the effective
landscape of the spiking system, we also wanted to set up the inter-module cou-
plings such that the spiking network as a whole would share static properties of
the state space with an Hopfield attractor network. This need motivated us to
develop a procedure that determines the inter-modular couplings such that the
network's activity generates a prescribed pattern of pairwise spatial correlations.
The procedure involves a new algorithm that extends the domain of appli-
17
cability of Boltzmann learning, and uses Rprop learning; as already remarked,
we believe this approach has a value beyond the specific purpose it served in
the present work.
It rests on the intuition that between excitatory synaptic connection strength
and neuronal activity correlation a monotone relationship should hold. On a
more general level, this is just an instance of a strategy (which is also a human
ability) to identify the relevant variables in a problem and code them in such
a way that they have a conditionally monotone relationship with the relevant
observables or, more specifically, with the statistic deemed sufficient for the
problem under exam [29]. Such ability, coupled with the proven effectiveness
in a variety of contexts of optimization algorithms that depend only on the
sign of the derivative of the function to be optimized (as Rprop)[30, 31], can
make the idea behind the proposed algorithm robustly generalizable to a wide
array of problems, dealing both with static and dynamic properties of neuronal
networks, e.g. by taking into account spatial as well as temporal correlations.
Finally we went back to dynamics. A natural step was to substitute the
original multidimensional time series with the corresponding 'symbolic dynam-
ics' of centroid labels, and to ask whether the resulting paths in the centroid
space would allow to extract information on the system's original dynamics
that would be difficult to directly expose. A case in point, in our context,
was to inspect how the sequences of transitions between centroids in the spik-
ing modular network would depend on the strength of SFA. SFA introduces
'memory' in the dynamics, and higher SFA makes the original time series more
history-dependent, and the corresponding symbolic dynamics of centroid labels
is expected to be less Markovian. As discussed in the text, the pattern of tran-
sitions between centroids results from an interplay, in the original time series,
between noise, spatial overlaps between attractor states of the multi-modular
network, and SFA-dependent effects.
To quantify the memory-related complexity of the network activity, we de-
fined a measure based on Lempel-Ziv (LZ) complexity (inspired to previous work
in various scientific domains, including neuroscience): for different time scales
of SFA, we compared the LZ complexity of the centroid time series with surro-
gate Markov sequences with the same transition probabilities, finding that, as
expected, longer timescales of SFA correspond to less complex (and less Marko-
vian) centroid sequences. We also provided insight into such SFA-dependent
non-Markovian nature by studying the occurrence probability of selected sub-
sequences.
Again, while we illustrated in some detail this reduction to symbolic dynam-
ics and the analysis of its complexity in the specific case under consideration,
its value rests with its generic applicability to multidimensional time series.
A few closing remarks, to facilitate comparison with other approaches to
the characterization of multidimensional time series in neuroscience (including
Hidden Markov Models, that recently have been frequently used in the analysis
of neural data (see e.g.)).
First, the proposed state-space approach is free from bias towards spherical
clusters and from a pre-defined number of clusters. Such freedom is inherent in
the approach taken here, which is also quite easy to implement.
Second, for large data sets the density-based clustering approach can become
computationally expensive, and of course it is meaningful to try to speed it up.
A recent successful attempt was made in [6], where a preliminary distance-based
18
selection procedure excludes outliers (points with low local density) from the
iterative procedure. The validity and performance of the approach is tested in
a variety of benchmark data sets, including the Olivetti face dataset [32]. In
comparing to the present work, we remark that on the one hand we did not
make an effort towards computational efficiency; rather, we wanted to show the
potential of density-based clustering for the analysis of neural data from multiple
simultaneous recordings. On the other hand, we tried the method proposed in
[6], and checked that it performs poorly in several representative situations, due
to the high level of noise. A reasonable strategy would probably be a mixed
method which first performs a number of iterations (dependent on the noise
level) of the mean-shift algorithm (to 'clean up' enough the data distribution in
the configuration space), followed by the faster procedure described in [6].
4 Materials and Methods
4.1 Clustering in the state space: a modified Mean Shift
Algorithm
We start with a data set {x1, x2, . . . xM } consisting of M (real valued) vectors in
a N -dimensional space. The Mean Shift (MS) algorithm [Fukunaga Hostetler][7]
iteratively picks a random i (1 ≤ i ≤ M ) and at step t updates the vector xi
according to the rule:
xt+1
j)(cid:17)
j w(cid:16)dist(xt
i = P1,M
i, xt
j6=i xt
j)(cid:17)
j6=i w(cid:16)dist(xt
P1,M
i, xt
;
(5)
i, xt
i is replaced by a weighted average of the other points xt
where w(·) is a non increasing function of its argument, dist(xt
j) is a distance
measure between vectors, and the original vector set is {x0
i }. In other words,
at each step xt
j : points
that are closer to xt
i are given more weight, thus defining a "soft" neighborhood
of xt
i (see Fig. 2). Therefore the update rule effectively moves xi towards regions
of the N dimensional space where the (local) density of points is higher at step
t. The other M − 1 vectors are instead left unchanged (xt+1
j for j 6= i);
though a parallel, deterministic update rule is in principle possible applying
Eq. 5 to all the data points at the same time, the random, sequential update
just described proved more robust in practice.
j = xt
The iteration terminates upon a threshold condition: when all xi(t) belong
to a discrete space, the iteration terminates when the fraction of cases (out of the
last M ) in which xi(t + 1) 6= xi(t) is below a fixed threshold; in the continuous
case, an additional threshold on a minimal displacement is required. Clusters
are finally determined as sets of initial points x0
i absorbed by a same final point
(their number we call the 'mass' of the centroid); each final point can be taken
as representative of a different cluster (its 'centroid'.)
The choice of the weighting function w(·) is central to the MS algorithm.
Such function can in general depend both on the step t and the configuration of
the neighborhood, thus adapting to different conditions. We used a step weight
function:
w(x) = (cid:26)1 if x ≤ r
0 if x > r
.
(6)
19
The choice of the radius r is critical; on the one hand, a small radius can
make the algorithm too sensitive to local variation of density (e.g. due to noisy
data), leading to many non-representative clusters; on the other, as the radius
increases, the reduced sensitivity to local fluctuations can lead to merging differ-
ent meaningful clusters. We therefore resorted to an adaptive heuristic criterion
to select r. At each step t, we compute the standard deviation σ(n) of the
distances between xt
i, the point selected for update, and its n nearest neighbors.
We then take nmin = arg minn σ(n), and set r as the distance between xt
i and its
nmin-th nearest neighbors. We found such heuristics to work well in our case of
discrete (hyper cubic) space, for different data sets and for a wide distribution
of cluster sizes.
This criterion for determining the radius, together with the additional MS
clustering performed on the centroids found in a first run (see Section 2), consti-
tute a modification of the original MS algorithm, which we found advantageous.
In the MS algorithm the initial ordering of data points is irrelevant, as they
are treated as samples of a static distribution. For time series this means that
the time structure of the data is completely ignored and points that are distant
in time can end up being clustered together, e.g. as the result of subsequent
'jumps' of the system in a same region of the state-space. Nonetheless, as we
saw in Section 2.4, segmenting the original time series in terms of transitions
from one cluster to another can highlight relevant dynamical features (non-
stationarities, memory effects) in the data.
4.2
Inference
Specifically, we will assume for the system under consideration an effective Ising-
like energy function:
H[σ] = −
1
N
N
Xij=1
σi Jij σj .
(7)
We propose to decompose the effective coupling matrix J in a set of weighted
Hopfield-like terms cµ
i cµ
j (see Eq. 2):
Jij =
1
N
C
Xµ=1
ωµcµ
i cµ
j ,
(8)
where the vectors cµ are the C centroids identified by the clustering procedure,
and the weights ωµ are to be found through an optimization procedure.
Notice that, when the ωs are roughly equal, they play a role similar to an
inverse temperature; indeed, we found that the value of β used in the Hopfield
simulations is very close to the ω values for the centroids corresponding to the
patterns.
For the optimization we use a recently introduced technique, Minimum Prob-
ability Flow (MPF) [33], with the advantage that it does not rely neither on
Monte Carlo (MC) runs (like Boltzmann learning does) nor on mean field ap-
proximations (like most Inverse Ising methods do [34]).
It is based on the
maximization of a function whose maxima are close to the Boltzmann Likeli-
hood ones, but with no needs of performing computationally expensive Gibbs
samplings.
20
Whilst Boltzmann learning aims to minimize the distance (more precisely
the Kullback-Liebler divergence) between the distribution of the data and the
distribution of the model, MPF aims to minimize the distance between the dis-
tribution of the data and the model distribution after an infinitesimal Monte
Carlo step, starting from the data. As the name of the algorithm suggests, this
amounts to minimizing the outflow, determined by the model, of probability
from the data. Since the latter is expressed only through the transition proba-
bilities, performing the Monte Carlo is not actually needed, and the gradients
of the objective function can be explicitly computed.
For completeness, we briefly summarize the main points in the MPF ap-
proach. The model probability density p, dependent on a set of parameters θ,
it is assumed to evolve according to the master equation:
The transition probabilities Γ are assumed to satisfy the detailed balance
condition:
p = Γ p
(9)
Γα→β p(∞)
α (θ) = Γβ→α p(∞)
β
(θ)
(10)
The Maximum Likelihood estimate θ of the parameters, given the data, is:
θML = argmin
DKL(P(0)P(∞)(θ))
θ
MPF proposes to consider instead, for infinitesimal ǫ:
θMPF = argmin
θ
DKL(p(0)p(ǫ)(θ))
It was proven in [33] that:
DKL(p(0)p(ǫ)(θ)) ≃
ǫ
NData Xα ∈Data
β /∈Data
Γα→β
The MPF strategy is therefore to search for:
min
θ Xα ∈Data
β /∈Data
Γα→β
(11)
(12)
(13)
(14)
4.3 Spiking network and Pseudo-Boltzmann learning
The single neuron obeys the sub-threshold dynamics:
Vi =
−Vi
τ
δ(t − tkj − dij) − gSF Aci + Ii
Jij Xkj
+Xj
+Xki
−ci
τSF A
ci =
δ(t − tki ),
(15)
with the additional condition that if Vi ≥ V th, the emission of a spike is recorded
and Vi is reset to a value H for the duration of a refractory period tref . In Eq. 15,
21
Vi is the membrane potential of neuron i, τ is the membrane time constant, Jij
is the synaptic efficacy from neuron j to neuron i, kj label spikes emitted by
neuron j, dij is the spike transmission delay from j to i. The neuron model
includes spike-frequency adaptation (SFA) in the form of a calcium-dependent
inhibitory (potassium) current: gSF A is the strength of SFA, ci represents the
intra-cellular calcium concentration, and τSF A is a characteristic time of calcium
dynamics. The ci variable acts as an integrator of the spiking activity of the
neuron i, such that SFA implements an activity-dependent self-inhibition. Ii
is an external current applied to the neuron, implemented as a Poisson process
with rate νext
.
i
E/I external spike trains through synapses with mean J ext
The network comprises 64 modules. Each module includes nE = 32 excita-
tory (E) neurons and nI = 16 inhibitory (I) neurons. E/I neurons within a
module are connected with probabilities cXY , X, Y = E/I; each E/I neuron
also receives Cext
E/I , and
with rate νext
init (init refers to the fact that external rates are set to an initial
value, then they are optimized during the learning process - see below). Delays
dij are drawn from an exponential distribution between dmin
XY , X, Y =
E/I, where dmin
XY = 0.1ms always. The values of these parameters are given in
Table 1. Inter-module synapses are only between excitatory neurons, and can
be positive or negative, which can can be viewed as an effective way to also
incorporate the excitatory input to inhibitory neurons; their mean efficacies are
determined through the learning procedure described below.
XY dmax
All the synapses in the network, given their (learnt or assigned) mean effi-
cacies, are drawn from a Gaussian distribution 25% relative variance.
The probability of connection between two excitatory neurons belonging to
different modules is 0.5 while the delays for inter-module communication have
dmax
EE = 50ms.
To choose the intra-module synaptic connectivity such that the isolated mod-
ule is approximately bistable, we used predictions from mean-field theory [8].
The bistable behavior and a narrow range of firing rates for the low and high
activity states proved to be robust against quenched noise due to different real-
izations of the synaptic matrices. This allowed us to define a threshold θbin on
the firing activity to binarize the dynamic states of the modules (the binariza-
tion is actually performed on a smoothed version of activity by averaging over
a 20 ms window). In this way, the dynamics of the network is represented by a
sequence of binary vectors corresponding to the high or low state of all modules.
The procedure to construct the inter-module connectivity (see below) preserves
to a large extent the modules' bistability for the interconnected network.
Numerical simulations were performed using the event-driven simulator we
described in [35].
The procedure to construct inter-module synaptic connectivity aims to "store"
prescribed patterns of activities using correlations between the binarized net-
work state vectors to build an effective Hopfield-like synaptic matrix. To store
patterns in the modular network we start from a Hopfield model with given
P stored patterns, from the simulation of which we estimate the mean single
spin magnetization mHopf
≡ hσpit, and spatial spin-spin correlation functions
cHopf
pq ≡ hσpσqit − hσpithσqit, where hit is the average over the whole time series.
, and the
corresponding quantities mp and cpq measured from the sequence of binarized
We then iteratively minimize the difference between mHopf
and cHopf
p
p
pq
22
Table 1: Parameters of the neural modules
Value
32 / 16
20 / 10 ms
2 / 1 ms
15 / 20 mV
0.5 / 0.5 / 1 / 1
0.72 / 1 / −2 / −0.0012 mV
21 / 21 / 1 / 1 ms
12 kHz
600 / 400
0.320 /0.111 mV
[62.5 − 4000] ms
25/τSF A
40 Hz
Parameter
nE / nI
τE / τI
tref
E / tref
I
H / V th
cEE / cIE / cEI / cII
JEE / JIE / JEI / JII
dmax
EE / dmax
IE / dmax
EI
/ dmax
II
νext
init
E / Cext
Cext
J ext
E / J ext
I
I
τSF A
gSF A
θbin
states of the multi-modular spiking network (such that p and q are module
labels). Such minimization is performed adopting an educated guess inspired
by Boltzmann learning that would be appropriate for a binary spin system, and
that we verify a posteriori. Specifically we implement the following procedure:
• Run a long simulation of the multi-modular system
• Binarize the activity of each module, obtaining a time-series of binary
state vectors, from which
• Measure the spatial connected correlations cpq and magnetizations mp
• Perform a step of the pseudo-Boltzmann iteration over the values of Jpq
, where Jpq is the average synaptic efficacy from neurons belonging
is the common
and νext
to module q to neurons belonging to module p; and νext
value of νext for all neurons in module p
p
p
∆J (t)
∆ν(t)
pq = −sign(c(t)
extp = −sign(m(t)
pq − cData
p − mData
) ∆(t)
pq
) ∆(t)
p
pq
p
(16)
with the learning rates ∆(t)
pq and ∆ν(t) determined by the Rprop algorithm:
∆(t)
pq =
pq
min(η+∆(t−1)
max(η−∆(t−1)
∆(t−1)
pq
pq
, ∆max)
, ∆min)
)(c(t)
)(c(t)
pq − cData
pq − cData
pq
pq
pq
pq − cData
pq − cData
if (c(t−1)
if (c(t−1)
otherwise
0 < η− < 1 < η+
pq
) > 0
) < 0
(17)
23
∆(t)
p =
p
min(η+∆(t−1)
max(η−∆(t−1)
∆(t−1)
p
p
, ∆max)
, ∆min)
)(m(t)
)(m(t)
p − mData
p − mData
p
p
p
p
p
− mData
− mData
if (m(t−1)
if (m(t−1)
otherwise
0 < η− < 1 < η+
p
) > 0
) < 0
(18)
In essence, Rprop adapts the learning rate for each parameter only using
the sign of the derivatives: when the sign of the derivative does not change
between two successive iteration steps, learning should speed up, while
when the sign of the derivative changes it should slow down.
We checked that, although every simulation during the training procedure
starts from the same exact condition, discarding the first 5 × τSF A steps of each
run is enough to ensure that no appreciable correlations are detectable between
different realizations.
4.4 Complexity measure of dynamics in the centroid space
After performing MS clustering on the time series generated by a simulation of
the multi-modular spiking network with given τSF A and gSF A, to each of the 64-
dimensional vectors containing the modules' firing activities we substitute the
label of the centroid that vector belongs to. In this way, the multi-dimensional
dynamics is converted into a symbolic sequence of centroid labels.
From such label sequence we measure the matrix of transition probabilities
between all pairs of centroids, and generate surrogate label sequences as Markov
processes with the same transition probabilities as the actual sequence. While,
by construction, the surrogate sequences are memoryless, this may not be the
case for the actual one. We focus on such possible memory effects, which are
in general interesting to uncover, and for this purpose we adopt a measure
typically used to evaluate the complexity of symbolic sequences, the Lempel-
Ziv (LZ) complexity, which is a measure of compressibility, suitably normalized
in order to eliminate trivial dependence on the length of the sequence.
The LZ complexity is computed as [24]
CLZ =
SLZ
S log(A)/ log(S)
(19)
where SLZ is the length of the compressed sequence, S is the length of the
sequence, A is the length of the alphabet.
And the relative complexity index is defined as:
R =
CMarkov
LZ
− CSample
LZ
CMarkov
LZ
(20)
where CSample
LZ
is the LZ complexity of the actual centroids sequence, and
is the average LZ complexity of the surrogate Markov centroids se-
CMarkov
LZ
quences.
24
5 Acknowledgements
We thank Maurizio Mattia and Cristiano Capone for a critical reading of a
previous version of the manuscript, and Giorgio Parisi for useful discussions in
the early stage of the work. We are grateful to Paolo Barucca for several con-
structive interactions during the initial phase of the work. Gabriel Baglietto
acknowledges financial support from CONICET. Paolo Del Giudice was sup-
ported in part by the EU Grant WaveScalES (EC FET Flagship HBP SGA1
720270)
References
[1] Stevenson IH, Kording KP (2011) How advances in neural recording affect
data analysis. Nature neuroscience 14: 139 -- 142.
[2] Schneidman E, Berry MJ, Segev R, Bialek W (2006) Weak pairwise cor-
relations imply strongly correlated network states in a neural population.
Nature 440: 1007 -- 1012.
[3] Hertz JA, Roudi Y, Tyrcha J (2013) Ising models for inferring network
structure from spike data. In: Quiroga RQ, Panzeri S, editors, The Oxford
Handbook of Innovation, CRC Press, chapter 27. pp. 527-546.
[4] Roudi Y, Dunn B, Hertz J (2015) Multi-neuronal activity and functional
connectivity in cell assemblies. Current opinion in neurobiology 32: 38 -- 44.
[5] Capone C, Filosa C, Gigante G, Ricci-Tersenghi F, Del Giudice P (2015)
Inferring synaptic structure in presence of neural interaction time scales.
PloS one 10: e0118412.
[6] Rodriguez A, Laio A (2014) Clustering by fast search and find of density
peaks. Science 344: 1492 -- 1496.
[7] Fukunaga K, Hostetler L (2014) The estimation of the gradient of a density
Information Theory,
function, with applications in pattern recognition.
IEEE Tran 21: 32 -- 40.
[8] Mattia M, Pani P, Mirabella G, Costa S, Del Giudice P, et al. (2013)
Heterogeneous attractor cell assemblies for motor planning in premotor
cortex. J Neurosci 33: 11155 -- 11168.
[9] Latimer KW, Yates JL, Meister ML, Huk AC, Pillow JW (2015) Single-trial
spike trains in parietal cortex reveal discrete steps during decision-making.
Science 349: 184 -- 187.
[10] Hopfield JJ (1982) Neural networks and physical systems with emergent
collective computational abilities. Proceedings of the National Academy of
Sciences 79: 2554-2558.
[11] Amit DJ (1992) Modeling brain function: The world of attractor neural
networks. Cambridge University Press.
25
[12] Abeles M, Bergman H, Gat I, Meilijson I, Seidemann E, et al. (1995) Corti-
cal activity flips among quasi-stationary states. Proceedings of the National
Academy of Sciences 92: 8616 -- 8620.
[13] Luczak A, Barth P, Marguet SL, Buzski G, Harris KD (2007) Sequential
structure of neocortical spontaneous activity in vivo. Proc Natl Acad Sci
U S A 104: 347 -- 352.
[14] Mazzucato L, Fontanini A, La Camera G (2015) Dynamics of multistable
states during ongoing and evoked cortical activity. The Journal of Neuro-
science 35: 8214 -- 8231.
[15] Abdollah-nia MF, Saeedghalati M, Abbassian A (2012) Optimal region
of latching activity in an adaptive potts model for networks of neurons.
Journal of Statistical Mechanics: Theory and Experiment 2012: P02018.
[16] Duarte RC, Morrison A (2014) Dynamic stability of sequential stimulus
representations in adapting neuronal networks. Frontiers in computational
neuroscience 8: 124.
[17] Huguet G, Rinzel J, Hup´e JM (2014) Noise and adaptation in multistable
perception: Noise drives when to switch, adaptation determines percept
choice. Journal of vision 14: 19 -- 19.
[18] Roach JP, Sander LM, Zochowski MR (2016) Memory recall and spike-
frequency adaptation. Physical Review E 93: 052307.
[19] Deco G, Rolls ET (2005) Sequential memory: a putative neural and synap-
tic dynamical mechanism. Journal of Cognitive Neuroscience 17: 294 -- 307.
[20] Theodoni P, Kov´acs G, Greenlee MW, Deco G (2011) Neuronal adaptation
effects in decision making. The Journal of Neuroscience 31: 234 -- 246.
[21] Mattia M, Sanchez-Vives MV (2012) Exploring the spectrum of dynamical
regimes and timescales in spontaneous cortical activity. Cognitive neuro-
dynamics 6: 239 -- 250.
[22] Gigante G, Mattia M, Del Giudice P (2007) Diverse population-bursting
modes of adapting spiking neurons. Physical Review Letters 98: 148101.
[23] Akrami A, Russo E, Treves A (2012) Lateral thinking, from the hopfield
model to cortical dynamics. Brain research 1434: 4 -- 16.
[24] Lempel A, Ziv J (1976) On the complexity of finite sequences. IEEE Trans-
actions on information theory 22: 75 -- 81.
[25] Kaspar F, Schuster H (1987) Easily calculable measure for the complexity
of spatiotemporal patterns. Physical Review A 36: 842.
[26] Ab´asolo D, Simons S, da Silva RM, Tononi G, Vyazovskiy VV (2015)
Lempel-ziv complexity of cortical activity during sleep and waking in rats.
Journal of neurophysiology 113: 2742 -- 2752.
[27] Amig´o JM, Szczepa´nski J, Wajnryb E, Sanchez-Vives MV (2004) Esti-
mating the entropy rate of spike trains via lempel-ziv complexity. Neural
Computation 16: 717 -- 736.
26
[28] Casali AG, Gosseries O, Rosanova M, Boly M, Sarasso S, et al. (2013) A
theoretically based index of consciousness independent of sensory process-
ing and behavior. Science translational medicine 5: 198ra105 -- 198ra105.
[29] Dawes RM (1979) The robust beauty of improper linear models in decision
making. American psychologist 34: 571.
[30] Riedmiller M, Braun H (1992) Rprop - a fast adaptive learning algorithm.
Technical report, Proc. of ISCIS VII), Universitat.
[31] Igel C, Husken M (2003) Empirical evaluation of the improved rprop learn-
ing algorithms. Neurocomputing 50: 105 -- 123.
[32] Samaria F, Harter A (1994) Parameterisation of a stochastic model for
human face identification.
In: Applications of Computer Vision, 1994.,
Proceedings of the Second IEEE Workshop on. pp. 138-142. doi:10.1109/
ACV.1994.341300.
[33] Sohl-Dickstein J, Battaglino PB, DeWeese MR (2011) New method for
parameter estimation in probabilistic models: Minimum probability flow.
Phys Rev Lett 107: 220601.
[34] Roudi Y, Tyrcha J, Hertz J (2009) Ising model for neural data: Model
quality and approximate methods for extracting functional connectivity.
Phys Rev E 79: 051915.
[35] Mattia M, Del Giudice P (2000) Efficient event-driven simulation of large
networks of spiking neurons and dynamical synapses. Neural Computation
12: 2305 -- 2329.
27
|
1504.07422 | 5 | 1504 | 2015-08-23T09:01:41 | Is there contextuality in behavioral and social systems? | [
"q-bio.NC",
"math.PR",
"quant-ph"
] | Most behavioral and social experiments aimed at revealing contextuality are confined to cyclic systems with binary outcomes. In quantum physics, this broad class of systems includes as special cases Klyachko-Can-Binicioglu-Shumovsky-type, Einstein-Podolsky-Rosen-Bell-type, and Suppes-Zanotti-Leggett-Garg-type systems. The theory of contextuality known as Contextuality-by-Default allows one to define and measure contextuality in all such system, even if there are context-dependent errors in measurements, or if something in the contexts directly interacts with the measurements. This makes the theory especially suitable for behavioral and social systems, where direct interactions of "everything with everything" are ubiquitous. For cyclic systems with binary outcomes the theory provides necessary and sufficient conditions for noncontextuality, and these conditions are known to be breached in certain quantum systems. We review several behavioral and social data sets (from polls of public opinion to visual illusions to conjoint choices to word combinations to psychophysical matching), and none of these data provides any evidence for contextuality. Our working hypothesis is that this may be a broadly applicable rule: behavioral and social systems are noncontextual, i.e., all "contextual effects" in them result from the ubiquitous dependence of response distributions on the elements of contexts other than the ones to which the response is presumably or normatively directed. | q-bio.NC | q-bio |
Is there contextuality in behavioral and social systems?
Ehtibar N. Dzhafarov1, Ru Zhang1 and Janne Kujala2
1Purdue University, [email protected]
2University of Jyvaskyla, [email protected]
Abstract
Most behavioral and social experiments aimed at revealing contextuality are confined to cyclic
systems with binary outcomes.
In quantum physics, this broad class of systems includes as
special cases Klyachko-Can-Binicioglu-Shumovsky-type, Einstein-Podolsky-Rosen-Bell-type, and
Suppes-Zanotti-Leggett-Garg-type systems. The theory of contextuality known as Contextuality-
by-Default allows one to define and measure contextuality in all such system, even if there are
context-dependent errors in measurements, or if something in the contexts directly interacts with
the measurements. This makes the theory especially suitable for behavioral and social systems,
where direct interactions of "everything with everything" are ubiquitous. For cyclic systems with
binary outcomes the theory provides necessary and sufficient conditions for noncontextuality, and
these conditions are known to be breached in certain quantum systems. We review several be-
havioral and social data sets (from polls of public opinion to visual illusions to conjoint choices to
word combinations to psychophysical matching), and none of these data provides any evidence for
contextuality. Our working hypothesis is that this may be a broadly applicable rule: behavioral and
social systems are noncontextual, i.e., all "contextual effects" in them result from the ubiquitous
dependence of response distributions on the elements of contexts other than the ones to which the
response is presumably or normatively directed.
Keywords:contextuality, cyclic systems, inconsistent connectedness
1
Introduction
Although the word is widely used in linguistics, psychology, and philosophy, the notion of contextuality
as it is used in this paper comes from quantum mechanics, where in turn it came from logic [1]. The
reason for the prominence of this notion in quantum theory is that classical-mechanical systems are
not contextual while some quantum-mechanical systems are. Contextuality is sometimes even presented
as one of the "paradoxes" of quantum mechanics.
In psychology, as it turns out, a certain variety
of (non)contextuality has been prominent too, but it is known under different name: selectiveness of
influences, or lack thereof (for details, see Refs. [2, 3]).
The term "contextuality" refers to properties of systems of random variables each of which can be
viewed (sometimes artificially) as a measurement of some "object" in some context. For instance, an
object q may be a question, and the context may be defined by what other question q(cid:48) it is asked in
combination with. Then the answer to this question is a random variable R(q,q(cid:48))
that can be interpreted
q
1
Dzhafarov, Zhang, Kujala
2
as the measurement of q in the context (q, q(cid:48)). If the same question q is then asked in combination with
some other question q(cid:48)(cid:48), then the measurement is a different random variable, R(q,q(cid:48)(cid:48))
. More generally,
context in which q is measured is defined by the conditions c under which the measurement is made,
yielding random variable Rc
q. This notation (or one of numerous variants thereof) is called contextual
notation for random variables: it codifies the idea that the identity of a measurement is defined both
by what is measured and by the conditions under it is measured [4 -- 11].
q
q
q(cid:48)
and R(q,q(cid:48))
Within each context the measurements are made "together", because of which they have an empiri-
cally defined joint distribution. Thus, in context (q, q(cid:48)) we have two jointly distributed random variables
R(q,q(cid:48))
. We call the set of all random variables jointly recorded in a given context a bunch (of
random variables, or of measurements). Two different bunches have no joint distribution, because there
is no empirically defined way of coupling the values of one bunch with those of another. We say that
they are stochastically unrelated. Thus, in
R(q,q(cid:48)) = (R(q,q(cid:48))
) and R(q,q(cid:48)(cid:48)) = (R(q,q(cid:48)(cid:48))
, R(q,q(cid:48)(cid:48))
, R(q,q(cid:48))
(1)
)
q
q(cid:48)(cid:48)
q
q(cid:48)
any component of R(q,q(cid:48)) is stochastically unrelated to any component of R(q,q(cid:48)(cid:48)), including R(q,q(cid:48))
R(q,q(cid:48)(cid:48))
.
q
q
and
This work is based on the theory of contextuality dubbed Contextuality-by-Default (CbD) [5,6,12 -- 17]
(for precursors of this theory, see Refs. [9 -- 11]). On a very general level, its main idea is that
a system of different, stochastically unrelated bunches of random variables can be character-
ized by considering all possible ways in which they can be coupled under well-chosen con-
straints imposed, for each object, on the relationship between the measurements of this object
in different contexts.
To couple different bunches simply means to impose a joint distribution on them. In the example above,
this means finding four jointly distributed random variables (A, B, X, Y ) such that, in reference to (1),
(A, B) ∼ R(q,q(cid:48)) and (X, Y ) ∼ R(q,q(cid:48)(cid:48)),
(2)
q
and R(q,q(cid:48)(cid:48))
∼ standing for "is distributed as". The quadruple (A, B, X, Y ) is then called a coupling for the bunches
R(q,q(cid:48)) and R(q,q(cid:48)(cid:48)). The "well-chosen constraints" is a key notion in the formulation above.
In our
example, these constraints should apply to A and X, the coupling counterparts of R(q,q(cid:48))
and R(q,q(cid:48)(cid:48))
measuring (answering) the same question q in two different contexts.
Intuitively, "noncontextuality" means "independence of context", and because of this it is tempting
to say that the system of two bunches in (1) is noncontextual if we can consider R(q,q(cid:48))
as
"one and the same" random variable, Rq. This may appear simple, but in fact it is logically impossible:
since R(q,q(cid:48))
are stochastically unrelated, they cannot be "the same". A random variable
cannot be stochastically unrelated to itself. The precise meaning here comes from considering couplings
(A, B, X, Y ) for the two bunches. Clearly, in every such a coupling A ∼ R(q,q(cid:48))
. We
can say that the measurement of q in the system is context-independent if among all possible couplings
(A, B, X, Y ) there is at least one in which Pr[A (cid:54)= X] = 0.
In this particular example, due to its
simplicity (only three random variables involved in two contexts) it can be shown that such a coupling
does exist, provided R(q,q(cid:48))
. In a more complex system, such a coupling may not exists even if
the system is consistently connected : which means that in this system the measurements of one and the
same "object" always have the same distribution.
and X ∼ R(q,q(cid:48)(cid:48))
∼ R(q,q(cid:48)(cid:48))
and R(q,q(cid:48)(cid:48))
q
q
q
q
q
q
q
q
q
The traditional approaches to contextuality were confined to consistent connectedness, but this con-
dition is too restrictive in quantum physics [14,16,18] and virtually inapplicable in social and behavioral
Dzhafarov, Zhang, Kujala
3
q
q
and R(q,q(cid:48)(cid:48))
sciences: almost always, a response to question (or stimulus) q will depend on the context in which it
is asked, which may translate into R(q,q(cid:48))
having different distributions. There is nothing
wrong in calling any such a case contextual, and this is done by many (see Sections 3 and 6 below).
It is, however, more informative to separate inconsistent connectedness from contextuality, and this is
what is done in the CbD theory. We use the term inconsistently connected for the systems that are not
necessarily consistently connected (but may be so, as a special or limit case).
The logic of the CbD approach is as follows. We first consider separately the random variables
measuring the same object in different contexts, in our example R(q,q(cid:48))
. We call this set of
random variables the connection (for the measured object, in our case q). Among all possible couplings
(A(cid:48), X(cid:48)) for the connection {R(q,q(cid:48))
∼
R(q,q(cid:48))
(cid:54)= X(cid:48)]. Then we look at the entire
system of the bunches, in our case (1), and among all possible couplings (A, B, X, Y ) for this system
we find the minimal value m for Pr[A (cid:54)= X]. It should be clear that m(cid:48) cannot exceed m, because in
every coupling (A, B, X, Y ) for (1) the part (A, X) forms a coupling for the connection {R(q,q(cid:48))
, R(q,q(cid:48)(cid:48))
}.
But they can be equal, m = m(cid:48), and then we say that the system is noncontextual. If m > m(cid:48), the
system is contextual. Again, due to its simplicity, the system consisting of the two bunches (1) cannot
be contextual, but this may very well be the case in more complex systems.
}, i.e., among all jointly distributed (A(cid:48), X(cid:48)) such that A(cid:48)
, we find the minimal value m(cid:48) of Pr[A(cid:48)
∼ R(q,q(cid:48)(cid:48))
and R(q,q(cid:48)(cid:48))
, R(q,q(cid:48)(cid:48))
and X(cid:48)
q
q
q
q
q
q
q
q
As an example of the latter, consider a system with two bunches
) and R(q(cid:48),q) = (R(q(cid:48),q)
R(q,q(cid:48)) = (R(q,q(cid:48))
, R(q,q(cid:48))
, R(q(cid:48),q)
(3)
in which there are only two "objects" q, q(cid:48), and the two contexts differ in the order in which these objects
are measured. We have two connections here,
, R(q(cid:48),q)
, R(q(cid:48),q)
(4)
q(cid:48)
q(cid:48)
)
q
q
{R(q,q(cid:48))
q
q
} and {R(q,q(cid:48))
q(cid:48)
q(cid:48)
}.
q R(q,q(cid:48))
q R(q(cid:48),q)
It easy to see that, across all possible couplings (A(cid:48), X(cid:48)) for {R(q,q(cid:48))
(cid:54)= X(cid:48)] is 0, and the same is true for the minimum value m(cid:48)
Let us assume the measurements are binary, with values +1 and −1 (e.g., corresponding to answers
Yes and No), and let us further assume that all four random variables are "fair coins", with equal
probabilities of +1 and -1. Then the distribution of the bunches R(q,q(cid:48)) and R(q(cid:48),q) in (3) are uniquely
defined by the product expected values (cid:104)R(q,q(cid:48))
}, the minimum value m(cid:48)
(cid:54)= Y (cid:48)] across all possible
of Pr[A(cid:48)
couplings (B(cid:48), Y (cid:48)) for {R(q,q(cid:48))
}. However, it follows from the general theory that across all possible
couplings (A, B, X, Y ) for the entire system (3) the values m1 of Pr[A (cid:54)= X] and m2 of Pr[B (cid:54)= Y ] cannot
q R(q(cid:48),q)
(cid:105) = (cid:104)R(q(cid:48),q)
be both zero unless (cid:104)R(q,q(cid:48))
(cid:105). The latter need not be the case: it may, e.g., very
q R(q(cid:48),q)
q R(q,q(cid:48))
well be that (cid:104)R(q,q(cid:48))
(cid:105) = −1 (perfect anti-correlation).
q(cid:48)
In this case m1 + m2 ≥ 1, whence either m1 > m(cid:48)
1 = 0 or m2 > m(cid:48)
2 = 0, indicating that the system is
q(cid:48)
q(cid:48)
q R(q,q(cid:48))
(cid:105) = 1 (perfect correlation) and (cid:104)R(q(cid:48),q)
(cid:105).
, R(q(cid:48),q)
2 of Pr[B(cid:48)
(cid:105) and (cid:104)R(q(cid:48),q)
, R(q(cid:48),q)
q(cid:48)
q(cid:48)
q(cid:48)
q(cid:48)
q(cid:48)
1
q
q
contextual.
As we show in this paper, the general rule for a broad spectrum of behavioral and social systems of
measurements seems to be that they are all noncontextual in the sense of CbD.
2 Cyclic systems of arbitrary rank
In this section and throughout the rest of the paper we assume that all our measurements are binary
random variables, with values ±1.
Dzhafarov, Zhang, Kujala
4
We apply the logic of the CbD theory to systems in which all objects are measured in pairs so that
each object belongs to precisely two pairs. We call such systems cyclic, because we can enumerate the
objects in such a system q1, . . . , qn and arrange them in a cycle
q1
/ q2
/ ···
/ qn−1
/ qn
(5)
in which any two successive objects form a context. The number n is referred to as the rank of the
system. Our last example in the previous section is a cyclic system of rank 2, the smallest possible.
qi
qi
and R(qi(cid:9)1,qi)
In accordance with our notation, each object qi in a cyclic system is measured by two random
variables: R(qi,qi⊕1)
, where the operations ⊕ and (cid:9) are cyclic addition and subtraction
(so that n ⊕ 1 = 1 and 1 (cid:9) 1 = n). Since there are no other random variables involved, we can
, measuring the first object in the context, by Vi, and R(qi(cid:9)1,qi)
simplify notation: we will denote R(qi,qi⊕1)
,
measuring the second object in the context, by Wi. As a result each bunch in a cyclic system has the
form (Vi, Wi⊕1); e.g., the bunch of measurements for (q1, q2) is (V1, W2), for (qn, q1) the bunch is (Vn, W1),
etc.
qi
qi
Now we can represent a cyclic system of measurements in the form of a V − W cycle:
V1
W2
V2
W3
···
Vn
W1
(6)
where solid lines indicate bunches (joint measurements) and point lines indicate connections (measure-
ments of the object in different contexts).
It is proved in Refs. [14, 16, 17] that such a system is noncontextual if and only if its bunches satisfy
the following inequality:
∆C = s1((cid:104)V1W2(cid:105), . . . ,(cid:104)Vn−1Wn(cid:105),(cid:104)VnW1(cid:105)) − (n − 2) −
n(cid:88)
i=1 (cid:104)Vi(cid:105) − (cid:104)Wi(cid:105) ≤ 0,
(7)
where (cid:104)·(cid:105) denotes expected value, and the s1-part is the maximum of all linear combinations ±(cid:104)V1W2(cid:105)±
. . .±(cid:104)Vn−1Wn(cid:105)±(cid:104)VnW1(cid:105) with the proviso that the number of minuses is odd. Note that the criterion is
written entirely in terms of the expectations of Vi, Wi and of the products Vi, Wi⊕1 (i = 1, . . . , n). This
means that the information about a cyclic system we need can be presented in the form of the diagram
(cid:104)V1W2(cid:105)
(cid:104)W2(cid:105)
(cid:104)V1(cid:105)
(cid:104)V2W3(cid:105)
(cid:104)V2(cid:105)
(cid:104)VnW1(cid:105)
(cid:104)W1(cid:105)
(cid:104)Vn(cid:105)
(8)
···
We will use such diagrams to discuss experimental data in the subsequent sections.
This criterion of noncontextuality is generally breached by quantum-mechanical systems. Thus,
for consistently connected systems, for n = 3, the inequality reduces to Suppes-Zanotti-Leggett-Garg
inequality [19, 20], for n = 4 it acquires the form of the Clauser-Horn-Shimony-Holt inequalities for the
Einstein-Podolsky-Rosen-Bell paradigm [21 -- 23], and for n = 5 (with an additional constraint) it becomes
what is known as Klyachko-Can-Binicioglu-Shumovsky inequality [24]. All of them are predicted by
quantum theory and supported by experiments to be violated by some quantum-mechanical systems.
For n = 3, using the criterion (7), violations are also predicted for inconsistently connected systems [18];
and for n = 5 violations of (7) were demonstrated experimentally [25] (as analyzed in Ref. [16]).
By contrast, we find no violations of (7) in all known to us behavioral and social experiments aimed
at revealing contextuality: ∆C never exceeds zero.
In the subsequent sections we demonstrate this
"failure to fail" the noncontextuality criterion on several experimental studies, for cyclic systems of rank
2, 3, and 4.
/
/
/
/
k
k
Dzhafarov, Zhang, Kujala
5
3 Question order effect (cyclic systems of rank 2)
Wang, Solloway, Shiffrin, and Busemeyer [26] considered 73 polls in which two questions, A and B
(playing the role of "objects" q1, q2 being measured), were asked in two possible orders, A → B and
B → A (forming two contexts). The possible answers to each question, random variables
V1 = RA→B
A
, W2 = RA→B
, V2 = RB→A
, W1 = RB→A
A
B
B
,
(9)
were binary: +1 (Yes) or −1 (No). For instance, in the Gallup poll results used in Ref. [27], one pair of
questions was (paraphrasing)
A: Do you think many white people dislike black people?
B: Do you think many black people dislike white people?
with the resulting estimates of joint and marginal probabilities
A → B Yes to B
.3987
Yes to A
.5599
.4161
N .=500
.5391
N .=500
Yes to B B → A
.4012
Yes to A
.4609
We translate "Yes to A" into V1 = 1 in A → B and into W1 = 1 in B → A; correspondingly, "Yes to
B" translates into W2 = 1 in A → B and into V2 = 1 in B → A. Using the notation (8), we deal here
with the system
(cid:104)V1(cid:105) (cid:104)V1W2(cid:105) (cid:104)W2(cid:105)
−.1678
.1198
.6428
(cid:104)V2W1(cid:105)
(cid:104)W1(cid:105)
(cid:104)V2(cid:105)
=
.0782
.6048
−.0782
To make sure the calculations are clear, for any ±1 random variables X, Y ,
(cid:104)XY (cid:105) = Pr[X = Y ] − Pr[X (cid:54)= Y ] = 4 Pr[X = 1, Y = 1] − 2 Pr[X = 1] − 2 Pr[Y = 1] + 1.
(cid:104)X(cid:105) = 2 Pr[X = 1] − 1,
The noncontextuality criterion (7) for cyclic systems of rank 2 specializes to the form
∆C = (cid:104)V1W2(cid:105) − (cid:104)V2W1(cid:105) − ((cid:104)V1(cid:105) − (cid:104)W1(cid:105) + (cid:104)V2(cid:105) − (cid:104)W2(cid:105)) ≤ 0.
(10)
For the values in the diagram above, ∆C = −0.406, so there is no evidence the system is contextual.
Ref. [26] contains analysis of 73 such pairs of questions, including 66 taken from PEW polls (with
N ranging from 125 to 927), four taken from Gallup polls reported by Moore [27] (with N about 500),
and three pairs of questions with N ranging from 106 to 305. (The data were kindly provided to us by
the authors of Ref. [26]; our computations based of these data are shown in supplementary file S1.)
The analysis is simplified if we accept the empirical regularity discovered by Wang and Busemeyer [28]
and convincingly corroborated in Ref. [26]: using our notation, the discovery is that for vast majority
of question pairs,
while
(cid:104)V1W2(cid:105) = (cid:104)V2W1(cid:105),
(cid:104)V1(cid:105) − (cid:104)W1(cid:105) + (cid:104)V2(cid:105) − (cid:104)W2(cid:105) (cid:54)= 0.
(11)
(12)
The last inequality is what traditionally called the question order effect [27], and (11) is dubbed by Wang
and Busemeyer the quantum question (QQ) equality. Wang and Busemeyer [28] theoretically justify
Dzhafarov, Zhang, Kujala
6
the QQ equality by positing that the process of answering two successive questions can be modeled by
orthogonally projecting a state vector ψ twice in a succession in a Hilbert space. Denoting the projectors
corresponding to response Yes to the questions A and B by P and Q, respectively, we have P 2 = P ,
Q2 = Q. The orthogonal projectors corresponding to response No to the same two questions are then
I − P and I − Q, with I denoting the identity operator. We have, for the question order A → B,
= (cid:107)QP ψ(cid:107)2 + (cid:107)(I − Q)(I − P )ψ(cid:107)2 = (cid:104)(P QP + (I − P )(I − Q)(I − P ))ψ ψ(cid:105),
1 + (cid:104)V1W2(cid:105)
2
and it is readily shown that
P QP + (I − P )(I − Q)(I − P ) = I − (P + Q) + (P Q + QP ).
As P and Q enter in this expression symmetrically, the expression is precisely the same for
1 + (cid:104)V2W1(cid:105)
2
= (cid:107)P Qψ(cid:107)2 + (cid:107)(I − P )(I − Q)ψ(cid:107)2.
The empirical QQ effect now follows from the assumption that the operators P, Q do not vary across
respondents (being determined by the questions alone), whereas the mixture of the initial states ψ has
the same distribution in any two large groups of respondents. At the same time, the question order
effect follows from the fact that (cid:107)QP ψ(cid:107)2 is not the generally the same as (cid:107)P Qψ(cid:107)2.
The QQ equality trivially implies (10), i.e., lack of contextuality. Therefore, to the extent the QQ
equality can be viewed as an empirical law (and Ref. [26] demonstrates this convincingly for 72 out of
73 question pairs), the criterion of noncontextuality should be satisfied for any (cid:104)V1(cid:105),(cid:104)W1(cid:105),(cid:104)V2(cid:105),(cid:104)W2(cid:105).
We can confirm and complement the statistical analysis presented in Ref. [26] of the 72 questions by
pointing out that the overall chi-square test of the equality (11) over all of them yields p > 0.35, df = 72.
The singled out pair of questions that violates the QQ equality is taken from the Gallup poll study
reported in Ref. [27]: paraphrasing,
A: Should Pete Rose be admitted to the baseball hall of fame?
B: Should shoeless Joe Jackson be admitted to the baseball hall of fame?
Refs. [26, 28] provide an explanation for why the double-projection model should not apply to this
particular pair of questions, but we need not be concerned with it. The diagram of the results for this
pair is
(cid:104)V1(cid:105) (cid:104)V1W2(cid:105) (cid:104)W2(cid:105)
.3241
.6190
−.2886
(cid:104)V2W1(cid:105)
(cid:104)W1(cid:105)
(cid:104)V2(cid:105)
=
,
−.0346
.3162
.0780
and it is readily seen to violate the equality (cid:104)V1W2(cid:105) = (cid:104)V2W1(cid:105) (p < 10−7, chi-square test with df = 1).
At the same time the diagram yields ∆C = −0.422, no evidence of contextuality. This example serves
as a good demonstration for the fact that while the QQ equality is a sufficient condition for lack of
contextuality, it is by no means necessary.
Considering the question pairs one by one, all but six ∆C values out of 73 are negative. In five
of these six cases, the QQ equality (cid:104)V1W2(cid:105) − (cid:104)V2W1(cid:105) = 0 cannot be rejected with p-values ranging
from 0.06 to 0.47. Therefore (10) cannot be rejected either. In the remaining case, p-value for the QQ
equality is 0.008, and ∆C = 0.063. While this case is suspicious, we do not think it warrants a special
Dzhafarov, Zhang, Kujala
7
Figure 1: Shroder's staircases used in the experiments reported in Ref. [29]
investigation: using conventional significance values, say, 0.01, for 73 similar cases we get the probability
of at least one rejection inflated to 0.52.
Note that in the literature cited, including Refs. [26, 28], the term "contextual effect" is used to
designate the question order effect (12). This meaning of contextuality corresponds to what we call here
inconsistent connectedness (or violations of marginal selectivity), and it should not be confused with
the meaning of contextuality as defined in Sections 1 and 2 and indicated by the sign of ∆C.
4 Schroder's staircase illusion (a cyclic system of rank 3)
Asano, Hashimoto, Khrennikov, Ohya, and Tanaka [29] studied a cyclic system of rank 3, using as
"objects" q1, q2, q3 Shroder's staircases tilted at three different angles, θ = 40, 45, 50 degrees, as shown
in Figure 1. In fact, these three angles formed the middle part of a set of 11 angles ranging from 0 to 90
degrees and presented either in the descending order (context c1), or in the ascending order (context c2),
or else in a random order (context c3). Each context involved a separate set of about 50 participants,
and each participant in response to each of 11 angles had to indicate whether she/he sees the surface
A in front of B (+1) or B in front of A (−1). From these 11 responses, in each context, the authors
selected two. In context c1 the selected responses where those to θ = 40, 45 deg, so, formally, c1 can
be identified with (q1, q2); in contexts c2 and c3 the selected responses were those to θ = 45, 50 deg
and to θ = 50, 40 deg, respectively, making c2 = (q2, q3) and c3 = (q3, q1). It is irrelevant to the logic
of the analysis that each context in fact contained all three tilts q1, q2, q3, as well as eight other tilts.
(Ref. [29] includes a variety of other combinations of three objects and three contexts extracted from
the experiment in question. The data set for the combination described here was kindly made available
to us by the authors of Ref. [29].)
The results of the experiment are shown in the diagram of expected values below:
(cid:104)V1(cid:105) (cid:104)V1W2(cid:105) (cid:104)W2(cid:105)
.708
.417
.625
(cid:104)W1(cid:105)
(cid:104)V3W1(cid:105)
(cid:104)V2W3(cid:105)
(cid:104)V2(cid:105)
=
.382
(cid:104)V3(cid:105)
(cid:104)W3(cid:105)
−.333
.127
.625
−.345
−.625
θ
Dzhafarov, Zhang, Kujala
The criterion of noncontextuality for a rank 3 cyclic system has the form
∆C = s1((cid:104)V1W2(cid:105),(cid:104)V2W3(cid:105),(cid:104)V3W1(cid:105)) − 1 −
3(cid:88)
i−1 (cid:104)Vi(cid:105) − (cid:104)Wi(cid:105) ≤ 0
8
(13)
where
s1(x, y, z) = max(x + y − z, x − y + z,−x + y + z,−x − y − z).
The calculation shows ∆C = −1.233, no evidence for contextuality.
Search for contextuality is the specific goal of Ref. [29], but the meaning of the concept there is
different from ours: there, it means violations of the Suppes-Zanotti-Leggett-Garg inequality (which is
the consistently connected case of (13)), irrespective of whether these violations are due to inconsistent
connectedness or due to contextuality in our sense.
5 Conjoint choices: Animals and sounds they make (a cyclic
system of rank 4)
Aerts, Gabora, and Sozzo [30] present results of an experiment in which each of 81 participants had to
choose between two animals and between two animal sounds, under four conditions c1, c2, c3, c4 (con-
texts), as shown below:
c1
Horse
Bear
V1
c2
Tiger
Cat
W3
W2
Growls Whinnies
.049
.259
.308
.630
.062
.692
V2
Growls Whinnies
.778
.086
.864
.086
.049
.135
.679
.321
.864
.135
c4
Horse
Bear
V4
Snorts Meows
.593
.296
.889
.025
.086
.111
c3
Tiger
Cat
W4
Snorts Meows
.148
.099
.247
.086
.667
.753
.618
.382
W1
.234
.766
V3
The "objects" to be measured here are the choices offered:
q1 = Horse or Bear? q2 = Growls or Whinnies?
q3 = Tiger or Cat?
q4 = Snorts or Meows?
Each of the four contexts corresponds to a pair of these objects,
c1 = (q1, q2), c2 = (q2, q3), c3 = (q3, q4), c4 = (q4, q1),
and the choices made are binary measurements (random variables)
c1
c2
c3
c4
(V1, W2)
(V2, W3)
(V3, W4)
(V4, W1)
.
The table of the results above translates into the diagram of expected values
Dzhafarov, Zhang, Kujala
9
(cid:104)V1(cid:105) (cid:104)V1W2(cid:105) (cid:104)W2(cid:105)
.358 −.778
−.384
(cid:104)W1(cid:105)
(cid:104)V4W1(cid:105)
(cid:104)V4(cid:105)
(cid:104)V2(cid:105)
.236
(cid:104)V2W3(cid:105)
=
.358
(cid:104)W3(cid:105)
.778
.728
.655
.728
(cid:104)V3W4(cid:105)
(cid:104)W4(cid:105)
(cid:104)V3(cid:105)
−.506
.630
−.532
The noncontextuality criterion for rank 4 has the form
∆C = s1((cid:104)V1W2(cid:105),(cid:104)V2W3(cid:105),(cid:104)V3W4(cid:105),(cid:104)V4W1(cid:105)) − 2 −
where
4(cid:88)
i=1 (cid:104)Vi(cid:105) − (cid:104)Wi(cid:105) ≤ 0,
s1(w, x, y, z) = max(w + x + y − z,w + x − y + z,w − x + y + z, − w + x + y + z).
The value computed from the data is ∆C = −3.357, providing no evidence for contextuality.
Ref. [30] reports that contextuality in this data set is present because
s1((cid:104)V1W2(cid:105),(cid:104)V2W3(cid:105),(cid:104)V3W4(cid:105),(cid:104)V4W1(cid:105)) − 2 > 0,
(14)
(15)
i.e., the data violate the classical CHSH inequalities [22, 23]. As pointed out in Ref. [31], the CHSH
inequalities are predicated on the assumption of consistent connectedness (marginal selectivity). With-
out this assumption they cannot be derived as a necessary or sufficient condition of noncontextuality,
and this assumption is clearly violated in the data. Aerts [32] has developed a theory which allows
for inconsistent connectedness, but it is unclear to us how this justifies the use of CHSH inequalties in
Ref. [30].
6 Word combinations and priming (cyclic systems of rank 4)
Bruza, Kitto, Ramm, and Sitbon [33] studied ambiguous two-word combinations, such as "apple chip".
One can understand this word combination to refer to an edible chip made of apples or to an apple
computer component. It is even possible to imagine such meanings as a piece chipped off of an apple
computer, or a computer component made of apples. In the experiments referred to the participants
were asked to explain how they understood the first and the second word in a combination: one meaning
of each word (e.g., the fruit meaning for "apple", the edible product meaning for "chip", etc.) can be
taken for +1, any other meaning being classified as −1. The meanings were determined by asking the
participants to explain how they understood the words. For each two-word combination the experi-
menters used one of four pairs of priming words presumably affecting the meanings. For the "apple
chip" combination, the priming words could be
q1 = banana
q2 = potato
q3 = computer q4 = circuit
,
Dzhafarov, Zhang, Kujala
forming four contexts
10
c1 = (banana, potato)
c3 = (computer, circuit) c4 = (circuit, banana)
c2 = (potato, computer)
.
The order in which we list the words in a context is chosen to create a cycle: (q1, q2), (q2, q3), etc.
Although this is not intuitive, formally, the measured "objects" here are the priming words q1, q2, q3, q4,
while the measurements are binary random variables indicating in what meaning (±1) the participant
understood "apple" and "chip". In (V1, W2) and (V3, W4) the V 's are meanings of "apple" and W 's
the meanings of "chip"; in (V2, W3) and (V4, W1) it is vice versa. (This is no more than a notational
convention, purely for the purposes of using the cyclic indexation.)
Ref. [33] presents data on 23 word combinations preceded by priming words (each combination in
each context being shown to each of 61-65 participants). In all 23 cases the computed values of ∆C are
negative, ranging from -2.882 to -0.418 (for the "apple chip" example the value is -1.640). We conclude,
once again, that there is no evidence in favor of contextuality. (The authors of Ref. [33] kindly provided
to us the word pairs and priming words, with the computed values of s1 and equivalents of (cid:104)Vi(cid:105)−(cid:104)Wi(cid:105)
(i = 1, . . . , 4), for all 23 word combinations; they are presented, with permission, in the supplementary
file S2, with the computation of ∆C added.)
The aim of Ref. [33] was not to study contextuality. Rather they were interested in the property called
compositionality, defined, in our terms, as consistent connectedness together with lack of contextuality.
Violations of this condition therefore amount to either inconsistent connectedness or, if connectedness
is consistent, to contextuality in our sense.
7 Psychophysical matching (cyclic systems of rank 4)
All experiments discussed so far use participants as replicants: the estimate of Pr[V = v, W = w]
in a given context is the proportion of participants who responded (v, w), v = ±1, w = ±1. In the
question order effect and Schroder's staircase illusion studies different groups of people participated
in different contexts, whereas the conjoint choices and word combinations studies employed repeated
measures design: each participant made one choice in each of the four contexts.
In our laboratory, we searched for possible contextual effects in a large series of psychophysical
experiments where each of very few (usually, three) participants were repeatedly "measuring" the same
four "objects" in the same four contexts. In each of the seven experiments the number of replications
per participant was 1000-2000, evenly divided between different contexts.
The logic of an experiment was as follows. The participant was shown two stimuli, target one
In each trial, the values
(T ) and adjustable one (A), both completely specified by two parameters.
α and β of these parameters (real numbers) in the target stimulus T (α, β) are fixed at one of several
values, each pair of values determining a context; in the adjustable stimulus the two parameters can be
simultaneously or (in some experiments) successively changed by the participant rotating a trackball.
At the end of each trial the participant reaches some values X and Y of these parameters that she/he
judges to make A(X, Y ) match (i.e., look the same as) T (α, β). In most experiments α and β vary on
several levels each (or even quasi-continuously within certain ranges), and we always choose four specific
values or subranges of their values: q1, q3 for α and q2, q4 for β. They form four contexts that can be
cyclically arranged as (q1, q2), (q2, q3), (q3, q4), (q4, q1), and for each of them we get empirical distributions
of X and Y : (X12, Y12) for context (q1, q2), (X41, Y41) for context (q4, q1), etc. In this notation, of the
two objects qi, qj, the random variable Xij "measures" the q with an odd index (1 or 3), whether i or j;
analogously, Yij "measures" the q with the even index.
Dzhafarov, Zhang, Kujala
11
Figure 2: Stimuli used in the matching experiments. The left panels show pairs of stimuli at the
beginning of a trial, the right panels show an intermediate stage in the matching process. Top panels:
in Experiments 1a-b there participants adjusted the position of the dot within a lower-right circle to
match a fixed position of the target dot in the upper-left circle. Middle panels:
in Experiments 2a-c
they adjusted the radii of two concentric circles on the right to match two fixed concentric circles on
the left. Bottom panels: in Experiments 3a-b they adjusted the amplitudes of two Fourier harmonics of
a floral shape on the right to match a fixed floral shape on the left. For details, see the supplementary
file S3.
The values of X and Y are then dichotomized in the following way: we choose a value xi and a value
yj (i = 1, 3, j = 2, 4) and define
The values of (x1, x3, y2, y4) can be chosen in a variety of ways, and for each choice we apply to the
obtained V and W variables the criterion (14).
As an example,
in one of the experiments the stimuli T and A were two dots in two circles,
like the ones shown in Figure 2, top, with a dot's position within a circle described in polar co-
ordinates (α and X denoting distance from the center in pixels, β and Y denoting angle in de-
grees measured counterclockwise from the horizontal rightward radius-vector). We extract from this
(cid:26) +1 if Xi,i⊕1 > xi
(cid:26) +1 if Xi(cid:9)1,i > xi
−1 if Xi,i⊕1 ≤ xi
−1 if Xi(cid:9)1,i ≤ xi
Vi =
Wi =
(cid:26) +1 if Yj,j⊕1 > yj
(cid:26) +1 if Yj(cid:9)1,j > yj
−1 if Yj,j⊕1 ≤ yj
−1 if Yj(cid:9)1,j ≤ yj
,
Vj =
, Wj =
.
.
(16)
(17)
Dzhafarov, Zhang, Kujala
12
Figure 3: Results for four contexts α (px)×β (deg) = {q1 = 53.67, q3 = 71.55}×{q2 = 63.43, q4 = 26.57}
extracted from Experiment 1a, participant P3, about 200 replications per context.
Dzhafarov, Zhang, Kujala
13
experiment a 2 × 2 subdesign as shown in Figure 3. Then we choose a value of x1 as any inte-
ger (in pixels) between max[minX12, minX41] and min[maxX12, maxX41], we choose y2 as any integer
(in degrees) between max[minY12, minY23] and min[maxY12, maxY23], and analogously for x3 and y4.
This yields 25 × 23 × 21 × 79 quadruples of (x1, x3, y2, y4), and the corresponding number of cyclic
systems of binary random variables (V1, W2, V2, W3, V3, W4, V4, W1). Consider, e.g., one such choice:
(x1, x3, y2, y4) = (72 px, 67 px, 60 deg, 23 deg). The diagram of this system is
(cid:104)V1(cid:105) (cid:104)V1W2(cid:105) (cid:104)W2(cid:105)
−.989
.211
−0.2
(cid:104)W1(cid:105)
(cid:104)V4W1(cid:105)
(cid:104)V4(cid:105)
(cid:104)V2(cid:105)
(cid:104)V2W3(cid:105)
=
(cid:104)W3(cid:105)
−0.902
.016
−0.006
0.300
.301
0.960
(cid:104)V3W4(cid:105)
(cid:104)W4(cid:105)
(cid:104)V3(cid:105)
0.167
.158
0.991
and the value of ∆C = −2.137, no evidence of contextuality. In fact negative values of ∆C are obtained
for all 25 × 23 × 21 × 79 dichotomizations. Clearly, different dichotomizations of the same random
variables are not stochastically independent, but there is no mathematical reason for ∆C to be of the
same sign in all of them.
In the supplementary file S3 we describe in detail how the dichotomizations were made, their number
ranging from 3024 to 11,663,568 per 2 × 2 (sub)design in each experiment for each participant. The
outcome is: not a single case with positive ∆C observed.
8 Conclusion
The empirical data analyzed above suggest that the noncontextuality boundaries, that are generally
breached in quantum physics, are not breached by behavioral and social systems. This may seem a
disappointing conclusion for some. With the realization that quantum formalisms may be used to con-
struct models in various areas outside physics [34 -- 37], the expectation was created that human behavior
should exhibit contextuality, perhaps even on a greater scale than allowed by quantum theory. However,
if the no-contextuality conclusion of the present paper is proved to be a rule for a very broad class of
behavioral and social systems, it is rather fortunate for behavioral and social sciences. Noncontextuality
means more constrained behavior, and constraints allow one to make predictions. The power of quantum
mechanics is not in that quantum systems breach the classical-mechanical bounds of noncontextuality,
but in the theory that imposes other, equally strict constraints. Presence of contextuality, in the absence
of a general theory like quantum mechanics, translates into unpredictability.
It must be noted that absence of contextuality in behavioral and social systems does not mean that
quantum formalisms are not applicable to them. A good argument for why this conclusion would be
groundless is provided by the question order effect discussed in Section 3: it is precisely the applicability
of a quantum-mechanical model in the question order effect analysis [26, 28] that allows one to predict
the lack of contextuality in this paradigm.
When discussing contextuality, one should be aware of the likelihood of purely terminological confu-
sions. It is clear that in the behavioral and social systems a context generally influences the measurement
of an object within it. For instance, the distribution of answers to a question depends on a question
Dzhafarov, Zhang, Kujala
14
asked and answered before it. One could call this contextuality, and many do. This is, however, a trivial
sense of contextuality, on a par with the fact that the distribution of answers to a question depends on
what this question is. One should not be surprised that other factors (such as temperature in the lab or
questions asked and answered previously) can influence this distribution too. We call this inconsistent
connectedness, and we offer a theory that distinguishes this ubiquitous feature from contextuality in a
different, one could argue more interesting meaning.
Acknowledgments
This research has been supported by NSF grant SES-1155956, AFOSR grant FA9550-14-1-0318, and A.
von Humboldt Foundation. We are grateful to the authors of Refs. [29], [33], and [26] for kindly providing
data sets for our analysis. We have benefited from discussions with Jan-Ake Larsson and Victor H.
Cervantes (who pointed out a mistake in an earlier version of the paper). The computations discussed
in Sections 3 and 6 are presented in the supplementary files S1 and S2, respectively. The original
data sets are available from the authors of Refs. [33] and [26]. Details of the experiments discussed in
Section 7 are presented in the supplementary file S3; the data sets are available as "Contextuality in
Psychophysical Matching", http://dx.doi.org/10.7910/DVN/OJZKKP, Harvard Dataverse, V1.
References
[1] Specker EP. 1960/1975 The logic of propositions which are not simultaneously decidable. In The
Logico-Algebraic Approach to Quantum Mechanics, The University of Western Ontario Series in
Philosophy of Science No. 5a, (ed. C. A. Hooker) pp. 135 -- 140. Netherlands: Springer.
[2] Dzhafarov EN, Kujala JV. 2012 Quantum entanglement and the issue of selective influences in
psychology: An overview. Lect. Notes Comput. Sc. 7620, 184-195.
[3] Dzhafarov EN, Kujala JV. 2012 Selectivity in probabilistic causality: Where psychology runs into
quantum physics. J. Math. Psychol. 56, 54-63.
[4] Dzhafarov EN, Kujala JV. 2013 All-possible-couplings approach to measuring probabilistic context.
PLoS One 8(5): e61712.
[5] Dzhafarov EN, Kujala, JV. 2014 Contextuality is about identity of random variables. Phys. Scripta
T163, 014009.
[6] Dzhafarov EN, Kujala JV, Larsson J-A, Cervantes VH. in press Contextuality-by-Default: A brief
overview of ideas, concepts, and terminology. Lect. Notes Comput. Sc. (available as arXiv:1412.4724.)
[7] Khrennikov A. 2005 The principle of supplementarity: A contextual probabilistic viewpoint to com-
plementarity, the interference of probabilities, and the incompatibility of variables in quantum me-
chanics. Found. Phys., 35, 1655 - 1693.
[8] Khrennikov A. 2009 Contextual Approach to Quantum Formalism. Berlin, Germany: Springer.
[9] Larsson J-A. 2002 A Kochen-Specker inequality. Europhys. Lett., 58, 799 -- 805.
[10] Svozil K. 2012 How much contextuality? Nat. Comput. 11, 261-265.
Dzhafarov, Zhang, Kujala
15
[11] Winter A. 2014 What does an experimental test of quantum contextuality prove or disprove? J.
Phys. A-Math. Theor. 47, 42403.
[12] Dzhafarov EN, Kujala JV. 2014 Embedding quantum into classical: Contextualization vs condi-
tionalization. PLoS One 9(3): e92818.
[13] Dzhafarov EN, Kujala JV. in press Conversations on contextuality. In Dzhafarov EN, Jordan JS,
Zhang R, Cervantes VH. (Eds) Contextuality from Quantum Physics to Psychology. New Jersey:
World Scientific. (available as arXiv:1508.00862.)
[14] Dzhafarov EN, Kujala JV, Larsson J-ı¿oe. 2015 Contextuality in three types of quantum-mechanical
systems. Found. Phys., 7, 762-782.
[15] Kujala JV, Dzhafarov EN. in press Probabilistic Contextuality in EPR/Bohm-type systems with
signaling allowed. In Dzhafarov EN, Jordan JS, Zhang R, Cervantes VH. (Eds) Contextuality from
Quantum Physics to Psychology. New Jersey: World Scientific.
[16] Kujala JV, Dzhafarov EN, Larsson J-A. 2015 Necessary and sufficient conditions for maximal
noncontextuality in a broad class of quantum mechanical systems. arXiv:1412.4724.
[17] Kujala JV, Dzhafarov EN. 2015 Proof of a conjecture on contextuality in cyclic systems with binary
variables. arXiv:1503.02181.
[18] Bacciagaluppi G. 2015 Leggett-Garg inequalities, pilot waves and contextuality. Int. J. Quantum
Found. 1, 1-17.
[19] Leggett A, Garg, A. 1985 Quantum mechanics versus macroscopic realism: Is the flux there when
nobody looks? Phys. Rev. Lett. 54, 857.
[20] Suppes P, Zanotti M. 1981 When are probabilistic explanations possible? Synthese 48, 191.
[21] Bell J. 1964 On the Einstein-Podolsky-Rosen paradox. Physics 1, 195-200.
[22] Clauser JF, Horne MA, Shimony A, Holt RA. 1969 Proposed experiment to test local hidden-
variable theories. Phys. Rev. Lett. 23, 880 -- 884.
[23] Fine A. 1982 Hidden variables, joint probability, and the Bell inequalities. Phys. Rev. Lett. 48,
291 -- 295.
[24] Klyachko AA, Can MA, Binicioglu S, Shumovsky AS. 2008 A simple test for hidden variables in
spin-1 system. Phys. Rev. Lett. 101, 020403.
[25] Lapkiewicz R, Li P, Schaeff C, Langford NK, Ramelow S, Wie´sniak M, Zeilinger A. 2011 Experi-
mental non-classicality of an indivisible quantum system. Nature 474, 490 -- 493.
[26] Wang Z, Solloway T, Shiffrin RM, Busemeyer JR. 2014 Context effects produced by question orders
reveal quantum nature of human judgments. Proc. Natl. Acad. Sci. 111, 9431-9436.
[27] Moore DW. 2002 Measuring new types of question-order effects. Public Opin. Quart. 66, 80-91.
[28] Wang Z, Busemeyer JR. 2013 A quantum question order model supported by empirical tests of an
a priori and precise prediction. Top. Cogn. Sci. 5, 689 -- 710.
Dzhafarov, Zhang, Kujala
16
[29] Asano M, Hashimoto T, Khrennikov A, Ohya M, Tanaka T. 2014 Violation of contextual general-
ization of the Leggett-Garg inequality for recognition of ambiguous figures. Phys. Scripta T 163,
014006.
[30] Aerts D, Gabora L, Sozzo S. 2013 Concepts and their dynamics: A quantum-theoretic modeling of
human thought. Top. Cogn. Sci. 5, 737-772.
[31] Dzhafarov EN, Kujala JV. 2014 Selective influences, marginal selectivity, and Bell/CHSH inequal-
ities. Top. Cogn. Sci. 6, 121 -- 128.
[32] Aerts D. 2014 Quantum and concept combination, entangled measurements, and prototype theory.
Top. Cogn. Sci. 6, 129-137.
[33] Bruza PD, Kitto K, Ramm BJ, Sitbon L. 2015 A probabilistic framework for analysing the com-
positionality of conceptual combinations. J. Math. Psychol. 67, 26-38.
[34] Busemeyer JR, Bruza PD. 2012 Quantum Cognition and Decision. Cambridge, UK: Cambridge
University Press.
[35] Haven E, Khrennikov A. 2012 Quantum Social Science. Cambridge, UK: Cambridge University
Press.
[36] Khrennikov A. 2010 Ubiquitous quantum structure: from psychology to finance. Heidelberg-Berlin-
New York: Springer.
[37] Ohya M, Volovich I. 2011 Mathematical Foundations of Quantum Information and Computation
and its Applications to Nano- and Bio-systems. Heidelberg-Berlin-New York: Springer.
s1--‐value<V1>--‐<W1>+<V2>--‐<W2>Delta
CDelta
C
>
0?QQ
chi^2QQ
p<0.05?QQ
p<0.01?Data
set
#order
A
--‐>
Border
B
--‐>
Aorder
A
--‐>
Border
B
--‐>
A<V1><W1><W2><V2><V1*W2><V2*W1>0.21050.0482--‐0.1100--‐0.14910.53590.47810.05780.2014--‐0.143500.9830001PEW0.25840.20810.00960.04070.56940.58820.01890.0814--‐0.062500.1150002PEW0.33980.27050.13590.09840.62140.58200.03940.1069--‐0.067500.2736003PEW--‐0.0484--‐0.0324--‐0.2823--‐0.21460.58060.57490.00570.0837--‐0.077900.0123004PEW0.12390.1367--‐0.11930.03250.61930.65290.03370.1646--‐0.130900.4268005PEW0.01360.1198--‐0.1900--‐0.06450.66060.64060.02010.2318--‐0.211700.1531006PEW0.04110.0206--‐0.2055--‐0.15120.63700.59450.04250.0748--‐0.032300.8474007PEW--‐0.1442--‐0.0946--‐0.3759--‐0.26340.59810.66750.06940.1620--‐0.092601.6353008PEW--‐0.0494--‐0.0847--‐0.2565--‐0.20370.64240.61560.02680.0881--‐0.061300.2559009PEW--‐0.0432--‐0.1231--‐0.2815--‐0.35880.63040.56330.06710.1571--‐0.090002.02350010PEW--‐0.2070--‐0.1544--‐0.3815--‐0.23540.62590.51390.11200.1987--‐0.086603.69790011PEW0.35150.3293--‐0.4387--‐0.46220.16620.11180.05440.04570.008710.52720012PEW0.30990.1750--‐0.3568--‐0.36500.27700.37000.09300.1431--‐0.050001.99500013PEW0.19790.2245--‐0.3298--‐0.28570.35620.38780.03160.0707--‐0.039100.22270014PEW0.17150.0633--‐0.3487--‐0.33210.39500.48940.09450.1248--‐0.030402.88340015PEW0.17890.0691--‐0.4252--‐0.43620.33720.33510.00210.1207--‐0.118500.00090016PEW0.06510.0815--‐0.4943--‐0.37180.38700.41150.02460.1389--‐0.114300.18380017PEW0.0805--‐0.0056--‐0.5862--‐0.55620.31030.29210.01820.1161--‐0.097900.06420018PEW0.02910.0132--‐0.5058--‐0.51450.37210.31400.05810.02460.033510.69170019PEW0.1358--‐0.0171--‐0.2741--‐0.38880.50120.46210.03910.2676--‐0.228500.40580020PEW0.0074--‐0.0076--‐0.4604--‐0.45820.46290.40250.06030.01720.043211.78960021PEW0.0030--‐0.0877--‐0.5137--‐0.48540.42250.35670.06580.1191--‐0.053300.85540022PEW--‐0.0620--‐0.0392--‐0.5682--‐0.51960.43420.31070.12350.07150.052013.48200023PEW0.02780.0997--‐0.4889--‐0.43940.40560.31000.09560.1215--‐0.025901.91470024PEW0.0290--‐0.0261--‐0.5411--‐0.47740.38160.36820.01350.1187--‐0.105300.04410025PEW0.0230--‐0.03000.0138--‐0.0700--‐0.4286--‐0.30000.12860.1369--‐0.008301.98700026PEW--‐0.0876--‐0.0729--‐0.0046--‐0.0729--‐0.2258--‐0.16670.05910.0829--‐0.023800.37190027PEW0.0486--‐0.0960--‐0.1243--‐0.1818--‐0.2000--‐0.12630.07370.2021--‐0.128401.06970028PEW--‐0.0611--‐0.1761--‐0.1444--‐0.21590.0056--‐0.03980.04530.1865--‐0.141200.36580029PEW--‐0.1940--‐0.1705--‐0.2239--‐0.17050.13430.06500.06930.0770--‐0.007700.93380030PEW--‐0.1279--‐0.2046--‐0.0444--‐0.1124--‐0.0548--‐0.11820.06330.1447--‐0.081400.73740031PEW--‐0.2185--‐0.3063--‐0.2185--‐0.18480.0231--‐0.02280.04590.1215--‐0.075600.41330032PEW--‐0.2204--‐0.40180.08630.1239--‐0.1757--‐0.11780.05790.2190--‐0.161100.55150033PEW--‐0.3199--‐0.3651--‐0.0428--‐0.0370--‐0.0428--‐0.07940.03650.0510--‐0.014400.25970034PEW--‐0.4142--‐0.3803--‐0.1420--‐0.21690.04140.08170.04030.1088--‐0.068500.28200035PEW--‐0.2845--‐0.2044--‐0.0264--‐0.2044--‐0.1261--‐0.08990.03620.2581--‐0.221900.23440036PEW--‐0.3433--‐0.3008--‐0.1164--‐0.0786--‐0.03880.00810.04690.0803--‐0.033400.38780037PEW--‐0.3598--‐0.4525--‐0.2691--‐0.18440.0255--‐0.05030.07580.1775--‐0.101701.02080038PEW--‐0.2978--‐0.2213--‐0.2089--‐0.1762--‐0.1022--‐0.04510.05710.1091--‐0.052000.76980039PEW--‐0.4000--‐0.3265--‐0.3158--‐0.1877--‐0.0526--‐0.14650.09390.2017--‐0.107801.71200040PEW--‐0.4677--‐0.4707--‐0.4160--‐0.29770.0801--‐0.03820.11830.1213--‐0.003102.72890041PEW0.05760.0335--‐0.02620.11960.68590.68420.00170.1699--‐0.168200.00050042PEW--‐0.03430.0554--‐0.02180.03210.76320.79010.02680.1435--‐0.116700.30150043PEW--‐0.1736--‐0.1304--‐0.1460--‐0.12370.77410.76590.00820.0654--‐0.057200.02750044PEW0.0360--‐0.04550.46760.5649--‐0.2230--‐0.22080.00220.1787--‐0.176500.00080045PEW0.40670.3800--‐0.2575--‐0.00190.00890.08880.08000.2824--‐0.202501.74990046PEW0.08460.07770.17410.33680.63180.53370.09820.1695--‐0.071401.43720047PEW0.14810.1743--‐0.06170.06420.69140.77980.08850.1521--‐0.063701.61980048PEW0.38380.23940.16160.17370.67680.82160.14480.1565--‐0.011704.91241049PEW--‐0.5987--‐0.6622--‐0.1973--‐0.18240.30430.25000.05430.0784--‐0.024000.47590050PEW--‐0.0236--‐0.1360--‐0.0866--‐0.29600.59060.61600.02540.3218--‐0.296300.06410051PEW--‐0.3111--‐0.4415--‐0.4833--‐0.58470.52780.59900.07130.2318--‐0.160501.44920052PEW0.20000.3135--‐0.01730.0618--‐0.1654--‐0.13040.03500.1926--‐0.157600.26360053PEW0.53710.56010.28000.33140.06860.05570.01290.0744--‐0.061500.02860054PEW--‐0.4947--‐0.5130--‐0.5211--‐0.44040.23680.19170.04510.0989--‐0.053700.40880055PEW0.06840.0957--‐0.0769--‐0.05220.63250.69570.06320.05200.011110.82810056PEW0.28740.3117--‐0.1494--‐0.09310.34870.35220.00360.0807--‐0.077100.00180057PEW--‐0.8069--‐0.6839--‐0.8414--‐0.78710.73100.61290.11810.1773--‐0.059203.81660058PEW0.31730.35480.0481--‐0.02150.28850.36560.07710.1071--‐0.030000.65620059PEW--‐0.6582--‐0.59600.04080.1010--‐0.1276--‐0.31310.18560.12240.063217.13091160PEW--‐0.10480.1181--‐0.4516--‐0.33070.36290.36220.00070.3439--‐0.343200.00010061PEW--‐0.0797--‐0.08420.37680.5421--‐0.2536--‐0.29970.04600.1697--‐0.123700.32880062PEW0.26000.1906--‐0.0680--‐0.06500.48220.52910.04690.0723--‐0.025401.34660063PEW--‐0.1441--‐0.0091--‐0.7205--‐0.74550.30130.22730.07400.1599--‐0.085900.66150064PEW--‐0.3260--‐0.2665--‐0.0313--‐0.06210.06450.04210.02240.0903--‐0.067900.13080065PEW--‐0.2343--‐0.35590.1033--‐0.0056--‐0.0730--‐0.00560.06740.2306--‐0.163200.85420066PEW0.06940.17590.33330.52310.55700.55090.00610.2964--‐0.290300.01190067Gallup--‐0.0481--‐0.18300.42790.37400.19450.18830.00620.1889--‐0.182700.00810068Gallup--‐0.16780.07820.1198--‐0.07820.64270.60490.03780.4440--‐0.406200.55150069Gallup--‐0.37930.07410.0862--‐0.22220.29310.44440.15130.7618--‐0.610501.48450070see
PNAS
2014
paper0.05080.18870.28810.39620.59320.67920.08600.2459--‐0.159900.69630071see
PNAS
2014
paper0.67920.2131--‐0.03750.65900.24230.42300.18061.1626--‐0.982005.48461072see
PNAS
2014
paper0.41110.56060.61660.33330.16800.15370.01430.8656--‐0.8513025.41641173Gallup:
Rose--‐JacksonResp
to
AResp
to
BWang,
Z.,
Solloway,
T.,
Shiffrin,
R.M.,
Busemeyer,
J.R.
2014and
Wang,
Z.,
Busemeyer,
J.R.
2013.The
73d
data
set
is
the
"Rose--‐Jackson"
question
pair
showing
a
significant
violation
of
QQ
equality.
Computed,
with
permission,
from
the
collection
of
data
sets
analyzed
by
Supplementary
Information
S1:
Question
Order
Effect word
combinationprime
q1prime
q3prime
q2prime
q4s1--‐value<V1>--‐<W1><V2>--‐<W2><V3>--‐<W3><V4>--‐<W4>NDelta
Cboxer bat dog fighter ball vampire 0.7400.3500.6760.2800.31664.000-2.882table file chair chart nail folder 0.3300.1160.2280.4700.22663.000-2.710star suit moon movie vest law 1.1800.6160.1080.3260.11662.000-1.986mole pen dig face pig ink 1.1800.2500.1260.0420.60063.000-1.838crane hatch lift bird door egg 1.9200.2820.2980.5920.46663.000-1.718stag yarn party deer story wool 1.7700.7500.2080.4380.09061.000-1.716apple chip banana computer potato circuit 2.1100.5000.5880.2280.43465.000-1.640web bug spider internet beetle computer 2.0000.4200.5920.1340.30663.000-1.452bank log money river journal tree 2.1300.1100.6760.1840.51465.000-1.354port vessel harbour wine ship bottle 1.5600.2120.2260.1700.23665.000-1.284count watch number dracula time look 1.4000.3900.0220.1260.12665.000-1.264rock strike stone music hit union 2.0100.3760.6260.2340.02664.000-1.252fan post football cool mail light 2.1300.7000.0500.2500.37663.000-1.246match bowl flame contest disk throw 1.7500.2740.1500.5000.04464.000-1.218seal pack walrus envelop leader suitcase 2.1400.1660.3240.4260.44264.000-1.218spring plant summer coil leaf factory 2.0200.5880.0000.2660.34664.000-1.180slug duck snail punch quack dodge 1.8300.1920.2660.3060.05263.000-0.986bill scale phone pelican weight fish 1.6300.1620.1080.2260.10864.000-0.974toast gag jam speech choke joke 1.2300.0000.0360.0160.05263.000-0.874net cap gain volleyball limit hat 1.8600.0700.1180.1840.35065.000-0.862battery charge car assault volt prosecute 2.0100.1340.2340.0960.24063.000-0.694club bar member golf pub handle 2.2800.2660.2500.0000.27664.000-0.512poker spade card fire ace shovel 2.1500.2720.0000.0700.22665.000-0.418Supplemntary
Information
S2:
Word
Pairs
with
PrimingReproduced
with
permission.and
coversion
of
probabilities
into
expectationsResults
of
the
experiment
byBruza,
P.D.,
Kitto,
K.,
Ramm,
B.J.,
&
Sitbon,
L.,analyzed
in
Section
6
of
the
main
text;all
the
computations
are
made
by
Bruza
et
al.,
except
for
the
last
column SupplementaryInformationS3:DetailsoftheMatchingExperiment1ParticipantsAlltheparticipantswerestudentsatPurdueUniversity.Thesecondauthorofthispaper,labeledasP3,participatedinalltheexperiments.Twopersons(P1andP2)participatedinExperiments1(a)and2(a),andtwootherpersons(P4andP5)inExperiments1(b),2(b),2(c),3(a),and3(b).Allparticipantswereagedaround25andhadnormalorcorrectedtonormalvision.2StimuliandProcedureVisualstimuliconsistingofcurvesand(sometimes)dotswerepresentedonaflat-panelmonitor.Theyweregrayish-whiteonacomfortablylowintensitybackground.Thediameterofthedotsandthewidthofthecurveswas5pixels(px).Theparticipantsviewedthestimuliindarknessusingachinrestwithaforeheadsupportatthedistanceof90fromthemonitor,making1screenpixelapproximately62secarc.Ineachtrialtheparticipantswereaskedtomatchafixedstimulusbyadjustingavariablestimulusbyrotatingatrackballusingtheirdominanthand.Oncearesponsewasmadetotheparticipant'ssatisfaction,sheorheclickedabuttononthetrackballdevicetoendthistrial,andanewstimulusappearedhalfasecondlater.Eachexperimenttookseveraldays,eachofwhichconsistedofabout200trialsconductedwitha10-minbreakinthemiddle;eachsuchsessionwasprecededbyapracticeseriesof10trials(whichwerenotrecorded).1 2.1Experiment1(a)Eachtrialbeganwithpresentingtwocircleswithadotinthefirstquadrantofeachcircle(asshowninFigure1,toppanels).Theradiusofeachcirclewas160px.Thedotintheupperleftcirclewasfixedatoneofrandomlychosensixpositions.Usingthecenterofitscircleastheorigin,theycanberepresentedequivalentlyusingtherectangularcoordinates:{(24px,48px),(32px,32px),(32px,64px),(48px,24px),(64px,32px),(64px,64px)}orthepolarcoordinates:{(53.67px,63.43deg),(45.25px,45deg),(71.55px,63.43deg),(53.67px,26.57deg),(71.55px,26.56deg),(90.51px,45deg)}.Hencetheexperimentaldesigncontaineda2⇥2"rectangular"subdesign,{32px,64px}⇥{32px,64px},anda2⇥2"polar"subdesign{53.67px,71.55px}⇥{63.43deg,26.57deg}.Thepositionofthedotinthebottomrightcirclewascontrolledbythetrackball,untilitslocationmatchedthatofthefixedone.Oncearesponsewasmade,theprogramrecordedthelocationsofthetargetdotandthematchingdotinbothrectangularcoordinatesandpolarcoordinates.Therewere1200trialsoverallwithapproximately200trialspertreatment.2.2Experiment1(b)Thehorizontalcoordinateandverticalcoordinateofthetargetdotwererandomintegersdrawnbeforeeachtrialfromthetherectangle[20px,80px)⇥[20px,80px).ThisCartesianrectanglecontainedthepolar-coordinaterectangle[40px,90px)⇥[30deg,60deg),allowingustoanalyzethedatafallingwithinitseparately.Theoverallnumberoftrialswas1800,ofwhich900fellwithinthepolar-coordinaterectangle.Inallotherrespect,theprocedurewasthesameasinExperiment1(a).2.3Experiment2(a)EachtrialbeganasshowninFigure1,middleleftpanel.Thetargetfigure,ontheleft,consistedoftwoconcentriccirclestogetherwiththeircenter.Theradiiofcircle1andcircle2wererandomlychosenfromthesets{16px,56px,64px}and{48px,72px,80px},respectively,ina3⇥3factorial2 design.Ontheright,inthebeginningofthetrial,therewasadotlocatedat(250px,0px)relativetothecenterofthetargetfigure.Byrotatingthetrackballtheparticipantaimedatmatchingthetargetfigureby"blowingup"twocirclesfromthedotontheright,onebyone.Oncethefirstmatchingcirclewasproduced(innerorouter,thepersoncouldchoose),theparticipantclickedabuttononthetrackballtostabilizethiscircleandthentheprogramenabledhimorherto"blow"theothercircle.Afterthesecondmatchwasmade,thetrialwasterminatedbyclickingthesamebuttononthetrackball.Theprogramrecordedtheradiiofthetargetandmatchingconcentriccirclesineachtrial.Therewere1800trialsoverall,approximately200trialspertreatment.2.4Experiment2(b)Experiment2(b)wasidenticaltoExperiment2(a)exceptthatineachtrialtheradiiofthetargetcircle1andcircle2wererandomlychosenfromfourpossibilities{12px,24px}⇥{18px,30px}.Therewere1600trialsoverall,about400trialspertreatment.2.5Experiment2(c)Experiment2(c)wasidenticaltoExperiment2(a)exceptthatineachtrialtheradiusofthetargetcircle1wasanumberrandomlychosenfromtheuniformdistributionontheinterval[18px,48px)andtheradiusofthetargetcircle2wasrandomlychosenfromtheinterval[56px,86px).Therewere1800trialsoverall.2.6Experiment3(a)ExamplesoftwofloralshapestogetherwiththeircentersareshowninFigure1,bottompanels.Twosuchconfigurationswerepresentedsimultaneouslyineachtrial.Thetargetonewasontheleft,thevariableoneontheright.Thefloralshapewasgeneratedusingthefunction3 x=cos(.02⇡ )[70+↵cos(.06⇡ )+ cos(.1⇡ )],(1)y=sin(.02⇡ )[70+↵cos(.06⇡ )+ cos(.1⇡ )],wherex,ystandfortherectangularcoordinates.Amplitude↵andamplitude oftheleftfloralshapewererandomlychosenfromthesets{-18px,10px,14px}and{-16px,-12px,20px},respectively. wasvariedfrom0to99withanincrementof1ateachstep.Ateachstep,apointwithcoordinates(x,y)wasdrawntothescreenandeachfloralshapewascomposedof100suchpoints.The↵and fortheshapeontherightwascontrolledbyrotatingthetrackball.Theprogramconvertedthehorizontal(vertical)componentoftherotationtothechangeof↵(respectively, ).Theinitialvaluesfortheseamplitudeswererandomlypickedfromtheinterval[-35px,35px).Therewere1800trialsoverall,about200trialspertreatment.2.7Experiment3(b)Experiment3(b)wasidenticaltoExperiment3(a)exceptthatthetwoamplitudesforthetargetshapewererandomlychosennumbersfromtheinterval[-30px,30px).3ResultsThedataareavailableas"ContextualityinPsychophysicalMatching"dataset(Excelfiles)inhttp://dx.doi.org/10.7910/DVN/OJZKKP,HarvardDataverse,V1.4AnalysisIneachexperimentwedeletedoutliers,defined,ratherinformally,asthematchingvaluesthatweretoofarfromthetargetvalues.Theoutliersmadelessthan1%ofalldata.(Notethatthefilesintherepositorydonothavetheoutliersdeleted.)InExperiment2bthedesignwas2⇥2.InExperiment1athereweretwo2⇥2subdesigns,the4 "rectangular"and"polar"ones.InExperiment2aand3athedesignwas3⇥3andtheanalysiswasmadeforeachoftheninepossible2⇥2subdesigns.InExperiments1b,2c,3bthevaluesofthetargetstimuluswerefirstdichotomizedintobelow-medianandabove-medianvalues,forminga2⇥2factorialdesignineachofthem.Oncea2⇥2designwasformed,theresponses(matchingvaluesofthevariablestimuli)weredichotomizedasdescribedinSection7ofthemaintext,bychoosingallpossiblecombinationsoffourintegervaluesintheintervalsx12{max[minX12,minX41],min[maxX12,maxX41]},x32{max[minX23,minX34],min[maxX23,maxX34]},y22{max[minY12,minY23],min[maxY12,maxY23]},y42{max[minY41,minY34],min[maxY41,maxY34]}Theanalysisafterwardsconsistedincomputingthevalueof C.Thenumberofdichotomizationsineach2⇥2(sub)designwasbetween3024and11,663,568.5 |
1808.08820 | 1 | 1808 | 2018-08-27T12:40:31 | Vanadium dioxide circuits emulate neurological disorders | [
"q-bio.NC"
] | Information in the central nervous system (CNS) is conducted via electrical signals known as action potentials and is encoded in time. Several neurological disorders including depression, Attention Deficit Hyperactivity Disorder (ADHD), originate in faulty brain signaling frequencies. Here, we present a Hodgkin-Huxley model analog for a strongly correlated VO2 artificial neuron system that undergoes an electrically-driven insulator-metal transition. We demonstrate that tuning of the insulating phase resistance in VO2 threshold switch circuits can enable direct mimicry of neuronal origins of disorders in the central nervous system. The results introduce use of circuits based on quantum materials as complementary to model animal studies for neuroscience, especially when precise measurements of local electrical properties or competing parallel paths for conduction in complex neural circuits can be a challenge to identify onset of breakdown or diagnose early symptoms of disease. | q-bio.NC | q-bio | Vanadium dioxide circuits emulate neurological disorders
Authors: J. Lin1,2, S. Guha1,2, and S. Ramanathan*3,4
Affiliations:
1Center for Nanoscale Materials, Argonne National Laboratory, Lemont, IL 60439 USA
2Institute for Molecular Engineering, University of Chicago, Chicago, IL 60615 USA
3School of Materials Engineering, Purdue University, West Lafayette, IN 47907 USA
4School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN 47907,
USA
*Correspondence to:
Prof. S. Ramanathan
Addr.: 701 W. Stadium Ave, Armstrong Hall of Engineering, Purdue University, IN 47907
Phone: 765-496-0546
Article type:
Original Research
Abstract:
Information in the central nervous system (CNS) is conducted via electrical signals known as
action potentials and is encoded in time. Several neurological disorders including depression,
Attention Deficit Hyperactivity Disorder (ADHD), originate in faulty brain signaling frequencies.
Here, we present a Hodgkin-Huxley model analog for a strongly correlated VO2 artificial neuron
system that undergoes an electrically-driven insulator-metal transition. We demonstrate that tuning
of the insulating phase resistance in VO2 threshold switch circuits can enable direct mimicry of
neuronal origins of disorders in the central nervous system. The results introduce use of circuits
based on quantum materials as complementary to model animal studies for neuroscience,
especially when precise measurements of local electrical properties or competing parallel paths for
conduction in complex neural circuits can be a challenge to identify onset of breakdown or
diagnose early symptoms of disease.
I. Introduction
Action potentials (AP) are generated in neurons and propagated to other neurons via synapses
(Hodgkin and Huxley, 1952; Kandel, 2012). The frequency of the spikes carries information and
is critical for brain function. How frequently neurons spike for a given stimulus and whether or
not they are able to travel without losing signal strength dictate normal vs. abnormal brain function
(Bartzokis, 2005; Salinas and Sejnowski, 2001; Wulff et al., 2009). Alteration in neural oscillations
caused by abnormal excitation of action potential has been found to play an important role in a
number of neurological disorders. Many research works on molecular neurophysiology have
suggested the correlation of pathologically altered action potential excitability. Various
neurological diseases are briefly summarized as follows.
Wu. et. al. reported that the dysfunctional calcium channel in mutant mouse model is associated
with the hypokalemic periodic paralysis which is a form of paroxysmal weakness that occurs in
motor neuron disease (Wu, 2012). Research has shown that the Alzheimer's disease can occur due
to disruption of neuronal excitability. As examples, Chakroborty et. al. showed that increase of
frequency and amplitude of AP due to certain protein channel dysregulation results in excitability
impairment, with which the Triple Tg expression model is developed (Chakroborty et al., 2009;
Santos et al., 2010). Drug addiction is also strongly related to abnormal excitation of action
potential. Kourrich et. al. revealed the relationship between drug addiction and brain activity. They
found that neurons subjected to certain dose of Cocaine will fire about 30% faster at high input
current, and 200% faster at low input current (Kourrich et al., 2015). Global Burden of Disease
Study revealed that major depression was the second largest cause of disability (estimated by the
loss of productivity from the disease) and it affected approximately 300 million people worldwide
in 2010 (Vos et al., 2012). Friedman et. al. showed that the midbrain dopamine neurons have
played important in certain depressions. When the dopamine neuron (in mice) fire rate increases
by 50% (from 1.6 Hz to 2.4 Hz), the social interaction (measured by a special experiment, see
reference) dropped by about 60% (Friedman et al., 2014). Bipolar disorder, also known as manic
depression, has been studied in monkeys and it is found that the mental disorder is related to the
prefrontal cortical neurons firing and signaling at the molecular level (Birnbaum et al., 2004).
Another example in the context of understanding and curing neurological syndromes is the
research on neuropathic pain. Researchers sought for treatment for pain by treatments to tune the
neuron oscillation frequency (Campbell and Meyer, 2006). These studies and references contained
in them strongly indicate the crucial role of controlling the spiking frequencies and action potential
generation in order-disorder transitions in biological neural circuits. Besides these examples, other
neurological disorders that are resulted from pathologically-altered brain signaling frequencies
also include neuromuscular diseases (Hutchison et al., 2004; Nelson and Valakh, 2015; Younger,
1999), ADHD (Brennan and Arnsten, 2008) and etc..
Understanding their origins and the mechanisms to minimize damage to neural pathways is a
principal area of study in neuroscience. However, diagnosis of neurological disorder at the
molecular level is challenging (Brown et al., 2004). One widely adopted method for
neurophysiological measurements is the multiple-electrode recording of the electrical signal of AP
spikes in brain tissue (Brown et al., 2004). To-date, neural recording experiments usually involve
invasive probing (Kinney et al., 2015) and the in vivo measurements are mostly carried out on
small animals such as mice (Barry, 2015; Schulz et al., 2014). Artificial circuits that mimic desired
signal propagation characteristics along neurons and can provide parametric information on
normal-abnormal signaling transitions from electrical properties of circuit components could be
valuable in evaluating or directing animal studies. Here, we propose understanding electrical
behavior of neurons and neurological disorders via synthetic circuits comprised of a strongly
correlated oxide VO2 that undergoes an electrically-driven insulator-metal transition (IMT).
Oxides have been studied for electronic devices such as resonant tunneling diodes, single-electron
transistors, and steep slope switches (Mannhart and Schlom, 2010; Vitale et al., 2015). Among
these emerging oxide-based electronic device concepts, phase changing artificial neurons has
primarily focused on applications in neuromorphic computing to mimic the leaky-integrate-fire
function (Lin et al., 2016; Pickett et al., 2013; Tuma et al., 2016). Here, we present a Hodgkin-
Huxley (HH) model analog for the intrinsic properties of a solid-state material, VO2. The strongly
correlated VO2 artificial neuron system can undergo an electrically driven insulator-metal
transition akin to the excitable membrane in the biological neuron. Changes in composition of the
material synergistically modifies the ground state resistivity, IMT strength defined as resistance
ratio in the two phases as well as the threshold voltage required for initiating a phase change. Such
material property is designed to capture the Intrinsic Membrane Excitability (IME) in biological
neurons, which refers to a neuron's propensity for generating action potential at a given input.
Building on this fundamental concept, we demonstrate neuronal function mimicking a vast range
of neuron types found in animal brains and simulate an archetypal monosynaptic circuit (e.g. the
knee-jerk reaction). Long term, our results may help in creating artificial systems to generate
knowledge about thresholds for onset for brain disorders due to neuronal malfunction.
II. Materials and Methods
VO2 thin films of 200 nm thickness were deposited on SiO2/Si by reactive sputtering at 775 K.
The stoichiometry and IMT transition strength in VO2 is controlled by the oxygen partial pressure
in the sputtering chamber. The insulator-metal transition occurs at a critical voltage Tc. IMT
transition strength (Rins/Rmet) is defined by the ratio of high resistance state (Rins, measured at room
temperature) and low resistance state (Rmet, measured at above critical transition temperature). In
VO2, Tc is 67oC. The low resistance state is taken at 120oC that is significantly higher than Tc. Our
film growth experiments have shown controllable thermal IMT strength variation from Rins/Rmet
>105 to Rins/Rmet=1 (complete loss of IMT characteristic) (Ha et al., 2013; Lin et al., 2016; Zhou
and Ramanathan, 2015). Rins and Rmet are respectively the resistivity for the insulating state and
metallic state, and is characterized by temperature-dependent Hall measurement.
Three VO2 samples were used for device fabrication and artificial neuron circuit testing. Their
IMT strength is respectively 20, 750, and 2x105. For device fabrication, electron-beam lithography
(EBL) is first used to define the length of the VO2 device, L, as shown in Fig. 1A. L is in the
channel direction and it varies from 50 nm to 17 m. Ti/Au of thickness 5/100 nm were evaporated
to form electrical contacts to the VO2. A "neck-down" design for the contact were used as
illustrated in Fig. 1B. The neck-down device drives only a small volume of the VO2 into transition
and reduces the voltage required to trigger the transition (Lin et al., 2017).
All experiments were carried out at room temperature. The DC sweeping and current-clamp are
both performed using Keysight B1500A. Waveforms for the current-clamp experiment were
acquired by Keysight Digital Oscilloscope DSO9104A. For DC sweeping, a current compliance
is set to 5 mA. For current-clamp response, the IMT device is protected by a series resistance
through a circuit board so that the measurement can be repeated reliably by avoiding excess heating
and burnout.
Fig. 1A shows the schematic of the VO2 devices under DC probing and Fig. 1B is the top view of
the lateral VO2 device with a "neck-down" layout. The "neck-down" layout is used to minimize
the volume of VO2 that undergoes transition (Lin et al., 2017). It leads to lower critical transition
voltage (Vc) and lower power in the switching operation. The smallest spacing between two
contacts is L= 200 nm for the device being studied in this work. Fig. 1C shows a measured current
versus voltage characteristic under DC condition. A hysteresis sweep is performed and the switch
between insulator-state and metal-state is reversible if the operation satisfies the safe criteria
introduced in (Lin et al., 2017).
Excessive bias stress to the VO2 device can result in non-reversible damage to the material which
is manifested in a permeant change in critical transition voltage under DC measurement. This can
happen when the device is subjected to a bias outside the safe operating criteria. Two forms of
non-reversible damages are shown in Fig. 2. Two forward DC sweeps (in the positive direction)
are carried out consecutively in one device. Fig. 2A shows an increase in Vc caused by an increase
of the HRS resistance (+R). The current drops over the whole range of applied voltage in the
second sweep. Fig. 2B shows a reduction in Vc which is the indication of a drop in the HRS
resistance (-R). The current is higher over the span of applied voltage.
(A)
(B)
V
I
Contact
VO2
L
Insulating substrate
VO2
Contact
500 nm
)
A
m
(
I
6
4
2
0
0.0
Experiment
VR
0.2
V (V)
VC
0.4
0.6
Fig. 1. Experimental details for the VO2 device. (A) The schematic of the VO2 devices under DC probing. (B)
The top view of the lateral VO2 device with a "neck-down" layout. The "neck-down" layout leads to lower
forward critical transition voltage (Vc) and lower power in the switching operation. (C) The typical current versus
voltage characteristic in DC measurement. The voltage sweep is in the sequence of forward (0 to 0.6 V) and
reverse (0.6 to 0 V) direction. The switch is reversible. The forward critical voltage and reverse voltage are
denoted.
(A)
)
A
(
I
100
10-1
10-2
10-3
10-4
10-5
0
Experiment
1
1
Experiment
VC
2
(B)
)
A
(
I
100
10-1
10-2
10-3
10-4
10-5
0.0
VC
2
2
3
V (V)
1
0.5
V (V)
1.0
1.5
Fig. 2. VO2 device under excessive DC stress experiencing non-reversible change in critical transition voltage.
Two consecutive sweeps are applied to the VO2 device with the first sweep stresses the device. (A) Increase in
Vc is caused by an increase of HRS resistance, +R. (B) Reduction in Vc is the indication of a drop in HRS
resistance, -R.
III. Results
Fig. 3 shows schematic of a biological neuron and an analogous VO2 neuron. The membrane of
the biological neuron (Fig. 3A) comprises of an insulating phospholipid bilayer that separates the
intracellular and extracellular fluids, and protein channels that control the permeation of various
ions. As described in the Hodgkin-Huxley model, the neuron membrane is equivalent to a parallel
combination of membrane capacitance, Cm, and transmembrane conductance, Gm. Gm is the sum
of various ion channels conductance and it can go through a reversible insulator-to-metal
transition depending on the voltage across the membrane. An input stimulus can trigger a train of
action potentials (AP) that is a temporary reversal of the polarity across the neuron membrane. The
AP propagates along the axon through which information is transmitted. In the central nervous
system (CNS) such as in the brain and spinal cord, the neuron is myelinated -- the myelin sheath
surrounds the axon of the neuron cells and promotes rapid signal transmission. Experimentally,
the neuron can be stimulated with an input current Iin. The output AP waveforms depend on the
input current, the electrical and geometric parameters of the cell, and the environment such as
temperature. The HH model and its parameters are described later in Sec. II.A.
The VO2 device and analogous VO2 neuron circuit are shown in Fig. 3B. The VO2 is well-known
for its reversible insulator-metal transition proximal to room temperature. Joule heating in two-
terminal devices can locally drive the phase change rapidly. This property can be exploited to
demonstrate highly nonlinear switches (Lin et al., 2017; Son et al., 2011) and artificial neurons
(Lin et al., 2016; Pickett et al., 2013; Tuma et al., 2016). The VO2 artificial neuron is a circuit that
comprises, a minimum of, only two components, the capacitor Co and the conductor (i.e. resistor)
GVO as shown in Fig. 3B. When the input stimulus Iin starts, the VO2 neuron exhibits an oscillatory
behavior similar to that in the biological neuron. The model and experiment for standalone VO2
devices are discussed respectively in Sec. II.B and Sec. II.C. The VO2 neuron circuit is described
in Sec. II.D.
(A)
Dendrite
Soma
Iin
Vm
Intracellular electrode
Myelin sheath
Axon terminal
Axon
Phospholipid
bilayer
Protein channel
Intracellular fluid
Extracellular fluid
Vm
GM
Vm
IION
Gm
CM
time
Fig. 3. The biological neuron and analogous VO2 neuron. (A) The membrane of the biological neuron can be
viewed as a parallel connection of a membrane capacitance (CM) and a membrane conductance (GM) that can go
through insulator-to-metal transition under stimulus. The membrane is polarized at the resting potential due to
the different ionic concentration in the intracellular and extracellular fluids. When the neuron is subjected to a
steady current clamping, a continuous action potential (AP) is generated, in which the trans-membrane potential
(VM) and the membrane conductance (GM) oscillate. The AP can propagate along the axon and transmit signal
to the other connected neurons. The myelin sheath surrounding the axon of some neuron cells can enhance the
speed at which impulses propagate. (B) A lateral VO2 device and a capacitor are used to construct the VO2
artificial neuron circuit. The VO2 material exhibits a reversible electrothermal insulator-to-metal transition. This
state change is used to mimic the biological neuron. At constant current input, the VO2 neuron output node and
VO2 conductance oscillate, similar to that of the biological neuron. The insulating-state resistance can be
changed when VO2 degrades, and this feature is utilized to model spike-timing related neural disorders. Here
+R represents an increase of resistance and -- R represents a drop in resistance.
A. Hodgkin-Huxley (HH) Model for Biological Neuron
The complete circuit schematic for a patch of the neuron membrane with the HH model is
illustrated in Fig. 4. The conduction occurs via three channels: The Na+ channel, the K+ channel
and the leakage channel.
The basic mechanisms in the HH model contains ion transport and transmission lines for Action
Potential (AP) propagation. The key equations are shown in Eqs. 1-3. Eq. 1 relates the membrane
current density IIN to membrane potential Vm. The area-normalized membrane capacitance is Cm.
Two ion channels with the leaky conductance are included in the model. Their conductance is
denoted as GNa, GK and GL. The Nernst equilibrium potential in Eq. 2 relates extracellular and
intracellular ion concentrations, respectively denoted as Ci and Co. Through Eq. 2, the Nernst
equilibrium potentials VNa and VK can be obtained for the given Na+ and K+ concentrations. The
molar gas constant R and Faraday's constant F are physical constants. Finally, the propagation of
AP along z direction is described by the core conductor equation in Eq. 3. It couples the voltage
and current along a cylindrical cell where the resistances per unit length inside and outside the cell
are respectively ri and ro. The cylindrical cell has diameter a. The baseline values of the parameters
in the HH model are listed in Table 1. The HH neuron model is constructed in Matlab.
Eq. 1
Eq. 2
Eq. 3
Fig. 4. Full schematic of biological neuron that contains two ion channels, leaky capacitive membranes.
The equations the form the Hodgkin-Huxley model are shown in the Eq. 1-3.
Table 1: Baseline input parameters for the Hodgkin-Huxley model
Fig. 5 shows the results from the HH model with a current clamp. The input is two discrete current
pulses. The current-clamped neuron is subjected to the deposition of charge from each of the pulse.
The deposited charge of one pulse is merely enough to trigger one neuron firing. The solid red
lines are the results for pulse width of 3 ms, and the dashed black lines are for pulse width of 6 ms.
The resultant AP profiles are identical for different pulse width, even the long pulse (dashed black)
has deposited twice as many charges as the short pulse (solid red). The additional charges are not
integrated because it falls into the "refractory period". A new integration cycle starts only after the
neuron resets itself.
(A)
)
2
m
c
/
A
(
n
i
I
6
4
2
0
(B)
)
V
m
(
m
V
50
25
0
-25
-50
-75
refractoriness
0
10
reset
20
t (ms)
30
40
Fig. 5. Simulated results from the Hodgkin-Huxley model. (A) The neuron is clamped to pulse current input,
and (B) cross membrane potential. The red solid lines and black dot lines are two different inputs that result in
the same firing patterns. The extra current inputs are not being integrated to the membrane capacitance due to
the existence of post-refractory period.
B. Model for VO2 Device
The model for electrothermal insulator-metal transition is discussed in this section. The basic form
of heat equation is a parabolic partial differential equation (Eq. 4) that describes the relationship
of temperature variation in a given volume over time. Eq. 4 assumes an isotropic and homogeneous
medium in a 3-dimensional space and zero heat flux. The 3D heat transport is simplified by the
quasi-1D assumption that the temperature variation perpendicular to the current transport direction
along y and z is much smaller than that in the transport direction along x. This is illustrated in Fig.
6 with the current flowing in x direction, and the heat equation is converted to Eq. 5. Under this
assumption, the temperature for a given segment at location x and time t is obtained as T(x, t). The
parameters describing the IMT property are given as follows, specific heat capacity Cth, density
o, and thermal conductivity K.
I
Metal
IMT
Metal
Fig. 6. Full schematic of an IMT device with length L between two metal contacts. The device is
discretized into segment of dx, etch with its own temperature and resistivity. Heat conduction is along x
direction while convective heat loss through the side wall.
Table 2. Input parameters for the coupled electrical-thermal model IMT model.
(Eq. 4)
(Eq. 5)
There are two origins of heat flux. Firstly, Joule heating results in incoming heat flux to the
medium. The power generated by Joule heat follows Ohm's law, and for a unit volume it is:
(Eq. 6)
where I is the total current through the IMT, A is the cross sectional area, and r is the temperature-
dependent resistivity of the IMT. Secondly, the outgoing heat flux is generated by convective heat
loss, modeled by the effective convective heat transfer coefficient h. Here h is assumed to be a
constant and is independent of the IMT temperature. Besides h, the heat loss through convection
for a unit volume is also related to the ambient temperature Ta, and IMT's surface to volume ratio
Lp/A where Lp is the cross sectional perimeter. The power dissipated through side wall heat
convection is:
Taking into account the heat fluxes, the differential equation for heat transfer is shown in Eq. 8.
(Eq. 7)
(Eq. 8)
As shown in Fig. 6, the IMT with length L is connected to two metal contacts with length Lc. As
the boundary condition, Lc is assumed to be long enough so that the value of Lc has negligible
impact -- Lc should be significantly longer than the heat diffusion length. Eq. 8 is solved using a
numerical method: forward difference for the time domain and central difference for spatial
domain. In the spatial domain, the IMT bar is discretized into segment of length dx. Each segment
has its resistivity r(x). The total IMT resistance RIMT is obtained by integrating the resistance of
all segments:
The current through the IMT can then be obtained by
(Eq. 9)
(Eq. 10)
Where RS is the series resistance. The IMT resistivity as a function of temperature follows a look-
up table of resistivity versus temperature as measured in the experimental VO2 device. A typical
example of the resistivity versus temperature is shown in Fig. 3B, which is characterized by the
insulator-state resistivity H, the metal-state resistivity L, and the critical transition temperature
(Tc). The baseline values of other physical parameters are listed in Table 2.
C. VO2 Neuron Circuit
The complete VO2 neuron circuit is shown in Fig. 7. This is one of the simplest artificial neuron
circuits that has been reported, which comprises of only two or three elements. Despite its
simplicity, it exhibits striking similarity to the biological neuron (Fig. 3A). To investigate the many
unexplored functions of the artificial neuron, particularly its connection to neurological diseases,
a physical neuron model is imperative.
The model is focuses on the material properties of VO2 that emulate biological neuron functions.
A series resistance Rs is added in series with the VO2 for two reasons. First, it limits the current
when the VO2 device transitions to the metallic state, and ensures reliable switching. The safe
operating design follows the theoretical guideline derived in (Lin et al., 2017). Second, it converts
the output current to an output voltage which is measured by the oscilloscope. Each spiking event
includes the following four steps: integration, fire, refractoriness and reset as discussed in Fig. 8.
One example of the simulation result is shown in Fig. 9. Two discrete current pulses are fed to the
input node of the neuron circuits. The deposited charge from each pulse is enough to fire the neuron
once. The pulse duration is 0.9 s for the red solid lines and 1.8 s for the black lines (Fig. 9A).
After each neuron firing, the VO2 stays in low resistance state for a finite period (Fig. 9B). During
this period, the input current are directly drained to ground through the VO2. The charge is not
integrated. As a result, inputs with two pulse durations generates the same firing patterns (Fig. 9C).
This results show similar post-firing refractoriness as the biological neuron (Fig. 5).
Post-firing refractoriness is another important feature in both biological and VO2 neurons. A
second AP is difficult to be produced immediately following the occurrence of an AP when the
cell is regarded to be refractory (Weiss, 1996). Following each firing, the VO2 element remains at
a temperature above the critical temperature for a time, th+el, where th and el are the thermal
and electrical time constants respectively. th is related to the thermal mass and heat dissipation.
For the electrical time constantel, it is given by RmetCo+RsCo. Rmet is the metallic-state resistance
and Rs is the series resistance. Usually, Rmet is much greater than Rs in normal operation (Lin et
al., 2017). Therefore, the refractory period is the same as the pulse width (w) and is given by RsCo.
During this period the VO2 element remains in metallic state and new input charge is continuously
discharged without being integrated in the capacitor Co. The quantity r therefore defines the
restitution time. Our coupled electrothermal model has captured this process and can be used to
design the restitution time in the VO2 neuron circuit. Subsequently, the VO2 element resets and
starts another integrate-and-fire cycle. The steps exactly mimic the electrically excitable
membrane in neuron cells.
IIN
Vo
RVO
Co
Rs
IO
Fig. 7. The complete VO2 neuron circuit. The whole circuit contains three elements: a capacitor as well as the
VO2 device with a sensing resistor in series. The output current is sensed by the sensing resistor. The voltage at
the capacitor node is denoted as Vo. It is also the input node for the injected current.
(A)
HRS
LRS
LRS
HRS
1. Integrate
2. Fire
3. Refractoriness
4. Reset
(B)
1.0
)
V
(
0.5
o
V
0.0
0
Exp.
10
2
1
3 4
20
30
t (s)
(C)
)
A
m
(
o
I
40
50
6
4
2
0
0
Exp.
10
2
3
4
1
20
30
t (s)
40
50
Fig. 8. The four stages in one spike cycle in the VO2 neuron and the corresponding experimental output
waveforms. At the integration stage, current is integrated to the capacitance. Very small current goes through
the resistor at the right branch as the VO2 is at HRS. Voltage at across the capacitor Vo is increasing. When Vo
reaches Vc, VO2 become metallic and it discharges the capacitor. An instantaneous large current spike appears
at the output. Vo drops sharply. This is stage 2, fire, which is followed by stage 3, refractoriness (refractory
period). In stage 3 the VO2 remains in its LRS for some time. Any input current will be drained to ground without
integrating to the capacitor. After the refractory period, the neuron reset and is ready for another spike cycle.
(A)
)
A
m
(
N
I
I
(B)
)
(
T
M
R
I
(C)
)
A
m
(
o
I
2.0
1.5
1.0
0.5
0.0
103
102
101
25
20
15
10
5
0
Refractory
period
0
1
2
3
4
t (s)
5
6
7
Fig. 9. Simulated waveform for current-clamp VO2 neuron. (A) Input current, (B) VO2 resistance and (C)
Output current. The red solid lines and black dot lines are two different inputs that result in the same firing
patterns. The extra current inputs are not being integrated to the capacitor due to the existence of post refractory
period. This property directly mimics the biological neuron.
IV. Discussion
A. Healthy vs Degenerative Neurons, and Their VO2 Analogy
In biological neurons, the inter-spiking interval (ISI) is defined as the time interval between two
adjacent spikes (Fadool et al., 2011; Okubo et al., 2015). The spiking frequency is the reciprocal
of ISI. The AP recorded as a function of time is shown for a healthy neuron (Fig. 10A) along with
two abnormal neurons (Figs. 10 B and C), simulated with the HH model. A pathological change
of the action potential firing frequency can lead to neurological and psychological disorders. For
instance, a decrease in AP frequency is tied to CNS depression and cognitive dysfunction
(Friedman et al., 2014). In contrary, an increase in firing frequency is responsible for seizures,
pain, ADHD, and anxiety (Wulff et al., 2009). Therapeutic treatment can be designed to restore
AP frequency according to the dysfunction mechanisms.
Similar characteristics can be observed in the VO2 neurons. The simulation results for three VO2
neurons are shown in Fig. 10. Fig. 10D is the VO2 neuron for baseline reference, and Figs. 10 E
and F are the cases where the HRS resistance is modified. ISI for the VO2 neuron is defined in the
same way as for the case of a biological neuron. The AP frequency reduces if the VO2 undergoes
a +R degradation, and vice versa. Analytically, the value for ISI can be derived from the VO2
neuron parameters as tISI=CoVc/Iin where Vc is the critical voltage to trigger an insulator-to-metal
transition in the VO2 device under DC I-V measurement and Iin is the input current. Vc is related
to the HRS resistance of the VO2 device. Definition of Vc is illustrated in Fig. 2. Experimentally,
we have observed the change of Vc (Vc) due to electrical-stress-induced resistance degradation
in the VO2 (Fig. 7). Vc be positive or negative depending on the degradation mechanism. Positive
Vc indicates an increase in the VO2 HRS resistance, and vice versa.
(A) Healthy bio. neuron
50
ISI
(B) Degenerative, slow
firing bio. neuron
50
0
-50
(C) Degenerative, fast
firing bio. neuron
50
0
-50
)
V
m
m
V
(
0
-50
-100
0
100m
10m
1m
100µ
)
A
o
(
I
50
100
t (ms)
(D)
VO2 neuron
150
200
-100
0
150
200
-100
0
50
100
t (ms)
50
100
t (ms)
150
200
(E) Degenerative VO2
neuron with +R
100m
10m
1m
100µ
(F) Degenerative VO2
neuron with -R
100m
10m
1m
100µ
10µ
0
20
40
60
10µ
0
20
40
60
10µ
0
20
40
60
t (s)
t (s)
t (s)
Fig. 10. Healthy vs degenerative fast-firing and slow-firing neurons, and VO2 analogy. (A) A healthy
biological neuron generates AP under a constant current stimulus. The time between two adjacent spikes is
termed as the inter-spike interval (ISI). (B) A degenerated neuron stimulated by the same input generates AP at
shorter ISI. (C) A degenerated neuron generates AP at longer ISI. (D) Simulation using the VO2 neuron model,
the intact VO2 neuron exhibits oscillatory behavior at a constant input current. (E) Simulation of a case where
the VO2 device is degraded and its insulating state resistance increases (+R). Such degenerative VO2 neuron
results in longer ISI. (F) Simulation of a case where the VO2 device is degraded by decreasing its insulating state
resistance increases (-R). The leakier VO2 neuron results in shorter ISI. (A-C) are simulated by the HH model.
(D-F) are simulated from the VO2 neuron model.
The pathologically-altered spike timing is linked to other serious degenerative diseases. For
example, certain neuromuscular disorders and motor neuron disease (MND) are resulted from the
ionic leakage of degenerating membrane and increase of rest conductance (Priori et al., 2002;
Younger, 1999). The electrical breakdown of the myelin sheath is one origin for a leaky membrane.
Excessive leakage in the membrane makes weak muscles (Wu, 2012). The HH model for
biological neuron shows that a neuron fails to fire when conductance is significantly increased
(Figs. 11 A and B). AP generation in a healthy neuron is accompanied by an insulator-to-metal
transition in the membrane. However, no transition can be observed in a leaky membrane at the
same current input since the charge cannot be integrated effectively.
The increase of conductance in a degenerative, leaky neuron can be modeled in a straightforward
manner in the VO2 circuit. The resistance versus temperature of the VO2 device is normalized to
the low resistance state, Rmet. Figs. 11 C and D show two neurons at a fixed Iin: (a), VO2 neuron
with Rins/Rmet=105 fires regularly and demonstrates an insulator-to-metal transition (b) while the
neuron with Rins/Rmet=103 fails to fire. The change of VO2 properties is shown in Fig. 11E.
To provide a systematical perspective on the design of VO2 neurons to mimic the corresponding
neural disorder, Fig. 11F collectively illustrates the impact of HRS resistance and input current on
neuron functions. The value of inter-spike interval depends on Rins and Iin and is shown as a contour
plot. The dark red color is the case where the combination of low input current and small Rins
results in failure in spike generation (tISI ∞). At a given Iin, tISI decreases as Rins drops. When
Rins drops to a critical value, the VO2 neuron fails to fire (Fig. 11G).
50
0
-50
)
V
m
m
V
(
-100
40
30
20
10
)
2
m
c
/
S
(
m
G
Conducting Insulating
Gm
GNa
GK
GLeak
Conducting
-100
40
30
20
10
)
2
m
c
/
S
(
m
G
(B) Degenerative bio.
neuron
(A) Healthy bio. neuron
50
0
-50
)
V
m
m
V
(
100
10
1
100m
10m
10
100m
1m
(C) VO2 neuron
(D)
100
(D) Degenerative VO2
neuron
10
1
100m
10m
10
100m
1m
)
A
m
(
o
I
)
S
(
O
V
G
)
A
m
(
o
I
)
(
S
O
V
G
Gm
Gna
GK
GLeak
0
0
5
15
10
t (ms)
20
25
0
0
5
15
10
t (ms)
20
25
10µ
52
56
58
54
t (s)
10µ
52
54
56
58
t (s)
(E)
106
Intact VO2
(F)
105
105
Neuron fires
t
e
M
R
R
/
104
102
100
20
Degenerative
VO2
40
60
T (oC)
80
100
t
o
i
t
a
r
f
f
104
104
e
m
O
R
n
O
T
M
s
n
R
/
i
I
-
T
Neuron fails to fire
103
103
100 200 300 400 500 600
1
2
5
3
4
Input Current (A)
Iin(A)
6
x 10-4
tISI (s)
50
50
45
40
40
35
30
30
25
20
20
15
10
10
5
0
(G)
20
)
s
(
I
S
I
t
15
10
5
0
103
104
Rins/Rmet
105
Fig. 11. Healthy vs degenerative leaky neurons, and VO2 artificial neurons. (A). Healthy biological neuron
shows an insulator-to-metal transition during one spike event. GNa, GK, and GLeak are respectively the Na+
conductance, K+ conductance and leakage conductance through the membrane, while GM is the the sum of the
conductance. (B) Degenerative biological neuron with excess leakage GLeak, while GK and GNa remain
unchanged. No AP spike is observed. (C) In one spike of the VO2 neuron, the VO2 device goes through an
insulator-to-metal transition. (D) Degenerative VO2 neuron with excess leakage (-R). (E) The resistance as a
function of temperature normalized to the metallic state resistance (Rins/Rmet) illustrates the reduction of
resistance of the insulating state by ~100. (F) Contour of ISI shows its dependency on material properties
(Rins/Rmet) and input stimulus (Iin). Reduction of insulating-state resistance narrows the neuron operating region
for a given input stimulus. (G) Three cut lines across Iin=200, 300 and 400 A in the contour plot (F). (A-B) are
simulated by the HH model. (C-G) are the simulated results from the VO2 neuron model.
The AP pulse width is another distinctive characteristic related to timing in different kinds of
mammalian central neurons (Bean, 2007). A short pulse and a long pulse in biological neurons are
respectively illustrated in Figs. 12 A and B, with the tw change by 10X. The pulse width can be
simulated in the VO2 neuron by altering the LRS resistance, Rmet, in the VO2 devices according to
tw=CoRmet. The VO2 neuron with short pulses of 0.2 s and 2 s are shown in Figs. 12 C and D.
Fig. 12E compares the biological neurons and VO2 neurons (both simulations and experiments).
The AP pulse width can range from a few 100 s to 10 ms which can be matched by the VO2
neuron with appropriate capacitance.
Short pulse in
bio. neuron
(A)
50
0
-50
)
V
m
m
V
(
-100
0
10
20
30
t (ms)
40
50
(B)
)
V
m
m
V
(
50
0
-50
Height
Long pulse in
bio. neuron
Pulse width,
tv
(C)
)
A
m
(
o
I
6
4
2
0
-2
(D)
10
5
)
A
m
(
t
u
o
I
Simulation
Short pulse in
VO2 neuron
Experiment
0
4
2
t (s)
6
8
Simulation
Long pulse in
VO2 neuron
Experiment
-100
0
10
20
30
t (ms)
40
50
0
-2
0
2
4
t (s)
6
8
(E)
10m
1m
)
s
(
h
t
100µ
Dopamine neuron
CA1 pyramidal neuron
Granule neuron
Purkinje neuron
i
d
w
e
s
u
P
l
10µ
1µ
VO2 neuron experiment
100n
1n
VO2 neuron simulation
10n 100n
1µ
C (F)
10µ 100µ 1m
Fig. 12. Diversity in AP pulse width across biological neurons, and VO2 artificial neuron analogy. (A). AP
with short pulse width in a biological neuron. The value of tw is taken at full width half maximum. (B) AP with
long pulse width. (C) VO2 neuron with short pulse of 0.2 s. Experiment and simulation show good agreement.
The pulse width control is achieved by changing resistance in the circuit (D) VO2 neuron with long pulse of 2
s. (E) Pulse width of the VO2 neuron and the range spans biological neuron studies reported in the neuroscience
literature.
B. Monosynaptic Neuron Circuit
We further extend this concept to two-stage cascading neuron circuits in Fig. 13. The circuit is a
modeling system for the monosynaptic motor neuron in muscular tissue that is responsible for
certain motion responses such as the knee-jerk reaction that is a model system in neuroscience.
Neuron 2 (in red) is the receptive neuron that is driven by Neuron 1 (in blue). Neuron 2 can either
fire or not fire depending on the output waveform of Neuron 1 as well as the synaptic resistor Rx.
Fig. 13. VO2 monosynaptic neuron circuit.
We emulate a monosynaptic circuit that corresponds to the well-known knee-jerk reaction used to
monitor responses in nerves (Kandel, 2012) as shown in Fig. 14. In the two cases for Figs. 14 A
and B, Neuron 2 is the same while the HRS resistance of VO2 in Neuron 1 are different. Both
neurons are initially at rest. Their temperatures are at equilibrium with the environment and is
below the critical transition temperature Tc. At t=0, a current is injected to Neuron 1 (see Fig. 9).
The neuron 1 in case A is intact with high Rins/Rmet, The output current as a function of time in
Fig. 14A shows the spike events for Neurons 1 and 2. The separation of spikes indicate the reaction
time (tdiff=0.6 s) for the signal to propagate between the two VO2 neurons. The neuron 1 in case
B has a lower HRS resistance. Premature spike in Neuron 1 results in a weak spike and it cannot
trigger a spike in Neuron 2 (Fig. 14B).
Degenerative VO2 for neuron 1,
Rins/Rmet=5x104 (Rins)
(B)
0.6
0.3
)
A
m
(
o
I
0.0
0.6
0.3
)
A
m
(
Neuron 1
Neuron 2
(A)
0.6
)
A
m
0.3
(
o
I
0.0
0.6
)
A
m
0.3
(
o
I
0.0
3
Intact VO2 for neuron 1
Rins/Rmet=105
Neuron 1
tdiff=0.6 s
Neuron 2
4
t (s)
5
6
o
I
0.0
3
4
5
6
t (s)
Fig. 14. Demonstration of degenerative Neuron 1 leading to the failed signal reception for Neuron 2 in a
monosynaptic circuit. (A) Output current as a function of time shows the spike events for the intact case. (B)
Premature spike in Neuron 1 when its HRS resistance is reduced and Neuron 2 fails to spike.
V. CONCLUSION
VO2 based circuits can emulate neuronal function and disorders. By carefully varying the electrical
properties of the ground state resistance of the artificial neuron, we can precisely identify
thresholds for firing and signal propagation that present an analogy to neuronal activity in the
brain. While the present study has focused on VO2 as a model system, a vast range of threshold
switching Mott semiconductors can further be explored in the future.
References
Barry, J.M., 2015. Axonal activity in vivo: technical considerations and implications for the
exploration of neural circuits in freely moving animals. Front. Neurosci. 9.
https://doi.org/10.3389/fnins.2015.00153
Bartzokis, G., 2005. Brain Myelination in Prevalent Neuropsychiatric Developmental Disorders.
Adolesc. Psychiatry 29, 55 -- 96.
Bean, B.P., 2007. The action potential in mammalian central neurons. Nat. Rev. Neurosci. 8,
451 -- 465. https://doi.org/10.1038/nrn2148
Birnbaum, S.G., Yuan, P.X., Wang, M., Vijayraghavan, S., Bloom, A.K., Davis, D.J., Gobeske,
K.T., Sweatt, J.D., Manji, H.K., Arnsten, A.F.T., 2004. Protein Kinase C Overactivity
Impairs Prefrontal Cortical Regulation of Working Memory. Science 306, 882 -- 884.
https://doi.org/10.1126/science.1100021
Brennan, A.R., Arnsten, A.F.T., 2008. Neuronal Mechanisms Underlying Attention Deficit
Hyperactivity Disorder. Ann. N. Y. Acad. Sci. 1129, 236 -- 245.
https://doi.org/10.1196/annals.1417.007
Brown, E.N., Kass, R.E., Mitra, P.P., 2004. Multiple neural spike train data analysis: state-of-
the-art and future challenges. Nat. Neurosci. 7, 456 -- 461. https://doi.org/10.1038/nn1228
Campbell, J.N., Meyer, R.A., 2006. Mechanisms of Neuropathic Pain. Neuron 52, 77 -- 92.
https://doi.org/10.1016/j.neuron.2006.09.021
Chakroborty, S., Goussakov, I., Miller, M., Stutzmann, G., 2009. Deviant ryanodine
receptormediated calcium release resets synaptic homeostasis in presymptomatic 3xTg-
AD mice. J Neurosci 30, 9458 -- 9470.
Fadool, D.A., Tucker, K., Pedarzani, P., 2011. Mitral Cells of the Olfactory Bulb Perform
Metabolic Sensing and Are Disrupted by Obesity at the Level of the Kv1.3 Ion Channel.
PLOS ONE 6, e24921. https://doi.org/10.1371/journal.pone.0024921
Friedman, A.K., Walsh, J.J., Juarez, B., Ku, S.M., Chaudhury, D., Wang, J., Li, X., Dietz, D.M.,
Pan, N., Vialou, V.F., Neve, R.L., Yue, Z., Han, M.-H., 2014. Enhancing Depression
Mechanisms in Midbrain Dopamine Neurons Achieves Homeostatic Resilience. Science
344, 313 -- 319. https://doi.org/10.1126/science.1249240
Ha, S.D., Zhou, Y., Fisher, C.J., Ramanathan, S., Treadway, J.P., 2013. Electrical switching
dynamics and broadband microwave characteristics of VO2 radio frequency devices. J.
Appl. Phys. 113, 184501. https://doi.org/10.1063/1.4803688
Hodgkin, A.L., Huxley, A.F., 1952. A quantitative description of membrane current and its
application to conduction and excitation in nerve. J. Physiol. 117, 500 -- 544.
Hutchison, W.D., Dostrovsky, J.O., Walters, J.R., Courtemanche, R., Boraud, T., Goldberg, J.,
Brown, P., 2004. Neuronal Oscillations in the Basal Ganglia and Movement Disorders:
Evidence from Whole Animal and Human Recordings. J. Neurosci. 24, 9240 -- 9243.
https://doi.org/10.1523/JNEUROSCI.3366-04.2004
Kandel, E.R., 2012. Principles of Neural Science, 5th ed. McGraw-Hill Education.
Kinney, J.P., Bernstein, J.G., Meyer, A.J., Barber, J.B., Bolivar, M., Newbold, B., Scholvin, J.,
Moore-Kochlacs, C., Wentz, C.T., Kopell, N.J., Boyden, E.S., 2015. A direct-to-drive
neural data acquisition system. Front. Neural Circuits 9, 46.
https://doi.org/10.3389/fncir.2015.00046
Kourrich, S., Calu, D.J., Bonci, A., 2015. Intrinsic plasticity: an emerging player in addiction.
Nat. Rev. Neurosci. 16, 173 -- 184. https://doi.org/10.1038/nrn3877
Lin, J., Alam, K., Ocola, L.E., Zhang, Z., Datta, S., Ramanathan, S., Guha, S., 2017. Physics and
Technology of Electronic Insulator-to-Metal Transition (E-IMT) for Record High On/Off
Ratio and Low Voltage in Device Applications, in: IEDM Tech. Dig. Presented at the
IEDM Tech. Dig., p. 23.4.
Lin, J., Annadi, A., Sonde, S., Chen, C., Stan, L., Narayanachari, K., Ramanathan, S., Guha, S.,
2016. Low-Voltage Artificial Neuron using Feedback Engineered Insulator-to-Metal-
Transition Devices, in: IEDM Tech. Dig. pp. 862 -- 865.
Mannhart, J., Schlom, D.G., 2010. Oxide Interfaces -- An Opportunity for Electronics. Science
327, 1607 -- 1611. https://doi.org/10.1126/science.1181862
Nelson, S.B., Valakh, V., 2015. Excitatory/Inhibitory balance and circuit homeostasis in Autism
Spectrum Disorders. Neuron 87, 684 -- 698. https://doi.org/10.1016/j.neuron.2015.07.033
Okubo, T.S., Mackevicius, E.L., Payne, H.L., Lynch, G.F., Fee, M.S., 2015. Growth and splitting
Pickett, M.D., Medeiros-Ribeiro, G., Williams, R.S., 2013. A scalable neuristor built with Mott
of neural sequences in songbird vocal development. Nature 528, 352 -- 357.
https://doi.org/10.1038/nature15741
Priori, A., Cinnante, C., Pesenti, A., Carpo, M., Cappellari, A., Nobile‐Orazio, E., Scarlato, G.,
memristors. Nat. Mater. 12, 114 -- 117. https://doi.org/10.1038/nmat3510
Barbieri, S., 2002. Distinctive abnormalities of motor axonal strength -- duration properties
in multifocal motor neuropathy and in motor neurone disease. Brain 125, 2481 -- 2490.
https://doi.org/10.1093/brain/awf255
Salinas, E., Sejnowski, T.J., 2001. Correlated neuronal activity and the flow of neural
information. Nat. Rev. Neurosci. 2, 539 -- 550. https://doi.org/10.1038/35086012
Santos, S.F., Pierrot, N., Octave, J.-N., 2010. Network excitability dysfunction in Alzheimer's
disease: insights from in vitro and in vivo models. Rev. Neurosci. 21, 153 -- 171.
Schulz, A., Walther, C., Morrison, H., Bauer, R., 2014. In Vivo Electrophysiological
Measurements on Mouse Sciatic Nerves. J. Vis. Exp. JoVE.
https://doi.org/10.3791/51181
Son, M., Lee, J., Park, J., Shin, J., Choi, G., Jung, S., Lee, W., Kim, S., Park, S., Hwang, H.,
2011. Excellent Selector Characteristics of Nanoscale VO2 for High-Density Bipolar
ReRAM Applications. IEEE Electron Device Lett. 32, 1579 -- 1581.
https://doi.org/10.1109/LED.2011.2163697
Tuma, T., Pantazi, A., Le Gallo, M., Sebastian, A., Eleftheriou, E., 2016. Stochastic phase-
change neurons. Nat. Nanotechnol. 11, 693 -- 699. https://doi.org/10.1038/nnano.2016.70
Vitale, W.A., Moldovan, C.F., Tamagnone, M., Paone, A., Schüler, A., Ionescu, A.M., 2015.
Steep-Slope Metal Insulator-Transition VO2 Switches With Temperature-Stable High
ION. IEEE Electron Device Lett. 36, 972 -- 974.
https://doi.org/10.1109/LED.2015.2454535
Vos, T., Flaxman, A., Naghavi, M., et. al, 2012. Years lived with disability (YLDs) for 1160
sequelae of 289 diseases and injuries 1990 -- 2010. A systematic analysis for the Global
Burden of Disease Study 2010. Lancet. 30, 2163 -- 2196.
Weiss, T.F., 1996. Cellular Biophysics, Vol. 1: Transport. A Bradford Book, Cambridge, Mass.
Wu, F., 2012. Calcium channel mutant mouse model of hypokalemic periodic paralysis. J. Clin.
Invest. 122, 4580 -- 4591.
Wulff, H., Castle, N.A., Pardo, L.A., 2009. Voltage-gated potassium channels as therapeutic
targets. Nat. Rev. Drug Discov. 8, 982 -- 1001. https://doi.org/10.1038/nrd2983
Younger, D.S., 1999. Motor Disorders. Lippincott Williams & Wilkins.
Zhou, Y., Ramanathan, S., 2015. Mott Memory and Neuromorphic Devices. Proc. IEEE 103,
1289 -- 1310. https://doi.org/10.1109/JPROC.2015.2431914
Acknowledgments:
This work was performed, in part, at the Center for Nanoscale Materials, a U.S. Department of
Energy Office of Science User Facility under Contract No. DE-AC02-06CH11357. Aspects of the
device work was supported by the National Science Foundation under grant 1640081, and the
Nanoelectronics Research Corporation (NERC), a wholly owned subsidiary of the Semiconductor
Research Corporation (SRC), through Extremely Energy Efficient Collective Electronics
(EXCEL), an SRC-NRI Nanoelectronics Research Initiative under Research Task ID 2698.001.
S.R. acknowledges the support by ARO W911NF-16-1-0289 and ONR N00014-16-1-2398. The
authors acknowledge K. V. L. V. Achari for providing vanadium dioxide film samples.
|
1703.09347 | 3 | 1703 | 2018-03-21T11:35:37 | Adaptive Scales of Spatial Integration and Response Latencies in a Critically-Balanced Model of the Primary Visual Cortex | [
"q-bio.NC",
"math.DS"
] | The brain processes visual inputs having structure over a large range of spatial scales. The precise mechanisms or algorithms used by the brain to achieve this feat are largely unknown and an open problem in visual neuroscience. In particular, the spatial extent in visual space over which primary visual cortex (V1) performs evidence integration has been shown to change as a function of contrast and other visual parameters, thus adapting scale in visual space in an input-dependent manner. We demonstrate that a simple dynamical mechanism---dynamical criticality---can simultaneously account for the well-documented input-dependence characteristics of three properties of V1: scales of integration in visuotopic space, extents of lateral integration on the cortical surface, and response latencies. | q-bio.NC | q-bio |
Adaptive Scales of Spatial Integration and Response Latencies in a
Critically-Balanced Model of the Primary Visual Cortex
Keith Hayton, Dimitrios Moirogiannis, and Marcelo Magnasco
Laboratory of Integrative Neuroscience, The Rockefeller University
(Dated: September 25, 2018)
The primary visual cortex (V1) integrates information over scales in visual space, which have
been shown to vary, in an input-dependent manner, as a function of contrast and other visual
parameters. Which algorithms the brain uses to achieve this feat are largely unknown and an open
problem in visual neuroscience. We demonstrate that a simple dynamical mechanism can account
for this contrast-dependent scale of integration in visuotopic space as well as connect this property
to two other stimulus-dependent features of V1: extents of lateral integration on the cortical surface
and response latencies.
INTRODUCTION
Stimuli in the natural world have quantitative char-
acteristics that vary over staggering ranges. Our ner-
vous system evolved to parse such widely-ranging stim-
uli, and research into how the nervous system can cope
with such ranges has led to considerable advances in
our understanding of neural circuitry. For example at
the sensory transduction level, the physical magnitudes
encoded into primary sensors, such as light intensity,
sound pressure level and olfactant concentration, vary
over exponentially-large ranges, leading to the Weber-
Fechner law [1]. As neuronal firing rates cannot vary
over such large ranges, the encoding process must com-
press physical stimuli into the far more limited ranges
of neural activity that represent them. These observa-
tions have stimulated a large amount of research into the
mechanisms underlying nonlinearly compression of phys-
ical stimuli in the nervous system . Of relevance to our
later discussion is the nonlinear compression of sound in-
tensity in the early auditory pathways [2, 3], where it has
been shown that poising the active cochlear elements on
a Hopf bifurcation leads to cubic-root compression.
But other characteristics besides the raw physical mag-
nitude still vary hugely. The wide range of spatial ex-
tents and correlated linear structures present in visual
scenery [4 -- 6] leads to a more subtle problem, if we think
of the visual areas as fundamentally limited by corre-
sponding anatomical connectivity. Research into this
problem has been focused on elucidating the nature of
receptive fields of neurons in the primary visual cortex
(V1) [7 -- 12]. Studies have found that as the contrast of
a stimulus is decreased, the receptive field [13, 14] size
or area of spatial summation in visual space increases
(Fig 1) [11, 12, 15, 16]. As an example of contextual
modulation of neuronal responses, this problem has nat-
urally received theoretical attention [17 -- 19]. However,
current literature does not describe this phenomenon as
structurally integral to the neural architecture but rather
either highlight a different set of features or the contex-
tual modulations are explicitly written in an ad hoc fash-
ion. Our aim is to develop a model which displays this
phenomenon structurally, as a direct consequence of the
In our proposed models, multiple
neural architecture.
length scales emerge naturally without any fine tuning
of the system's parameters. This leads to length-tuning
curves similar to the ones measured in Kapadia et al.
over the entire range (Fig 1) [11].
The findings of Kapadia et al. demonstrate that recep-
tive fields in V1 are not constant but instead grow and
shrink, seemingly beyond naive anatomical parameters,
according to stimulus contrast. The "computation" being
carried out is not fixed but is itself a function of the input.
Let us examine this distinction carefully. There are nu-
merous operations in image processing, such as Gaussian
blurs or other convolutional kernels, whose spatial range
is fixed.
It is very natural to imagine neural circuitry
having actual physical connections corresponding to the
nonzero elements of a convolutional kernel, and in fact a
fair amount of effort has been expended trying to identify
actual synapses corresponding to such elements [20, 21].
There are, however, other image-processing operations,
such as floodfill (the "paint bucket") whose spatial ex-
tent is entirely dependent on the input; the problem of
"binding" of perceptual elements is usually thought about
in this way, and mechanisms posited to underlie such
propagation dynamics include synchronization of oscilla-
tions acting in a vaguely paint-bucket-like way [22 -- 24].
This dichotomy is artificial because these are only the
two extremes of a potentially continuous range. While
the responses of neurons in V1 superficially appear to be
convolutional kernels, their strong dependence on input
characteristics, particularly the size of the receptive field,
demonstrates a more complex logic in which spatial ex-
tent is determined by specific characteristics of the input.
What is the circuitry underlying this logic?
Neurons in the primary visual cortex are laterally con-
nected to other neurons on the cortical surface and derive
input from them. Experiments have shown that the spa-
tial extent on the cortical surface from which neurons
derive input from other neurons through such lateral in-
teractions varies with the contrast of the stimulus [25].
In the absence of stimulus contrast, spike-triggered trav-
eling waves of activity propagate over large areas of cor-
tex. As contrast is increased, the waves become weaker
in amplitude and travel over increasingly small distances.
These experiments suggest that the change in spatial
summation area with increasing stimulus contrast may
be consistent with the change in the decay constants of
the traveling wave activity. However, no extant experi-
ment directly links changes in summation in visual space
to changes in integration on the cortical surface, and no
explicit model of neural architecture has been shown to
simultaneously account for, and thus connect, the input-
dependence of spatial summation and lateral integration
in V1. The latter one is our aim, and a crucial clue will
come from the input-dependence of latencies.
Figure 1. Reprinted from [11] with permission: Copy-
right (1999) National Academy of Sciences, U.S.A.
Measurements of single-neuron responses in the V1
area of monkeys to optimally oriented bars of light
of different lengths and contrasts. Panels a and b are
measurements from two distinct neurons. The units of length
along the horizontal axis are in minutes of arc. The solid,
dotted, and dashed curves represent bars of light of 50% con-
trast, 15% contrast, and 50% contrast embedded in a textured
background, respectively. The dashed curves are irrelevant to
the focus of this paper.
Recently, a critically-balanced network model of cor-
tex was proposed to explain the contrast dependence of
functional connectivity [26].
It was shown that in the
absence of input, the model exhibits wave-like activity
with an infinitely-long ranged susceptibility, while in the
presence of input, perturbed network activity decays ex-
ponentially with an attenuation constant that increases
with the strength of the input. These results are in direct
agreement with Nauhaus et al. [25].
We will now demonstrate that a similar model also
leads to adaptive scales of spatial integration in visual
space. Our model makes two key assumptions. The first
is a local, not just global, balance of excitation and in-
hibition across the entire network; all eigenmodes of the
network are associated with purely imaginary eigenval-
ues.
It has been shown that such a critically-balanced
configuration can be achieved by simulating a network of
neurons with connections evolving under an anti-Hebbian
rule [27]. The second key assumption is that all interac-
tions in the network are described by the connectivity
matrix; nonlinearities do not couple distinct neurons in
the network.
There are a number of examples of dynamical criti-
cality in neuroscience, including experimental studies in
2
motor cortex [28], theoretical [29] and experimental stud-
ies [30] of line attractors in oculomotor control, line at-
tractors in decision making [31], Hopf bifurcation in the
auditory periphery [32] and olfactory system [33], and
theoretical work on regulated criticality [34]. More re-
cently, Solovey et al. [35] performed stability analysis of
high-density electrocorticography recordings covering an
entire cerebral hemisphere in monkeys during reversible
loss of consciousness. Performing a moving vector au-
toregressive analysis of the activity, they observed that
the eigenvalues crowd near the critical line. During loss
of consciousness, the numbers of eigenmodes at the edge
of instability decrease smoothly but drift back to the crit-
ical line during recovery of consciousness.
We also examine the dynamics of the system and show
that its activity exponentially decays to a limit cycle over
multiple timescales, which depend on the strength of the
input. Specifically, we find that the temporal exponential
decay constants increase with increasing input strength.
This result agrees with single-neuron studies which have
found that response latencies in V1 decrease with in-
creasing stimulus contrast [11, 36 -- 38]. We now turn to
describing our model.
METHODS
Let x ∈ CN be the activity vector for a network of
neurons which evolve in time according to the normal
form equation:
(cid:88)
xi =
Aijxj − xi2xi + Ii(t)
(1)
j
In this model, originally proposed by Yan and Magnasco
[26], neurons interact with one another through a skew-
symmetric connectivity matrix A. The cubic-nonlinear
term in the model is purely local and does not couple
the activity states of distinct neurons, while the external
input I(t) ∈ CN to the system may depend on time and
have a complex spatial pattern.
The original model considered a 2-D checkerboard
topology of excitatory and inhibitory neurons. For theo-
retical simplicity and computational ease, we will instead
consider a 1-D checkerboard layout of excitatory and in-
hibitory neurons which interact through equal strength,
nearest neighbor connections (Fig 2). In this case, Aij =
(−1)js(δi,j+1 +δi,j−1), where i, j = 0, 1, ..., N −1 and s is
the synaptic strength. Boundary conditions are such that
the activity terminates to 0 outside of the finite network.
We are specifically interested in the time-asymptotic
response of the system, but explicitly integrating the stiff,
high-dimensional ODE in (1) is difficult. Fortunately, we
can bypass numerical integration methods by assuming
periodic input of the form I(t) = F eiωt, where F ∈ CN
and look for solutions X(t) = Zeiωt, where Z ∈ CN.
Substituting these into (1), we find that:
0 = (A − iω)Z − Z2Z + F
(2)
3
Figure 2. Simplest connectivity matrix A. A finite line of
excitatory and inhibitory neurons. White nodes represent ex-
citatory neurons. Black nodes are inhibitory. All connections
have strength of the same magnitude.
And define g(Z) to be equal to the right hand side of
(2).
the multivariable Newton-Raphson method in CN:
The solution of (2) can be found numerically by using
(cid:101)Z → (cid:101)Z − J((cid:101)Z)−1(cid:101)g(Z)
where (cid:101)Z and(cid:101)g are the concatenations of the real and
bian of(cid:101)g with respect to
imaginary parts of Z and g, respectively. J is the Jaco-
(3)
∂(cid:101)gi
∂(cid:101)zj
Jij(z) =
RESULTS
To test how the response of a single neuron in the net-
work varies with both the strength and length of the in-
put, we select a center neuron at index c and then calcu-
late, for a range of input strengths, the response of the
neuron as a function of input length around it. Formally,
for each input strength level B ∈ R, we solve (3) for:
Fk(B, l) =
Bvk
0
if k ∈ [c − l, c + l]
otherwise
(4)
(cid:40)
Figure 3. Length-response curves for different eigen-
frequencies. In each panel, which corresponds to a different
eigenfrequency of A, we plot the response of a neuron (with
index N/2) as a function of input length for a group of expo-
nentially distributed input strengths. The blue arrow in the
first plot indicates the direction of increasing input strength.
The length of the input is recorded in number of neurons and
the response is taken as the modulus of the amplitude of the
time-asymptotic stable limit cycle.
where k = 0, ..., N − 1, v ∈ CN describes the spatial
shape of the input, and 2l + 1 is the length of the in-
put in number of neurons. The response of the center
neuron is taken as the modulus of Zc, and we focus on
the case where ω is an eigenfrequency of A and v the
corresponding eigenvector.
The results for a 1-D checkerboard network of 64 neu-
rons is shown in Fig 3. Here we fix a center neuron and
sweep across a small range of eigenfrequencies ω of A.
The curves from bottom to top correspond to an as-
cending order of base-2 exponentially distributed input
strengths C = 2i. For all eigenfrequencies, the peak of
the response curves shift towards larger input lengths as
the input strength decreases. In fact, for very weak in-
put, the response curves rise monotonically over the en-
tire range of input lengths without ever reaching a max-
imum in this finite network. This is in contrast to the
response curves corresponding to strong input, which al-
ways reach a maximum but, depending on the eigenfre-
quency, exhibit varying degrees of response suppression
beyond the maximum. This is consistent with variability
of response suppression in primary visual cortex studies
[11, 12]. In Fig 3, eigenfrequencies ω = 1.92, 1.96, 1.99
show the greatest amount of suppression while the oth-
ers display little to none.
To understand why certain eigenfrequencies lead to
suppression, we fix the eigenfrequency to be ω = 1.92 and
examine the response curves of different center neurons.
The response of four center neurons (labeled by network
position) and the modulus of the eigenfrequency's cor-
responding eigenvector are plotted in Fig 4. The center
neurons closest to the zeros of the eigenvector experience
the strongest suppression for long line lengths. Neuron
38 closer to the peak of the eigenvector's modulus ex-
periences almost zero suppression. This generally holds
for all eigenvectors and neurons in the network as all
eigenvectors are periodic in their components with an
eigenvalue-dependent spatial frequency.
To strengthen the connection between model and neu-
rophysiology, one can consider a critically-balanced net-
the slow decay regime becomes dominant. The multiple
decay regimes is a surprising result which doesn't appear
in the case of a single critical Hopf oscillator.
4
Figure 4. Length-response curves of different neurons.
The top plot depicts the modulus of the eigenvector corre-
sponding to eigenfrequency ω = 1.92. The 4 panels below are
plots of the length-response curves for 4 different neurons in
the network. The position of the neurons relative to the shape
of the eigenvector are noted with the gray bars and arrows.
work with an odd number of neurons so that 0 is now
an eigenfrequency of the system. In our model, input as-
sociated with the 0-eigenmode represents direct current
input to the system which is what neurophysiologists uti-
lize in experiments; the visual input is not flashed [11, 12].
Contrary to the even case, long range connections must
be added on top of the nearest neighbor connectivity in
order to recover periodic eigenvectors and hence suppres-
sion past the response curves maximums.
Next, we show that the network not only selectively
integrates input as a function of input strength but also
operates on multiple time scales which flexibly adapt to
the input. This behavior is not surprising given that
in the case of a single critical Hopf oscillator, the half
width of the resonance, the frequency range for which
the oscillator's response falls by a half, is proportional
to the forcing strength of the input, Γ ∝ F 2
3 where Γ
is the half-width F the input strength [2]. Thus, decay
constants in the case of a single critical oscillator should
grow with the input forcing strength as F 2
3 .
Assuming input F eiωt, as described above, the network
activity x(t), given by (1), decays exponentially in time
to a stable limit cycle, X(t) = Zeiωt. This implies that
for any neuron i in the network, xi(t) = e−btf (t) + Zi
during the approach to the limit cycle. We therefore
plot log (xi(t) − Zi) over the transient decay period
and estimate the slope of the linear regimes. We do this
for a nearly network size input length (input length=29,
N = 32) and a range of exponentially distributed input
strengths. In Fig 5, we plot representative transient peri-
ods of a single neuron corresponding to 3 input strengths:
2−10, 2−4, and 22. For weak input there is a fast sin-
gle exponential decay regime (red) that determines the
system's approach to the stable limit cycle. As we in-
crease the input, however, the transient period displays
two exponential decay regimes: the fast decay regime
(red) which was observed in the presence of weak input
and a new slow decay regime (blue) immediately preced-
ing the stable limit cycle. For very large input strength,
Input-strength-dependent timescales. For
Figure 5.
neuron i in the network, we plot log (xi(t) − Zi) as a func-
tion of time for 3 different input strengths. Linear regions
correspond to exponential decay. In the presence of weak in-
put, a fast decay regime (red) guides the dynamics towards
a stable limit cycle. For the intermediate input strength, a
new, distinct slow decay regime appears (blue), which be-
comes dominant for strong input
We estimate the exponential decay constants as a func-
tion of input strength and plot them on a log-log scale
in Fig 6. The red circles correspond to the fast decay
regime, while the blue circles correspond to the slow de-
cay regime, which becomes prominent for large forcings.
We separately fit both the slow and fast decay regimes
with a best fit line. Unsurprisingly, the slopes of the
3. Thus, the decay
lines are equal and approximately 2
constants grow with the input as ∝ F 2
3 , where F is the
input strength. This implies that the system operates
on multiple timescales dynamically switching from one
to another depending on the magnitude of the forcing.
Larger forcings lead to faster network responses.
In this paper, we consider a line of excitatory and in-
hibitory neurons, but our results hold equally well for a
ring of neurons with periodic boundary conditions and
appropriately chosen long range connections. Ring net-
works have extensively been studied as a model of ori-
entation selectivity in V1 [39 -- 45].
In agreement with
recent findings [46], the critically-balanced ring network
exhibits surround suppression in orientation space when
long range connections are added on top of nearest neigh-
bor connectivity.
CONCLUSION
We have shown that a simple dynamical system poised
at the onset of instability exhibits an input-strength-
dependent scale of integration of the system's input and
input-strength-dependent response latencies. This find-
ing strongly complements our previous results showing
that a similar nonlinear process with fixed, nearest neigh-
bor network connectivity leads to input-dependent func-
5
tional connectivity. This system is thus the first pro-
posed mechanism that can account for contrast depen-
dence of spatial summation, functional connectivity, and
response latencies. In this framework, these three char-
acteristic properties of signal processing in V1 are intrin-
sically linked to one another.
Input-strength dependence of exponential
Figure 6.
decay constants. The temporal exponential decay constants
for a range of input strengths is depicted above. A fast decay
regime (red circles) is accompanied by a slow decay regime
(blue circles) at large input strengths. Each decay regime is
separately approximated by a least squares line (dashed lines)
in log-log space.
[1] Fechner G. 1966 Elements of Psychophysics. Howes, DH,
Boring, EC, Adler, HE Holt, Rinehart and Winston, New
York ((Translated from German) Originally published in
1860). 1860.
[2] EguÃluz VM, Ospeck M, Choe Y, Hudspeth AJ, Mag-
nasco MO. Essential nonlinearities in hearing. Physical
Review Letters. 2000 May 29;84(22):5232.
[3] Camalet S, Duke T, JÃOElicher F, Prost J. Auditory sen-
sitivity provided by self-tuned critical oscillations of hair
cells. Proceedings of the National Academy of Sciences.
2000 Mar 28;97(7):3183-8.
[4] Field DJ. Relations between the statistics of natural im-
ages and the response properties of cortical cells. Josa a.
1987 Dec 1;4(12):2379-94.
[5] Ruderman DL, Bialek W. Statistics of natural images:
Scaling in the woods. Physical Review Letters. 1994 Aug
8;73(6):814-817.
[6] Sigman M, Cecchi GA, Gilbert CD, Magnasco MO. On
a common circle: natural scenes and Gestalt rules. Pro-
ceedings of the National Academy of Sciences. 2001 Feb
13;98(4):1935-40.
[7] Kapadia MK, Ito M, Gilbert CD, Westheimer G. Im-
provement in visual sensitivity by changes in local con-
text: parallel studies in human observers and in V1 of
alert monkeys. Neuron. 1995 Oct 31;15(4):843-56.
[8] Zipser K, Lamme VA, Schiller PH. Contextual modula-
tion in primary visual cortex. Journal of Neuroscience.
1996 Nov 15;16(22):7376-89.
[9] Levitt JB, Lund JS. Contrast dependence of contex-
tual effects in primate visual cortex. Nature. 1997 May
1;387(6628):73.
[10] Polat U, Mizobe K, Pettet MW, Kasamatsu T, Nor-
cia AM. Collinear stimuli regulate visual responses de-
pending on cell's contrast threshold. Nature. 1998 Feb
5;391(6667):580-4.
[11] Kapadia MK, Westheimer G, Gilbert CD. Dynamics of
spatial summation in primary visual cortex of alert mon-
keys. Proceedings of the National Academy of Sciences.
1999 Oct 12;96(21):12073-8.
[12] Sceniak MP, Ringach DL, Hawken MJ, Shapley R. Con-
trast's effect on spatial summation by macaque V1 neu-
rons. Nature neuroscience. 1999 Aug 1;2(8):733-9.
[13] Hubel DH, Wiesel TN. Receptive fields, binocular inter-
action and functional architecture in the cat's visual cor-
tex. The Journal of physiology. 1962 Jan 1;160(1):106-54.
[14] Kuffler SW. Discharge patterns and functional organiza-
tion of mammalian retina. Journal of neurophysiology.
1953 Jan 1;16(1):37-68.
[15] DeAngelis GC, Robson JG, Ohzawa I, Freeman RD. Or-
ganization of suppression in receptive fields of neurons in
cat visual cortex. Journal of Neurophysiology. 1992 Jul
1;68(1):144-63.
[16] DeAngelis GC, Freeman RD, Ohzawa IZ. Length and
width tuning of neurons in the cat's primary visual cor-
tex. Journal of neurophysiology. 1994 Jan 1;71(1):347-74.
[17] Schwabe L, Obermayer K, Angelucci A, Bressloff PC.
The role of feedback in shaping the extra-classical recep-
tive field of cortical neurons: a recurrent network model.
Journal of Neuroscience. 2006 Sep 6;26(36):9117-29.
[18] Lochmann T, Ernst UA, Deneve S. Perceptual inference
predicts contextual modulations of sensory responses.
Journal of neuroscience. 2012 Mar 21;32(12):4179-95.
[19] Zhu M, Rozell CJ. Visual nonclassical receptive field ef-
fects emerge from sparse coding in a dynamical system.
PLoS computational biology. 2013 Aug 29;9(8):e1003191.
[20] Olshausen BA, Field DJ. Emergence of simple-cell recep-
tive field properties by learning a sparse code for natural
images. Nature. 1996 Jun;381(6583):607.
[21] Reid RC, Alonso JM. Specificity of monosynaptic con-
nections from thalamus to visual cortex. Nature. 1995
Nov;378(6554):281.
[22] Rosenblatt F. Principles of neurodynamics. perceptrons
and the theory of brain mechanisms. CORNELL AERO-
NAUTICAL LAB INC BUFFALO NY; 1961 Mar 15.
[23] Von der Malsburg C. The what and why of binding: the
modeler's perspective. Neuron. 1999 Sep 30;24(1):95-104.
[24] Lee TS, Mumford D. Hierarchical Bayesian inference in
the visual cortex. JOSA A. 2003 Jul 1;20(7):1434-48.
[25] Nauhaus I, Busse L, Carandini M, Ringach DL. Stimu-
lus contrast modulates functional connectivity in visual
cortex. Nature neuroscience. 2009 Jan 1;12(1):70-6.
[26] Yan XH, Magnasco MO. Input-dependent wave attenu-
ation in a critically-balanced model of cortex. PloS one.
2012 Jul 25;7(7):e41419.
[27] Magnasco MO, Piro O, Cecchi GA. Self-tuned critical
anti-Hebbian networks. Physical review letters. 2009 Jun
22;102(25):258102.
[28] Churchland MM, Cunningham JP, Kaufman MT, Fos-
ter JD, Nuyujukian P, Ryu SI, Shenoy KV. Neu-
ral population dynamics during reaching. Nature. 2012
Jul;487(7405):51.
[29] Seung HS. Continuous attractors and oculomotor control.
Neural Networks. 1998 Nov 30;11(7):1253-8.
[30] Seung HS, Lee DD, Reis BY, Tank DW. Stability of
the memory of eye position in a recurrent network of
conductance-based model neurons. Neuron. 2000 Apr
30;26(1):259-71.
[31] Machens CK, Romo R, Brody CD. Flexible control of
mutual inhibition: a neural model of two-interval dis-
crimination. Science. 2005 Feb 18;307(5712):1121-4.
[32] Choe Y, Magnasco MO, Hudspeth AJ. A model for am-
plification of hair-bundle motion by cyclical binding of
Ca2+ to mechanoelectrical-transduction channels. Pro-
ceedings of the National Academy of Sciences. 1998 Dec
22;95(26):15321-6.
[33] Freeman WJ, Holmes MD. Metastability, instability, and
state transition in neocortex. Neural Networks. 2005 Aug
31;18(5):497-504.
[34] Bienenstock E, Lehmann D. Regulated criticality
in the brain?. Advances in complex systems. 1998
Dec;1(04):361-84.
[35] Solovey G, Alonso LM, Yanagawa T, Fujii N, Magnasco
MO, Cecchi GA, Proekt A. Loss of consciousness is as-
6
sociated with stabilization of cortical activity. Journal of
Neuroscience. 2015 Jul 29;35(30):10866-77.
[36] Carandini M, Heeger DJ. Summation and division by
neurons in primate visual cortex. Science-AAAS-Weekly
Paper Edition-including Guide to Scientific Information.
1994 May 27;264(5163):1333-5.
[37] Gawne TJ, Kjaer TW, Richmond BJ. Latency: another
potential code for feature binding in striate cortex. Jour-
nal of neurophysiology. 1996 Aug 1;76(2):1356-60.
[38] Albrecht DG, Geisler WS, Frazor RA, Crane AM. Visual
cortex neurons of monkeys and cats: temporal dynamics
of the contrast response function. Journal of Neurophys-
iology. 2002 Aug 1;88(2):888-913.
[39] Ben-Yishai R, Bar-Or RL, Sompolinsky H. Theory of ori-
entation tuning in visual cortex. Proceedings of the Na-
tional Academy of Sciences. 1995 Apr 25;92(9):3844-8.
[40] Hansel D, Sompolinsky H. 13 Modeling Feature Selectiv-
ity in Local Cortical Circuits.
[41] Shriki O, Hansel D, Sompolinsky H (2003) Shriki O,
Hansel D, Sompolinsky H. Rate models for conductance-
based cortical neuronal networks. Neural computation.
2003 Aug;15(8):1809-41.
[42] Ermentrout B. Neural networks as
spatio-temporal
pattern-forming systems. Reports on progress in physics.
1998 Apr;61(4):353.
[43] Bressloff PC, Bressloff NW, Cowan JD. Dynamical mech-
anism for sharp orientation tuning in an integrate-and-
fire model of a cortical hypercolumn. Neural computa-
tion. 2000 Nov;12(11):2473-511.
[44] Bressloff PC, Cowan JD, Golubitsky M, Thomas PJ,
Wiener MC. Geometric visual hallucinations, Euclidean
symmetry and the functional architecture of striate
cortex. Philosophical Transactions of the Royal So-
ciety of London B: Biological Sciences. 2001 Mar
29;356(1407):299-330.
[45] Dayan P, Abbott LF. Theoretical neuroscience. Cam-
bridge, MA: MIT Press; 2001.
[46] Rubin DB, Van Hooser SD, Miller KD. The stabilized
supralinear network: a unifying circuit motif underlying
multi-input integration in sensory cortex. Neuron. 2015
Jan 21;85(2):402-17.
|
1406.5096 | 3 | 1406 | 2015-01-15T01:21:08 | Nonlinear network dynamics under perturbations of the underlying graph | [
"q-bio.NC"
] | Many natural systems are organized as networks, in which the nodes (be they cells, individuals or populations) interact in a time-dependent fashion. The dynamic behavior of these networks depends on how these nodes are connected, which can be understood in terms of an adjacency matrix, and connection strengths. The object of our study is to relate connectivity to temporal behavior in networks of coupled nonlinear oscillators. We investigate the relationship between classes of system architectures and classes of their possible dynamics, when the nodes are coupled according to a connectivity scheme that obeys certain constrains, but also incorporates random aspects.
We illustrate how the phase space dynamics and bifurcations of the system change when perturbing the underlying adjacency graph. We differentiate between the effects on dynamics of the following operations that directly modulate network connectivity: (1) increasing/decreasing edge weights, (2) increasing/decreasing edge density, (3) altering edge configuration by adding, deleting or moving edges.
We discuss the significance of our results in the context of real life networks. Some interpretations lead us to draw conclusions that may apply to brain networks, synaptic restructuring and neural dynamics. | q-bio.NC | q-bio | Nonlinear network dynamics under perturbations of the
underlying graph
Anca Radulescu∗,1, Sergio Verduzco-Flores2
1 Department of Mathematics, SUNY New Paltz, NY 12561
2 Department of Psychology, University of Colorado at Boulder, CO 80309
Abstract
Many natural systems are organized as networks, in which the nodes (be they cells, individuals
or populations) interact in a time-dependent fashion. The dynamic behavior of these networks
depends on how these nodes are connected, which can be understood in terms of an adjacency
matrix, and connection strengths. The object of our study is to relate connectivity to temporal
behavior in networks of coupled nonlinear oscillators. We investigate the relationship between
classes of system architectures and classes of their possible dynamics, when the nodes are cou-
pled according to a connectivity scheme that obeys certain constrains, but also incorporates
random aspects.
We illustrate how the phase space dynamics and bifurcations of the system change when per-
turbing the underlying adjacency graph. We differentiate between the effects on dynamics of
the following operations that directly modulate network connectivity: (1) increasing/decreasing
edge weights, (2) increasing/decreasing edge density, (3) altering edge configuration by adding,
deleting or moving edges.
We discuss the significance of our results in the context of real life networks. Some interpreta-
tions lead us to draw conclusions that may apply to brain networks, synaptic restructuring and
neural dynamics.
5
1
0
2
n
a
J
5
1
]
.
C
N
o
i
b
-
q
[
3
v
6
9
0
5
.
6
0
4
1
:
v
i
X
r
a
1Assistant Professor, Department of Mathematics, State University of New York at New Paltz; New York, USA;
Phone: (845) 257-3532; Email: [email protected]
1
The study of a dynamical system with interconnected nodes tends to gain little insight
from the graph-theoretical properties of the underlying graph. Brain networks show
properties such as small-world connectivity and repeated motifs, but the computa-
tional impact of those properties remains unclear. An avenue pursued in this paper in
order to investigate the relation between a network's structure and its dynamics is to
find relations between the network's underlying graph, and the behavior of this system.
We perform a computational study of dynamics under different forms of connectivity
between two densely connected modules. Using phase diagrams and a probabilistic
extension of bifurcation diagrams in the parameter space, we find several properties
of the network dynamics. Among them, we find that the spectrum of the adjacency
matrix is a poor predictor of dynamics when using nonlinear nodes, that increasing
the number of connections between the two nodes is not equivalent to strengthening a
few connections, and that there is no single factor among those we tested that governs
the stability of the system.
1
Introduction
A large body of literature over the past decade has been dedicated to the study of networks and
their applications to understanding the behavior of social, neural and biological systems. One of the
particular points of interest has been the question of how the hardwired structure of the network
(its underlying graph) affects its function, for example in the context of information storage or
transmission between nodes along time [2]. There are two key coupling aspects that govern dynamic
function in such networks: the underlying graph (characterized by its adjacency matrix) and the
connection strengths. Understanding the effects of configuration (which is another term we'll use for
the adjacency matrix) on coupled dynamics is of great importance for a wide variety of applications.
There are not many previous studies dealing with the direct effect of configuration on a dynam-
ical system. Naquib et al [1] found that by varying the location of synthetic sites for two different
chemical species (one excitatory and one inhibitory) that diffuse across a one-dimensional ring, they
could find many dynamic behaviors, including fixed points, out-of-phase oscillations, quasiperiod-
icity, and chaos. An interesting aspect of this study is that the structure of the system (i.e. the
location and identity of the synthetic sites) acts as a bifurcation parameter. We also explore the
idea of having structure as a bifurcation parameter; a structure found in the adjacency matrix of
the network.
The differential equations that model dynamical systems consisting of interconnected nonlinear
nodes do not usually admit closed form analytical solutions.
In this situation, the qualitative
behavior of the system may still be grasped through bifurcation analysis, which for complex systems
is usually carried out using numerical continuation methods. These methods are appropriate to
study a single system, where the graph describing the connections among its nodes is fixed. For
the aim of this paper, however, we want to understand how the system's dynamics change as
the adjacency matrix that describes its underlying graph experiences variations. This involves
evaluating a large number of systems, each with a different adjacency matrix. Because of this,
bifurcation analysis with numerical continuation methods becomes computationally expensive (the
number of possible adjacency matrices increases exponentially with the number of nodes), and it
would be unclear how to interpret a large number of bifurcation diagrams, each one for a different
underlying graph.
We propose a simple approach to visualize the qualitative behavior of nonlinear dynamical
sytems with an underlying graph structure. The approach starts by discretizing the values in the
bifurcation diagram to obtain a finite number of points in parameter space. For each point in
2
parameter space we take a sample of the adjacency matrices, and for each one, find whether the
system expresses the dynamical behaviors of interest (e.g. bistability, or oscillations). For each
dynamical behavior of interest, we can create a diagram where each point in parameter space is
associated with the fraction of adjacency matrices causing the behavior to appear in the system.
This diagram expresses, for each point in parameter space, what is the probability that a given
dynamical behavior will appear in the absence of information about the system's configuration. If
the sampling of adjacency matrices is restricted to those satisfying a particular constraint (e.g. a
fixed number of ones in the adjacency matrix), then the diagram will express an approximation to
the corresponding conditional probabilities. We use the term of (dynamic) behavior frequency plots
to refer to diagrams produced this way, which can be applied for any specific behavior.
To focus in a particular direction, we place and interpret our results in the context of brain
architecture and dynamics. Understanding the way in which various parts of the brain (from the
micro-scale of neurons to the macro-scale of functional regions) are wired together is one of the great
scientific challenges of the 21st century, currently being addressed by large-scale research collabo-
rations, such as the Human Connectome Project [24]. Many recent studies (e.g.,
[21, 23, 13]) have
used a combination of dynamical systems and graph theoretical approaches to investigate general
organizational principles of brain networks. With nodes and edges defined according to imaging
modality appropriate scales, empirical studies have found certain generic topological properties of
the human brain architecture, such as modularity, small-worldness, the presence of rich clubs and
hubs, and other connectivity patterns [5, 18, 13, 22].
Purely empirically-based analyses cannot, however, explain in and off themselves the mecha-
nisms by which connectivity patterns may actually act to change the system's dynamics, and thus
the observed behavior. Substantial research efforts are being directed towards constructing un-
derlying network models that are tractable theoretically or numerically, and which could therefore
be used in conjunction with the data towards interpreting the empirical results, and for making
further predictions. To this aim, the theoretical dependence of dynamics on connectivity (e.g., in
the context of stability and synchronization in networks of coupled neural populations) has been
investigated both analytically and numerically, in a variety of contexts, from biophysical models [11]
to simplified systems [20]. These analyses revealed a rich range of potential dynamic regimes and
transitions [4], shown to depend as much on the coupling parameters of the network as on the
arrangement of the excitatory and inhibitory connections [11]. Understanding and teasing apart
the different effects of these dependences is the central goal of this work.
We will start our study by considering a type of architecture already used in previous work [17,
19]: an oriented graph composed of two interconnected cliques (fully connected subgraphs), module
X and module Y , so that all nodes {xk}k=1,N within X are mutually connected by "excitatory"
edges with equal positive weights gxx, and all nodes {yk}k=1,N within Y are mutually connected by
excitatory edges with positive weights gyy (Figure 1). The connectivity patterns from X-to-Y and
Y -to-X can be described by two binary N ×N blocks A = (Akp) and B = (Bkp), representing which
of the nodes in X are cross-connected to nodes in Y (with equal excitatory, positive weights gxy) and
conversely, which of the nodes in Y are connected to nodes in X (with inhibitory, negative weights
−gyx). This graph structure was chosen in previous work as a very simple framework for studying
the excitatory/inhibitory feedback interaction in a control system composed of two brain regions
(in our case the amygdala and the prefrontal cortex), with the nodes representing hemodynamic
oscillators. The set-up can be used, however, at other spacial and temporal scales, or for any
bimodular network defining a similar feedback loop. One can easily adapt it to incorporate more
than two modules, or can prune out the dense intra-modular connections to obtain more realistic
conditions, while keeping it simple enough to address numerically or analytically, for sufficiently
large numbers of nodes.
In a previous paper [19], we had focused primarily on the properties of the graph underlying a
neural network, and discussed how factors such as changes in density or other edge restructuring
3
Figure 1: Schematic representation of the network for N = 4 nodes per module. Module X
is shown on the left; module Y is shown on the right; they are both fully-connected, local sub-graphs of
the oriented graph corresponding to the whole network. The thick blue arrow shows that there are Mxy
connections from X to Y , and the thick green arrow that there are Myx connections from Y to X (the edges
are not shown in their specific positions, making this a general representation of any configuration with the
given edge densities). The coupling weights gxx, gxy, gyx, gyy are marked on the corresponding edges.
may affect the spectrum of the adjacency matrix. In this paper, our attention is directed towards
further relating adjacency properties to the system's temporal behavior, and understanding the
subsequent changes they trigger in the coupled dynamics. More precisely, we are interested in
varying the number of active inter-modular edges Mxy and Myx (i.e., Mxy is the number of 1
entries in A and Myx is the number of 1 entries in B, both ranging from zero to the theoretical
maximum N 2), but also in changing the edge configuration for a fixed pair (Mxy, Myx) (which we
will call the density type of the graph, for the remainder of this paper).
We investigate the consequences that each of these two aspects has on the overall dynamics of a
system of coupled nodes, where we identify each node with a continuous-time nonlinear oscillator.
From a vast collection of such models, we drew our inspiration from the simple and traditional
Wilson-Cowan equations, a system conceived and used historically to model interaction of excita-
tory and inhibitory neural populations [25]. This two-dimensional system was shown to exhibit
interesting dynamic behavior in the two-dimensional phase-space, with Hopf and fold bifurcations
between stable equilibria and stable limit cycles (including bistability windows). In our study, we
work within the parameter ranges proposed in the original Wilson-Cowan paper [25] as weel as in
subsequent work in higher dimensions [3, 6], thus placing the system in the vicinity of the inter-
esting phase-plane phenomena. We study how the phase-plane dynamics and the parameter-plane
transitions change when perturbing the underlying coupling graph.
Some of the results we obtained for this system were intuitive, but others were rather unex-
pected. For example, we established that the eigenvalues of the adjacency matrix do not determine
the dynamic behavior, which is not surprising. Conversely however, the dynamic behavior seems
to determine the eigenvalues. The intermodular connection weights can put the system in sensitive
regimes where changes in the adjacency matrix is more likely to affect the dynamic behavior. (At
least in the low dimensional case, these regimes appear for the gxy, gyx values near bifurcation
curves for individual configurations.) In fact sparse connectivity (smaller Mxy, Myx values) pro-
motes a higher variety of dynamic behaviors, accessible by changes in either weights of adjacency
4
xk = −xk + (1 − xk) · Sbx,θx
yk = −yk + (1 − yk) · Sby,θy
− N(cid:88)
N(cid:88)
p=1
gyxakpyp +
gxxxp + P
N(cid:88)
N(cid:88)
p=1
configuration. Quite surprisingly, we found that the network does not experience chaotic dynamics.
The simplicity of the system makes it ideal for analytical and numerical investigations. However,
its tight intramodular coupling (leading to a high degree of synchronizations in the nodes' activity)
and its lack of aperiodic behavior make it unrealistic as a model of real worls networks, which are
typically more complex and may spend considerable time in chaotic regimes. To address this, we
considered in Section 3 an extended model of coupled Wilson-Cowan pairs. While this system also
illustrates the tight relationshops between connectivity and dynamics, its behavior is much richer;
one can easily produce desynchonization and/or tune the system to aperiodic behavior by changes
in its parameters.
2 Coupled nonlinear oscillators
We consider the following 2N-dimensional system of nonlinear oscillators (whose architecture is
illustrated in Figure 1 when N = 4):
gxybkpxp +
gyyyp + Q
(1)
p=1
p=1
with 1 ≤ k ≤ N . Each node is driven by external sources (P for the nodes xk in the module X,
and Q for the nodes yk in the module Y ). In addition, each node receives input from all other
nodes that are connected to it through incoming edges, with weights g, indexed as described in
the previous section and in Figure 1. The coefficients akp, bkp ∈ {0, 1} are the binary entries of the
adjacency blocks A and B. The effective input to each node is the sum of all such external and
internal sources, modulated by the sigmoidal:
Sb,θ[Z] =
1
1 + exp(−b[Z − θ])
−
1
1 + exp(bθ)
(2)
with parameters b = bx and θ = θx when the target node is in module X, and b = by and θ = θy
when the target node is in module Y . Throughout our analysis, we fixed: bx = 1.3, by = 2, θx = 4,
θy = 3.7, gxx = 16/N , gyy = 3/N , P = 1.5, Q = 0. We allowed the range [0, 30] for the X-to-Y
and Y -to-X connectivity strengths gxy and gyx. The form of the equations and the parameters are
typical for Wilson-Cowan dynamics.
A comprehensive study of parameter dependence for such a system would be almost intractable
(as it would be for any system attempting to model real world, complex phenomena affected by
a wide collection of factors). Let us notice, for example, that perturbing the individual node dy-
namics (e.g., the logistic function) has its own -- distinct -- effect on the temporal evolution of the
coupled system. The best one can do is to analyze the sensitivity of the system with respect to
one or two parameters of interest at the time, and eventually use this information to quantify and
directly compare the effects of each factor on the system's behavior.
To continue, we will first consider a small network size (N = 2), and inspect the dynamic behavior of
the system for every possible theoretical configuration (adjacency matrix) corresponding to a fixed
pair of edge densities. In Section 2.1, we discuss the cases (Mxy, Myx) = (3, 3) and (Mxy, Myx) =
(2, 3), but a similar analysis can be carried for any density pair. We study how small changes
in the graph (such as adding/deleting an edge, or moving an edge by a sequence of add/delete
5
operations) influence the system's dynamics, and we try to understand in which scenario these
dynamics are most sensitive to weight changes. Furthermore, we are interested in finding whether
structural changes (edge shifting) may have comparable effects with varying the weights, or under
which circumstances this may be true.
Our specific interest remains, however, in studying what happens for higher network sizes N .
That is because natural systems are likely to be formed, even at a macroscopic level, of hundreds or
thousands of node-units. Since the number of configurations increases extremely fast (combinato-
rially squared) with the size N , it is no longer ideal, for high N values, to describe each individual
configuration in this large set; we propose a probabilistic approach to be more appropriate.
In
Section 2.2, we define behavior frequency plots, quantifying the statistical likelihood of the system
(over the distribution of all possible configurations corresponding to a fixed density type) to exhibit
a certain dynamics at a fixed combination of edge weights. While in this paper we only establishe
a proof of principle, by inverstigating small sizes (N = 4, leading to thousands of configurations for
each density type), the methods can be applied to higher network sizes by using increased resources,
or by concentrating the search on more specific aspects.
Through the following sections, we will use the notation DMxy,Myx for the collection of all adjacency
matrices with density type (Mxy, Myx).
2.1 Low dimensional dynamics. Bifurcation diagrams
In the low dimensional case of N = 2 nodes per module, the system (1) becomes:
x1 = −x1 + (1 − x1) · Sbx,τx[gxx(x1 + x2) − gyx(a11y1 + a12y2) + P ]
x2 = −x2 + (1 − x2) · Sbx,τx[gxx(x1 + x2) − gyx(a21y1 + a22y2) + P ]
y1 = −y1 + (1 − y1) · Sby,τy [gxy(b11x1 + b12x2) − gyy(y1 + y2) + Q]
y2 = −y2 + (1 − y2) · Sby,τy [gxy(b21x1 + b22x2) − gyy(y1 + y2) + Q]
(3)
Intuitively, we expect the dynamics to be influenced by the flow/dissipation of the information in
the system, i.e., by the average length of the minimal path that connects any two nodes. While
the density type (Mxy, Myx) strongly influences the dynamics of the system, it clearly does not
completely determine temporal behavior in and off itself, and the dynamics are only partly encoded
in the density type, or in the adjacency spectrum. One common sense expectation is that, for a fixed
density type (Mxy, Myx), two adjacency configurations with the same eigenspectrum can produce
significantly different phase-space dynamics. We verify this conjecture and try to better describe
the correspondence between adjacency and dynamics, but we also propose that other options for
measuring the properties of the graph may capture better the system's dynamic complexity.
For a phase-plane analysis of a nonlinear dynamical system, one typically starts by establishing
the position and stability of equilibria, searching for invariant sets (e.g. cycles, invariant tori, etc)
and for potential aperiodic/chaotic behavior. Since, due to the nonlinearity of the system, describ-
ing these objects precisely is quite challenging, we use numerical algorithms to approximate the
attractors' position and shape, establish their stability and study their change under perturbation
of parameters. Throughout this study, we keep all other system parameters fixed, and only vary
the between-module connection strengths gxy and gyx, and the system's underlying geometry (by
allowing various configurations for the binary matrices A =
and B =
(cid:21)
(cid:20) a11 a12
a21 a22
(cid:20) b11
b21
(cid:21)
).
b12
b22
This choice is motivated by our aim to understand and compare the different effects on dynamics
of three distinct ways of altering inter-modular connectivit: (1) by changing the edge density type
(Mxy, Myx), (2) by changing the node-node edge configuration (the positions of the 1 entries in the
binary matrices A and B) and (3) by changing the inter-modular edge weights (gxy and gyx).
6
In order to understand, for each individual adjacency configuration, the changes in dynamics
produced by varying gxy and gyx, we first use bifurcation diagrams in the (gxy, gyx) parameter plane.
Then, we observe how these diagrams change when perturbing the underlying adjacency graph. To
generate the bifurcation diagrams, we used the continuation algorithms provided by the Matcont
package [9], initialized in a region containing values of gxy and gyx corresponding to Hopf and saddle
node bifurcations in the classic Wilson-Cowan system. We investigated the Hopf and limit point
(saddle point) curves in our own coupled system, delimiting behaviors such as convergence to a
unique stable equilibrium versus oscillations towards a stable limit cycle (including bistability).
To illustrate our ideas, the two tables in the Appendix show all possible (gxy, gyx) parameter
planes that can be obtained for N = 2 and density types (Mxy, Myx) = (3, 3) and (Mxy, Myx) =
Interestingly, all 16 combinatorial configurations in D3,3 produce only four
(2, 3), respectively.
distinct dynamic parameter planes, which we will call the dynamic classes for (Mxy, Myx) = (3, 3),a
dn which we show in Table 1. Notice that all 4 dynamic classes can be obtained by fixing A to
any configuration and considering all 4 cases for Bs, but also by fixing B and considering all
4 configurations for A. Similarly, all 24 combinatorial configurations in D2,3 produce only six
dynamic classes, distinct than the ones in D3,3, shown in Table 2, all obtainable by fixing B and
varying A.
The presence of bifurcations for all dynamic classes implies that, when fixing the network,
changing one of the weights gxy or gyx can push the system over a bifurcation curve, placing it in
a different regime. This change may consist for example of switching between "rest" (convergence
to a stable equilibrium) and "oscillations" (convergence to a limit cycle) when crossing a Hopf
bifurcation, or of sharply switching attractors (when crossing a limit point curve).
Aternatively, looking across all dynamic classes for each (Mxy, Myx), one may easily note that
the changes in dynamics triggered by changes in configuration are rather localized to certain regions
in the parameter plane. That is, the behavior of the system might be, between classes, very different
or very similar at different points in the (gxy, gyx) plane. This suggests that the system's sensitivity
to the network geometry depends on the actual connection weights. There appears to be a critical
(gxy, gyx) locus were the system is most sensitive to geometry: deleting or shifting one edge can
push the system from a stable equilibrium (in one panel) to oscillations (in a different panel). Away
from this region, there is a more topographic correspondence between parameter planes (i.e., the
dynamic classes have qualitatively more consistent, or even identical behavior between panels).
We say that two configurations are in the same adjacency class if they have the same eigenspectra.
When investigating the relationship between the adjacency configuration and the dynamic behavior
of the system, a natural question to ask is whether dynamic classes may be predicted simply by
looking at the adjacency spectrum. We conjecture that the correspondence dynamics → adjacency
classes is well defined, but clearly not bijective. That is: a specific dynamic class can't be obtained
from two different configurations, but a single adjacency class may lead to different dynamics.
While in general, for high dimensions, proving this relationship may be quite difficult, for low
dimensions it is easy to illustrate. For example, Table 1 shows each configuration in D3,3 together
with its adjacency and dynamic class. In this case, there are three distinct adjacency eigenspectra
(designated by letters A through C), each class containing respectively 8, 4 and 4 of the total of
(2N )2 = 16 configurations. In counterpart, there are four distinct dynamics classes (designated by
indices i through iv). With this convention, the table shows that no dynamics can be obtained
from multiple adjacency classes, but that some adjacency classes can lead to multiple dynamics.
Similarly, Table 2 shows how the 6 dynamic classes are mapped to the 4 adjacency classes in the
case of D2,3.
This suggests that, while the adjacency class, together with the density type, clearly have a
contribution to dynamics, they cannot be directly used either to predict these dynamics. In our
current work, we are investigating whether other descriptions of the adjacency matrix are better
7
choices to help predict the dynamics or the complexity of a network's evolution. Node degree dis-
tribution, connectivity coefficient, number of particular motifs may be finer network measures than
edge density, or adjacency spectrum, and therefore more efficient in classifying dynamic complexity.
Figure 2: Bifurcation diagrams versus search algorithm. A. Bifurcation diagram in the (gxy, gyx)
plane for the dynamic class (ii) in D2,3, created with the Matcont extension algorithms. Hopf curves are
shown in blue, limit point curves in green, and codimension two points are shown as stars: green (cusp points),
red (generalized Hopf points), purple (Bogdanov-Takens points). B. Dynamic regimes in the (gxy, gyx) plane
for the same dynamic class, obtained using our numerical search for different behaviors within the system:
the locus corresponding to a unique stable equilibrium is in black, the locus for multiple stable equilibria is in
red, the locus for a unique stable cycle is in orange, and the locus where the stable equilibrium and the stable
cycle coexist is in white.
It is becoming clear that, even for small N , the system has many dynamic possibilities (depending on
configuration), thus making undesirable an individual descriptive approach to each configuration-
specific parameter space. A statistical approach seems more appropriate, bearing in mind that
some dynamic classes may be more substantial than others, and thus have a stronger contribution
to driving these statistics. While these are ideas that we elaborate more in the following sections,
here we set the grounds for this path by describing the numerical methods used and by illustrating
how they work on a simple N = 2 example.
For the each (gxy, gyx) parameter point we took a sample of adjacency matrices with a given density.
For each adjacency matrix in the sample we ran simulations and analyzed each one in order to find
the range of dynamic behaviors it could produce when starting from different initial conditions.
Our search algorithms could detect six types of behaviors: 1) a single fixed point, 2) multiple
fixed points, 3) periodic oscillations, 4) non periodic oscillations, 5) both a single fixed point and
periodic oscillations, and 6) both multiple fixed points and periodic oscillations. We only analyzed
the second half of the simulations in order to remove the transient part of the activity.
For each pair of connection weights and for each adjacency matrix we explored the space of
8
initial conditions using a basic Particle Swarm Optimization (PSO) algorithm. The utility function
used by the PSO algorithm depend on which behaviors had already been found. If only a single
fixed point had been found then the initial condition with the largest utility was the one with
largest amplitude in its oscillations. This amplitude was determined as the difference between the
mean value of each node's response and its largest value in the last quarter of the simulation. The
largest amplitude among all nodes was selected. If no fixed points had been found, then the utility
function was set to one minus the utility of the previous case.
Detecting fixed points was done using the same amplitude that constituted the utility function
for the PSO algorithm. When this amplitude was below a threshold, a fixed point was detected.
To detect multiple fixed points, for each initial condition where a fixed point was found the average
value of the response for the first node was stored.
If the difference between the largest stored
average value and the smallest one was above a threshold, multiple fixed points were detected.
Detecting periodic oscillations was done using a basic algorithm that convolved the time-
discretized response of a node with with a time inverted version of itself. Intuitively, this response
could be conceived as a vector, and the convolution as an inner product between a part of this vec-
tor and a shifted version of itself. The reason why this algorithm works is that when the sections
of the vectors participating in the inner product are normalized, the inner product will attain its
maximum value when the response vector and its time shifted version are the same. This happens
when the response signal is periodic.
A non periodic oscillation was detected when the response was not a fixed point, but our
algorithm could not detect periodic behavior. Whenever aperiodic behavior was detected, the sim-
ulation was extended for a longer period of time and then analyzed again in order to prevent false
detections due to transient properties of the response.
and B =
parameter plane for the configuration A =
We first illustrate the efficiency of this search algorithm by applying it to find all behaviors in the
in D2,3, which was found by
Matcont to be of class (ii). Figure 2 compares side by side the diagrams obtained in this particular
case: via the Matcont software on the left, and via our search algorithm on the right.
We then used the search algorithm by itself, to illustrate the likelihood for each behavior at
each parameter point (gxy, gyx), over all configurations in D3,3. We will call the parameter loci for
different behaviors -- p-bifurcations of the system. Each panel in Figure 3 illustrates the likelihood
for an arbitrary configuration in D3,3 to exhibit one of the following attracting sets: a globally
stable equilibrium (Figure 3a), multiple stable equilibria (Figure 3b), a globally stable limit cycle
(Figure 3c), or a coexisting stable equilibrium and stable cycle (Figure 3d). Our search algorithm
found only artifactual aperiodic behavior, which, upon inspection, was clearly due to a slower initial
transient phase of the solution, mistakenly labeled by our code as aperiodic behavior.
In higher dimensions, one may expect the system's attractors to transcend simple limit cycles
(for example, a paper by Borisyuk et al [3] found a similar four-dimensional, coupled system to
additionally exhibit symmetric, antisymmetric and nonsymmetric invariant tori), which are hard
(and computationally rather expensive) to track down. One of our goals is to investigate the
presence of aperiodic behavior in our system for higher dimensions, which we do in the next section
for the case N = 4.
2.2 P-bifurcations
In this section, we focus on constructing and understanding behavior frequency plots. Each point
(gxy, gyx) in the parameter plane may correspond or not, for each adjacency configuration in
DMxy,Myx , to a specific dynamic behavior.
In other words the point will be on one side ver-
sus the other of some bifurcation curve in (gxy, gyx), with a specific probability (over the whole
9
(cid:21)
(cid:20) 1 0
1 0
(cid:21)
(cid:20) 0 1
1 1
Figure 3: Behavior frequency plots, showing the number (out of all 16 configurations in D3,3) which
exhibit: A. one stable equilibrium; B. multiple stable equilibria; C. one stable cycle; D. a coexisting stable
cycle and equilibrium.
configuration distribution DMxy,Myx ). This represents in a sense a probabilistic extension of the
concept of bifurcation,
As shown before, one can define the p-bifurcation diagram of a system for any particular dynamic
behavior. Fix the size N and the density type (Mxy, Myx). For each pair of edge weights (gxy, gyx),
we can calculate (or estimate numerically) the fraction P(gxy, gyx) of adjacencies in DMxy,Myx
which, for weights (gxy, gyx), lead to a specific dynamic behavior. E.g., by estimating the fraction
of configurations which lead to coexistence of a stable equilibrium and a stable cycle, one can
establish the locus in the parameter plane where there exist configurations with equilibrium/cycle
bistability, evaluate how likely it is to randomly pick a configuration with such bistable behavior,
and observe what is needed to push the system from a regime of likely bistability into a purely
oscillatory or quiet regime.
For Mxy = Myx = N 2, there is only one possible configuration, and the behavior loci are
delimited by regular bifurcation curves. When DMxy,Myx (cid:54)= 1, the transition is smooth, so that
there is a region where 0 < P < 1, which corresponds to a "smeared" bifurcation curve.
For example: in Figure 3 we show, for N = 2, the four nontrivial behavior frequency plots for
D3,3. These look as one would expect from "overlapping" the four dynamic classes in D3,3 (shown
in Table 1). Due to the similarities and differences between the Hopf and limit point bifurcation
curves across configurations, the resulting frequency plots are a "smeared" version of the diagrams
for individual classes. We conjecture that the profile of a frequency plot, as well as the degree of
smearing (i.e., the width of the region with values transitioning between P = 0 and P = 1) depends
on the pair (Mxy, Myx). Since there are such few different behaviors, one can still distinguish the
contours of the individual bifurcation diagrams in the p-diagrams, which is no longer the case
for higher N (see Figures 4 and 5). For larger DMxy,Myx, there are more configurations, and
more dynamic behaviors/classes. Since the number of configurations in DMxy,Myx increases with N
extremely fast, when studying the same phenomena for larger values of N , it is more convenient to
investigate the behavior distribution based on a sample probability.
Figures 4 and 5 illustrate the behaviors we found numerically in a variety of systems with
N = 4. It is interesting to notice that we did not find aperiodic oscillations, or multiple stability.
Each row shows, for one density type (Mxy, Myx), the frequency plots for the remaining four
behaviors. Figures 4 illustrates how these loci change when the two densities Mxy and Myx are
identical, but increase from very low (Mxy = Myx = 4), to medium (Mxy = Myx = 8), to high
(Mxy = Myx = 12). Figures 5 illustrates two cases of uneven densities, one with Mxy > Myx and
the other with Myx > Mxy.
Broadly, one can notice that in some regions in the parameter plane the weights are a strong de-
10
Figure 4: Behavior frequency plots for N = 4, for equal densities Mxy and Myx. We show the
fraction of all configurations in (i) D4,4; (ii) D8,8; (iii) D12,12 which exhibit: A. one stable equilibrium;
B. one stable cycle; C. multiple stable equilibria; C. a coexisting stable cycle and equilibrium. No other
behaviors were found. The simulations are based on a randomly generated sample of S = 200 configurations
in the respective DMxy,Myx (hence the color bar represents numbers from 0 to 200).
terminant of the potential dynamics, while in other regions only a very large jump in the (gxy, gyx)
parameter plane would significantly infuence the likely dynamics. The same applies to the sensi-
tivity to weight changes: some regions are consistent between corresponding panels, showing that
a switching from one density type to another would have almost no effect, while other regions are
very sensitive to density and to configuration changes.
It is also interesting to notice that higer densities create sharper transitions, which is hardly
surprising:
if there are more edges, a small global change in the weights is more likely to have
a substantial effect on dynamics. It follows that, for lower densities, higher weights are required
to place the system in an oscillatory (stable cycle) rather than quiet (stable equilibrium) regime.
Moreover, the smoother spread of the plots for lower densities means that the dynamics is more
susceptible to perturbations in configuration, even when the low densitites are fixed.
Remark 1. Statistically speaking, the approach is appropriate when comparing behaviors within
11
Figure 5: Behavior frequency plots for N = 4, for non-equal densities Mxy and Myx. We show
the fraction of all configurations in (i) D6,10 and (ii) D10,6 which exhibit: A. one stable equilibrium; B. one
stable cycle; C. multiple stable equilibria; D. a coexisting stable cycle and equilibrium. No other behaviors
were found. The simulations are based on a sample of S = 200 configurations.
one single DMxy,Myx, where each pair (gxy, gyx) has the same number of corresponding configura-
tions . When comparing behaviors between distributions DMxy,Myx for different values of Mxy and
Myx, we tried to be more careful, and verified the validity of our sample-based method by comput-
ing the standard deviations over the chosen samples, to ensure that the results are not biased by
using the same sample size for different size distributions.
Remark 2. In the case of this simple system, the full-connectedness of the moduli maintains the
moduli synchronized, so that, looking at the time evolution of one node, one can visualize with good
approximation the temporal behavior of the whole module. This presents the advantage of behavior
simple enough to be easily tractable even in higher dimensions. For example, contrary to what one
might have expected, our numerical searches did not find any parameter set for which the system
exhibits aperiodic behavior. However, this is not a situation expected to occur biologically, and we
use it only as a starting point. In the following section, we present an extension of this model which
is a better candidate for biophisical and connectivity modeling in the brain, and which exhibits
richer and more plausible behavior. We use the same techniques to investigate this extension.
3 Coupled Wilson-Cowan pairs
We describe a more realistic scenario, where inhibition is implemented through a separate collection
of nodes. This may be an appropriate representation of a brain network in which inhibition is
performed via a hidden layer of neurons, different than the target cells that ultimately need to be
inhibited. For example, the prefrontal cortex (PFC) projects excitatory fibers on the inter-neurons
12
in ITC (an amygdala nucleus), which in turn inhibit the cells in the basal amygdala, the functional
area considered to be responsible for emotion regulation. Hence the overall effect of the PFC
on arousal reactions controlled by the amygdala produces "fear extinction" (closing the negative
feedback loop that regulates arousal).
To represent this situation, we consider the following model (see also Figure 6):
Figure 6: A schematic representation of the coupled Wilson-Cowan system for N = 4 pairs
of nodes in each module, XY and respectively U V . Each (+)/(-) pair is coupled according to the original
Wilson-Cowan model. In addition, each module has full (-) to (+) connectivity (i.e., each inhibitory unit is
connected with all excitatory units within its module). A fraction Mxu of the (+) units in module XY are
connected with (+) units in module U V , and a fraction Muy of (+) units in module U V are connected with
(-) units in module XY .
(cid:105)
(cid:105)
(cid:88)
giexyp + I
guyAkpup
(cid:105)
gxuBkpxp
gieuvp +
(cid:104)
gxxk − giyyk −(cid:88)
(cid:104)−gyyk + gexxk +
(cid:88)
(cid:104)
guuk − givvk −(cid:88)
= −xk + (1 − xk) · Sbe,θe
= −yk + (1 − yk) · Sbi,θi
= −uk + (1 − uk) · Sbe,θe
= −vk + (1 − vk) · Sbi,θi [−gvvk + geuuk]
τe
τi
τe
dxk
dt
dyk
dt
duk
dt
dvk
dt
τi
(4)
where Sb,θ(Σ) = (1 + exp[−b(Σ − θ)])−1, and we fixed the following Wilson-Cowan parameters:
be = 1.3, bi = 2, θe = 4, θi = 3.7, I = 1.5. Connectivity parameters: gx = gu = 16, gy = gv = 3,
gex = geu = 15, giy = giv = 12, giex = gieu = 5/N , gxu = guy = 10/N .
While the modules retain a form of full (-) to (+) connectivity, the dynamics of this system
is much more complex than that of the original model of simple coupled oscillators. Here, the
network is spending most of its time in complex oscillatory regimes, in which the nodes are no
longer synchronized within each module. We want to investigate whether this system exhibits
aperiodic behavior, and what types of changes in the network configuration can throw the system
from periodic oscillations into chaotic behavior.
13
Figure 7: Transitions in dynamics when altering connectivity and configuration. All panels show
solutions for the Wilson-Cowan coupled pairs, for N = 4 pairs of nodes in each module: the evolution of
the nodes xk and yk is shown on top in blue and red, respectively, and the nodes uk and vk are shown
on the bottom in green and purple. The simulations were performed for the parameters given in the text,
and each for an arbitrary single configuration A. of density type (Mxy, Myx) = (8, 8); B. of density type
(Mxy, Myx) = (14, 8); C. of density type (Mxy, Myx) = (16, 8); D. of density type (Mxy, Myx) = (8, 8), for a
different adjacency configuration than that used in A; E. of density type (Mxy, Myx) = (8, 8), for a different
configuration than those in A and D.
Figure 7 shows, on the left, how the oscillatory regime can be affected by changes in density
type. While for (Mxy, Myx) = (8, 8) the nodes typically perform aperiodic oscillations, increasing
Mxy gradually introduces more structure (for Mxy = 14) and renders them purely periodic (for
Mxy = 16). On the right, the figure illustrate how, for the same density pair, the oscillations can
be tuned (singular "spikes" versus periodic "bursts," versus sustained aperiodic oscillations) by
altering only the configuration.
As in the case for the simple coupled oscillators model, we aim to understand better the types
of behaviors accessible to the system, and how changes in weights, densities or configuration may
be used to swap between these behaviors. In Figure 8, we show the frequency plots in D8,8 and
D14,8.
14
Figure 8: Behavior frequency plots for the system of coupled Wilson-Cowan oscillators for
N = 4, qnd density type D8,8 (top) and D14,8 (bottom). The illustrations are based on samples of size S = 50.
Each panel shows the number (out of the total of 50) of configurations leading to one of the following behaviors:
A. globally stable equilibrium; B. multiple equilibria; C. globally stable cycle; D. aperiodic oscillations; E.
coexistance of a stable equilibrium with a stable cycle; F. coexistence of multiple stable equilibria and cycles.
One imediately notices, in all both cases, the increased complexity of this system's dynamics
compared with the simple model: all six behaviors appear in each of the density types, with large
parameter loci allowing periodic and aperiodic oscillations. As in the previous system, however,
richer behavior seems to correspond to lower densities (the higher the two densitites the more likely
it is for the system to fall into simple periodic oscillations, as already suspected from Figure 7).
In Figure 9, we illustrate one way of tracking changes in the system's dynamics when fixing the
weights and density type and only changing the configuration by adding/deleting edges. The figure
shows the evolution of the system's approximate entropy (estimated from the system's solutions
according to an algorithm proposed by Pincus [14]) along two network "paths" from one initial state
(of relatively low entropy) to a final state (with higher entropy). More precisely, we considered an
initial state in which only one unit in module X is cross-connected to all units in module Y (i.e.,
the block matrix A has ones on the first row, and the block matrix B has ones on the first column),
and a final state in which the units are connected bijectively (both A and B are the identity).
We constructed two paths in the adjacency graph from the initial to the final states, by defining
each step to be a 0/1 flip (a 1 swaps with a 0 at a neighboring position in the adjacency matrix,
corresponding to an edge deletion and then addition in a proximal position). We want to suggest
that there are many ways in which a system can evolve from low to high entropy through a sequence
of slight edge perturbations (without the network even changing density type). We suspect that
this is possible even if we additionally require the paths to be of monotonely increasing entropy.
The states along these paths can be seen as states that the system will have to take provided it
evolves along the respective path.
15
Figure 9: Evolution of the system's approximate entropy from h0 = 0.0581 to h1 = 0.0701 along
two distinct paths in the set of adjacency graph configurations.
4 Discussion
4.1 Strengthening versus restructuring
In our paper, we focused on understanding a few aspects of how dynamic behavior in a network
depends on its underlying adjacency matrix. To do this, we used an underlying graph with simple
bimodular architecture, with each of the interconnected modules fully connected. We discovered
that different temporal effects are to be expected when perturbing different aspects of the network
connectivity. We compared the effects of globally increasing the weights between the two intercon-
nected modules versus increasing the number of edges between the modules. While both actions
lead to "increasing connectivity" between the two modules, they produced qualitatively different
effects on dynamics.
We noticed that, while certain regimes are robust to perturbations (local changes in weights or
in adjacency don't produce qualitative effects on dynamics), other parameter regions tend to be
very sensitive to such changes. Furthermore, when in sensitive regimes, small local perturbations in
the network wiring (e.g., locally modifying the adjacency matrix by adding or deleting edges) may
have dramatic effects on the system's dynamics, more substantial than those obtained by a global
change in the system's weights (recall that our weight parameters gxy, gyx affect all the connections
from one module to another).
Perhaps the most important question here is how the three hardware components (edge density,
position and strength) act differently on the temporal behavior of the system, and how they work
together to tune the network's dynamic complexity. This question is extremely important in the
context of understanding a variety of real world networks.
For example, one could think of what type of adjustments should be performed by the system
in order to shift its dynamics most efficiently from a quiet to an oscillatory regime or vice-versa. If
the state of the network is, to begin with, in a region sensitive to weight changes in the (gxy, gyx)
parameter plane, the system may perform the "phase transition" via a small change in the weights.
Otherwise, if operating away from such regions, only a large, global change in the overall values
of the weights can significantly increase the probability of the system to switch regimes. On the
other hand, a small change in the graph structure could produce instead the desired dynamic
change, pushing the system over into a more complex, or more stiff range of functioning. To help
16
us illustrate this phenomenon, in Figure 10 we show the frequency plots for fixed weights (gxy, gyx),
with respect to the two densities Mxy and Myx, veiwed as system parameters.
Consider for instance the situation in Figure 4II(b), where it is clear that, for Mxy = Myx = 8,
only a rather large change in weights would increase the potential for the quiet system gxy = gyx =
15 to oscillate. Figure 10III (corresponding to gxy = gyx = 15) shows, however, that the density
point (Mxy, Myx) = (8, 8) is in a sensitive region, where changes in the densities can strongly affect
its likelihood to change behaviors, and changes in configuration can also be used to switch between
available behaviors. For another example consider the point gxy = 5, gyx = 10, for which the
system with Mxy = 10 and Myx = 6 (shown in Figure 5II) performs oscillations with extremely
high likelihood (for almost all configurations). Figure 10I (for gxy = 5, gyx = 10) show that the
point Mxy = 10 and Myx = 6 is in a range where the dynamics is much more sensitive to changes
in density and configuration.
Figure 10: Behavior frequency plots in the parameter plane of densities (Mxy, Myx). I. for
gxy = 5, gyx = 10; II. for gxy = 5, gyx = 60; III. for gxy = 15, gyx = 15.
Finally, recall that each density type generates a large number of distinct configurations. The
dynamics of the system may experience a whole collection of dynamic modes (some qualitatively dis-
tinct, some equivalent) over the whole distribution of possible configurations. While configuration-
triggered changes in dynamics are important (and are in fact more likely in intermediate density
type regimes), the dynamics seem, however, more robust to perturbations in configuration than to
those involving a change in density type. This robustness, previously noticed in [17], may be partly
17
explained by the robustness of the adjacency spectrum when exploring all configurations of fixed
density type [19]. However, in Section 2.2, we have shown that adjacency spectrum classes are not
in bijection with dynamics classes. Part of our current work is aimed towards understanding the
theoretical bases of this robustness.
4.2 Applications to learning and the brain connectome
As discussed in a previous paper [19], these choices of the mechanisms used to trigger changes in
dynamics are extremely important for networked systems like the brain, in order to maintain their
adequate function of performing complex simultaneous tasks. There are many different models
describing the synaptic restructuring that occurs in a network of neurons during processes like
learning, or memory formation, most likely involving a combination of weight changes of existing
synapses, and creating/deleting connections. In terms of our model, this means that not only the
edge weights, but also the edge distribution is likely to exhibit both short and long-term changes
during learning. Knowledge of the geometry of the network is therefore very important when
determining which connectivity schemes are plausible to use for models of learning.
A lot of effort has been invested recently towards developing and using graph-theoretical network
measures in conjunction with statistical methods, in order to identify the effects of abnormal connec-
tivity patterns (measured as structural connectivity, for anatomical links; functional connectivity,
for undirected statistical dependencies; and effective connectivity, for directed causal relationships
among distributed responses [13]) on the efficiency of brain function. By applying graph theoreti-
cal measures of segregation (e.g., clustering coefficient, motifs, modularity, rich clubs), integration
(e.g., distance, path length, efficiency) and influence (e.g., node degree, centrality) these studies
have been investigating the sensitivity of systems to removing/adding nodes or edges at different
locations in the underlying network.
Working with empirical data, such measures have been used to understand behavioral impair-
ments in subjects with compromised connectivity due to existing lesions [8], or group differences
between healthy controls and patients with mental illnesses associated with deficient feedback cir-
cuitry. In our previous work with fMRI data [17], we ourselves used a simple graph-theoretical
model as a formal framework to study how network density can affect the complexity of signal
outputs, measured by the log-log slope of their power spectra (power spectrum scale invariance,
PSSI). Indeed, for sufficiently large networks, the log-log spectra were close to linear within certain
frequency bands, and the PSSI slopes were found to vary as a function of both input type (excita-
tory, inhibitory) and input density (mean number of long-range connections), with comparatively
insignificant dependence on the node-specific geometric distribution.
Without attempting to understand the source of either dependence on density or robustness to
specific configuration, we focused on the possible interpretations and applications. We suggested
a testable framework for interpreting the empirical data in conjunction with the model, to deliver
a connectivity-based hypothesis for the difference in functional regimes corresponding to different
levels of anxiety. Individuals with average emotional reactivity had experimental PSSI values in
the pink noise range for amygdala and prefrontal regions, corresponding to well-regulated control
systems, with well balanced excitatory and inhibitory projections. Individuals at the anxious end of
the spectrum, showed experimentally white noise primarily for the amygdala, and were predicted by
our model to have relatively weaker inhibitory inputs from the prefrontal cortex (producing weaker
feedback). Individuals at the stress resilient end of the spectrum, showed white noise primarily
for the prefrontal cortex, and were predicted by our model to have relatively stronger excitatory
inputs from the amygdala (producing stronger feedback). This last simulation result may seem
surprising, but in fact produces a reasonable hypothesis: enhanced projections from the amygdala
to prefrontal cortex effectively lower the threshold for inhibitory feedback, thereby suppressing
all but the strongest stimuli. Broadly speaking, we saw as very promising the fact that such a
18
simple and general setup may yet inform successfully our human imaging results in a circuit as
important as the one regulating human emotion. That is because its simplicity allows us to study
and understand (analytically or numerically) the sources that drive different aspects of the system's
behavior (thus producing the different regimes of function); its generality opens such the model
(with minor modifications) to possible applications other than emotion regulation.
The results in this paper (which used an identical network structure in its analysis) explain some
of the more important (although perhaps counterintuitive) features observed computationally in
Radulescu et al [17]. Among these are the robustness of the coupled dynamics to certain changes in
the network architecture and its vulnerability to others, as well as the differences between updating
connection strengths versus perturbing connection density or geometry.
In developing future iterations of this model with possible applications to learning mechanisms,
it will also be important to explore how the learning process itself shapes the connectivity scheme,
with possible emerging structures in which modularity is purposefully broken into hub-like sub-
networks [20]. Understanding the source and limits of a network's robustness and vulnerability to
perturbations may be an instrument that could help us investigate in the future many aspects of
brain circuitry: from determining which architectures favor convergence under particular learning
algorithms, and which not, to classifying cognitive deficits and psychiatric illnesses.
References
[1] Faisal Naqib, Thomas Quail, Louai Musa, Horia Vulpe, Jay Nadeau, Jinzhi Lei, and Leon
Glass. Tunable oscillations and chaotic dynamics in systems with localized synthesis. Phys-
ical Review E, 85(4):046210, 2012.
[2] Stefano Boccaletti, Vito Latora, Yamir Moreno, Martin Chavez, and D-U Hwang. Complex
networks: Structure and dynamics. Physics reports, 424(4):175 -- 308, 2006.
[3] Galina N Borisyuk, Roman M Borisyuk, Alexander I Khibnik, and Dirk Roose. Dynamics
and bifurcations of two coupled neural oscillators with different connection types. Bulletin
of Mathematical Biology, 57(6):809 -- 840, 1995.
[4] Nicolas Brunel. Dynamics of sparsely connected networks of excitatory and inhibitory
spiking neurons. Journal of computational neuroscience, 8(3):183 -- 208, 2000.
[5] Ed Bullmore and Olaf Sporns. Complex brain networks: graph theoretical analysis of
structural and functional systems. Nature Reviews Neuroscience, 10(3):186 -- 198, 2009.
[6] Shannon Campbell and DeLiang Wang. Synchronization and desynchronization in a net-
work of locally coupled wilson-cowan oscillators. Neural Networks, IEEE Transactions on,
7(3):541 -- 554, 1996.
[7] P Collet and Eckmann J-P. Iterated maps on the interval as dynamical systems. Progess in
Physics, Birkhauser, Boston, 1980.
[8] Maurizio Corbetta. Functional connectivity and neurological recovery. Developmental psy-
chobiology, 54(3):239 -- 253, 2012.
[9] Kuznetsov Y Dhooge A, Govaerts W. Matcont: a matlab package for numerical bifurcation
analysis of odes. ACM Transactions on Mathematical Software.
[10] Adrien Douady. Topological entropy of unimodal maps. In Real and complex dynamical
systems, pages 65 -- 87. Springer, 1995.
19
[11] Richard T Gray and Peter A Robinson. Stability and structural constraints of random
brain networks with excitatory and inhibitory neural populations. Journal of computational
neuroscience, 27(1):81 -- 101, 2009.
[12] C Hauptmann, H Touchette, and MC Mackey. Information capacity and pattern formation
in a tent map network featuring statistical periodicity. Physical Review E, 67(2):026217,
2003.
[13] Hae-Jeong Park and Karl Friston. Structural and functional brain networks: from connec-
tions to cognition. Science, 342(6158):1238411, 2013.
[14] Steven M Pincus. Approximate entropy as a measure of system complexity. Proceedings of
the National Academy of Sciences, 88(6):2297 -- 2301, 1991.
[15] Anca Radulescu. The connected isentropes conjecture in a space of quartic polynomials.
Discrete and Continuous Dynamical Systems, Series B, 19(1):139 -- 175, 2007.
[16] Anca Radulescu. Computing topological entropy in a space of quartic polynomials. Journal
of Statistical Physics, 130(2):373 -- 385, 2008.
[17] Anca Rdulescu and Lilianne R Mujica-Parodi. Network connectivity modulates power spec-
trum scale invariance. NeuroImage, 90:436 -- 448, 2013.
[18] Mikail Rubinov and Olaf Sporns. Complex network measures of brain connectivity: uses
and interpretations. Neuroimage, 52(3):1059 -- 1069, 2010.
[19] Anca Radulescu. Neural network function, density or geometry? Preprint arXiv:1304.5232.
[20] Benoit Siri, Mathias Quoy, Bruno Delord, Bruno Cessac, and Hugues Berry. Effects of
hebbian learning on the dynamics and structure of random networks with inhibitory and
excitatory neurons. Journal of Physiology-Paris, 101(1):136 -- 148, 2007.
[21] Olaf Sporns. The human connectome: a complex network. Annals of the New York Academy
of Sciences, 1224(1):109 -- 125, 2011.
[22] Olaf Sporns and Rolf Kotter. Motifs in brain networks. PLoS biology, 2(11):e369, 2004.
[23] Olaf Sporns and Giulio Tononi. Classes of network connectivity and dynamics. Complexity,
7(1):28 -- 38, 2001.
[24] Arthur W Toga, Kristi A Clark, Paul M Thompson, David W Shattuck, and John Darrell
Van Horn. Mapping the human connectome. Neurosurgery, 71(1):1, 2012.
[25] Hugh R Wilson and Jack D Cowan. Excitatory and inhibitory interactions in localized
populations of model neurons. Biophysical journal, 12(1):1 -- 24, 1972.
20
Appendix: Adjacency versus dyamics classes for N = 2
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
0
1
1
1
1
0
1
1
1
1
0
0
1
1
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
1
0
1
0
1
1
1
1
0
1
0
1
1
1
(Aiii)
(Bii)
(Biv)
(Ci)
(Bii)
(Aiii)
(Ci)
(Biv)
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
1
1
0
1
0
1
1
1
1
0
1
0
1
1
1
0
1
1
1
1
1
1
0
0
1
1
1
0
1
1
1
0
1
1
1
1
0
1
1
1
1
0
1
0
1
1
1
(Biv)
(Ci)
(Aiii)
(Bii)
(Ci)
(Biv)
(Bii)
(Aiii)
Table 1: Adjacency and dynamics classes for N=2, density type (Mxy,Myx)=(3,3). Adjacency
classes are designated by letters (A − D) and dynamics classes by subscripts (i − iv). The four possible
parameter planes are shown on the right, with Hopf curves in blue, limit point curves in green and codimension
two bifurcations marked with stars: cusp (green), Bautin (red) and Bogdanov-Takens (purple).
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
(Av)
(Ai)
(Civ)
(Diii)
(Bii)
(Bvi)
1
0
1
0
1
1
0
0
1
0
0
1
0
1
1
0
0
0
1
1
0
1
0
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
(Bvi)
(Ai)
(Diii)
(Civ)
(Bii)
(Av)
1
0
1
0
1
1
0
0
1
0
0
1
0
1
1
0
0
0
1
1
0
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
(Av)
(Bii)
(Diii)
(Civ)
(Ai)
(Bvi)
1
0
1
0
1
1
0
0
1
0
0
1
0
1
1
0
0
0
1
1
0
1
0
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
0
1
1
1
(Bvi)
(Bii)
(Civ)
(Diii)
(Ai)
(Av)
1
0
1
0
1
1
0
0
1
0
0
1
0
1
1
0
0
0
1
1
0
1
0
1
Table 2: Adjacency and dynamics classes for N=2, density type (Mxy,Myx)=(2,3).
|
1606.04344 | 1 | 1606 | 2016-06-14T13:29:08 | Estimating direction in brain-behavior interactions: Proactive and reactive brain states in driving | [
"q-bio.NC"
] | Conventional neuroimaging analyses have revealed the computational specificity of localized brain regions, exploiting the power of the subtraction technique in fMRI and event-related potential analyses in EEG. Moving beyond this convention, many researchers have begun exploring network-based neurodynamics and coordination between brain regions as a function of behavioral parameters or environmental statistics; however, most approaches average evoked activity across the experimental session to study task-dependent networks. Here, we examined on-going oscillatory activity and use a methodology to estimate directionality in brain-behavior interactions. After source reconstruction, activity within specific frequency bands in a priori regions of interest was linked to continuous behavioral measurements, and we used a predictive filtering scheme to estimate the asymmetry between brain-to-behavior and behavior-to-brain prediction. We applied this approach to a simulated driving task and examine directed relationships between brain activity and continuous driving behavior (steering or heading error). Our results indicated that two neuro-behavioral states emerge in this naturalistic environment: a Proactive brain state that actively plans the response to the sensory information, and a Reactive brain state that processes incoming information and reacts to environmental statistics. | q-bio.NC | q-bio | Neuro-behavioral state modulations in driving 1
Estimating direction in brain-behavior interactions: Proactive and reactive brain states in
driving
Javier O. Garcia1, Justin Brooks1, Scott Kerick1, Tony Johnson5, Tim Mullen2, & Jean M. Vettel1,3,4
1U.S. Army Research Laboratory, Aberdeen Proving Ground, MD
2Qusp Labs., San Diego, CA
3University of California, Santa Barbara, CA
4University of Pennsylvania, Philadelphia, PA
5DCS Corporation, Alexandria, VA
Highlights
• Traditional neuroscience studies investigate localized task-evoked responses
• Our approach examines continuous tracking of brain-behavior interactions in oscillatory activity
• Brain leads behavior in a Proactive state, while brain follows behavior in a Reactive state
• Reactive states are largely carried by alpha activity in regions sensitive to environmental
• Proactive states rely more on a diffuse delta-beta network, particularly when linked with
statistics
steering behavior
Keywords: EEG, driving, neuro-behavioral analysis, source analysis,
Running Head: Neuro-behavioral state modulations in driving
*Corresponding Author:
Javier O. Garcia
Email: [email protected]
RDRL-HRS-C (ARL/HRED/TNB)
459 Mulberry Point Road
Aberdeen Proving Ground, MD 21005
Neuro-behavioral state modulations in driving 2
Abstract
Conventional neuroimaging analyses have revealed the computational specificity of localized
brain regions, exploiting the power of the subtraction technique in fMRI and event-related potential
analyses in EEG. Moving beyond this convention, many researchers have begun exploring
network-based neurodynamics and coordination between brain regions as a function of
behavioral parameters or environmental statistics; however, most approaches average evoked
activity across the experimental session to study task-dependent networks. Here, we examined
on-going oscillatory activity and use a methodology to estimate directionality in brain-behavior
interactions. After source reconstruction, activity within specific frequency bands in a priori regions
of interest was linked to continuous behavioral measurements, and we used a predictive filtering
scheme to estimate the asymmetry between brain-to-behavior and behavior-to-brain prediction.
We applied this approach to a simulated driving task and examine directed relationships between
brain activity and continuous driving behavior (steering or heading error). Our results indicated
that two neuro-behavioral states emerge in this naturalistic environment: a Proactive brain state
that actively plans the response to the sensory information, and a Reactive brain state that
processes incoming information and reacts to environmental statistics.
Graphical Abstract
Introduction
The brain is composed of roughly 160 billion neural and non-neural support cells that
coalesce into dynamic, neuronal assemblies of coordinated activity (Azevedo et al., 2009), and
neuroscientists have developed a battery of neuroimaging and analysis techniques to study the
local specialization of neuronal populations at the macro-scale level of organization. In EEG, brain
activity has often been explored with event-related potential (ERP) analyses that reveal localized
peak responses in scalp electrodes that differentiate experimental conditions (for review, see
Luck, 2014) or by averaging EEG data organized into epochs and reconstructed in source space
to examine condition-specific effects in more localized regions of interest in a 3D head model
(Michel et al., 2004). Likewise, conventional fMRI analyses have subtracted whole-brain activation
between experimental conditions to identify localized regions of task-specific computational
processing or analyzed a priori regions of interest, defined functionally or anatomically, to quantify
their sensitivity to a variety of stimuli or task demands (for review, see Huettel et al., 2004).
Collectively, these imaging approaches have produced a rich understanding about segregated
areas of the brain and local specialization within brain regions. Recently, however, there has been
increased interest in examining how the brain coordinates activity across these spatially disperse
regions (Alivisatos et al., 2012).
Neuro-behavioral state modulations in driving 3
Researchers have developed several new methods to investigate entire brain networks,
including diverse approaches such as independent component analysis (e.g., Calhoun et al.,
2009), encoding and decoding algorithms (e.g., Serences and Saproo, 2012) and graph
theoretical approaches (e.g., Bassett and Bullmore, 2006; Bullmore and Sporns, 2009). These
advances have allowed us to investigate the symphony of neural processing rather than a
compartmentalized snapshot of the brain's dynamic response, and research on brain connectivity
among regions continues to increase within the field (e.g., Li et al., 2009; Sakkalis, 2011). Overall,
connectivity-based neuroimaging methodologies show promise
for augmenting our
understanding of how dynamic changes in brain networks support millisecond fluctuations in
behavior (Alivisatos et al., 2012; Friston, 1994; Sporns, Chialvo, Kaiser, & Hilgetag, 2004).
In particular, our research effort presupposes that functional network connectivity across
disparate brain regions yields brain activity that underlies cognition and interaction in a complex
world. We are interested in quantifying ongoing brain network dynamics that underlie the concept
of a brain state, or the "fundamental algorithm by which cognition arises" (Gilbert and Sigman,
2007). Often termed state dependency, some researchers have investigated how resting state
brain activity modifies incoming information (Wörgötter et al., 1998) or disrupts behavioral
performance (Silvanto et al., 2008). Synchronized frequency oscillations are posited as a
mechanism to form transient networks that can integrate information across local, specialized
brain regions (He et al., 2015; Klimesch et al., 2007), and global brain dynamics characterized by
specific frequency oscillations appear to have functional consequences (Buzsáki and Draguhn,
2004). Researchers have identified brain network responses sensitive to task demands (DeSalvo
et al., 2014), stimulus properties (Stansbury et al., 2013), and biomarker development for disease
(Bassett et al., 2008). In this study, we are interested in the relationship between distributed brain
activity and continuous task performance, and we use EEG to study whole-brain oscillatory
activity and capitalize on its temporal precision (Buzsáki and Draguhn, 2004; Engel et al., 2001;
Steriade, 2001).
Here, we investigate temporal dynamics of brain activity and continuous behavioral
performance in a simulated driving task. Driving is a complex visuo-motor task that requires
interaction among cognitive systems to successfully navigate from one location to another, keep
a safe distance from other vehicles, and maintain a consistent lane location while in motion.
Despite the complexity of the task, experienced drivers successfully perform this task, with ease,
often in a near-automatic fashion. Studies have investigated the underlying neural mechanisms
in both real (Sandberg et al., 2011) and simulated (Calhoun et al., 2002; Spiers and Maguire,
2007) driving environments, explored networks that produce task failures (Simon et al., 2011),
and used neural measures to predict vehicle parameters (Lin et al., 2005). Thus, a simulated
driving task affords the opportunity to study relationships between continuous behavioral
measurements and brain dynamics in a naturalistic, everyday task.
Our neuro-behavioral analysis method calculates time-varying asymmetries between
fluctuations in oscillatory activity and two measures of driving behavior. Oscillatory activity is
estimated for four common frequency bands within 12 cortical regions of interest (ROI)
determined a priori from previous driving research (Calhoun et al., 2002; Spiers and Maguire,
2007). The two behavioral measures, steering wheel angle and vehicle heading error, were
chosen because previous research has suggested that heading error of the vehicle is used to
determine the steering response and tightly coupled with brain dynamics (Hildreth et al., 2000; Li
and Cheng, 2011). In this framework, the heading error is a kinematic variable used to scale a
steering response, and the steering response relates back to the heading error by a dynamic
transfer function that accounts for vehicle speed and current heading, among other parameters.
In this manner, examining the neural dynamics related to heading error and steering wheel angle
permit the investigation of the brain's "closed loop" control of the vehicle. From this analysis, we
identify two distinct neuro-behavioral brain states: a Proactive state where the brain activity
predominantly causes behavior and a Reactive state where the brain activity is predominantly
Neuro-behavioral state modulations in driving 4
caused by behavior. We apply analysis of variance to investigate the effects of ROI, frequency
band, and experimental factors on the proportion of time spent in each state, as well as the
transition probability within and between states. Our analysis suggests that the Proactive state
actively plans the response to the sensory information and the Reactive state processes incoming
information and reacts to statistics of the environment.
Method
Twenty-eight neurologically healthy volunteers participated in this experiment. This study
was conducted in accordance with IRB requirements (32 CFR 219 and DoDI 3216.02). Upon
arrival to the lab, participants were introduced to the driving environment and instructed how to
perform the task. Subjects were asked to maintain the vehicle in the center of the rightmost lane
of a four-lane highway (two lanes in each direction) and to maintain consistent vehicle speed at
45 mph as precisely as possible (See Figure 1 for a diagram of the display). Lateral perturbations
resembling wind gusts were periodically imposed on the vehicle causing changes in its heading,
and the participants were instructed to counter them by steering the vehicle back into the center
of the rightmost lane as quickly and accurately as possible. Training on the task consisted of
participants driving for 10-15 min until asymptotic performance in steering and speed control was
demonstrated. They were then outfitted and prepped for the EEG acquisition. Following
completion of the training and experimental setup, the participants proceeded to drive in a 45-min
experimental condition where traffic density was manipulated (sparse, heavy). Vehicle
perturbations ('wind gusts') were also presented in blocks of either high (every 8-10 s) or low
(every 24-30 s) rates. These manipulations were introduced to make the driving experience more
naturalistic and to investigate whether either factor imposed a modulation on the measured neuro-
behavioral states.
Figure 1: Experimental display and behavioral measurements. A) In this experiment, subjects were asked to maintain
course in the far-right lane whilst on-going traffic and perturbations were introduced. A speedometer reading in the
center bottom of the display indicated current speed. B) Continuous steering deviation, the absolute angular difference
from the stationary angle (deg). Lower time course is the steering deviation from approximately 3 min of the experiment
for one participant. C) Heading error, the absolute angular deviation the vehicle was positioned from the center of the
Neuro-behavioral state modulations in driving 5
right lane (deg). Lower time course is the heading error from approximately 3 min of the experiment from one participant.
Neuro-behavioral Analysis
An overview of the analysis steps are graphically described in Figures 2 and 3 and
succinctly introduced here. First, standard preprocessing of EEG was completed on the raw
signal, and continuous behavioral measures were temporally resampled and synchronized with
the EEG signal. Next, cortical current source density (CSD) was estimated using cortically
constrained low resolution electrical tomographic analysis (cLORETA), and the mean CSD was
obtained for 12 regions of interest (ROI) defined a priori from previous literature. For each ROI,
the Hilbert transform was applied to obtain spectral analytic amplitude within four common
frequency bands (delta, theta, alpha, and beta). For each band, time-varying dependency
between spectral amplitude and each behavioral measurement (steering and heading error) was
inferred using generalized partial directed coherence (GPDC). To obtain the GPDC, we fit non-
stationary multivariate autoregressive (MVAR) models to the data using dual extended Kalman
Filtering. From GPDC estimates, a measure of asymmetry was obtained and used to assign
Proactive vs. Reactive state labels to each time point. Finally, we statistically analyzed the effects
of cortical ROI, frequency band, and experimental traffic manipulations (traffic density and vehicle
perturbation frequency) on the proportion of time spent in each state as well as on the transition
probability within and between states. Each step of this pipeline is described in further detail
below.
Data Acquisition and Preprocessing
EEG measurements were made using a 64-channel (1024Hz sampling rate) Biosemi
ActiveTwo System (Biosemi Instrumentations, The Netherlands). Raw EEG measurements were
pre-processed using in-house software in Matlab (Mathworks, Inc.) and the EEGLAB toolbox
(Delorme and Makeig, 2004). The pre-processing pipeline largely follows the PREP approach
(Bigdely-Shamlo et al., 2015) and contains five steps: (1) resampling the raw EEG to 250Hz, (2)
line noise removal via a frequency-domain (multi-taper) regression technique to remove 60Hz
and harmonics present in the signal, (3) a robust average reference with a Huber mean, (4) artifact
subspace reconstruction to remove residual artifact (the standard deviation cutoff parameter was
set to 8), and (5) a piece-wise detrending algorithm to remove low frequency drift in the signal
(window size = 312ms, step size= 20ms).
We computed two measures of driving behavior: steering deviation as the absolute
angular rotation of the steering wheel (measured in degrees), and heading error as the absolute
angular deviation of the vehicle's motion trajectory and a line parallel with the simulated,
indefinitely straight road (measured in degrees). These synchronized behavioral measures
(100Hz sampling rate) were recorded using a distributed architecture, in which multiple data
streams were recorded by different CPU's via an Arduino-based system (Brooks and Kerick,
2015; Jaswa et al., 2012). Each computer in the system produced data logs that included the
common sync marker, and synchronization was performed in post-processing. The measured
jitter within the system was confirmed to be below the resolution of the analysis (50Hz). These
behavioral measures were then converted to degrees, and the absolute value was taken for the
neuro-behavioral analysis.
Neuro-behavioral state modulations in driving 6
Figure 2: Regions of interest, channel locations, and atlas demarcation. Standard 64 channel EEG channel positions
and a headmodel created from the Colin brain in MNI space was used to transform the EEG channel data into cortical
current source density (CSD) via cLORETA. For subsequent analysis, CSD was averaged within 12 ROIs: the posterior
and anterior cingulate cortices (ACC/PCC) as well as bilateral regions in middle frontal gyrus (MFG), supplemental
motor area (SMA), parietal cortex, motor cortex, and lateral occipital cortex.
Distributed Source Reconstruction
From the pre-processed EEG data, we estimated current source density over a 5003-
vertex cortical mesh. A boundary element method (BEM) forward model was derived from the
'Colin 27' anatomy (Holmes et al., 1998) and transformed into MNI305 space (Evans et al., 1993)
using standard electrode positions fit to the Colin 27 head surface in BrainStorm (Tadel et al.,
2011). The BEM solution was computed using OpenMEEG (Gramfort et al., 2010; Kybic et al.,
2005), and the cLORETA approach was used for inverse modeling as described in detail in
(Mullen et al., 2015) and implemented in the BCILAB (Kothe and Makeig, 2013) and Source
Information Flow (SIFT)(Mullen, 2014) toolboxes.
Using averaged CSD from appropriate vertices of the cortical mesh, the functional activity
in 12 a priori ROIs selected from previous driving studies (Calhoun et al., 2002; Spiers and
Maguire, 2007) were estimated from the 5cm (496 parcel) subparcellation of the Desikan-Killiany
atlas (Desikan et al., 2006). These ROIs were anterior cingulate cortex (ACC), posterior cingulate
cortex (PCC), and bilateral regions corresponding to middle frontal gyrus (MFG), supplementary
motor area (SMA), parietal cortex (portions of the inferior and superior parietal lobe), motor cortex
(dorsal precentral gyrus), and lateral occipital regions. ROI locations on the mesh are visualized
and labeled in Figure 2.
Power Spectral Estimation
Continuous, time-varying measures of spectral power within delta (2-3Hz), theta (4-7Hz),
alpha (8-12Hz), and beta (13-25Hz) frequency bands were obtained for each ROI. For each
frequency band, the time-series were filtered with a zero-phase FIR band-pass filter with 6dB
attenuation, as implemented in EEGLAB (Delorme and Makeig, 2004), then Hilbert transformed
to extract the complex analytic amplitude, and finally the magnitude-squared instantaneous power
was obtained. Since power and behavioral measures change relatively slowly, to improve
computational efficiency of subsequent modeling steps and reduce model complexity, these
measures were downsampled to 50Hz prior to modeling.
Neuro-behavioral Relationships
As graphically outlined in Figure 3, time-varying dependencies between ROI power and
behavioral measures were inferred using an effective connectivity measure related to Granger-
Geweke causality (Geweke, 1982; Granger, 1969). When time series data are fit by a multivariate
autoregressive (MVAR) model by minimizing the error terms, the model coefficients provide
information about time lag influences between the signals and capture the causal dependencies
between two or more time-series. However, brain and behavioral dynamics are typically non-
Neuro-behavioral state modulations in driving 7
causal estimate (𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑛𝑛𝑛𝑛𝑛𝑛).
stationary (Boashash et al., 2000; Ku and Kawasumi, 2007).To account for this, we modeled non-
stationary brain-behavior dynamics with a locally-linear MVAR dynamical model estimated using
dual extended Kalman filtering (DEKF) (Wan and Nelson, 1997).This method has previously been
used to model non-stationary causal influences in EEG data (Omidvarnia et al., 2011).
Using the SIFT toolbox, a 5th order time-varying MVAR model was fit to each pair of
normalized bandpower and behavioral measure time-series using DEKF. The DEKF forgetting
factor was set to .01, allowing an effective time window of 2 seconds. The MVAR model
coefficients were then used to estimate generalized partial directed coherence (GPDC, Baccalá
and de Medicina, 2007). This was integrated over all frequencies to yield a time-domain GPDC
The GPDC has the important property of being scale-invariant, ensuring that the inferred
strength of the granger causal influence is independent of the relative scale of data time-series.
Furthermore, since MVAR modeling assumes homoscedastic (equal variance) data, power and
behavioral time-series were temporally normalized using an adaptive z-scoring method within 2
sec windows prior to model fitting. Levene's test for homoscedasticity was applied within a 2 sec
sliding window (step = 1.5 sec) to confirm equal variance between modeled time-series. Results
showed that on average, across subjects, frequency bands, and behavioral measures, only 1.8%
of windows showed significant differences in variance between power and behavior time-series
pairs (p < .05, FDR corrected) confirming the effectiveness of the adaptive z-scoring procedure.
Figure 3: Graphical display of analysis. Power in each frequency band was obtained for each brain region. Pairwise
MVAR models were fit to each pair of EEG frequency band and behavioral measure (Model Coefficients) using a dual
extended Kalman filter (DEKF) and Generalized PDC was calculated (GPDC Calculation). Asymmetry was defined as
Neuro-behavioral state modulations in driving 8
the difference divided by the sum of the GPDC measures from brain activity to behavior and vice versa (Asymmetry
Values). The final step was to compare neuro-behavioral states that significantly deviated from a null distribution
obtained by shuffling the timecourses for behavior and brain activity (Proportion of Time). This allowed us to create a
measure that corresponded to the amount of time each participant's brain was significantly in a Proactive or Reactive
brain state.
GPDC estimates of casual influence were calculated for each time-series pair (12 regions,
4 bands, 2 behaviors = 96 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑛𝑛𝑛𝑛𝑛𝑛 time courses per subject). We then summarized the granger
causal influence of brain activity (BA) on behavior (B) (𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵𝐵𝐵→𝐵𝐵) or relative to that of behavior
on brain activity (𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵→𝐵𝐵𝐵𝐵) using the asymmetry ratio:
𝐴𝐴=𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵𝐵𝐵→𝐵𝐵−𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵→𝐵𝐵𝐵𝐵
𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵𝐵𝐵→𝐵𝐵+𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵→𝐵𝐵𝐵𝐵
Thus, when A is greater than zero, the relative influence of brain activity on behavior is greater
than the influence of behavior on brain activity (Figure 3, Asymmetry row). We label these states
as Proactive since the neural activity can be thought of as controlling the driving behavior.
Conversely, when A is less than zero, behavior predominantly influences brain activity, and we
label these states Reactive since the neural activity can be thought of as occurring in response
to actions taken in the driving task.
Next, for each frequency band, ROI, brain state, and behavioral measure, we calculated
the proportion of asymmetry values larger in magnitude than expected under a null hypothesis of
granger causal independence of brain and behavioral measures. For each pair of time-series, an
asymmetry null distribution, Anull, was constructed. This destroys the temporal dependency
between time-series by circularly shifting the continuous time-series of each behavioral measure
relative to each corresponding brain measure. This was followed by GPDC and asymmetry ratio
calculation as described above. Asymmetry values lying outside the central 95% of Anull were then
deemed significant at the level of p<0.05. Example null distributions are depicted in Figure 3,
bottom row.
Analyses of variance (ANOVAs) were applied to proportional values to quantify the effect
of experimental and neurophysiological factors on the proportion of time spent in Proactive and
Reactive states for steering and heading error. The factors examined were (1-2) two experimental
traffic manipulations, traffic density and perturbation frequency,(3) neuronal frequency band and
(4) cortical ROI.
Finally, we investigated the relative stationarity of Proactive and Reactive states for each
behavioral measure and frequency band by estimating the transition probability between and
within each state: Proactive-to-Proactive, Proactive-to-Reactive, Reactive-to-Reactive, Reactive-
to-Proactive. The transition probability 𝑝𝑝𝑖𝑖𝑖𝑖=𝐺𝐺(𝑆𝑆𝑛𝑛=𝑗𝑗 𝑆𝑆𝑛𝑛−1=𝑖𝑖) , where𝑆𝑆𝑛𝑛 is the state at time t and
𝑖𝑖,𝑗𝑗∈{Proactive,Reactive},was obtained using the maximum likelihood estimate 𝑝𝑝𝑖𝑖𝑖𝑖=𝑛𝑛𝑖𝑖𝑖𝑖 ∑𝑛𝑛𝑖𝑖𝑖𝑖
𝑖𝑖⁄
where 𝑛𝑛𝑖𝑖𝑖𝑖 is the number of sequential transitions from state 𝑖𝑖 to 𝑗𝑗. We then submitted these
transition probabilities to an ANOVA with 12 regions, 4 frequency bands, 2 behaviors, and 4 state
transitions as factors.
Post-hoc paired t-tests were used to investigate simple effects driving main effects and
interactions found for each ANOVA. All t-tests are reported with significance corrected for multiple
comparisons with the false discovery rate procedure (FDR, Benjamini and Yekutieli, 2001).
Results
In this experiment, we study the relationship between the temporal dynamics of EEG
oscillatory activity and continuous fluctuations in two behavioral measures of driving performance
in a simulated environment, the participant's steering behavior and vehicle's heading error
(Hildreth et al., 2000; Li and Cheng, 2011). Classical granger causal estimates of connectivity
assume a stationary and linear representation of multichannel or multisource EEG activity, an
assumption that is often inaccurate when modeling EEG data. Here, we estimate the cortical
,
Neuro-behavioral state modulations in driving 9
source activity, parcellate mesh vertices using the Desikan-Killiany atlas (Desikan et al., 2006),
and select 12 a priori regions of interest from previous driving (Calhoun et al., 2002; Spiers and
Maguire, 2007). We then use a time-varying model, a dual extended Kalman filter (DEKF), to link
brain activity to behavior, and we calculate generalized partial directed coherence, GPDC, to
estimate the granger causal influence for each ROI-behavior pair in four frequency bands (delta,
theta, alpha, and beta). Thus, this analysis computes a total of 96 GPDC time courses per
participant from 12 a priori regions, 4 frequency bands, and 2 continuous driving behaviors.
Our neuro-behavioral analysis uses an asymmetry measure to emphasize the
directionality of the casual GPDC brain-behavior relationship during the driving task. When values
label these time intervals as a Proactive brain state since the neural activity can be thought of as
are higher for 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵𝐵𝐵→𝐵𝐵, the brain activity precedes and predicts behavior performance, and we
controlling the driving behavior. Conversely, when values are higher for 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐵𝐵→𝐵𝐵𝐵𝐵, the behavioral
permuted null distribution of the brain-behavior values, 𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝐺𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛. This asymmetry measure is
performance precedes and predicts brain activity, and we label these time intervals as a Reactive
brain state since the neural activity occurs in response to actions needed in the driving task. The
significant time intervals of the brain state in these time intervals is computed by comparing to a
used for two of the result sections: the first identifies the dominant brain state by quantifying the
distribution of asymmetry values, and the other estimates the transition probability of switching
between brain states.
For the other three result sections, we calculate the proportion of time each of 96 GPDC
time courses was in the Proactive or Reactive brain states to examine differential relationships
among regions, frequency bands, and the two driving performance measures. These proportion
values were used to examine (1) the effect of two experimental traffic manipulations, traffic density
and perturbation frequency, on detected brain states, (2) the dependence of brain states on
particular frequency band oscillations, and (3) the regions that make up the brain networks that
underlie the Proactive and Reactive brain states in this simulated driving task.
Subjects are primarily in a reactive state
First, we examine histograms of the asymmetry measure to assess the relative frequency
of Proactive and Reactive states in the simulated driving task (Figure 4). We observe that group-
averaged histograms for both steering and heading error are left-skewed. This suggests subjects
were primarily in a Reactive brain state with behavior predominantly influencing brain activity, for
each subject (mean = .84 vs .16, SD = .03, across subjects). Further, for a subset of the subjects,
there was an effect between driving measures within the Proactive brain state. Figure 4 shows a
larger proportion of time spent in the Proactive state (right side of histogram, positive asymmetry
values) for steering (green) versus heading error (blue).This likely reflects demand characteristics
of the driving measurement since steering is arguably a more proactive behavioral estimate than
heading error.
Neuro-behavioral state modulations in driving 10
Figure 4: Histograms of asymmetry values across the experiment, averaged over all subjects (light shaded regions
denote 1 standard deviation across subjects).
Effects of task manipulations on neuro-behavioral states
Although the Reactive state is most common in this simulated driving task, we ask whether
the proportion of time spent in Reactive and Proactive states reflect task demands in driving. The
experimental design incorporated two major task manipulations during the 45-min highway drive:
traffic density was blocked as intervals of sparse and heavy traffic, and the frequency of vehicle
perturbations ('wind gusts') occurred in blocks of either high (every 8-10 s) or low (every 24-30 s)
rates. These manipulations were introduced to make the driving experience more naturalistic and
to investigate whether either factor imposed a modulation on brain and behavioral measures. For
example, more frequent perturbation events during driving may require a more frequently
Proactive state to ensure efficient correction to lane deviations.
To examine the effect of the traffic manipulations, we submitted the proportion of time
participants are within each neuro-behavioral state to an ANOVA with brain state, traffic density,
and perturbation frequency as factors. First, only a main effect of brain state was found (F(1,222)
= 17.6, p < .001), confirming the analysis of histogram values (Figure 4) that identified a
predominance of the Reactive brain state in this simulated driving task.
Second, two significant interaction effects were found: the first was a state by traffic density
interaction (F(1,222) = 4.5, p = .03), and the second was a brain state by perturbation frequency
interaction (F(1,222) = 4.5, p = .04). As Figure 5 shows, the state by traffic density interaction was
driven by a reversal of effects for the traffic manipulation: in the high traffic density condition, there
is more time spent in the Reactive state and less time spent in the Proactive state, whereas the
amount of time is similar across the two brain states in the low traffic density condition. The brain
state by perturbation frequency was driven by an increase in amount of time in a Reactive state
between low and high perturbation frequency conditions, whereas the amount of time in a
Proactive state is slightly decreased between the low and high perturbation frequency.
Collectively, these interaction effects suggest that the Proactive brain state is more sensitive to
these experimental conditions than the Reactive state.
Neuro-behavioral state modulations in driving 11
Figure 5: Effect of perturbation frequency on neuro-behavioral state. Lines represent the mean proportional time for
the Proactive (left panel) and Reactive (right panel) brain states within two experimental conditions: perturbation
frequency (columns, x axis) and traffic density (lines). Error bars are SEM.
Switching between states
We have shown that subjects are primarily in a Reactive state within this experiment and
the Proactive state is more malleable to task demands. But how often do the participants switch
to each brain state and what are the temporal dynamics of this transition between brain states?
The asymmetry values from the right parietal cortex are shown in Figure 6A for approximately 3
minutes for a single subject's steering behavior with each of the 4 frequency bands. This
visualization highlights our general observation that state transitions are infrequent but statistically
significant timeframes occur for each neuro-behavioral state.
We quantify the transition probability between brain states using statistically significant
asymmetry values. On average, the transition probability within states is.48 (Proactive-to-
Proactive and Reactive-to-Reactive), whereas between states (Proactive-to-Reactive and
Reactive-to-Proactive) is .02. These results indicate that transitions between Proactive and
Reactive states are infrequent (Figure 6B). An ANOVA with driving behavior, state transition
(Proactive-to-Reactive, Reactive-to-Proactive, Reactive-to-Reactive, Proactive-to-Proactive), and
frequency band as factors showed main effects of state transition (F(3,10364) = 6.9X105, p <
.001), reflecting the large difference between within state stationarity (.48) compared to the rare
state transition (.02). The ANOVA also revealed two significant interactions, one between
behavior and state transition (F(3,10364) = 39.9, p < .001) and another between frequency band
and state transition (F(9,10358) = 220, p < .001).
As shown in Figure 6C, the behavior by state transition interaction was driven by the
increased within state probability from Reactive-to-Reactive for heading error compared to
steering deviation and an increase in transition probability for the steering behavior when
inspecting between-state
(Reactive-to-Proactive, Proactive-to-Reactive). This
suggests that the brain stays in a reactive state in conjunction with the heading error measure
more frequently than with the steering angle. This result confirms the histogram analysis (Figure
4) that suggested the demand characteristics of steering may require the Proactive state
compared to heading error.
To further examine the between state transitions (R-P and P-R) observed only in
conjunction with the steering measure, we investigated frequency band by state transition
interaction, and Figure 6D displays the means for the between state transitions. In FDR-corrected
paired t-tests (q < .05) between each frequency pair, we find that all pair wise comparisons are
significantly different from each other with alpha showing the highest probability of Proactive-to-
Reactive and Reactive-to-Proactive state transitions.
Collectively, these results confirm the dominance of the Reactive brain state, particularly
for heading behavior, and the largest transitions between the two brain states occur in the alpha
band in conjunction with steering behavior which is an arguably more Proactive behavioral
transitions
measure.
Neuro-behavioral state modulations in driving 12
Figure 6: A) A 3-minute segment of continuous asymmetry values between parietal ROI power and steering for delta,
theta, alpha, and beta bands. Significant asymmetry values lie outside the shaded regions, which denote the central
95th percentiles of corresponding null distributions. B) Mean transition probabilities reveal that transitions between
Proactive and Reactive states are relatively infrequent. C) Mean difference (SEM across subjects) in neuro-behavioral
state transition probability between the two behavioral measures (heading error – steering deviation) is shown. Positive
values indicate greater transition probability for heading error, while negative values indicate greater transition
probability for steering. For within state transitions, heading error has the most Reactive-Reactive transitions, while
steering shows increased rate of transition between states for both Proactive-Reactive and Reactive-Proactive. D) The
between state transition probabilities for steering behavior, averaged over ROIs, are shown for each frequency band.
The probability of a state transition is significantly greater for Alpha than any other frequency band (q <.05, FDR
corrected).
Frequency interactions with neuro-behavioral state
We continue to examine the proportion of time spent in Proactive and Reactive states as
Neuro-behavioral state modulations in driving 13
a function of frequency band. To quantify frequency effects, the proportions were submitted to an
ANOVA with 12 regions, 4 frequency bands, 2 behaviors, and 2 brain states as factors. Main
effects of frequency band (F(3, 5180) = 70.2, p < 0.001), behavior (F(1,5182) = 82.5, p < 0.001),
and brain state (F(1,5182) = 660.2, p < 0.001) were found, with several significant interactions.
Of note, behavior by brain state (F(1, 5182) = 155.0, p < 0.001) and frequency band by brain state
(F(3, 5180) = 336.7, p < 0.001) were both significant.
To understand these effects, we performed post-hoc t-tests to examine pairwise
differences in the mean proportion of time spent in each state, averaged over all ROIs and
subjects, for all frequency bands, states, and behavioral measures (Figure 7). Significance tests
were corrected for multiple comparisons using FDR (q < .05). The analysis revealed a number of
significant differences in means. For the Proactive brain state, each of the four frequency bands
showed statistically significant within-band differences in means between heading error and
steering, further arguing that steering deviation is a more proactive measure. Furthermore, within
each behavioral measure, the means were significantly different for all pairs of frequency bands.
Conversely, for the Reactive state no significant differences in means were found within or
between behavioral measures, although the alpha and delta comparison for steering is marginally
significant (p < .05, uncorrected). The means for each frequency in the Reactive brain state in
Figure 7 suggest that both heading error and steering deviation have similar frequency band
profiles with a dominant role in the alpha band.
Figure 7:Proportion of time spent in Proactive and Reactive states, averaged over all ROIs and subjects, for each
frequency band (delta, theta, alpha, beta) and behavioral measure (heading error, steering deviation). Error bars denote
SEM. (*) denotes statistically significant differences in means at the FDR-corrected level of q < 0.05. For the Proactive
state, the means are significantly different between all pairs of frequency bands. For the Reactive state, there were no
significant differences in means between frequency bands.
Regional contributions to neuro-behavioral state
Finally, we investigate the regional contributions to Proactive and Reactive states as a
function of frequency band. In Figure 8A, we first re-plot the frequency effects for steering that
were shown in Figure 7 by placing a colored orb for each of our 12 ROIs (Figure 2), and its size
represents the proportion of time measurement.
In Figure 8B, results from our final analysis are shown. Here, the orbs are scaled within
Neuro-behavioral state modulations in driving 14
those three frequency bands (mean subtracted) to reveal what regions are the strongest
contributors for steering behavior across the two task states. For the Proactive state, steering
behavior is dependent on delta activity across a diffuse set of brain regions consisting of SMA,
motor, frontal, and PCC. For beta activity, steering behavior depends on right lateralized parietal,
occipital, and frontal regions with substantial contribution from the motor cortex. For the Reactive
state, steering behavior predominantly influences alpha activity in right parietal, motor, and frontal
brain regions.
Figure 8: Data from Figure 7 (right panel; steering deviation), plotted as orbs representing the 12 ROIs. A) Orbs plotted
at the centroid of each of our 12 ROIs and size scaled across frequencies based on proportion of time within each task
state. B) Orb size now scaled by regional contribution to proportion of time in each task state for each frequency, as
indicated by the colored boxes and labels.
Discussion
We investigated the directionality in brain-behavior interactions during a simulated driving
task, and identified two distinct neuro-behavioral states. The Proactive state consists of task
intervals when brain activity precedes behavior and seen as the cause of the behavior. The
Proactive state is flexible and dependent on task demands, is most predictive within the delta and
beta bands, and may reflect motor execution and error determination. The Reactive brain state,
on the other hand, appears inflexible to task demands, is most predictive by posterior parietal-
motor-frontal alpha band activity, and may reflect the monitoring of environmental statistics. More
generally, our results indicate the power of this neuro-behavioral method that is not constrained
by segregation and averaging over experimental trials; instead, this neuro-behavioral analysis
reveals the relationship between neural signals and continuous behavioral measurements,
enabling the study of dynamic fluctuations in task performance dependent on idiosyncratic
Neuro-behavioral state modulations in driving 15
changes in neuro-behavioral state.
The Proactive Neuro-behavioral State
The Proactive state, where brain activity precedes and predicts behavior, is characterized
by diffuse anterior-posterior delta and beta oscillatory activity, and it seems to reflect a state which
actively plans the response to the sensory information. Subjects were rarely in this state within
the experiment, most likely due to the fact that little response was necessary to maintain course.
However, this state was more predominant for steering behavior than heading error, and this likely
reflects demand characteristics of the driving measurement since steering is arguably a more
proactive behavioral estimate than heading error. This result was captured in the histogram
analysis with a larger number of windows for significant timeframes in the Proactive state as well
as the analysis of transition probabilities. We observed increased transitions from Proactive-to-
Reactive and Reactive-to-Proactive states in conjunction with steering behavior. These between
state transitions may suggest that subjects quickly switch to the Proactive brain state when an
action is needed. The flexibility of this state is also supported by the interaction between state and
our two naturalistic driving conditions that revealed more flexibility in the Proactive state
dependent on environmental statistics.
Regions and frequency bands also support the notion that the Proactive state is, indeed,
one of action. Numerous studies have associated beta band activity generated in motor cortex
and surrounding areas with preparatory action (Alegre et al., 2003; Baker, 2007; Tan et al., 2013;
Tzagarakis et al., 2015). Beta activity of this sort is also coherent with electromyographic activity
(Baker et al., 1999) and is significantly linked with BOLD fluctuations within motor and pre-motor
regions (Ritter et al., 2009). Although delta band activity is often associated with sleep (Amzica
and Steriade, 1998), it has also been prominent in the decision making literature, including making
judgments to discriminate stimuli in auditory oddball tasks (Başar-Eroglu et al., 1992; Schürmann
et al., 1995). Moving beyond these behavioral links, the topographic distribution of delta band
activity is often diffuse across scalp electrodes, and previous research has interpreted this diffuse
pattern as consistent with a "distributed response system" (Başar et al., 2001). Our source
analysis also finds a diffuse anterior-posterior network that complements this previous literature.
Collectively, our results support the notion that the Proactive brain state is one of preparation and
action, including a role in deciding and planning the response.
The posterior cingulate cortex was a strong contributor of the Proactive state within the
delta band, but it may play a more regulating role within this network. It is the primary hub in the
default mode network, a network of brain regions that show reliable deactivation during a variety
of cognitive tasks (Raichle et al., 2001), and recent research has shown it to play a more active
role in regulating cognition (Gilbert and Sigman, 2007; Hampson et al., 2006; Leech et al., 2012;
Pearson et al., 2011). In the nonhuman primate, the PCC has also been shown to signal
environmental change and the need to alter behavior (Hayden et al., 2010), so this role would
also be relevant within the driving task.
These results raise an interesting question about whether these two brain networks,
separable by frequency bands and regional contribution, communicate with one another to
accomplish task aims. Though beyond the scope of the current study, there is substantial
evidence that relationships between fast and slow rhythms in the brain are linked to behavioral
action (or inaction) under varying levels of motivation (Putman, 2011; Schutter and Van Honk,
2005).
The Reactive Neuro-behavioral State
Participants spent most of the 45 minute, simulated driving session in a Reactive state,
where driving performance behavior predicts brain activity. The Reactive state seems to process
incoming information and react to environmental statistics. We attribute the predominance of this
state to high monotony in the task. Consistent with a reactive interpretation, this neuro-behavioral
Neuro-behavioral state modulations in driving 16
state was predominantly associated with alpha activity in posterior parietal, motor, and frontal
regions. Alpha band activity (8-12Hz) is an intrinsic brain oscillation of fervent study due to its
prominence in resting EEG and sensitivity to various task demands. Several hypotheses have
been proposed ascribing a functional role to its presence in EEG. The first was proposed by
Adrian and Mathews (1934) who found that the power within the alpha band increases when
subjects are awake with eyes closed. They interpreted this as alpha band activity reflecting a brain
state of inactivity, priming the brain for incoming information. This theory has been expanded and
revised to more clearly represent 'cortical idling' (Pfurtscheller et al., 1996), and modern
extensions of this have shown that even at a shorter temporal scale, alpha activity may gate
perceptual information (Jensen and Mazaheri, 2010; van Dijk et al., 2008). More recently,
however, these theories have been further developed, proposing that alpha activity represents
controlled access to a knowledge system, constrained by the limits of attention (Klimesch et al.,
2007). Within the context of this driving task, a reactive state where behavior predicts brain activity
appears to be consistent with these theories about alpha oscillations, and we interpret its role as
gating perceptual information, allowing the mind to wander until a salient perceptual event that
requires a behavior response occurs in the environment and requires the brain to react and correct
performance.
Results from both the asymmetry values and transition probabilities demonstrated that this
Reactive state was sustained, and our analysis of the two naturalistic driving conditions revealed
inflexibility in the Reactive brain state since it is not dependent on environmental statistics. The
underlying network consisted mostly of sensory regions and attentional regions in parietal cortex.
In our experiment, these regions contributed equally to this neuro-behavioral state and suggests
an interesting network of attentional sensory gating. In particular, the occipital and parietal regions
are regularly implicated in visual perception and attention tasks, and they have been consistently
shown to be sources of alpha band activity(e.g., Laufs et al., 2003).
Together, these findings may characterize a Reactive neuro-behavioral state associated
with an alpha-band, sensory gating network with attentional constraints to indicate environmental
change. Future research applying this method to more dynamic environments and a diverse set
of tasks will determine the flexibility of this sensory-driven brain state. It would also be interesting
to examine the connectivity between the regions supporting this neuro-behavioral state.
Advancements in linking behavior with brain activity
Although this analysis focused on behavior-brain dynamics while driving, this approach
can be used more generally to study relationships between functional brain networks and
continuous task performance in other domains. Examining the directional relationship between
brain and behavior operationalizes the concept of a brain state, emphasizing the study of large-
scale oscillatory activity from EEG data to investigate cross-region communication and whole-
brain dynamics. Thus, it can test the hypothesis that synchronized frequency oscillations provide
a mechanism to form transient networks that can integrate information across local, specialized
brain regions (He et al., 2015; Klimesch et al., 2007), and it can reveal how specific oscillations
result in behavioral consequences and dynamic fluctuations in task performance. Here, we
identified relationships between the temporal dynamics of EEG oscillatory activity and continuous
fluctuations in two behavioral measures of driving performance in a simulated environment, the
participant's steering behavior and vehicle's heading error. The directional influence reveals
insights about the role of neuro-behavioral states, indicating when an individual is actively
planning a course of action versus timeframes when the person is merely reacting to salient
events in the environment. Future research will determine whether brain-behavior interactions
found within additional task domains also show controlled transitions among brain states in
conjunction with ongoing task demands.
Neuro-behavioral state modulations in driving 17
Acknowledgements
We would like to thank Piotr Franaszczuk and Nima Bigdely Shamlo for input on analyses and
earlier drafts of this work and Patrick Connolly for data collection. Research was sponsored by
the U.S. Army Research Laboratory, including work under Cooperative Agreement Numbers
W911NF-10-2-0022. The views and conclusions contained in this document are those of the
authors and should not be interpreted as representing the official policies, either expressed or
implied, of the Army Research Laboratory or the U.S. Government.
Neuro-behavioral state modulations in driving 18
References
Adrian, E.D., Matthews, B.H.C., 1934. The interpretation of potential waves in the cortex. J.
Physiol. 81, 440–471.
Alegre, M., Gurtubay, I.G., Labarga, A., Iriarte, J., Malanda, A., Artieda, J., 2003. Alpha and
beta oscillatory changes during stimulus-induced movement paradigms: effect of
stimulus predictability. Neuroreport 14, 381–385.
doi:10.1097/01.wnr.0000059624.96928.c0
Alivisatos, A.P., Chun, M., Church, G.M., Greenspan, R.J., Roukes, M.L., Yuste, R., 2012. The
brain activity map project and the challenge of functional connectomics. Neuron 74,
970–4. doi:10.1016/j.neuron.2012.06.006
Amzica, F., Steriade, M., 1998. Electrophysiological correlates of sleep delta waves1.
Electroencephalogr. Clin. Neurophysiol. 107, 69–83. doi:10.1016/S0013-4694(98)00051-
0
Azevedo, F.A.C., Carvalho, L.R.B., Grinberg, L.T., Farfel, J.M., Ferretti, R.E.L., Leite, R.E.P.,
Filho, W.J., Lent, R., Herculano-Houzel, S., 2009. Equal numbers of neuronal and
nonneuronal cells make the human brain an isometrically scaled-up primate brain. J.
Comp. Neurol. 513, 532–541. doi:10.1002/cne.21974
Baccalá, L.A., de Medicina, F., 2007. Generalized Partial Directed Coherence, in: 2007 15th
International Conference on Digital Signal Processing. Presented at the 2007 15th
International Conference on Digital Signal Processing, pp. 163–166.
doi:10.1109/ICDSP.2007.4288544
Baker, S.N., 2007. Oscillatory interactions between sensorimotor cortex and the periphery. Curr.
Opin. Neurobiol. 17, 649–655. doi:10.1016/j.conb.2008.01.007
Baker, S.N., Kilner, J.M., Pinches, E.M., Lemon, R.N., 1999. The role of synchrony and
oscillations in the motor output. Exp. Brain Res. 128, 109–117.
Başar, E., Başar-Eroglu, C., Karakaş, S., Schürmann, M., 2001. Gamma, alpha, delta, and theta
oscillations govern cognitive processes. Int. J. Psychophysiol. 39, 241–248.
doi:10.1016/S0167-8760(00)00145-8
Başar-Eroglu, C., Başar, E., Demiralp, T., Schürmann, M., 1992. P300-response: possible
psychophysiological correlates in delta and theta frequency channels. A review. Int. J.
Psychophysiol. 13, 161–179. doi:10.1016/0167-8760(92)90055-G
Bassett, D.S., Bullmore, E., 2006. Small-world brain networks. Neurosci. Rev. J. Bringing
Neurobiol. Neurol. Psychiatry 12, 512–523. doi:10.1177/1073858406293182
Bassett, D.S., Bullmore, E., Verchinski, B.A., Mattay, V.S., Weinberger, D.R., Meyer-
Lindenberg, A., 2008. Hierarchical organization of human cortical networks in health and
schizophrenia. J. Neurosci. Off. J. Soc. Neurosci. 28, 9239–9248.
doi:10.1523/JNEUROSCI.1929-08.2008
Benjamini, Y., Yekutieli, D., 2001. The control of the false discovery rate in multiple testing
under dependency. Ann. Stat. 29, 1165–1188. doi:10.1214/aos/1013699998
Bigdely-Shamlo, N., Mullen, T., Kothe, C., Su, K.-M., Robbins, K.A., 2015. The PREP pipeline:
standardized preprocessing for large-scale EEG analysis. Front. Neuroinformatics 9, 16.
doi:10.3389/fninf.2015.00016
Boashash, B., Carson, H., Mesbah, M., 2000. Detection of seizures in newborns using time-
frequency analysis of EEG signals, in: Proceedings of the Tenth IEEE Workshop on
Statistical Signal and Array Processing, 2000. Presented at the Proceedings of the
Tenth IEEE Workshop on Statistical Signal and Array Processing, 2000, pp. 564–568.
doi:10.1109/SSAP.2000.870188
Brooks, J., Kerick, S., 2015. Event-related alpha perturbations related to the scaling of steering
Neuro-behavioral state modulations in driving 19
wheel corrections. Physiol. Behav. 149, 287–293. doi:10.1016/j.physbeh.2015.05.026
Bullmore, E., Sporns, O., 2009. Complex brain networks: graph theoretical analysis of structural
and functional systems. Nat. Rev. Neurosci. 10, 186–198. doi:10.1038/nrn2575
Buzsáki, G., Draguhn, A., 2004. Neuronal Oscillations in Cortical Networks. Science 304, 1926–
1929. doi:10.1126/science.1099745
Calhoun, V.D., Liu, J., Adalı, T., 2009. A review of group ICA for fMRI data and ICA for joint
inference of imaging, genetic, and ERP data. NeuroImage, Mathematics in Brain
Imaging 45, S163–S172. doi:10.1016/j.neuroimage.2008.10.057
Calhoun, V.D., Pekar, J.J., McGinty, V.B., Adali, T., Watson, T.D., Pearlson, G.D., 2002.
Different activation dynamics in multiple neural systems during simulated driving. Hum.
Brain Mapp. 16, 158–167. doi:10.1002/hbm.10032
Delorme, A., Makeig, S., 2004. EEGLAB: an open source toolbox for analysis of single-trial EEG
dynamics including independent component analysis. J. Neurosci. Methods 134, 9–21.
doi:10.1016/j.jneumeth.2003.10.009
DeSalvo, M.N., Douw, L., Takaya, S., Liu, H., Stufflebeam, S.M., 2014. Task-dependent
reorganization of functional connectivity networks during visual semantic decision
making. Brain Behav. 4, 877–885. doi:10.1002/brb3.286
Desikan, R.S., Ségonne, F., Fischl, B., Quinn, B.T., Dickerson, B.C., Blacker, D., Buckner, R.L.,
Dale, A.M., Maguire, R.P., Hyman, B.T., Albert, M.S., Killiany, R.J., 2006. An automated
labeling system for subdividing the human cerebral cortex on MRI scans into gyral based
regions of interest. NeuroImage 31, 968–980. doi:10.1016/j.neuroimage.2006.01.021
Engel, A.K., Fries, P., Singer, W., 2001. Dynamic predictions: oscillations and synchrony in top-
down processing. Nat. Rev. Neurosci. 2, 704–716. doi:10.1038/35094565
Evans, A.C., Collins, D.L., Mills, S.R., Brown, E.D., Kelly, R.L., Peters, T.M., 1993. 3D statistical
neuroanatomical models from 305 MRI volumes, in: Nuclear Science Symposium and
Medical Imaging Conference, 1993., 1993 IEEE Conference Record. Presented at the
Nuclear Science Symposium and Medical Imaging Conference, 1993., 1993 IEEE
Conference Record., pp. 1813–1817 vol.3. doi:10.1109/NSSMIC.1993.373602
Friston, K.J., 1994. Functional and effective connectivity in neuroimaging: A synthesis. Hum.
Brain Mapp. 2, 56–78. doi:10.1002/hbm.460020107
Geweke, J., 1982. Measurement of Linear Dependence and Feedback between Multiple Time
Series. J. Am. Stat. Assoc. 77, 304–313. doi:10.1080/01621459.1982.10477803
Gilbert, C.D., Sigman, M., 2007. Brain States: Top-Down Influences in Sensory Processing.
Neuron 54, 677–696. doi:10.1016/j.neuron.2007.05.019
Gramfort, A., Papadopoulo, T., Olivi, E., Clerc, M., 2010. OpenMEEG: opensource software for
quasistatic bioelectromagnetics. Biomed. Eng. OnLine 9, 45. doi:10.1186/1475-925X-9-
45
Granger, C., 1969. Investigating Causal Relations by Econometric Models and Cross-spectral
Methods. Econometrica 37, 424–438. doi:10.2307/1912791
Hampson, M., Driesen, N.R., Skudlarski, P., Gore, J.C., Constable, R.T., 2006. Brain
connectivity related to working memory performance. J. Neurosci. Off. J. Soc. Neurosci.
26, 13338–13343. doi:10.1523/JNEUROSCI.3408-06.2006
Hayden, B.Y., Smith, D.V., Platt, M.L., 2010. Cognitive Control Signals in Posterior Cingulate
Cortex. Front. Hum. Neurosci. 4. doi:10.3389/fnhum.2010.00223
He, Y., Gebhardt, H., Steines, M., Sammer, G., Kircher, T., Nagels, A., Straube, B., 2015. The
EEG and fMRI signatures of neural integration: An investigation of meaningful gestures
and corresponding speech. Neuropsychologia 72, 27–42.
doi:10.1016/j.neuropsychologia.2015.04.018
Hildreth, E.C., H, M., Boer, E.R., Royden, C.S., 2000. From vision to action: Experiments and
models of steering control during driving. J. Exp. Psychol. Hum. Percept. Perform. 26,
1106–1132. doi:10.1037/0096-1523.26.3.1106
Neuro-behavioral state modulations in driving 20
Holmes, C.J., Hoge, R., Collins, L., Woods, R., Toga, A.W., Evans, A.C., 1998. Enhancement of
MR images using registration for signal averaging. J. Comput. Assist. Tomogr. 22, 324–
333.
Huettel, S.A., Song, A.W., McCarthy, G., 2004. Functional magnetic resonance imaging.
Sinauer Associates Sunderland.
Jaswa, M., Kellihan, B., Cannon, M., Hairston, D., Gordon, S., Lance, B., 2012. A general-
purpose low-cost solution for high-resolution temporal synchronization in human-subject
experimentation. Presented at the Society for Neuroscience Annual Meeting, New
Orleans, LA.
Jensen, O., Mazaheri, A., 2010. Shaping Functional Architecture by Oscillatory Alpha Activity:
Gating by Inhibition. Front. Hum. Neurosci. 4. doi:10.3389/fnhum.2010.00186
Klimesch, W., Sauseng, P., Hanslmayr, S., Gruber, W., Freunberger, R., 2007. Event-related
phase reorganization may explain evoked neural dynamics. Neurosci. Biobehav. Rev.
31, 1003–1016. doi:10.1016/j.neubiorev.2007.03.005
Kothe, C.A., Makeig, S., 2013. BCILAB: a platform for brain-computer interface development. J.
Neural Eng. 10, 056014. doi:10.1088/1741-2560/10/5/056014
Ku, Y.G., Kawasumi, M., 2007. Nonstationary EEG Analysis using random-walk model, in:
Magjarevic, R., Nagel, J.H. (Eds.), World Congress on Medical Physics and Biomedical
Engineering 2006, IFMBE Proceedings. Springer Berlin Heidelberg, pp. 1067–1070.
Kybic, J., Clerc, M., Abboud, T., Faugeras, O., Keriven, R., Papadopoulo, T., 2005. A common
formalism for the Integral formulations of the forward EEG problem. IEEE Trans. Med.
Imaging 24, 12–28. doi:10.1109/TMI.2004.837363
Laufs, H., Kleinschmidt, A., Beyerle, A., Eger, E., Salek-Haddadi, A., Preibisch, C., Krakow, K.,
2003. EEG-correlated fMRI of human alpha activity. NeuroImage 19, 1463–1476.
doi:10.1016/S1053-8119(03)00286-6
Leech, R., Braga, R., Sharp, D.J., 2012. Echoes of the Brain within the Posterior Cingulate
Cortex. J. Neurosci. 32, 215–222. doi:10.1523/JNEUROSCI.3689-11.2012
Li, K., Guo, L., Nie, J., Li, G., Liu, T., 2009. Review of methods for functional brain connectivity
detection using fMRI. Comput. Med. Imaging Graph. 33, 131–139.
doi:10.1016/j.compmedimag.2008.10.011
Li, L., Cheng, J.C.K., 2011. Heading but not path or the tau-equalization strategy is used in the
visual control of steering toward a goal. J. Vis. 11, 20–20. doi:10.1167/11.12.20
Lin, C.-T., Wu, R.-C., Jung, T.-P., Liang, S.-F., Huang, T.-Y., 2005. Estimating driving
performance based on EEG spectrum analysis. EURASIP J. Adv. Signal Process. 2005,
3165–3174.
Luck, S.J., 2014. An Introduction to the Event-Related Potential Technique. MIT Press.
Michel, C.M., Murray, M.M., Lantz, G., Gonzalez, S., Spinelli, L., Grave de Peralta, R., 2004.
EEG source imaging. Clin. Neurophysiol. 115, 2195–2222.
doi:10.1016/j.clinph.2004.06.001
Mullen, T.R., 2014. The Dynamic Brain: Modeling Neural Dynamics and Interactions From
Human Electrophysiological Recordings. UNIVERSITY OF CALIFORNIA, SAN DIEGO.
Mullen, T.R., Kothe, C.A.E., Chi, Y.M., Ojeda, A., Kerth, T., Makeig, S., Jung, T.-P.,
Cauwenberghs, G., 2015. Real-time neuroimaging and cognitive monitoring using
wearable dry EEG. IEEE Trans. Biomed. Eng. 62, 2553–2567.
doi:10.1109/TBME.2015.2481482
Omidvarnia, A.H., Mesbah, M., Khlif, M.S., O'Toole, J.M., Colditz, P.B., Boashash, B., 2011.
Kalman filter-based time-varying cortical connectivity analysis of newborn EEG. Conf.
Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. IEEE Eng. Med. Biol. Soc. Annu. Conf.
2011, 1423–1426. doi:10.1109/IEMBS.2011.6090335
Pearson, J.M., Heilbronner, S.R., Barack, D.L., Hayden, B.Y., Platt, M.L., 2011. Posterior
cingulate cortex: adapting behavior to a changing world. Trends Cogn. Sci. 15, 143–151.
Neuro-behavioral state modulations in driving 21
doi:10.1016/j.tics.2011.02.002
Pfurtscheller, G., Stancák, A., Neuper, C., 1996. Event-related synchronization (ERS) in the
alpha band--an electrophysiological correlate of cortical idling: a review. Int. J.
Psychophysiol. Off. J. Int. Organ. Psychophysiol. 24, 39–46.
Putman, P., 2011. Resting state EEG delta–beta coherence in relation to anxiety, behavioral
inhibition, and selective attentional processing of threatening stimuli. Int. J.
Psychophysiol. 80, 63–68. doi:10.1016/j.ijpsycho.2011.01.011
Raichle, M.E., MacLeod, A.M., Snyder, A.Z., Powers, W.J., Gusnard, D.A., Shulman, G.L.,
2001. A default mode of brain function. Proc. Natl. Acad. Sci. 98, 676–682.
doi:10.1073/pnas.98.2.676
Ritter, P., Moosmann, M., Villringer, A., 2009. Rolandic alpha and beta EEG rhythms' strengths
are inversely related to fMRI-BOLD signal in primary somatosensory and motor cortex.
Hum. Brain Mapp. 30, 1168–1187. doi:10.1002/hbm.20585
Sakkalis, V., 2011. Review of advanced techniques for the estimation of brain connectivity
measured with EEG/MEG. Comput. Biol. Med., Special Issue on Techniques for
Measuring Brain Connectivity 41, 1110–1117. doi:10.1016/j.compbiomed.2011.06.020
Sandberg, D., Anund, A., Fors, C., Kecklund, G., Karlsson, J.G., Wahde, M., Åkerstedt, T.,
2011. The Characteristics of Sleepiness During Real Driving at Night-A Study of
Driving Performance, Physiology and Subjective Experience. SLEEP.
doi:10.5665/sleep.1270
Schürmann, M., Başar-Eroglu, C., Demiralp, T., Başar, E., 1995. A Psychophysiological
Interpretation of Theta and Delta Responses in Cognitive Event-Related Potential
Paradigms, in: Kruse, D.P., Stadler, P.D.M. (Eds.), Ambiguity in Mind and Nature,
Springer Series in Synergetics. Springer Berlin Heidelberg, pp. 357–388.
Schutter, D.J.L.G., Van Honk, J., 2005. Electrophysiological ratio markers for the balance
between reward and punishment. Brain Res. Cogn. Brain Res. 24, 685–690.
doi:10.1016/j.cogbrainres.2005.04.002
Serences, J.T., Saproo, S., 2012. Computational advances towards linking BOLD and behavior.
Neuropsychologia 50, 435–446. doi:10.1016/j.neuropsychologia.2011.07.013
Silvanto, J., Muggleton, N., Walsh, V., 2008. State-dependency in brain stimulation studies of
perception and cognition. Trends Cogn. Sci. 12, 447–454. doi:10.1016/j.tics.2008.09.004
Simon, M., Schmidt, E.A., Kincses, W.E., Fritzsche, M., Bruns, A., Aufmuth, C., Bogdan, M.,
Rosenstiel, W., Schrauf, M., 2011. EEG alpha spindle measures as indicators of driver
fatigue under real traffic conditions. Clin. Neurophysiol. 122, 1168–1178.
doi:10.1016/j.clinph.2010.10.044
Spiers, H.J., Maguire, E.A., 2007. Neural substrates of driving behaviour. NeuroImage 36, 245–
255. doi:10.1016/j.neuroimage.2007.02.032
Sporns, O., Chialvo, D.R., Kaiser, M., Hilgetag, C.C., 2004. Organization, development and
function of complex brain networks. Trends Cogn Sci 8, 418–25.
doi:10.1016/j.tics.2004.07.008
Stansbury, D.E., Naselaris, T., Gallant, J.L., 2013. Natural Scene Statistics Account for the
Representation of Scene Categories in Human Visual Cortex. Neuron 79, 1025–1034.
doi:10.1016/j.neuron.2013.06.034
Steriade, M., 2001. Impact of Network Activities on Neuronal Properties in Corticothalamic
Systems. J. Neurophysiol. 86, 1–39.
Tadel, F., Baillet, S., Mosher, J.C., Pantazis, D., Leahy, R.M., Tadel, F., Baillet, S., Mosher,
J.C., Pantazis, D., Leahy, R.M., 2011. Brainstorm: A User-Friendly Application for
MEG/EEG Analysis, Brainstorm: A User-Friendly Application for MEG/EEG Analysis.
Comput. Intell. Neurosci. Comput. Intell. Neurosci. 2011, 2011, e879716.
doi:10.1155/2011/879716, 10.1155/2011/879716
Tan, H.-R.M., Leuthold, H., Gross, J., 2013. Gearing up for action: attentive tracking dynamically
Neuro-behavioral state modulations in driving 22
tunes sensory and motor oscillations in the alpha and beta band. NeuroImage 82, 634–
644. doi:10.1016/j.neuroimage.2013.04.120
Tzagarakis, C., West, S., Pellizzer, G., 2015. Brain oscillatory activity during motor preparation:
effect of directional uncertainty on beta, but not alpha, frequency band. Front. Neurosci.
9. doi:10.3389/fnins.2015.00246
van Dijk, H., Schoffelen, J.-M., Oostenveld, R., Jensen, O., 2008. Prestimulus Oscillatory
Activity in the Alpha Band Predicts Visual Discrimination Ability. J. Neurosci. 28, 1816–
1823. doi:10.1523/JNEUROSCI.1853-07.2008
Wan, E.A., Nelson, A.T., 1997. Neural dual extended Kalman filtering: applications in speech
enhancement and monaural blind signal separation, in: Neural Networks for Signal
Processing [1997] VII. Proceedings of the 1997 IEEE Workshop. Presented at the
Neural Networks for Signal Processing [1997] VII. Proceedings of the 1997 IEEE
Workshop, pp. 466–475. doi:10.1109/NNSP.1997.622428
Wörgötter, F., Suder, K., Zhao, Y., Kerscher, N., Eysel, U.T., Funke, K., 1998. State-dependent
receptive-field restructuring in the visual cortex. Nature 396, 165–168.
doi:10.1038/24157
|
1906.07777 | 1 | 1906 | 2019-06-18T19:25:08 | Cortical computations via metastable activity | [
"q-bio.NC"
] | Metastable brain dynamics are characterized by abrupt, jump-like modulations so that the neural activity in single trials appears to unfold as a sequence of discrete, quasi-stationary states. Evidence that cortical neural activity unfolds as a sequence of metastable states is accumulating at fast pace. Metastable activity occurs both in response to an external stimulus and during ongoing, self-generated activity. These spontaneous metastable states are increasingly found to subserve internal representations that are not locked to external triggers, including states of deliberations, attention and expectation. Moreover, decoding stimuli or decisions via metastable states can be carried out trial-by-trial. Focusing on metastability will allow us to shift our perspective on neural coding from traditional concepts based on trial-averaging to models based on dynamic ensemble representations. Recent theoretical work has started to characterize the mechanistic origin and potential roles of metastable representations. In this article we review recent findings on metastable activity, how it may arise in biologically realistic models, and its potential role for representing internal states as well as relevant task variables. | q-bio.NC | q-bio | Cortical computations via metastable activity
Giancarlo La Camera,1,2,∗ Alfredo Fontanini,1,2 and Luca Mazzucato3
1Department of Neurobiology and Behavior and
2Graduate Program in Neuroscience, State University of New York at Stony Brook, Stony Brook, NY 11794
3Departments of Biology and Mathematics and Institute of Neuroscience, University of Oregon, Eugene, OR 94703
∗corresponding author
ABSTRACT: Metastable brain dynamics are characterized by abrupt, jump-like modulations so that the neural
activity in single trials appears to unfold as a sequence of discrete, quasi-stationary 'states.' Evidence that
cortical neural activity unfolds as a sequence of metastable states is accumulating at fast pace. Metastable
activity occurs both in response to an external stimulus and during ongoing, self-generated activity. These
spontaneous metastable states are increasingly found to subserve internal representations that are not locked
to external triggers, including states of deliberations, attention and expectation. Moreover, decoding stimuli
or decisions via metastable states can be carried out trial-by-trial. Focusing on metastability will allow us to
shift our perspective on neural coding from traditional concepts based on trial-averaging to models based on
dynamic ensemble representations. Recent theoretical work has started to characterize the mechanistic origin
and potential roles of metastable representations. In this article we review recent findings on metastable activity,
how it may arise in biologically realistic models, and its potential role for representing internal states as well as
relevant task variables.
Emails: giancarlo dot lacamera at stonybrook.edu, alfredo dot fontanini at stonybrook.edu,
luca dot mazzucato at uoregon dot edu
9
1
0
2
n
u
J
8
1
]
.
C
N
o
i
b
-
q
[
1
v
7
7
7
7
0
.
6
0
9
1
:
v
i
X
r
a
Contents
1 Introduction
2 Observing metastable activity
2.1 Consistent features of metastable activity
2.2 Metastability of ongoing cortical activity
2.3 Metastable sequences as a substrate for internal computations
3 Spiking network models of metastable activity
3.1 Metastable activity in clustered networks of spiking neurons
3.2 Predictions of the clustered spiking network
3.3 Variations on the clustered architecture
4 Conclusions
Highlights
2
2
2
4
4
6
6
6
8
8
• Metastable activity is increasingly being observed in many cortical areas and in a variety of tasks.
• Metastable activity is correlated with sensory and cognitive processes and may subserve state-dependent
computations.
• Metastable activity can emerge spontaneously in spiking networks with clustered architecture.
• Metastable activity is a candidate neural substrate for coding both external events and internal represen-
tations.
-- 1 --
1
Introduction
Brain circuits consist of large neural networks, where populations of neurons are recurrently coupled via synap-
tic connections. These circuits can be interpreted as dynamical systems with many coupled degrees of freedom
and therefore capable of generating a wealth of dynamical behaviors on very diverse timescales [1]. Those in-
clude transient relaxations towards a point or line attractor, oscillatory patterns, chaotic dynamics or metastable
activity [2, 3]. These dynamical behaviors have been implicated in important brain functions. Transient re-
laxation to stable neural activity (such as a point or line attractor) may subserve memory [4, 5] and perceptual
decisions [6, 7]. Oscillatory dynamics may subserve respiration, locomotion, and other rhythmic forms of
behavior [8, 9]. Chaotic dynamics amplify random perturbations and, when successfully tamed by learning,
can generate complex computations [10, 11]. Metastable dynamics, initially characterized in the presence of
a sensory stimulus [12, 13], might underlie a range of internal computations during cognitive tasks [14 -- 17].
Relaxations to an attractor, oscillations and chaotic dynamics typically describe neural activity as smoothly
varying over time [18, 19]. In contrast, metastable dynamics are characterized by abrupt, jump-like modula-
tions so that neural activity appears to unfold as a sequence of discrete, quasi-stationary "states." Metastable
activity is being found in an increasing number of brain structures of different species engaged in a variety of
tasks, and is the focus of this opinion. The intuitive appeal of metastable activity is that it resonates with our
intuition that our thoughts and actions proceed along a sequence of distinct episodes, as we scan alternatives and
ponder potential options during everyday tasks. It is natural to think that such episodes are being represented in
transient but well-defined neural patterns in our brains. Recent models have started to clarify how metastable
activity could emerge spontaneously as a collective phenomenon via attractor dynamics in spiking networks
[20 -- 23]. These models have proved powerful tools to investigate the circuit origin of metastable states and
their potential role in sensory and cognitive processes. In this article we review recent progress in the analysis
of metastable state sequences, focusing on issues of detection, modeling and interpretation.
2 Observing metastable activity
Metastable states were first characterized in electrophysiological recordings from the prefrontal cortex of mon-
keys performing a delayed localization task [12, 24, 25]. Spike counts from simultaneously recorded neurons
were analyzed with a hidden Markov model (HMM), a popular latent variable model for analyzing sequential
data arising e.g. in time series analysis, speech recognition [26], DNA sequence analysis [27], behavioral anal-
ysis [28], and many other problems [29, 30]. Under the most common application of HMM to neural data,
population activity from simultaneously recorded neurons unfolds through a sequence of "hidden" states, each
state being a set of ensemble firing rates (Fig. 1). Stochastic transitions between states occur at random times
according to an underlying Markov chain, and neurons discharge as Poisson processes with state-dependent
firing rates, though the model can be extended to include refractory periods and other history-dependent factors
[31]. The number of hidden states may be selected using information criteria [15, 32, 33], cross-validation [14],
or nonparametric Bayesian methods [34, 35]. Several measures have been used to assess the goodness-of-fit
of inferred sequences, including variance explained [14], comparison to shuffled datasets [16], and comparison
of single-neuron transitions to ensemble state transition [22]. The strength of HMM is to allow a principled,
unsupervised method for segmenting neural activity into a sequence of discrete metastable states, and its use
has led to intriguing new observations in ensemble dynamics, as we discuss next.
2.1 Consistent features of metastable activity
HMM and related procedures for detecting hidden states have established a number of consistent features across
datasets, animals and tasks. Ensemble activity unfolds as a sequence of metastable states, each lasting from a
-- 2 --
Figure 1. Observing metastability in neural populations . a: Schematics illustration of a hidden Markov model (HMM) with
three states. Hidden states (squares) are collections of firing rates across neurons; arrows indicate transition rates among the
states (thicker arrows denote larger transition rates). b: HMM applied to electrophysiology recordings from the gustatory
cortex of a behaving rat. The top panels show spike rasters from 10 simultaneously recorded neurons segmented via HMM
analysis. Two trials are shown, one in which sucrose (left) and one in which quinine (right) was administered to the rat.
HMM states were assigned to each bin of data when their posterior probability exceeded 0.8 (dashed horizontal line). The
bottom panels show the state sequences in 4 different trials with either sucrose (left) or quinine (right) as a taste stimulus.
Transitions may occur at variable times across trials, but the state sequences are reliable. Panel (b) modified from [13].
few hundred ms to a second or more, with sharp transitions among the states [12, 13, 15, 22] (Fig. 1b). The
transitions are typically one order of magnitude faster than the state durations, are close to their theoretically
observable lower bound, and are not an artifact of the HMM [12, 13, 15]. When the start of the sequence is
aligned to an external event such as stimulus onset, specific state sequences are observed (Fig. 1b), raising the
possibility that stimuli are coded by dynamic state sequences [12, 13]. Among the most significant early results
of applying HMM to the analysis of neural data were the demonstration that (i) the hidden states identified in
response to a stimulus tend to recur during most of the recorded activity, not just in response to the stimulus,
and (ii) pairwise correlations among simultaneously recorded neurons depend on an underlying global activity
state, and not just on neural connectivity [12, 24]. These two observations are very relevant for current issues
in neuroscience, including the origin and role of ongoing activity [36, 37], the importance of the connectome
[38, 39], and the meaning of correlations [40 -- 42]. It's worth remarking that the ability of HMM analysis to
segment the neural activity into distinct states in an unsupervised manner -- i.e., without knowledge of external
triggers such as stimulus timings, -- has been crucial to obtaining the results reviewed above. It is also important
to point out that the HMM analysis on ensemble data allows to decode the neural activity on a trial-by-trial basis,
avoiding the necessity of averaging across trials. Most findings discussed in this review would go undetected if
data were averaged across trials, because ensemble transitions occur at different times in each trial. This latter
feature also allows the time-warping of state sequences, explaining why, for example, the method is so useful
for speech recognition and for the analysis of neural activity during birdsongs [43]. This could also explain
part of the trial-to-trial variability observed in cortex after aligning spike trains across different trials to e.g..
-- 3 --
abneuron #P(state data)aOngoing activityEvoked activityStim.Neurons19P(state)10SucroseCitric acidOngoingEvokedShared1234# of firing rates060Ong.Ev.0**1secNeurons [%]60bdMultistabilityc₀Hidden Markov ModelStatesstate 1state 3state 2firing ratesneuron #statestate stimulus onset [13].
2.2 Metastability of ongoing cortical activity
One central question in contemporary neuroscience is why ongoing cortical activity is so rich in structure and
resembles so closely stimulus-evoked activity [36, 37, 44 -- 49]. The origin of ongoing activity and the nature
of its interaction with evoked activity can potentially reveal much about the way cortical networks are func-
tionally organized and how this may support flexible coding (see [36, 37] for two recent proposals). However,
quantifying precisely the similarities and differences between ongoing and evoked activity has proved difficult.
A recent HMM analysis of ongoing spiking activity in the gustatory cortex of behaving rats has demonstrated
that ongoing activity, similarly to evoked activity, unfolds as a sequence of metastable states [22]. Transitions
among states are characterized by a partial degree of coordination, so that ongoing activity is neither com-
pletely asynchronous nor completely synchronized. Ongoing and stimulus-evoked activity share most of the
same states, although some of the states occur mostly during ongoing activity and others during evoked activity
[33]. About 50% of the ensemble neurons exhibit three or more different firing rates across states, i.e., they are
"multi-stable" rather than bistable; and this fraction is substantially reduced by stimulation [22]. These findings
have all been captured by a spiking network model of metastability that might shed light into the link between
ongoing and evoked activity (see section Spiking network models of metastable activity).
2.3 Metastable sequences as a substrate for internal computations
Do metastable states have precise and specific meanings as neural correlates of external or internally generated
events? Visual [12], gustatory [13, 22] and vibro-tactile stimuli [15] seem represented by reliable sequences
of HMM states in primate and rodent cortical areas in the context of different tasks. Even more remarkably,
though, aspects of the metastable dynamics seem to subserve internal states of attention, expectation, and in-
ternal deliberations during decisions, and can be used to predict behavior. In monkeys performing a selective
attention task, the probability of detecting a change in the attended stimulus was significantly greater when
occurring during an "ON" state (characterized by vigorous multi-unit activity in area V4) than during an "OFF"
state (characterized by faint multi-unit activity) [14] (Fig. 2a). Similarly, different hidden states were preferen-
tially associated to different task conditions in a human working memory task, with occupancy rate in each state
predicting better performance in the corresponding task [35]. Moreover, changes in patterns of functional con-
nectivity across many brain areas co-occurred more reliably with state transitions than with external triggers,
suggesting that metastable states supported by flexible patterns of functional connectivity may reflect internal
representations of task demands.
A similar relationship between state occupancy rate and decisions has been found in monkeys performing
a choice task between differentially rewarded stimuli [17]. Ensemble activity in orbitofrontal cortex (OFC)
mostly switched between two states representing the values of the stimuli currently being offered, with larger
occupancy rate predicting the behavioral decision (Fig. 2b, left).
Intriguingly, slower decisions (reflecting
longer internal deliberation) occurred when the states had similar occupancy rates and were found on trials of
the same kind (thus reflecting subjective preferences rather than actual task difficulty; Fig. 2b, right). These
findings are reminiscent of those on dynamic changes of mind [50] and complement earlier intriguing results on
the possible meaning of state transitions, which may underlie the sudden realization that task rules have changed
[51] or reflect the difficulty of a vibro-tactile discrimination [15]. In the rat hippocampus, HMM states were
recently found to represent the position in a linear track and in an open field during a navigation task [16] (Fig.
2c), an alternative view to decoding spatial maps based on single neurons place fields [52]. The state sequences
were observed in area CA1 during hippocampal sharp wave ripples when the animal was not exploring the
track, and could be used to infer a spatial map of the environment without any reference to external locations
-- 4 --
Figure 2. Metastable activity and cognitive function. a: Left: Multiunit activity from monkey V4 during a selective
attention task alternates between ON (green) and OFF (pink) HMM states. Right: Behavioral performance improves when
the decision occurs during ON intervals compared to OFF intervals. b: Left: Probability of being in a state representing the
value of the chosen (red), unchosen (blue) and unavailable (gray) option from monkey OFC ensembles in a decision making
task. Right panels: Comparison of time-course of state probability in quick vs. deliberative decisions as judged by the
dynamics of eye movements (depicted at top). c: Left: Dynamics of HMM states decoded from hippocampal neural activity
(top) and position (bottom) of a rat running along a linear track (6 runs; only data during sharp wave ripples, comprising
2% of the session, were used). Right: Mapping latent state probabilities to associated animal positions yields latent-state
place fields describing the probability of each state for every position on the track. (a) panels modified from [14]; (b) panels
modified from [17]; (c) panels modified from [16].
[16]. In rat gustatory cortex, state sequences after stimulus delivery proceed faster when the delivery is expected
compared to when it is not expected [33], which may explain the faster decoding of expected stimuli found in
this area [53] (a spiking network model explanation of this phenomenon is discussed in the next section). The
examples discussed in this subsection make a convincing case that metastable sequences underlie a variety of
important cognitive processes, and more examples will likely be found across different cognitive domains not
yet explored with latent state models.
-- 5 --
112200 ms1 mmTimeChannel numberOn/Off phaseTime relative to the changed stimulus onset (ms)Detection probability-200-150-100-5000.60.70.8 On phase Off phaseaCh 2, UnCh 1Ch 4, UnCh 3040004000400bstate probabilitytime (ms)0.200.300.40ChUnChNAQuick decisionsDeliberationcdef0.250.350.45chosenunchosentime (ms)quick decisiondeliberation0 4000 4000 400 800cposstate20 x 100 ms bins01probabilitystateposition [cm]02501304lsPFs using RUNlatent-state place fields0 0.5 1 3 Spiking network models of metastable activity
In addition to statistical descriptions in terms of HMM and related latent state models, mechanistic accounts of
metastable activity based on biologically plausible networks of spiking neurons have also been given. We shall
mostly focus on efforts that have purposefully combined spiking network modeling with HMM [22, 33, 54, 55],
as we believe that such an approach is the most promising to understand the potential origin and function of
metastable activity in neural circuits. Metastable activity naturally occurs when multiple hidden states are
attractive fixed points of the neural dynamics which are either inherently unstable, or can be destabilized by
internal noise or external perturbations [56 -- 58]. Such fixed points would attract the dynamic trajectories of the
neural activity and force them to linger in their neighborhood for a finite amount of time. It is natural, therefore,
to look for models wherein metastability is caused by the coexistence of multiple attractor states.
3.1 Metastable activity in clustered networks of spiking neurons
Among the class of attractor neural networks with multiple attractor states, spiking network models allow to
build biologically plausible models capable of generating metastable state sequences. Recurrent networks with
strong synaptic connections can produce highly temporally fluctuating activity [59 -- 66], often the signature of
chaotic activity. However, to endow a network of spiking neurons with endogenously generated metastable
states, a specific partition of the network in subpopulations of neurons seems required
[20 -- 23]. The clustered architecture has emerged as an effective way to generate metastability. The models
studied so far comprise excitatory and inhibitory spiking neurons with (typically) the excitatory neurons orga-
nized in clusters, as seems the case in cortical circuits [67 -- 69]. The common feature of these models is that the
average synaptic strength (and/or connectivity) inside clusters is larger than between clusters (Fig. 3a, right).
A mean field analysis of these models [4, 22, 54] shows that multiple attractor states are possible depending on
the average intra-cluster synaptic weight (ICW; Fig. 3b). Above a critical ICW value (dashed green line in Fig.
3b), multiple configurations of the network emerge wherein neurons in clusters tend to occupy several states
with different values of mean firing rates (grey and red lines in Fig. 3b), where the firing rates depend on the
number of active clusters. In an infinite network, those configurations would be stable attractors; but in a finite
network, or in the presence of sufficient external noise, those configurations become metastable. Note that if the
ICW is large enough (e.g., full green line in Fig. 3b), the configuration with no active clusters is unstable even
in the infinite network, and at least one cluster must be active at any given time. It is unclear whether suitable
connectivity patterns exist such that metastable itinerant dynamics is possible also in a deterministic network of
infinite size [70]. In the following, we assume that fluctuations originate in finite size effects in a deterministic
network [20, 22], and later consider alternative models (see subsection Variations on the clustered architecture).
In finite clustered networks above the critical ICW value, a combination of erratic spiking activity and recurrent
inhibition can ignite an itinerant exploration of multiple configurations (as in Fig. 3a, right), whose number and
features depend on network parameters such as the ICW. When an HMM analysis is run on these networks,
segmentation of the spontaneous ongoing activity in discrete states is found (Fig. 3c). Such metastable activity
is modified, but not removed, by an external stimulus. Since the firing rates depend on the number of active
clusters, the neurons exhibit multiple firing rates across different HMM states, as found experimentally in rat
gustatory cortex [22]. Note that the neurons themselves can have graded activity, hence their multistability is
an emergent property of the collective network's dynamics and not an intrinsic property of the units.
3.2 Predictions of the clustered spiking network
Clustered spiking networks capture many statistical and dynamical features of the data, including a stimulus-
induced reduction of: (i) trial-to-trial variability [20, 21, 71], (ii) multistability of the firing rates across states
[22], and (iii) neural dimensionality [54], i.e., the minimal number of effective dimensions (between 1 and the
-- 6 --
Figure 3. Spiking network model of metastable activity . a: Spike rasters of a spiking network model of excitatory (E)
and inhibitory (I) integrate-and-fire neurons with two different architectures. Left: homogeneous network. Right: clustered
network. The homogeneous network has uniform connectivity among the E neurons.
In the clustered network, the E
neurons are organized in clusters, so that synaptic connectivity is stronger inside each cluster than among clusters. The
spiking activity in the clustered network is metastable. Insets: graphical description of the topology of each network. b:
The mean field theory of a clustered spiking network with 30 E clusters shows coexistence of several attractors for intra-
cluster synaptic weight (ICW) J+ beyond the critical value 4.2 (dashed vertical green line; J+ is in units of the baseline
synaptic weights outside clusters). The curves represent the firing rates of neurons in each cluster according to the number
of active clusters (numbers from 1 to 8). Grey and purple curves: E neurons; red curves: I neurons. A state on the purple
curve is "globally inactive" in the sense that no clusters are active and firing rates of E neurons remain low. c: Raster plot
from a typical simulation of the model in (b) with ICW corresponding to the full vertical green line. 30 E neurons (one for
each separate cluster) are shown, together with the segmentation of the rasters in HMM states. Panel (a) adapted from [20];
panel (b) adapted from [54]; panel (c) adapted from [22].
number of neurons) necessary to describe the ensemble dynamics [60, 72, 73]. It is worth noticing that while
the stimulus-driven reduction of trial-to-trial variability has been a catalyst for the development of the clustered
network model [20, 21], the reduction of multistability and dimensionality were found to be natural emergent
properties of the model which led to finding the corresponding properties in the data. These predictions capture
aspects of the data and thus help to corroborate the validity of the model. However, two other recent predictions
link metastability directly with sensory and cognitive functions. The first is that visiting metastable states
during ongoing activity will help to maintain learned stimulus responses and improve performance, suggesting
yet another role for ongoing activity. This prediction emerged from an attempt to understand how neural clusters
and metastable states can emerge during learning, and revealed the need to complement synaptic plasticity with
homeostatic mechanisms [74 -- 76]. The second prediction has led to the proposal that a state of expectation
could be understood as an acceleration of metastable dynamics [33]. When rats are trained to respond to taste
-- 7 --
abJ+10activeInhglobal inactive123456785.5804Mean field theoryFiring rate [Hz]Stimulus afferents0.5 secNeuronsVEP(state)01Fano factorTime [s]-10513Neurons [%]ong.ev.0****DATAMODEL601 sec20 mV1sNeurons130abClustered networkTrial-to-trial variabilityMultistabilitycgfdEnsemble size N0100d110 ong.ev. eDimensionalityICPExcInhJ+10activeInhglobal inactive123456785.5804Mean field theoryFiring rate [Hz]Stimulus afferents0.5 secNeuronsVEP(state)01Fano factorTime [s]-10513Neurons [%]ong.ev.0****DATAMODEL601 sec20 mV1sNeurons130abClustered networkTrial-to-trial variabilityMultistabilitycgfdEnsemble size N0100d110 ong.ev. eDimensionalityICPExcInhbcNeuron012d1 secneuron #neuron #Neuron0121 seccneuron #1 secP(state data)firing rate (spikes/s)INHEXCglobally inac.J+ stimuli, the identity of the stimuli can be decoded faster from the neural activity if stimulus delivery is expected
compared to when stimuli are randomly, and passively, administered [53]. In the model, the onset of a predictive
cue modifies the network's landscape of metastable configurations so as to make state transitions more frequent.
Faster state transitions, in turn, induce faster onset of stimulus-coding states after stimulus delivery, an effect
uncovered by an HMM analysis of the model and confirmed in the data.
3.3 Variations on the clustered architecture
Variations of the clustered network architecture and its generated dynamics can explain related kinds of metastable
activity such as perceptual bistability [77, 78], the emergence of specific slow oscillations known as "up" and
"down" states [79], or networks that can sustain bidirectional sequence propagation at slow and tunable speed
[80]. These types of metastable dynamics are different from those reviewed earlier in subtle, but significant,
ways. Clustered spiking networks have also been used to uncover subtle differences between mechanisms
for decision making. Specifically, spiking networks with multiple attractors built for decision making [7, 81]
can work in two different modes, the "ramping" mode, characterized by gradual stimulus-driven firing rate
increases, and the "jumping" mode [56, 57], wherein the stimulus ignites a sequence of metastable states. By
leveraging a mechanism reminiscent of stochastic resonance, the jumping mode can outperform the ramping
mode (underlying the so-called "integration-to-bound" models [82]) during perceptual decision-making in the
face of sensory noise. Finally, we briefly mention that models with endogenously generated activity are not
necessary to produce metastable activity. External fluctuations or the presence of firing rate adaptation both
can create instability and lead to metastability in a finite network with multiple attractors [83, 84] (including
"up" and "down" states at low firing rate, [85]), and sometimes both ingredients are required to reproduce the
statistics of the empirical data [85, 86]. Also, it is possible to obtain stimulus-evoked quenching of variability
in non-metastable networks, such as networks with chaotic attractors [59, 60, 87] or the supralinear stabilized
network (SSN) [88, 89]. These models capture the reduction in variability via different mechanisms and make
different predictions; for example, the SSN can capture the tuning-dependent modulation of variability observed
in primary visual cortex [90].
4 Conclusions
We have reviewed recent progress made in elucidating the nature and the role of metastable states in brain
function. Thanks to the availability of ensemble data from behaving animals, the study of metastable activity is
changing the way we think about neural coding for both external events and internal representations. We have
also reviewed a class of spiking network models that promise to provide a mechanistic understanding of the
origin and function of metastability. These models have demonstrated the importance of a clustered architecture
for the spontaneous generation of metastability, the coexistence of ongoing and evoked metastability, the bene-
fits of metastable sequences to decision making, and the explanation of faster coding of expected events via the
acceleration of metastable dynamics. Because of their biological plausibility, not only can these models shed
light on the mechanisms, but they can also be powerful tools for analyzing metastable states, predict causal
relationships to behavior, and help to design new experimental studies. In the future, it will be important to
establish a causal relationship between metastable states and behavior, by manipulating neural activity during
behavioral tasks while recording large ensembles across multiple sites. It will also be important to establish
the emergence of neural clusters and metastable dynamics though development or learning. Recurrent spiking
network models, in conjunction with latent-state models such as HMM, are likely to play an important role in
his endeavor.
-- 8 --
Conflict of interest statement
The authors declare no conflict of interest.
Acknowledgments
The authors acknowledge support from the NSF (grant IIS-1161852 to GLC), the NIH (NIH/NIDCD R01DC015234
to AF and NIH/NIDCD K25DC013557 to LM) and the Swartz Foundation (Award 66438 to LM) during the
development of some of the ideas and concepts discussed in this article.
References and recommended reading
Papers of particular interest, published within the period of review, have been highlighted as:
• of special interest
•• of outstanding interest
•• [20] (Litwin-Kumar and Doiron, 2012) This theoretical work demonstrated that a balanced spiking
network with clustered architecture can generate slow metastable activity via internally generated fluctuations.
•• [22] (Mazzucato et al, 2015) This study demonstrates that ongoing activity in rat gustatory cortex
unfolds as a sequence of metastable states that is compatible with the dynamics of a clustered spiking network
model. The model predicts key features of both ongoing and evoked activity, including a stimulus-induced
reduction of multistability in single neurons.
•• [23] (Schaub et al, 2015) This paper presents the link between metatable activity and the eigenvalue
spectrum of the synaptic matrix in a clustered spiking network. When the model is augmented with a hierarchi-
cal architecture, transitions among metatable states at higher level in the hierarchy occur on slower timescales
than transitions among the states at lower level of the hierarchy.
•• [17] (Rich and Wallis, 2016) In the OFC of monkeys, periods of deliberation during a choice task were
associated with sequences of states representing the values (rewards) associated to the options being offered
in the current choice. The longer the presence of one state compared to the others, the quicker the behavioral
decision, regardless of the presumed difficulty of the choice according to the similarity in values. It is also
shown that ensembles of OFC neurons, unlike single neurons, can represent the value of unchosen options.
•• [14] (Engel et al, 2016) In area V4 of behaving monkeys, ON and OFF metastable states are related
to global changes in cortical state (associated with arousal) as well as local selective attention demands, with
better performance when a stimulus change occurred during ON states compared to OFF states.
•• [16] (Maboudi et al, 2018) Spontaneous metastable states observed in CA1 during population burst
events (when the animal is idling) reconstruct a map of place fields evoked during locomotion.
•• [35] (Taghia et al, 2018) An HMM analysis of fMRI data in 122 human subjects performing a working
memory task with multiple task conditions reveals the existence of metastable states preferentially associated
to different task conditions. The hidden states are, in turn, associated to different patterns of functional connec-
tivity across many brain areas.
•• [33] (Mazzucato et al, 2019) This paper shows that expectation induces faster coding of taste stimuli
in rat gustatory cortex by modulating the timescale of ensemble metastable dynamics. A spiking network
model explains the phenomenon in terms of decreased energy barriers between metastable configurations of
the network?s activity. Shallower barriers cause faster transitions to stimulus-coding states identified with an
HMM analysis, accelerating stimulus coding.
-- 9 --
References
[1] A. Bernacchia, H. Seo, D. Lee, and X.-J. Wang, A reservoir of time constants for memory traces in cortical neurons,
Nature neuroscience 14 (2011), no. 3 366.
[2] M. I. Rabinovich, P. Varona, A. I. Selverston, and H. D. Abarbanel, Dynamical principles in neuroscience, Reviews
of modern physics 78 (2006), no. 4 1213.
[3] P. Miller, Dynamical systems, attractors, and neural circuits, F1000Research 5 (2016).
[4] D. J. Amit and N. Brunel, Model of global spontaneous activity and local structured activity during delay periods in
the cerebral cortex, Cereb Cortex 7 (1997), no. 3 237 -- 52.
[5] H. S. Seung, How the brain keeps the eyes still, Proceedings of the National Academy of Sciences 93 (1996), no. 23
13339 -- 13344.
[6] V. Mante, D. Sussillo, K. V. Shenoy, and W. T. Newsome, Context-dependent computation by recurrent dynamics in
prefrontal cortex, nature 503 (2013), no. 7474 78.
[7] X.-J. Wang, Probabilistic decision making by slow reverberation in cortical circuits, Neuron 36 (2002), no. 5
955 -- 968.
[8] X.-J. Wang, Neurophysiological and computational principles of cortical rhythms in cognition, Physiological
reviews 90 (2010), no. 3 1195 -- 1268.
[9] G. Buzsaki, Rhythms of the Brain. Oxford University Press, 2006.
[10] D. Sussillo and L. F. Abbott, Generating coherent patterns of activity from chaotic neural networks, Neuron 63
(2009), no. 4 544 -- 557.
[11] R. Laje and D. V. Buonomano, Robust timing and motor patterns by taming chaos in recurrent neural networks,
Nature neuroscience 16 (2013), no. 7 925.
[12] M. Abeles, H. Bergman, I. Gat, I. Meilijson, E. Seidemann, N. Tishby, and E. Vaadia, Cortical activity flips among
quasi-stationary states, Proc Natl Acad Sci USA 92 (1995) 8616 -- 8620.
[13] L. M. Jones, A. Fontanini, B. F. Sadacca, P. Miller, and D. B. Katz, Natural stimuli evoke dynamic sequences of
states in sensory cortical ensembles, Proc Natl Acad Sci U S A 104 (2007), no. 47 18772 -- 7.
[14] T. A. Engel, N. A. Steinmetz, M. A. Gieselmann, A. Thiele, T. Moore, and K. Boahen, Selective modulation of
cortical state during spatial attention, Science 354 (2016), no. 6316 1140 -- 1144.
[15] A. Ponce-Alvarez, V. Nacher, R. Luna, A. Riehle, and R. Romo, Dynamics of cortical neuronal ensembles transit
from decision making to storage for later report, J Neurosci 32 (2012), no. 35 11956 -- 69.
[16] K. Maboudi, E. Ackermann, L. W. de Jong, B. E. Pfeiffer, D. Foster, K. Diba, and C. Kemere, Uncovering temporal
structure in hippocampal output patterns, eLife 7 (2018) e34467.
[17] E. L. Rich and J. D. Wallis, Decoding subjective decisions from orbitofrontal cortex, Nature neuroscience 19 (2016),
no. 7 973.
[18] M. M. Churchland, J. P. Cunningham, M. T. Kaufman, J. D. Foster, P. Nuyujukian, S. I. Ryu, and K. V. Shenoy,
Neural population dynamics during reaching, Nature 487 (2012), no. 7405 51.
[19] O. Mazor and G. Laurent, Transient dynamics versus fixed points in odor representations by locust antennal lobe
projection neurons, Neuron 48 (2005), no. 4 661 -- 673.
[20] A. Litwin-Kumar and B. Doiron, Slow dynamics and high variability in balanced cortical networks with clustered
connections, Nat Neurosci 15 (2012), no. 11 1498 -- 505.
[21] G. Deco and E. Hugues, Neural network mechanisms underlying stimulus driven variability reduction, PLoS Comput
Biol 8 (2012), no. 3 e1002395.
-- 10 --
[22] L. Mazzucato, A. Fontanini, and G. La Camera, Dynamics of multistable states during ongoing and evoked cortical
activity, The Journal of Neuroscience 35 (2015), no. 21 8214 -- 8231.
[23] M. T. Schaub, Y. N. Billeh, C. A. Anastassiou, C. Koch, and M. Barahona, Emergence of slow-switching assemblies
in structured neuronal networks, PLoS computational biology 11 (2015), no. 7 e1004196.
[24] E. Seidemann, I. Meilijson, M. Abeles, H. Bergman, and E. Vaadia, Simultaneously recorded single units in the
frontal cortex go through sequences of discrete and stable states in monkeys performing a delayed localization task,
J Neurosci 16 (1996), no. 2 752 -- 68.
[25] I. Gat and N. Tishby, Statistical modeling of cell assemblies activities in associative cortex of behaving monkeys, in
Advances in neural information processing systems, pp. 945 -- 952, 1993.
[26] L. R. Rabiner, A tutorial on hidden markov models and selected applications in speech recognition, Proceedings of
the IEEE 77 (1989) 257 -- 286.
[27] R. Durbin, S. R. Eddy, A. Krogh, and G. Mitchison, Biological sequence analysis: probabilistic models of proteins
and nucleic acids. Cambridge university press, 1998.
[28] A. B. Wiltschko, M. J. Johnson, G. Iurilli, R. E. Peterson, J. M. Katon, S. L. Pashkovski, V. E. Abraira, R. P. Adams,
and S. R. Datta, Mapping sub-second structure in mouse behavior, Neuron 88 (2015), no. 6 1121 -- 1135.
[29] W. Zucchini, I. L. MacDonald, and R. Langrock, Hidden Markov models for time series: an introduction using R.
Chapman and Hall/CRC, 2016.
[30] P. Dymarski, Hidden Markov Models: Theory and Applications. BoD -- Books on Demand, 2011.
[31] S. Escola, A. Fontanini, D. Katz, and L. Paninski, Hidden markov models for the stimulus-response relationships of
multistate neural systems, Neural Comput 23 (2011), no. 5 1071 -- 132.
[32] B. F. Sadacca, N. Mukherjee, T. Vladusich, J. X. Li, D. B. Katz, and P. Miller, The behavioral relevance of cortical
neural ensemble responses emerges suddenly, Journal of Neuroscience 36 (2016), no. 3 655 -- 669.
[33] L. Mazzucato, G. La Camera, and A. Fontanini, Expectation-induced modulation of metastable activity underlies
faster coding of sensory stimuli, Nature neuroscience 22 (2019), no. 5 787.
[34] S. W. Linderman, M. J. Johnson, M. A. Wilson, and Z. Chen, A bayesian nonparametric approach for uncovering rat
hippocampal population codes during spatial navigation, Journal of neuroscience methods 263 (2016) 36 -- 47.
[35] J. Taghia, W. Cai, S. Ryali, J. Kochalka, J. Nicholas, T. Chen, and V. Menon, Uncovering hidden brain state
dynamics that regulate performance and decision-making during cognition, Nature communications 9 (2018), no. 1
2505.
[36] C. Stringer, M. Pachitariu, N. Steinmetz, C. B. Reddy, M. Carandini, and K. D. Harris, Spontaneous behaviors drive
multidimensional, brain-wide activity, cell 1 (2019), no. 10 100.
[37] G. Orb´an, P. Berkes, J. Fiser, and M. Lengyel, Neural variability and sampling-based probabilistic representations
in the visual cortex, Neuron 92 (2016), no. 2 530 -- 543.
[38] L. W. Swanson and J. W. Lichtman, From cajal to connectome and beyond, Annual Review of Neuroscience 39
(2016) 197 -- 216.
[39] C. I. Bargmann, Beyond the connectome: how neuromodulators shape neural circuits, Bioessays 34 (2012), no. 6
458 -- 465.
[40] S. Ostojic, N. Brunel, and V. Hakim, How connectivity, background activity, and synaptic properties shape the
cross-correlation between spike trains, Journal of Neuroscience 29 (2009), no. 33 10234 -- 10253.
[41] B. Doiron, A. Litwin-Kumar, R. Rosenbaum, G. K. Ocker, and K. Josi´c, The mechanics of state-dependent neural
correlations, Nature neuroscience 19 (2016), no. 3 383.
-- 11 --
[42] R. Rosenbaum, T. Tchumatchenko, and R. Moreno-Bote, Correlated neuronal activity and its relationship to coding,
dynamics and network architecture, Frontiers in computational neuroscience 8 (2014) 102.
[43] B. Florian, K. Sepp, H. Joshua, and H. Richard, Hidden markov models in the neurosciences, in Hidden Markov
Models, Theory and Applications. IntechOpen, 2011.
[44] A. Arieli, A. Sterkin, A. Grinvald, and A. Aertsen, Dynamics of ongoing activity: explanation of the large variability
in evoked cortical responses, Science 273 (1996), no. 5283 1868 -- 71.
[45] T. Kenet, D. Bibitchkov, M. Tsodyks, A. Grinvald, and A. Arieli, Spontaneously emerging cortical representations of
visual attributes, Nature 425 (2003), no. 6961 954 -- 6.
[46] G. Deco and V. K. Jirsa, Ongoing cortical activity at rest: criticality, multistability, and ghost attractors, J Neurosci
32 (2012), no. 10 3366 -- 75.
[47] A. Luczak, P. Bartho, and K. D. Harris, Spontaneous events outline the realm of possible sensory responses in
neocortical populations, Neuron 62 (2009), no. 3 413 -- 25.
[48] M. Tsodyks, T. Kenet, A. Grinvald, and A. Arieli, Linking spontaneous activity of single cortical neurons and the
underlying functional architecture, Science 286 (1999), no. 5446 1943 -- 6.
[49] D. L. Ringach, Spontaneous and driven cortical activity: implications for computation, Current opinion in
neurobiology 19 (2009), no. 4 439 -- 444.
[50] R. Kiani, C. J. Cueva, J. B. Reppas, and W. T. Newsome, Dynamics of neural population responses in prefrontal
cortex indicate changes of mind on single trials, Current Biology 24 (2014), no. 13 1542 -- 1547.
[51] D. Durstewitz, N. M. Vittoz, S. B. Floresco, and J. K. Seamans, Abrupt transitions between prefrontal neural
ensemble states accompany behavioral transitions during rule learning, Neuron 66 (2010), no. 3 438 -- 48.
[52] E. I. Moser, M.-B. Moser, and B. L. McNaughton, Spatial representation in the hippocampal formation: a history,
Nature neuroscience 20 (2017), no. 11 1448.
[53] C. L. Samuelsen, M. P. Gardner, and A. Fontanini, Effects of cue-triggered expectation on cortical processing of
taste, Neuron 74 (2012), no. 2 410 -- 422.
[54] L. Mazzucato, A. Fontanini, and G. La Camera, Stimuli reduce the dimensionality of cortical activity, Frontiers in
systems neuroscience 10 (2016) 11.
[55] P. Miller and D. B. Katz, Stochastic transitions between neural states in taste processing and decision-making, J
Neurosci 30 (2010), no. 7 2559 -- 70.
[56] P. Miller and D. B. Katz, Stochastic transitions between states of neural activity, Nature neuroscience 14 (2011),
no. 3 366.
[57] P. Miller, Itinerancy between attractor states in neural systems, Current opinion in neurobiology 40 (2016) 14 -- 22.
[58] E. A. Phillips, C. E. Schreiner, and A. R. Hasenstaub, Cortical interneurons differentially regulate the effects of
acoustic context, Cell reports 20 (2017), no. 4 771 -- 778.
[59] H. Sompolinsky, A. Crisanti, and H.-J. Sommers, Chaos in random neural networks, Physical review letters 61
(1988), no. 3 259.
[60] K. Rajan, L. Abbott, and H. Sompolinsky, Stimulus-dependent suppression of chaos in recurrent neural networks,
Physical Review E 82 (2010), no. 1 011903.
[61] S. Ostojic, Two types of asynchronous activity in networks of excitatory and inhibitory spiking neurons, Nature
neuroscience 17 (2014), no. 4 594.
[62] C. Huang and B. Doiron, Once upon a (slow) time in the land of recurrent neuronal networks?, Current opinion in
neurobiology 46 (2017) 31 -- 38.
-- 12 --
[63] O. Harish and D. Hansel, Asynchronous rate chaos in spiking neuronal circuits, PLoS computational biology 11
(2015), no. 7 e1004266.
[64] F. Mastrogiuseppe and S. Ostojic, Intrinsically-generated fluctuating activity in excitatory-inhibitory networks, PLoS
computational biology 13 (2017), no. 4 e1005498.
[65] J. Kadmon and H. Sompolinsky, Transition to chaos in random neuronal networks, Physical Review X 5 (2015),
no. 4 041030.
[66] S. Wieland, D. Bernardi, T. Schwalger, and B. Lindner, Slow fluctuations in recurrent networks of spiking neurons,
Physical Review E 92 (2015), no. 4 040901.
[67] S. Song, P. J. Sjostrom, M. Reigl, S. Nelson, and D. B. Chklovskii, Highly nonrandom features of synaptic
connectivity in local cortical circuits, PLoS biology 3 (2005), no. 3 e68.
[68] R. Perin, T. K. Berger, and H. Markram, A synaptic organizing principle for cortical neuronal groups, Proceedings
of the National Academy of Sciences (2011) 201016051.
[69] R. Kiani, C. J. Cueva, J. B. Reppas, D. Peixoto, S. I. Ryu, and W. T. Newsome, Natural grouping of neural responses
reveals spatially segregated clusters in prearcuate cortex, Neuron 85 (2015), no. 6 1359 -- 1373.
[70] B. Doiron and A. Litwin-Kumar, Balanced neural architecture and the idling brain, Frontiers in computational
neuroscience 8 (2014) 56.
[71] M. M. Churchland, B. M. Yu, J. P. Cunningham, L. P. Sugrue, M. R. Cohen, G. S. Corrado, W. T. Newsome, A. M.
Clark, P. Hosseini, B. B. Scott, D. C. Bradley, M. A. Smith, A. Kohn, J. A. Movshon, K. M. Armstrong, T. Moore,
S. W. Chang, L. H. Snyder, S. G. Lisberger, N. J. Priebe, I. M. Finn, D. Ferster, S. I. Ryu, G. Santhanam, M. Sahani,
and K. V. Shenoy, Stimulus onset quenches neural variability: a widespread cortical phenomenon, Nat Neurosci 13
(2010), no. 3 369 -- 78.
[72] R. C. Williamson, B. R. Cowley, A. Litwin-Kumar, B. Doiron, A. Kohn, M. A. Smith, and M. Y. Byron, Scaling
properties of dimensionality reduction for neural populations and network models, PLoS computational biology 12
(2016), no. 12 e1005141.
[73] P. Gao, E. Trautmann, M. Y. Byron, G. Santhanam, S. Ryu, K. Shenoy, and S. Ganguli, A theory of multineuronal
dimensionality, dynamics and measurement, bioRxiv (2017) 214262.
[74] G. K. Ocker, A. Litwin-Kumar, and B. Doiron, Self-organization of microcircuits in networks of spiking neurons with
plastic synapses, PLoS computational biology 11 (2015), no. 8 e1004458.
[75] A. Litwin-Kumar and B. Doiron, Formation and maintenance of neuronal assemblies through synaptic plasticity,
Nat Commun 5 (2014) 5319.
[76] F. Zenke, E. J. Agnes, and W. Gerstner, Diverse synaptic plasticity mechanisms orchestrated to form and retrieve
memories in spiking neural networks, Nature communications 6 (2015) 6922.
[77] R. Moreno-Bote, J. Rinzel, and N. Rubin, Noise-induced alternations in an attractor network model of perceptual
bistability, Journal of neurophysiology 98 (2007), no. 3 1125 -- 1139.
[78] R. Cao, A. Pastukhov, M. Mattia, and J. Braun, Collective activity of many bistable assemblies reproduces
characteristic dynamics of multistable perception, Journal of Neuroscience 36 (2016), no. 26 6957 -- 6972.
[79] H. Setareh, M. Deger, C. C. Petersen, and W. Gerstner, Cortical dynamics in presence of assemblies of densely
connected weight-hub neurons, Frontiers in computational neuroscience 11 (2017) 52.
[80] H. Setareh, M. Deger, and W. Gerstner, Excitable neuronal assemblies with adaptation as a building block of brain
circuits for velocity-controlled signal propagation, PLoS computational biology 14 (2018), no. 7 e1006216.
[81] X.-J. Wang, Decision making in recurrent neuronal circuits, Neuron 60 (2008), no. 2 215 -- 234.
-- 13 --
[82] R. Bogacz, E. Brown, J. Moehlis, P. Holmes, and J. D. Cohen, The physics of optimal decision making: a formal
analysis of models of performance in two-alternative forced-choice tasks., Psychological review 113 (2006), no. 4
700.
[83] M. Giugliano, G. La Camera, S. Fusi, and W. Senn, The response of cortical neurons to in vivo-like input current:
theory and experiment: Ii. time-varying and spatially distributed inputs, Biol Cybern 99 (2008), no. 4-5 303 -- 18.
[84] E. Russo and A. Treves, Cortical free-association dynamics: Distinct phases of a latching network, Physical Review
E 85 (2012), no. 5 051920.
[85] D. Jercog, A. Roxin, P. Bartho, A. Luczak, A. Compte, and J. De La Rocha, Up-down cortical dynamics reflect state
transitions in a bistable network, Elife 6 (2017) e22425.
[86] A. Shpiro, R. Moreno-Bote, N. Rubin, and J. Rinzel, Balance between noise and adaptation in competition models
of perceptual bistability, Journal of computational neuroscience 27 (2009), no. 1 37.
[87] A. Crisanti and H. Sompolinsky, Path integral approach to random neural networks, Physical Review E 98 (2018),
no. 6 062120.
[88] D. B. Rubin, S. D. Van Hooser, and K. D. Miller, The stabilized supralinear network: a unifying circuit motif
underlying multi-input integration in sensory cortex, Neuron 85 (2015), no. 2 402 -- 417.
[89] Y. Ahmadian, D. B. Rubin, and K. D. Miller, Analysis of the stabilized supralinear network, Neural computation 25
(2013), no. 8 1994 -- 2037.
[90] G. Hennequin, Y. Ahmadian, D. B. Rubin, M. Lengyel, and K. D. Miller, The dynamical regime of sensory cortex:
stable dynamics around a single stimulus-tuned attractor account for patterns of noise variability, Neuron 98 (2018),
no. 4 846 -- 860.
-- 14 --
|
1503.05485 | 1 | 1503 | 2015-03-18T16:58:26 | Petri Net Modeling of the Brain Circuit Involved in Aggressive Behavior | [
"q-bio.NC"
] | The purpose of this work in to demonstrate the initial results of a research project having as its goal to develop dynamic models of the brain network involved in aggressive behavior. In this way, the complex neural process correlated to basic anger emotions and resulting in aggressive behaviors is purportedly schematized by the use of Petri nets, a work-flow computational tool. Initially, the modeling technique is introduced taking into account the most recent and accepted notion of the neural substrates of emotions, particularly the brain structures involved in the aggression neural network, including their inputs, outputs, and internal connectivity. In order to optimally represent these structures, their connections, and temporal dynamics, the Petri net foundations employed in the simulation theory are defined. Finally, the model and dynamic simulation of the neural process associated with aggressive behavior is presented and evaluated as a feasible in silico experimental support of the Patterned-Process Theory. | q-bio.NC | q-bio | 1
Petri Net Modeling of the Brain Circuit Involved in
Aggressive Behavior
Modelo descriptivo del circuito cerebral involucrado en la
agresión mediante redes de Petri
Johann Heinz Martínez Huartos
Facultad de Administración, GIPE, NEUROS, Universidad del
Rosario. Bogotá Colombia.
Dirección Correspondencia: Dpt. Ciencia y Tecnología Aplicadas a
la I.T.A. E.U.I.T. Agrícola, UPM Ciudad Universitaria s/n 28040
Madrid, España. Tel: 91 336 3720
[email protected]
Carlos B. Moreno
Escuela de Medicina y Ciencias de la Salud, Grupo NEUROS
Universidad del Rosario. Bogotá Colombia.
[email protected]
José Luis Díaz
Departamento de Historia y Filosofía de la Medicina,
Facultad de Medicina, Universidad Autónoma de México.
México, D.F., México. [email protected]
Resumen
2
el
complejo
proceso
nervioso
El propósito de este trabajo es mostrar los resultados
iniciales de una línea de investigación que tiene como
objetivo desarrollar modelos dinámicos del circuito cerebral
implicado en la conducta agresiva y sus estructuras
asociadas. De esta forma, se pretende esquematizar por medio
de redes de Petri (una herramienta computacional tipo “work
flow”)
usualmente
correlacionado con emociones de rabia y que desemboca en
comportamiento agresivo. En primer término se presenta la
técnica de modelamiento tomando en cuenta la noción más
actual y aceptada del substrato neurobiológico de las
emociones, en especial de las estructuras involucradas en el
circuito de la agresión, sus entradas, salidas y
conectividad. A continuación se muestran los fundamentos
operativos para la posible simulación de sistemas dinámicos
basada en redes de Petri que aquí se utilizan para modelar
esas estructuras, sus conexiones y su dinámica en el tiempo.
Finalmente, se presenta el modelo resultante y la simulación
del proceso nervioso asociado con la conducta agresiva, lo
cual se propone que otorga un sustento experimental in
silico a la Teoría de los Procesos Pautados.
Palabras claves: Modelos de redes neurales, Redes de Petri
Sistemas dinámicos, Sustrato de la conducta agresiva.
3
Summary
The purpose of this work in to demonstrate the initial
results of a research project having as its goal to develop
dynamic models of the brain network involved in aggressive
behavior. In this way, the complex neural process correlated
to basic anger emotions and resulting in aggressive
behaviors is purportedly schematized by the use of Petri
nets, a work-flow computational tool. Initially, the
modeling technique is introduced taking into account the
most recent and accepted notion of the neural substrates of
emotions, particularly the brain structures involved in the
aggression neural network, including their inputs, outputs,
and internal connectivity. In order to optimally represent
these structures, their connections, and temporal dynamics,
the Petri net foundations employed in the simulation theory
are defined. Finally, the model and dynamic simulation of
the neural process associated with aggressive behavior is
presented and evaluated as a feasible in silico experimental
support of the Patterned-Process Theory.
Keywords: Aggressive behavior substrate, Dynamic systems,
Neural network modeling, Petri nets
Introducción
4
El término agresión designa a un conjunto de conductas
nocivas dirigidas por individuos usualmente incitados por la
emoción básica de enojo, rabia o furia hacia uno o varios
receptores que amenazan con producir o de hecho provocan
temor, dolor y lesión. Son conductas interactivas y sociales
que en contextos agonistas de confrontación se manifiestan
como amenazas, embates, persecuciones, prendimientos,
golpes, laceraciones o mordidas y que pueden ser respondidos
de formas tan diversas como defensa, contraataque, pelea,
lamento, protesta, sumisión, derrota, evasión o huida 1.
Dado su destacado papel en la territorialidad, la selección
y competencia sexual, la dominancia, o el control de los
recursos, la agresión es una conducta relevante en el
proceso evolutivo de la mayoría de las especies animales y
su desarrollo ha contribuido a la adaptación y preservación
de la especie humana 2,3. La agresión se ha estudiado
ampliamente en diversas disciplinas, desde aproximaciones
etológicas tanto cualitativas como cuantitativas en
diferentes especies 3,4, hasta estudios anatómicos y
funcionales de las estructuras cerebrales implicadas en su
expresión 5,6,7,8,9,10. El hallazgo y modelaje de los sistemas
cerebrales involucrados en este tipo de comportamientos ha
constituido una ruta de investigación creciente y promisoria
5
que
permite
similar
para comprender mejor las bases cerebrales de la conducta 10,
11, 12, 13, 14, 15, 16, 17, 18.
Entre las propuestas actuales para abordar la generación de
los correlatos cognitivos y cerebrales del comportamiento
está la teoría de los procesos pautados 19. Esta teoría
permite modelar el correlato cerebral de la agresión pues
propone herramientas epistemológicas y metodológicas para
establecer una relación funcional y dinámica entre los
procesos cerebrales, conductuales y mentales. Según la
teoría, estos tres fenómenos comparten una morfología
dinámica
utilizando
herramientas matemáticas de los sistemas complejos. Esto es
así porque un proceso pautado implica una transición de
estados en el espacio y en el tiempo que transporta y
transforma información relevante mediante un desarrollo
funcional entre las estructuras o condiciones involucradas.
La identificación de los elementos constitutivos del
sistema, el cambio de estados y el transporte de información
en un sistema abierto de tipo red permite modelarlos con
herramientas funcionales y representarlos en diagramas
dinámicos. En el caso del sistema nervioso, las pautas
inter-modulares de actividad cursarían entre los módulos
cerebrales de una forma semi-ordenada y se acoplarían de una
modelarlos
6
manera lo suficientemente compleja como para que ocurra un
correlato mental.
La herramienta propuesta inicialmente para modelar los
procesos pautados se denomina redes de Petri, grafos bi-
partitos que sirven para hacer análisis tanto empíricos como
teóricos de sistemas complejos concurrentes 20. Las redes de
Petri tienen dos tipos de nodos, llamados lugares y
transiciones. Los lugares o plazas P, se representan como
círculos separados por una transición T, que se representa
por líneas rectas cortas o rectángulos; los lugares se unen
a las transiciones por medios de vectores o Arcos. Este tipo
de gráficos tienen además la capacidad de almacenar en sus
lugares fichas o tokens K, marcadas como puntos rellenos que
constituyen la parte dinámica que simula las actividades
concurrentes en un sistema dinámico (Figura 1)21. La dinámica
temporal de la red se procesa mediante “disparos” por medio
de los cuales los tokens ubicados en una plaza o nodo se
transportan a otra plaza a través de una transición marcada
entre ellas. Este comportamiento dinámico es particularmente
relevante y ajustable a un circuito nervioso en el cual la
información pasa de un módulo a otro mediante las vías
aferentes y eferentes de conexión entre los módulos
involucrados de tal forma que, los tokens transferidos de
una plaza a otra representan a la información nerviosa
7
transmitida entre los módulos representados por las plazas y
al mismo tiempo el estado de activación de cada plaza entre
los disparos de la red.
Las redes de Petri se usan para visualizar procesos de
distribución de información y comunicación 22. Además, es
posible obtener ecuaciones que gobiernen transiciones de
estado de un sistema 23, 24. Las reglas para la dinámica de
las redes de Petri se resumen en tres: (1) Una transición t
está habilitada si todos los lugares de entrada tienen al
menos un token; (2) una transición habilitada se puede
disparar moviendo un token de cada lugar de entrada y
poniéndolo en algún lugar de salida y (3) cada disparo
genera un nuevo estado y por ende un cambio en el vector
marcador o de estados del sistema en una red.
A partir de estos desarrollos, de los modelos de sistemas
dinámicos y de la teoría de procesos pautados, se pretende
en el presente trabajo modelar la dinámica del sistema de
núcleos cerebrales involucrados en la conducta agresiva
utilizando redes de Petri. El trabajo es parte de un
proyecto que tiene como objetivo realizar una simulación
computacional que aplique la teoría de los procesos pautados
a sistemas cerebrales que tienen correlatos tanto
conductuales y expresivos como cognitivo-emocionales.
8
Materiales y métodos
Correlato nervioso de la agresión y modelo
funcional
Para especificar el sistema cerebral involucrado en la
génesis de la conducta agresiva se utilizó el modelo
propuesto por Nelson y Trainor 25 (Figura 2a). En este modelo
la actividad del circuito se origina en la corteza cerebral
en respuesta a diversos estímulos del medio exterior que
llegan a las zonas sensoriales primarias. Estas zonas sirven
a su vez como eferentes de una información que se bifurca
para alcanzar a la corteza órbito-frontal (OFC) y al núcleo
medial de la amígdala (Ame). Posteriormente, el Ame envía
señales hacia el hipotálamo anterior (HA) y el núcleo del
lecho de la estría terminalis (BNST), para que, a su vez,
estos envíen información a la sustancia gris periacueductal
(SGPA). El sistema incluye una señal inhibitoria que es
enviada desde COF a AMe para regular a este último núcleo.
(Figura 2a).
Con base en este modelo del sistema cerebral de la agresión
se realizó un modelo funcional primario constituido por
nodos con entradas y salidas (figura 2b) que especifican las
vías que conectan los diferentes núcleos. Dado que se trata
9
de un modelo funcional no se tomaron en cuenta la morfología
de las estructuras cerebrales involucradas o sus distancias
anatómicas; en cambio, se consideró cardinal el intercambio
encauzado de información excitatoria o inhibitoria realizado
por las vías aferentes y eferentes de conexión entre dichas
estructuras.
En el modelo funcional desarrollado (figura 3) se incluyen
dos nodos adicionales que representan las señales de
activación (Activ) y salida o expresión (Expre) del sistema.
Las líneas dobles recortadas que entran y salen de las
estructuras internas del circuito reflejan los eventos o
señales que ingresan al cerebro y salen de éste para
expresarse en los diversos comportamientos agresivos. El
conector del nodo COF al nodo AMe simula en la red de Petri
la conexión inhibitoria que ocurre según el modelo de Nelson
y Trainor 25.
En la Tabla 1 se comparan las características de las redes
de Petri y las condiciones del sistema cerebral definido
para la agresión. La primera columna muestra las
características básicas de una red de Petri. En la segunda
columna se da una breve explicación de la característica
asociada de la red Petri para ser acorde con el
comportamiento neurobiológico.
10
Modelación del sistema cerebral de la agresión
mediante una red de Petri
La simulación fue hecha con el software de implementación de
redes de Petri (PetriNet Simulator by Biogenerics. Inc.) y
los datos fueron analizados con un software especializado
(Mathematica 7.0, Wolfram Research).
De acuerdo con la teoría de procesos pautados, los núcleos o
módulos cerebrales se representan como los lugares o plazas
P de la red pues representan los sitios que generan las
activaciones o inhibiciones de cada nodo o estructura
cerebral. El conjunto de conexiones sinápticas entre dos
estructuras se representa por las transiciones T. Los tokens
son las unidades de transición entre nodos y representan
paquetes de información de la red neuronal de cada núcleo en
forma de una señal específica sea de activación o de
inhibición. Para modelar adecuadamente el circuito y contar
con suficientes tokens, se planteó una red de capacidad
infinita.
De acuerdo con el procesamiento funcional que normalmente
ocurre en redes neuronales in vivo, la información de tipo
concurrente y la retroinformación entre estructuras son
también factores necesarios para que funcione el circuito
artificial. De esta forma, a continuación se detallan
algunas variables funcionales del modelo que se eligieron
11
para emular de manera verosímil el funcionamiento del
circuito de la agresión. Además de los módulos básicos que
constituyen el circuito, se introdujeron en el modelo dos
plazas de input (ACTIV1 y ACTIV2) que representan a las
cortezas sensoriales que inician el procesamiento nervioso
que desemboca en el comportamiento agresivo. La plaza
ACTIV1, por ejemplo, es generatriz de un token en lapsos de
tiempos aleatorios, mientras que ACTIV2 es generatriz de
cantidades aleatorias de tokens en lapsos de tiempos
igualmente aleatorios. El objetivo es emular los ingresos de
información en forma de estímulos relevantes que pueden
acceder al cerebro en cualquier momento.
La figura 4 muestra en líneas no punteadas las conexiones
que van de los núcleos de activación hacia COF; por lo tanto
representan a las vías que convergen en la corteza órbito
frontal. Los módulos de entrada ACTIV1 y ACTIV2 envían
información a AMe (Figura 5). Además, COF tiene influencia
sobre AMe por medio del comparador TTL. Este comparador TTL
es un módulo ficticio requerido en el modelo (plazas de
recurso), usado para emular la inhibición de COF sobre AMe
mediante la comparación binaria de tokens en el tiempo. Es
decir, los tokens que modelan la información nerviosa
transmitida entre los módulos representados deben primero
pasar por la corteza orbito frontal y, dados los valores de
12
las latencias temporales en este nodo, son enviados luego a
AMe. Las latencias entre las estructuras COF y AMe generan
retrasos y por tanto desfases temporales entre los tokens de
los diferentes núcleos para simular las inhibiciones que COF
tiene sobre AMe y así poder modelar la concurrencia de la
información dentro de la red. En general, los tokens
enviados de COF a AMe y que simulan el control de la corteza
frontal sobre la amígdala son a su vez controlados por los
retardos que tienen lugar en el nodo controlador (TTL).
En general, las características funcionales de los nodos del
circuito se basan en datos neurofisiológicos 4, 25, excluyendo
los nodos no operacionales (ficticios, pero requeridos para
emular el sistema) que sirven como ajustes verosímiles
realizados en el modelo: nodos de activaciones, el
(o
comparador
reverberación) y de expresión.
Entre las características del modelo se incluye una conexión
entre la corteza orbitofrontal (COF) y el hipotálamo
anterior (HA) que se ha descrito en humanos 4, 18, 25 (figura
6). Como se explica a continuación, la plaza TTL es de tipo
recurso pues está asociada con tiempos de cada una de las
estructuras COF y AMe que permiten las latencias entre éstas
(figura 7).
recaptación
funciones
TTL,
las
de
13
El circuito reverberante en este modelo tiene la opción de
tomar alguno de los tres valores (-1, 0, 1), que muestran
los niveles de activación de TTL, la cual es alimentada
continuamente por COF y AMe. Estos valores representan un
retardo, concurrencia y anticipo de tokens de la siguiente
manera: -1 si un token que viene de OFC tiene un retardo
respecto a un token que venga de MeA; 0 para el flujo normal
del token en la red a través de ese nodo o para la
concurrencia espacio-temporal de tokens y 1 si un token que
viene de OFC esta adelantado respecto a un token que venga
de MeA.
La plaza de control Recap (figura 8) representa el proceso
de recaptura de tokens en una parte del circuito que permite
alimentar continuamente mecanismos de reverberación. El
grafo autocíclico representa la recaptura de tokens en la
transición T1 que integra la información que va desde ACTIV2
y TTL hasta AMe. Así, en aras de evitar pérdidas de
información, el nodo Recap ayuda a mantener la información
en forma de tokens dentro de toda la red y se ubicó en T1
porque a esta transición llega la información de todas las
activaciones iniciales.
Luego que la información ha pasado por el circuito
reverberante, al llegar a AMe es dirigida a las estructuras
LNET y HA (figura 9). Es indispensable que las activaciones
14
de estos núcleos (representadas por los tokens), se den en
forma concurrente y en consecuencia se dispara la transición
asociada con LNET y HA. De esta manera el proceso de
recaptura es usado en el diseño del experimento
computacional de forma global para alimentar continuamente
el circuito reverberante y poder mantener estable el flujo
de información en la red.
La activación de T5 envía tokens a SGPA que, al activarse,
genera finalmente la conducta agresiva. Cada una de las
plazas tiene tiempos de simulación diferentes de forma que
la generación de la conducta agresiva depende de que el
envío de información a SGPA sea de forma concurrente y
sincrónica, en lapsos de tiempo muy cortos e independientes
de los tiempos asociados con cada plaza. Finalmente, la
información pasa desde SGPA hasta la representación de la
plaza asociada al sistema motor EXPRE (figura 10). Lo
anterior se asume bajo la consideración que la expresión de
conductas agresivas involucra movimientos voluntarios,
pautas de actividad motriz y acciones coordinadas por lo que
debe
piramidales,
extrapiramidales y cerebelosos. El módulo EXPRE se
identifica con dichos sistemas motores de manera general y
no únicamente con la corteza motora, aunque en adelante
hagamos referencia solo a esta.
involucrar
motores
a
sistemas
15
Criterios operativos de la red
En el algoritmo de simulación los tiempos promedios de
respuesta y los tipos de conexiones entre núcleos se tomaron
como constantes globales. La entrada del programa permite
variar el tiempo total para la simulación en la red. Los
datos arrojados o “salidas” del programa son datos numéricos
del número de tokens en cada uno de los núcleos neuronales
que muestran la relación de la activación en la dinámica de
cada uno de los núcleos que conforman la red.
El modelo final requirió de latencias temporales asociadas
con algunas plazas y tiempos de latencia relacionados con
las activaciones y disparos de COF y AMe en el circuito
reverberante, para simular por medio de retardos las
inhibiciones de COF sobre AMe y su acción sobre HA. Se hizo
una regresión lineal para efectos de comparación entre los
tiempos biológicos de transferencia de información y los
tiempos de simulación. El promedio ponderado de la pendiente
de la regresión generó un factor de escalamiento temporal
para obtener la escala de tiempo. De la simulación se
extrajeron datos de la activación dada por la cantidad de
tokens en un tiempo específico de simulación en cada uno de
los núcleos representados por las plazas. Se graficaron
estos datos para ver el comportamiento de cada núcleo y para
16
observar el fenómeno de la concurrencia en ejes de
activación contra tiempo simulado.
El factor de escalamiento temporal
Dadas estas características se consideraron velocidades de
conducción entre 0.5m/s y 100m/s, lo que conlleva a tiempos
entre 2s y 0.01s respectivamente, que es el rango de los
tiempos obtenidos en experimentos neurofisiológicos y
denominados en este trabajo como rango de tiempos reales de
conducción. A su vez, para los tiempos de simulación se
usaron intervalos entre 1ms y 30ms asociados con algunas
estructuras neuronales, a saber: 30ms para COF, 20ms para
AMe, 5ms para HA, y 1ms para LNET. En la simulación estos
tiempos se usan para retener y procesar la información
correspondiente a los tokens en cada una de estas
estructuras. De la misma forma, para la simulación del
circuito reverberante se tuvieron en cuenta los tiempos de
latencia generados entre las estructuras COF y AMe mediante
el comparador, es decir, los valores que determinan cuánto
tiempo en ms le toma al comparador TTL dejar de estar
activado después de finalizar sus tareas u operaciones. Por
ejemplo, al comparador TTL le toma 5 ms inactivarse después
de haber estado enviando información a COF y 10 ms
inactivarse después de enviar información a AMe. Esto está
17
de acuerdo con el planteamiento inicial de la red, en la
cual COF debe estar menos tiempo inactivo que AMe para poder
controlar su activación y esto último se refiere al control
fisiológico que COF mantiene sobre AMe 25.
La necesidad de correlacionar los tiempos usados en la
simulación con el rango de tiempos reales de conducción de
información neuronal tiene como base la relación dada por la
siguiente expresión:
𝑡!=𝑛𝑡!
donde n representa el factor de escalamiento entre los
tiempos de simulación (ts) y los tiempos reales de
conducción (tr).
A continuación se normalizaron las unidades de los tiempos
en ms haciendo estimaciones de la escala lineal entre ellos.
La gráfica de la figura 11 se generó a partir de los datos
de los tiempos de simulación vs los tiempos reales de
conducción a los que se les hizo una regresión lineal para
obtener el factor de escalamiento entre el tiempo de
simulación y el tiempo real. Se obtuvo la relación lineal de
la siguiente ecuación:
𝑡!=0.0147938 𝑡!+0.447731 (𝑚𝑠)
18
y el factor de escalas temporales n con el valor asociado en
la siguiente igualdad:
𝑛=0.0147938
que representa la tasa a la que aumentan los tiempos de
simulación con respecto a los tiempos reales del sistema
biológico, es decir, ∆𝑡!≅0.01 ∆𝑡!. La tasa muestra que la
variación de los pasos computacionales es aproximadamente
una centésima de la variación de los tiempos biológicos
reales.
Resultados
Dinámica de la red
Las gráficas siguientes corresponden a dos variables
definidas como la cantidad de tokens en una estructura
específica (estado de activación), representada en el eje
vertical y los pasos de la simulación (paso computacional
marcado por los “disparos” de la red). La simulación se hizo
en un tiempo de 20000 ms a pasos de 20 ms para un total de
1000 repeticiones, -en adelante iteraciones-, en 20 segundos
(figura 12).
La gráfica muestra las activaciones de las plazas Activ 1 y
Activ 2 que tienen una función generatriz de tokens de tipo
19
aleatorio y periódico, respectivamente. Así, mientras la
plaza Activ 1 tenía una función generadora de tokens de tipo
aleatorio en las iteraciones, la plaza Activ 2 tenía una
función generadora de tipo aleatorio tanto en las
iteraciones como en la cantidad de tokens que pudiese
generar. De esta forma se representaron los ingresos de
información al circuito de la agresión pues, como se sabe,
el cerebro está sometido a una infinidad de estímulos que
ingresan con un patrón posible al propuesto en la
simulación.
Las activaciones de las plazas input envían información
conjuntamente a las plazas operacionales que representaban
los núcleos COF y AMe los que, dentro de un circuito
reverberante, presentan activaciones en términos de las
iteraciones (figura 13).
La dinámica de las activaciones en términos de las
iteraciones de la simulación muestra una baja probabilidad
de activaciones simultáneas en COF y AMe. De hecho, las
activaciones de COF producen caídas en las activaciones de
AMe lo que genera un comportamiento de tipo pulsátil entre
estas dos estructuras. Las activaciones de COF impiden las
de AMe, que es el módulo encargado de enviar información a
las demás estructuras, pero solo si COF lo permite.
20
Las activaciones e inhibiciones de AMe mediadas directamente
por COF son el resultado de las interacciones de la
información dentro del circuito reverberante. Para el diseño
del circuito reverberante se tuvo en cuenta el efecto de la
recaptura de información, es decir de la información
recobrada por reverberación, y el efecto de comparación de
información entre núcleos neuronales.
Cabe resaltar que la plaza que representa la recaptación o
recobro de información (figura 14), y la plaza que
representa el comparador TTL no son operacionales pues no
son estructuras del circuito de agresión, sino plazas de
control y de recurso respectivamente. El comparador TTL
mantuvo tres niveles de activación predeterminados donde sus
valores de 1, 0, −1 se asocian con niveles de uso. De esta
manera, como salida de la simulación y en un rango de 1000
iteraciones en 20 segundos el porcentaje de utilización del
comparador fue de 54.2%, lo que indica que la función de
comparar información entre módulos es fundamental para
generar la concurrencia necesaria para que se logre activar
la conducta agresiva.
Retomando la idea funcional del circuito propuesto de la
agresión, diversos estímulos externos llegan a las cortezas
auditiva o visual enviando información a la corteza orbito
frontal y a la amígdala media. Estas estructuras
21
intercambian información mediante un circuito reverberante
capaz de comparar información, lo cual genera latencias
entre la información que viaja desde COF hasta AMe y HA,
inhibiendo a estas estructuras e impidiendo la generación de
la conducta agresiva. Así, la clave para que se exprese la
conducta de agresión por SGPA depende de la concurrencia de
la información en las estructuras LNET y HA, pues no es
suficiente la activación de solamente una de ellas (figura
15).
La figura 15 muestra, de arriba hacia abajo las activaciones
en términos de los pasos de computo de LNET, HA, SGPA
divididas en una magnificación de las mismas en la parte
izquierda y su escala normal en la derecha. Es decir: los
óvalos que encierran las regiones de las activaciones de la
derecha son magnificadas en la izquierda para mostrar la
concurrencia espacio-temporal de tokens. El óvalo superior
proyecta en la izquierda un evento exacto en el que LNET con
cuatro tokens de activación concurre con una activación de
un token en HA a los 351 ms. El evento previo de
concurrencia en las plazas anteriores permite que la
transición T5 se active y dé paso exactamente un milisegundo
después a la activación de SGPA con un token. Este token
viajará después por medio de la transición T6 al nodo EXPRE,
el cual refleja la producción de la conducta agresiva. Se
22
hace énfasis del momento preciso en el que la activación de
LNET concurre con la activación de HA, dando paso a la
activación de SGPA. De esta forma se pone de manifiesto que
solo un proceso pautado debido a envíos de información por
estructuras cerebrales jerarquizadas en tiempo discreto, en
lapsos cortos y en forma concurrente puede generar la
conducta de agresión.
Discusión
La red de Petri usada para simular el circuito de agresión
es una red temporal discreta de tipo aleatorio-periódico,
dotada además de un grafo autocíclico. Es una red ordinaria
por tener todos los pesos de sus arcos con un valor de 1 y
de capacidad infinita debido a que la cantidad de tokens que
se pueden almacenar en cada nodo no tiene un límite
estipulado.
Aunque el algoritmo usado está de acuerdo con el sistema
fisiológico ya que toma en cuenta sus principales
características, el papel funcional que tiene cada una de
las estructuras del circuito relacionado con la conducta
agresiva es incierto ya que ninguna de ellas se involucra
sólo en este comportamiento, sino en muchas otras funciones
del cerebro. De forma aledaña, la dinámica del circuito
correlacionado con la conducta agresiva se puede considerar
de tipo emergente puesto que el sistema está compuesto de
23
un
tipo
de
revela
información
múltiples partes interconectadas en las que la información
adicional resultante se oculta al observador. En efecto, la
simulación
que
estructuralmente no era evidente. El comportamiento
emergente se debe a que a partir de sus interacciones la red
muestra una dinámica que no se podría explicar aislando sus
partes. La heterogeneidad en las conexiones, que es una de
las características de las redes complejas, se evidencia en
la presencia de nodos con muchas, pocas y medianas
conexiones. El comportamiento emergente requiere que las
entradas y salidas de la información en el sistema sean
comandadas por funciones lógicas, algunas de las cuales
pueden ser descritas y otras solamente se evidencian por el
comportamiento de la red entera y no por la causalidad
lineal de elementos aislados del sistema. De esta manera, el
circuito neurobiológico de agresión modelado en este trabajo
es un sistema complejo debido a que está en continuo
intercambio de energía o de información con el entorno,
posee un comportamiento no lineal e imprevisible y tiene
propiedades emergentes. La generación final de la conducta
agresiva aparece como resultado de un proceso holístico y
jerárquico porque se demuestra que la generación de la
expresión de la conducta de agresión es llevada a cabo
24
después de procesamientos de información en los diferentes
niveles.
De acuerdo con las evidencias de la neurociencia, este
modelo incorpora diversos circuitos de comunicación entre
módulos o cúmulos neuronales. Hay circuitos lineales como la
conexión entre AMe y el núcleo del lecho de la estría
terminal y circuitos convergentes en un módulo debido a la
actividad eferente de diferentes núcleos. Así, no se puede
generar la conducta de agresión sólo desde la plaza SGPA si
no concurren los impulsos de LNET y de HA al mismo tiempo en
la dinámica de la red. Se observa, también, un circuito
reverberante que permite que la información de un núcleo de
neuronas module la información que pasará a un posterior
grupo de neuronas al salir de este circuito.
El estado final del sistema modelado surge tanto de las
relaciones dentro la red neuronal como de las estructuras
involucradas y sus dependencias temporales a partir de
estímulos que ingresan del exterior. El modelo también
permite validar el efecto de un macroestado computacional 19
que se supone subyace a la cognición y a la conducta
organizada en función de las relaciones de tiempo y de
activaciones en curso de la red neuronal 26. Los resultados
de esta investigación permiten afirmar que la generación de
la conducta de agresión evidencia un comportamiento complejo
25
que, simulado mediante redes de Petri, representa una
variedad de sistemas dinámicos concurrentes.
El modelo presentado en esta investigación contribuye a
validar la teoría de procesos pautados mediante la
simulación del circuito nervioso de agresión debido a que su
dinámica muestra secuencia de pautas, transformación de
información, actividad cinética y periodicidad entre estados
nerviosos en secuencia. Según la teoría, los correlatos
cognitivos o conductuales del proceso cerebral ocurren por
una serie de procesos y eventos discretos y con lapsos de
tiempo cortos que se presentan y detectamos como si los
eventos fueran continuos 19. El estudio interdisciplinario de
pautas
formación,
características y evolución en el sistema nervioso desde la
perspectiva de la teoría de la complejidad puede llevar un
mejor conocimiento de los procesos nerviosos, conductuales y
mentales de los que se ocupa la neurociencia cognitiva.
funcionales
su
dinámicas,
de
1.
2.
3.
Bibliografía
Díaz JL. The psychobiology of aggression and violence:
Bioethical Implications. International Social Science Journal
(UNESCO) 2011; 200-201: 233-245.
Morris D. El hombre al desnudo. Barcelona: Círculo de
lectores; 1977.
Lorenz K. On Aggression. New York. NY, USA: Harcourt, Brace
and World; 1966.
4.
5.
6.
7.
8.
9.
26
Davison R, Putnam K, Larson, C. Dysfunction of the neural
circuitry of emotion regulation. A possible prelude to
violence. Science 2000: 289 (5479): 591-594.
Heimer L, Van Hoesen G. The limbic lobe and its output
channel: Implication for emotional functions and adaptive
behavior. Neurosci Biobehav Rev 2006; 30 (2): 126-147.
Gregg T, Siegel A. Brain structures and neurotransmitters
regulating aggression in cats: Implications for human
aggression. Prog. Neuro-Psychopharmacol. Biol. Psychiatry
2005; 25: 91-140.
Anderson S, Bechara A, Damasio H, Tranel D, Damasio A.
Impairment of social and moral behavior related to early
damage in human prefrontal cortex. Nat Neurosci 1999; 2
(11):1032-1037.
de Almeida RM, Ferrari PF, Parmigiani F, Miczek KA.
Escalated aggressive behavior: Dopamine, serotonin and GABA.
Eur J Pharmacol 2005; 526 (1-3): 51-64.
Klaus A, Fish, E. Monoamines GABA, Glutamate and Aggression.
En: Randy J. Nelson, Brian C. Trainor. Biology of Aggression.
Oxford, UK: Oxford University Press; 2006.
10. de Boer SF, Koolhaas JM. 5HT1A and 5-HT1B receptor agonists
and aggression: A pharmacological challenge of the serotonin
deficiency hypothesis. Eur J Pharmacol 2005; 526 (1-3): 125-
139.
11. Massachusetts Institute of Technology MediaLab. (04 de 02 de
2013). Affective Computing. Recuperado el 02 de 2013, de
http://affect.media.mit.edu/
12. Martínez J, Ávila C, Fayad R. (2011). Transporte de Señales
Eléctricas en Microtúbulos: Una Aproximación Cuántica. Rev
Col Fis 2011; 42 (2): 101-106.
13. Hammeroff S. Quantum computation in brain microtubules? The
Penrose-Hameroff ORQ model of consciousness. Phil Trans R Soc
Lond A 1998; 356: 1869-1896.
14. Tuszýnski J, Hawrylak P, Brown J. Dielectric polarization,
electrical conduction, Information processing and Quantum
27
computation in microtubules. Are they pausible? Phil Trans R
Soc Lond 1998; A 356: 1897-1926.
15. Llinás R, Paré D. The brain as a closed system modulated by
the senses. En: Llinás R, Churchland PS (eds). The mind-
brain continuum. Sensorial processes. USA: MIT Press; 1996.
16. Kelso J. Dynamic patterns. The self organization of brain and
behavior. 1th Ed. Boston, USA: MIT Press; 1999.
17. Buldú J, Bajo R, Maestú F. Reorganization of Functional
Networks in Mild Cognitive Impairment (Vol. 6). Madrid:
PlosOne; 2011.
18. Muñoz-Delgado J, Díaz JL, Moreno C. Agresión y Violencia.
Cerebro, comportamiento y bioética. México: Herder; 2010.
19. Díaz JL. La conciencia viviente. México, México: Fondo de
cultura económica; 2008.
20. Murata T. Petri nets. Properties, analysis and applications.
Proceedings of the IEEE, 1989; 77(4), 541-580
21. David R, Alla A. Discrete, continuous and hybrid Petri nets.
Berlin, German: Springer Berlin Heidelberg; 2005.
22. Velden K, Peccoud J. Modeling networks of molecular
interactions in the living cells: Structure, dynamics and
application. Petri Nets and Performance Models, Proceedings.
IEEE 10th International Workshop. 2-8, 2003.
23. Hass P. Sthocastics Petri net; Modelling, Stability,
Simulation. NY: Springer-Verlag; 2002.
24. David. R, Alla. H. Discrete, continuous and hybrid Petri
nets. USA: Springer Verlag Press; 2002.
25. Nelson R, Trainor B. Neural mechanisms of aggression. Nat Rev
Neurosci 2007; 8(7): 536-546.
26. Churchland PS. Hacia una neurobiología de la mente. En:
Llinás R, Churchland PS. Hacia una neurobiología de la mente.
Bogotá, Colombia: Ed. Universidad Nacional- Universidad del
Rosario; 2006.
Figuras
28
Figura 1: Grafo bi-partito en el que se muestran las vías o arcos
y las transiciones entre cada plaza o nodo. Los tokens
representan unidades dinámicas de intercambio entre las plazas o
nodos que comparten una transición y su ubicación define el
estado del sistema.
a. b.
Figura 2: a. Circuito funcional de estructuras relacionadas con
la conducta de la agresión (Modificado de referencia 28). b.
Primera aproximación al modelo, los signos + y – representan
activación e inhibición, respectivamente.
Figura 3: Imagen del estado dinámico de la red de Petri que
representa el circuito de agresión. Las latencias entre los
núcleos se representan por las plazas semi-sombreadas. La
cantidad de área sombreada que se va cubriendo en sentido de las
manecillas del reloj, representa la duración de la latencia de
los nodos al recibir tokens. Una plaza se considera activada con
al menos un token en ella como se muestra a manera de ilustración
en Recap, LNET y EXPRE.
Figura 4: Vías de entrada a COF desde los núcleos de activación
Activ 1 y Activ 2.
Figura 5: Vías de entrada a AME desde Activ 1 y Activ 2.
Figura 6: Vías que van desde COF hacia Ame y HA.
Figura 7: Circuito reverberante entre COF, Ame y el comparador
TTL representado por las líneas resaltadas.
29
Figura 8: Representación del circuito de recobro como un grafo
auto cíclico.
Figura 9: Vías que de AMe van hacia LNET y HA.
Figura 10: Vías finales del circuito en el caso de concurrencia
espacio-temporal de tokens en las plazas LNET y HA. La activación
de estas plazas permite que la transición T5 envíe un token a
SGPA y por ende a EXPRE.
Figura 11: Regresión lineal de los tiempos de simulación vs
tiempos reales de conducción de señales. El eje vertical
representa el rango de valores de tiempos de computo en los nodos
de la red usados en la simulación y el eje horizontal asocia los
rango de tiempos de conducción reales.
Figura 12: Dinámica de primera y segunda activación, Activ 1 y
Activ 2, respectivamente. En esta figura y las siguientes el eje
vertical representa la cantidad de tokens que están en cada nodo
en un determinado instante de la simulación que dura 20000 ms en
pasos de a 20 ms para un total de 1000 iteraciones. Nótese que la
actividad en Activ 1 es aleatoria temporalmente y siempre genera
un solo token, al tiempo que Activ 2 es aleatoria en el tiempo y
genera una cantidad de tokens aleatoriamente.
30
Figura 13: Dinámica de COF (Puntos negros) y AMe (Cuadrados
grises) en términos de los tiempos de los pasos de cómputo. Un
token es la máxima cantidad alojada en estos nodos. Nótese que la
actividad de AMe es mayor en comparación con la de COF y la
coincidencia espacio-temporal de la activación de estos nodos no
es recurrente.
Figura 14: Dinámica del circuito de recobro, el cual muestra que
en todo el tiempo de la simulación (variable horizontal) la
activación de tokens (variable vertical) es constante, continua y
siempre con un token en este nodo, lo que genera la zona gris.
Figura 15: Dinámica y detalle del fenómeno de concurrencia en las
activaciones de LNET, HA y SGPA (Arriba hacia abajo). La zona
izquierda es una magnificación de las activaciones de estos nodos
representadas en la derecha. Los óvalos enfatizan el momento
exacto de la concurrencia entre nodos de niveles jerarquizados
(óvalo grande) y su consecuente efecto de activación de SGPA
(óvalo pequeño) para que ocurra el fenómeno de la agresión. Las
dos líneas superiores que llegan a la magnificación finalizan en
la concurrencia espacio-temporal de LNET y HA, la línea inferior
muestra la consecuente activación de SGPA.
Tablas
Tabla 1: Comparación entre las características de las redes de
Petri y las condiciones de la red neuronal.
Ajuste lineal Tiempos reales
ms
30
25
20
S
t
15
10
5
0
0
500
1000
tR
1500
2000
|
1207.2816 | 1 | 1207 | 2012-07-12T00:30:41 | Formation of feedforward networks and frequency synchrony by spike-timing-dependent plasticity | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Spike-timing-dependent plasticity (STDP) with asymmetric learning windows is commonly found in the brain and useful for a variety of spike-based computations such as input filtering and associative memory. A natural consequence of STDP is establishment of causality in the sense that a neuron learns to fire with a lag after specific presynaptic neurons have fired. The effect of STDP on synchrony is elusive because spike synchrony implies unitary spike events of different neurons rather than a causal delayed relationship between neurons. We explore how synchrony can be facilitated by STDP in oscillator networks with a pacemaker. We show that STDP with asymmetric learning windows leads to self-organization of feedforward networks starting from the pacemaker. As a result, STDP drastically facilitates frequency synchrony. Even though differences in spike times are lessened as a result of synaptic plasticity, the finite time lag remains so that perfect spike synchrony is not realized. In contrast to traditional mechanisms of large-scale synchrony based on mutual interaction of coupled neurons, the route to synchrony discovered here is enslavement of downstream neurons by upstream ones. Facilitation of such feedforward synchrony does not occur for STDP with symmetric learning windows. | q-bio.NC | q-bio |
Formation of Feedforward Networks and
Frequency Synchrony by
Spike-timing-dependent Plasticity
Naoki Masuda
Amari Research Unit, RIKEN Brain Science Institute,
2-1 Hirosawa, Wako, Saitama 351-0198 Japan
Hiroshi Kori
Department of Mathematics, Hokkaido University,
Kita 10, Nishi 8, Kita-Ku, Sapporo, Hokkaido, 060-0810 Japan
Corresponding author: Naoki Masuda
Tel: +81-48-467-9664
Fax: +81-48-467-9693
Email: [email protected]
1
Abstract
Spike-timing-dependent plasticity (STDP) with asymmetric learning windows is
commonly found in the brain and useful for a variety of spike-based computations such
as input filtering and associative memory. A natural consequence of STDP is estab-
lishment of causality in the sense that a neuron learns to fire with a lag after specific
presynaptic neurons have fired. The effect of STDP on synchrony is elusive because
spike synchrony implies unitary spike events of different neurons rather than a causal
delayed relationship between neurons. We explore how synchrony can be facilitated by
STDP in oscillator networks with a pacemaker. We show that STDP with asymmetric
learning windows leads to self-organization of feedforward networks starting from the
pacemaker. As a result, STDP drastically facilitates frequency synchrony. Even though
differences in spike times are lessened as a result of synaptic plasticity, the finite time
lag remains so that perfect spike synchrony is not realized. In contrast to traditional
mechanisms of large-scale synchrony based on mutual interaction of coupled neurons,
the route to synchrony discovered here is enslavement of downstream neurons by up-
stream ones. Facilitation of such feedforward synchrony does not occur for STDP with
symmetric learning windows.
Keywords:
spike-timing-dependent plasticity, synchronization,
feedforward net-
works, complex networks
2
1
Introduction
In many neural circuits, synaptic plasticity depends on relative timing of presynaptic
and postsynaptic spikes, which is known as spike-time dependent plasticity (STDP)
(Gerstner et al., 1996; Bell et al., 1997; Markram et al., 1997; Bi and Poo, 1998;
Zhang et al., 1998). Specifically, long-term potentiation (LTP) ensues when a presy-
naptic neuron fires slightly before a postsynaptic neuron (of the order of 10 ms), whereas
long-term depression (LTD) is elicited in the opposite case (solid line in figure 1). This
STDP rule promotes causal relationship between a pair of neurons in the sense that the
strength of a synapse that contributes to generation of postsynaptic spikes is reinforced.
Computationally, STDP is useful
for synaptic competition (Kempter et al., 1999;
Song et al., 2000; van Rossum et al., 2000; Song and Abbott, 2001), coincidence de-
tection (Gerstner et al., 1996), spike-based associative memory (e.g. Lengyel et al.,
2005),
implementation of the synfire chain (Horn et al., 2000; Levy et al., 2001),
generation of reproducible spatiotemporal spike patterns (Izhikevich et al., 2004;
Izhikevich, 2006), selection of earlier inputs, to name a few (e.g. Gerstner and Kistler,
2002).
Related to neural computation, coincident firing of multiple neurons in the oscil-
latory regime is found in many parts of the brain and believed to play an important
role (Singer and Gray, 1995; Ritz and Sejnowski, 1997; Buzs´aki and Draguhn, 2004).
One could imagine that real neural networks learn to synchronize spikes of different
neurons by STDP-related synaptic plasticity, as suggested by some modeling studies.
However, contribution of STDP to spike synchrony may be limited. For example,
STDP can lead to division of a neural population into clusters in each of which neu-
rons fire in spike synchrony (Horn et al., 2000; Levy et al., 2001). This self-organizing
process actually necessitates homogeneous synaptic transmission delays for different
synapses and puts a strong restriction on the firing period. If there is just one cluster
3
(all neurons firing synchronously), the firing period has to be equal to the synaptic
transmission delay. Similarly, if there are two clusters, the first cluster excites a syn-
chronous volley in the second cluster after the synaptic transmission delay. The second
cluster reexcites the first cluster in a similar way. The firing period is equal to the
synaptic transmission delay multiplied by the number of clusters, which seems restric-
tive. Alternatively, coincident firing is achieved via STDP if the amount of LTP and
that of LTD that are caused by a presynaptic and postsynaptic spike pair are perfectly
balanced (Karbowski and Ermentrout, 2002). Evolution of coincident firing survives
heterogeneity in neurons and in the amount of plasticity. However, how coincident
firing is affected by the imbalance between LTP and LTD remains to be explored.
Coincident firing in recurrent neural networks may not be established through
STDP. In general, synchronous firing can be induced by sufficiently strong coupling be-
tween elements (Kuramoto, 1984; Pikovsky et al., 2001; Gerstner and Kistler, 2002).
By contrast, STDP cannot strengthen the synaptic weights between two neurons bidi-
rectionally. An increase of the synaptic weight in one direction implies a decrease in
the opposite direction, and stability requires that the net decrease and the net increase
are roughly balanced (Song et al., 2000; Song and Abbott, 2001). Therefore, STDP
does not necessarily enhance mutual interaction. Indeed, in recurrent neural networks,
STDP does not necessarily support synchronous firing (Masuda and Aihara, 2004). It
rather reinforces reproducible spatiotemporal spike patterns composed of causal spike
pairs of different neurons (Izhikevich et al., 2004; Izhikevich, 2006).
We examine possible mechanisms of STDP-induced synchrony in recurrent networks
of oscillatory elements. We distinguish two types of synchrony using the terminology of
coupled oscillators. One type is phase synchrony, which is equivalent to spike synchrony.
When neurons are in phase synchrony, they share spike timing. The other weaker notion
is frequency synchrony in which neurons possibly with different intrinsic firing rates
share a common firing rate. Frequency synchrony does not imply phase synchrony. The
4
spike time of the postsynaptic neuron can differ from that of the presynaptic neuron.
According to these definitions, the previous studies cited above, which relate STDP to
synchrony, regard phase synchrony.
STDP may be more relevant to frequency synchrony. For example, STDP pro-
motes frequency synchrony, but not phase synchrony, in a hybrid circuit of an aplysia
abdominal neuron and an emulated neuron (Nowotny et al., 2003). Unidirectional con-
nectivity from the emulated neuron to the aplysia neuron eventually forms. Numeri-
cal simulations of two coupled Hodgkin-Huxley neurons (Zhigulin et al., 2003) and of
large neural networks (Zhigulin and Rabinovich, 2004) also support the notion that
frequency synchrony is facilitated by STDP.
In the present work, we show that the standard STDP facilitates frequency syn-
chrony to a great extent, particularly when LTP and LTD are roughly balanced. To
examine how heterogeneous neurons interact to produce possible synchronization, we
analyze networks of oscillators with a pacemaker. The pacemaker has a distinct nat-
ural frequency, is not affected by other oscillators, and sets the rhythm to influence
other neurons. No matter whether a pacemaker is realized by a network or a sin-
gle neuron, existence of pacemaker neurons is suggested in, for example, the basal
ganglia (Plenz and Kitai, 1999) and respiratory networks in the pre-Botzinger com-
plex (Ramirez et al., 2004). Furthermore, many neurons (Hutcheon and Yarom, 2000)
and recurrent microcircuits (Jefferys et al., 1996) are intrinsically oscillatory (also see
e.g. Singer and Gray, 1995). and their rhythmic activities are resistant to perturbation.
These neural networks and single neurons can also serve as pacemakers. With frozen
and sufficiently strong synapses, oscillator networks with pacemakers allow frequency
synchrony (Kori and Mikhailov, 2004). We show that STDP considerably facilitates
frequency synchrony of pacemaker systems by establishing feedforward network struc-
ture whose root is the pacemaker.
For analytical tractability, we mostly deal with networks of phase oscillators in
5
which coupling strength evolves according to STDP. Coupled phase oscillators approx-
imate various natural systems composed of self-sustained oscillators with weak coupling
(Winfree, 1980; Kuramoto, 1984; Glass and Mackey, 1988),
including pulse-coupled
neurons (Kuramoto, 1991; Hansel et al., 1993; Hansel et al., 1995; Kori, 2003). We
introduce the model in section 2 and analyze simple cases of two connected oscillators
in section 3. We numerically analyze larger oscillator networks with STDP in section 4.
In section 5, we numerically simulate pulse-coupled pyramidal neuron models to show
that our results obtained for coupled phase oscillators qualitatively apply to spiking
neuron models.
2 Model
We analyze a network of n phase oscillators. One oscillator is assumed not to be dis-
turbed by the other n − 1 oscillators. We designate this special oscillator as pacemaker
and use the term oscillator to refer to the other n − 1 elements. The pacemaker has
natural frequency Ω and phase φ0 ∈ [0, 2π). The other oscillators are assumed to have
the identical natural frequency ω , and the phase of the i-th oscillator (1 ≤ i ≤ n − 1)
is denoted by φi ∈ [0, 2π). We identify φi = 0 and φi = 2π (0 ≤ i ≤ n − 1). We write
(i, j ) ∈ E if there is a synaptic connection from oscillator i to oscillator j . In other
words, E is the set of edges of the underlying neural network. As in real neural net-
works, connectivity is asymmetric in general so that (i, j ) ∈ E does not imply (j, i) ∈ E .
A pair of connected oscillators interact via sinusoidal coupling, which usually promotes
synchrony (Kuramoto, 1984).
Dynamics for fixed synaptic strengths are represented by:
φ0 = Ω,
(1)
φi = ω +
gj i sin(φj − φi ),
1
hki Xj :(j,i)∈E
where hki is the average number of incoming edges per oscillator. The coupling strength
6
(1 ≤ i ≤ n − 1)
(2)
gj i is associated with synapse (j, i). We note that gi0 , which is the synaptic weight from
an oscillator to the pacemaker, does not affect the network dynamics: the pacemaker
is not perturbed by external input. However, we will monitor gi0 to examine how this
connection evolves as synaptic plasticity goes on.
We assume that Ω > ω unless otherwise stated. By rescaling the timescale and the
coupling strength, we set Ω = ω +1 without losing generality. To set the values of Ω and
ω , we take care of two subtle factors. First, a small ω would yield backward rotation
by the effect of coupling. This is because the second term of the right-hand side of
equation (2) can be large negative to overwhelm the first term. Then, the condition
φi < 0 may be satisfied for long enough time to elicit backward firing. This is unrealistic
as a neuron model. Second, we avoid a pair of Ω and ω that accommodates the relation
M1Ω = M2ω with small integers M1 and M2(M1 6= M2 ). In such a situation, resonant
behavior appears when the pacemaker and the oscillators are decoupled through STDP
and has a pathological effect (see the explanation after equation (16) for more details).
The resonant firing is ruled out by dynamical noise in many real neural networks.
However, we have to carefully specify Ω and ω in the present work because we do not
assume noise for analytical tractability. Keeping these caveats in mind, we set Ω = 9.1
and ω = 8.1.
Spike time is defined to be the time when the φi crosses 0. Synaptic update based
on STDP takes place based on a pair of nearest presynaptic and postsynaptic spike
times, without paying attention to remote spike pairs (see arguments in e.g. Froemke
and Dan, 2002). We compare the upshot of two types of STDP rules for synapse gj i
((j, i) ∈ E ), namely, asymmetric STDP and symmetric STDP.
Asymmetric STDP is modeled as follows. LTP is induced if a presynaptic firing
(spike of oscillator j ) precedes a postsynaptic firing (spike of oscillator i).
In the
opposite case, LTD occurs. We denote the presynaptic (postsynaptic) spike time by
7
tpre (tpost). A spike-pair event modifies the synaptic weight: gj i → gj i + ∆gj i , where
∆gj i =
A+ exp (cid:16)− tpost−tpre
(cid:17) ,
tpre < tpost ,
τ
(cid:17) ,
−A− exp (cid:16)− tpre−tpost
tpre > tpost ,
τ
under the limitation gj i ∈ [0, gmax ]. A sample learning window is indicated by the solid
line in figure 1. The width of the learning window is specified by τ , which is known
(3)
to be of the order of 10–20 ms (Bi and Poo, 1998; Zhang et al., 1998). We confine
ourselves to the regime in which firing rates are not very large (5–20 Hz), as is true
for many pyramidal neurons. Then, τ is several times smaller than the characteristic
interspike interval T = 50–200 ms. We thus set
τ =
1
6 ×
2π
Ω ∼=
T
6
.
(4)
For completeness, we assume that tpre = tpost does not induce plasticity.
We assume that synaptic weights evolve so slowly that we can solve equation (2) by
regarding the synaptic weights as constant. This assumption is valid if A+Ω, A−Ω ≪
g 2 for the following reason. Because the relative phase relationship determines the
evolution of synaptic weights, we should compare the typical timescale of the relative
phase dynamics with that of synaptic plasticity. The former is the inverse of typical
synaptic weight g0 . By introducing dimensionless synaptic weight g/g0 , we find from
equation (3) that the timescale of dimensionless synaptic plasticity is the inverses of
A+Ω/g0 and A−Ω/g0 . The two timescales are separated if A+Ω/g0 , A−Ω/g0 ≪ g0 ,
0 ∼= g 2 . On the slow timescale of synaptic plasticity, we
which leads to A+Ω, A−Ω ≪ g 2
can set A− = 1 by rescaling the time, so that only the ratio A+/A− is relevant. For the
stability, A+/A− must be balanced. This ratio is assumed to be slightly smaller than
unity according to previous literature (Song et al., 2000; Song and Abbott, 2001).
Most of our theoretical efforts are invested in asymmetric STDP because many
pyramidal neurons show asymmetric STDP. However, symmetric STDP, in which
the synaptic update rule depends only on tpre − tpost , is also found in some experi-
8
ments. Particularly, the learning window is often shaped like a mexican hat in excita-
tory synapses; small (large) tpre − tpost induces LTP (LTD) (Nishiyama et al., 2000;
Abbott and Nelson, 2000; Shouval et al., 2002). Symmetric learning windows have
been found for inhibitory synapses (Woodin et al., 2003) and for the amount of LTD
in excitatory synapses (Dan and Poo, 1992). We numerically analyze networks with
symmetric STDP in section 4. We adopt superposition of two gaussian distributions as
the symmetric learning window, as depicted by the dotted line in figure 1. A spike-pair
event modifies the synaptic weight: gj i → gj i + ∆gj i , where
A+
(tpre − tpost)2
! −
! ,
exp −
exp −
A−
√2πσ+2
√2πσ−2
2σ+2
with σ+ < σ− . We set σ+ = 0.6τ and σ− = 2σ+ = 1.2τ so that the timescale of the
(tpre − tpost)2
2σ−2
∆gj i =
(5)
symmetric learning window is comparable to that of the asymmetric learning window
defined in equation (3). The values of A+ and A− are assumed to be the same as those
for asymmetric STDP so that gj i is bounded.
3 Analysis of Small Networks with Asymmetric
STDP
We begin with small networks of two oscillators with asymmetric STDP. For these
networks, how much initial coupling is necessary for synchrony can be analytically
evaluated.
3.1 One Pacemaker and One Oscillator
We deal with the case n = 2, namely, a network of one pacemaker and one oscillator.
Because the connection from the oscillator to the pacemaker does not affect the dy-
namics of the pacemaker, it suffices to consider the unidirectional case. The network
is schematically shown in figure 2(a). We write g = g01 to simplify the notation. The
9
short-term dynamics in which g is regarded to be constant are described by
φ0 = Ω,
φ1 = ω + g sin(φ0 − φ1).
With ψ ≡ φ0 − φ1 , equations (6) and (7) reduce to
ψ = Ω − ω − g sin ψ .
(6)
(7)
(8)
Based on the assumption that synaptic plasticity occurs much more slowly than firing,
we perform quasistatic analysis. For a frozen g , let us derive the average angular
frequency of the oscillator denoted by ω . If g ≥ Ω − ω , the pacemaker and the oscillator
are in frequency synchrony, i.e. ω = Ω. If 0 ≤ g < Ω − ω , equation (8) is equivalent to
dψ
Z
= Z dt.
Ω − ω − g sin ψ
Integration of equation (9) over a cycle yields
dψ
dt = Z 2π
= Z T
Ω − ω − g sin ψ
0
0
2π
=
,
q(Ω − ω )2 − g 2
2π
Ω − ω
(9)
(10)
which results in
ω = Ω − q(Ω − ω )2 − g 2 .
(11)
Note that ω ≤ ω < Ω.
The direction and the amount of synaptic plasticity induced by a single spike-pair
event is determined by tpost − tpre . We estimate tpost − tpre in terms of ψ as follows.
Suppose that the phase difference is equal to ψ when the pacemaker fires. Then, it
approximately takes tpost − tpre = ψ/ ω for the oscillator to fire.
is induced because the pacemaker is presynaptic to the oscillator. The pacemaker
In this case, LTP
and the oscillator can fire in the opposite order.
If the phase difference is ψ when
10
the oscillator fires prior to the pacemaker does, the pacemaker spends approximately
tpost − tpre = ψ/Ω before firing.
ourselves to the case in which Ω does not deviate so much from ω , we approximate
tpost − tpre ∼= ψ/Ω regardless of the order of firing.
In equation (4), we assumed that the decay rate of the learning window τ is suffi-
In this case, LTD is induced. Because we confine
ciently smaller than T /2, which corresponds to phase π . Therefore, the amount of LTP
is negligible for tpost − tpre ∼= π/ ω or larger, and the LTP rule is effective only when
0 < ψ ≤ π . By the same token, the LTD rule is effective only for −π < ψ < 0. Using
these approximations, we aim to describe the dynamics of synaptic plasticity in terms
of phase variables. The amount of plasticity given by equation (3) can be rewritten as
∆g =
A+ exp (cid:16)− ψ
Ωτ (cid:17) , 0 < ψ ≤ π , φ1 = 0
Ωτ (cid:17) , −π < ψ < 0, φ0 = 0
−A− exp (cid:16) ψ
where φ1 = 0 and φ0 = 0 indicate the postsynaptic spike time for an LTP event (the
spike time of the oscillator) and that of an LTD event (the spike time of the pacemaker),
(12)
respectively.
We denote by g (0) the initial synaptic weight. If g (0) ≥ Ω − ω , fast dynamics have
two steady states given by
ψ ∗ = arcsin Ω − ω
g (0) ! .
The solution with π/2 < ψ ∗ ≤ π is unstable, and hence the fast dynamics converges to
ψ ∗ satisfying 0 ≤ ψ ∗ ≤ π/2. Therefore, STDP induces potentiation of g . Then, ψ ∗ for
an altered g becomes even smaller, which induces further potentiation of g . Eventually,
(13)
g = gmax is achieved. In sum, if
g (0) ≥ gc ≡ Ω − ω ,
(14)
the pacemaker and the oscillator will synchronize quickly without plasticity. The STDP
does not break frequency synchrony. Note that STDP generally decreases the phase
11
difference ψ , but ψ does not tend to 0 (no phase synchrony) unless Ω = ω . Alternatively,
if gmax = ∞, g diverges, and ψ goes to 0.
Does asymmetric STDP facilitate frequency synchrony? If g (0) < gc , the oscillator
is initially not entrained by the pacemaker. Then, ψ slips. By averaging over many
spike-pair events, we represent the synaptic dynamics on a slow timescale by
ω
2π (cid:20)Zt such that 0<ψ<π
A−e−ψ/Ωτ dt(cid:21)(15)
A+ e−ψ/Ωτ dt − Zt such that −π<ψ<0
g =
A+
e−ψ/Ωτ
gc + g sin ψ ! dψ .
A−
∝ Z π
(16)
gc − g sin ψ −
0
Derivation of equation (15) requires the nonresonant situation. If M1Ω = M2ω holds
with small M1 and M2 , the dynamics become periodic with a rather small period when
the oscillator decouples from the pacemaker due to STDP. If this were the case, the
dependence on the initial condition does not vanish permanently. In other words, ψ
conditioned by φ0 = 0 or φ1 = 0 in equation (12) would take only limited values. Then
the distribution of ψ conditioned by a spike event would deviate from the unconditioned
distribution of ψ . With our choice of Ω = 9.1 and ω = 8.1, the effect of such a resonance
is very small.
In the region of g where g > 0 holds, the RHS of equation (16) increases monotoni-
cally with g . If g (0) is greater than the value that makes the RHS equal to zero, which
we denote by gc−stdp , we obtain g > 0. Under this condition, g continues to increase,
and ψ ∗ decreases. This makes g in equation (16) even larger. This positive feedback
lasts until g ≥ gc is eventually satisfied. As a result, frequency synchrony is elicited
by STDP. However, if g (0) < gc−stdp , g converges to the lower bound 0, so that the
oscillator is disconnected from the pacemaker.
We bound gc−stdp as follows:
RHS of equation (16)
e−ψ/Ωτ
A+
= Z π
gc
0
1
1 − g sin ψ
gc
−
gc
A−
1 −
g sin ψ
gc
1 + g sin ψ
gc
12
e−ψ/Ωτ
gc
A+
gc 1 +
gc ! −
A−
g sin ψ
g sin ψ
1
≥ Z π
1 −
1 + g
gc
0
gc
A− − A+
A−
g
Z π
Z π
A+ +
= −
1 + g
g 2
gc
0
0
c
gc
A+ + A−
(A− − A+ )Ωτ (1 − e−π/Ωτ )
Ω2 τ 2 (1 + e−π/Ωτ )
g
+
1 + g
g 2
1 + Ω2 τ 2
gc
c
gc
dψ
e−ψ/Ωτ sin ψdψ
e−ψ/Ωτ dψ +
= −
.
(17)
gc−stdp ≤
The value of g that makes the RHS of the above equation zero gives an upper bound
of gc−stdp . When A+ and A− are balanced (A+ ∼= A− ), we obtain
(1+Ω2 τ 2 )(1−e−π/Ωτ )
A−−A+
Ωτ (1+e−π/Ωτ )
A−+A+
(1+Ω2 τ 2 )(1−e−π/Ωτ )
1 − A−−A+
Ωτ (1+e−π/Ωτ )
A−+A+
Because Ωτ is assumed to be of the order of π (see equation (4)), Ωτ = O(1).
In
addition, when A+ ∼= A− , the inequalities in equations (17) and (18) nearly hold with
the equalities. In such a case,
(1 + Ω2 τ 2 )(1 − e−π/Ωτ )
Ωτ (1 + e−π/Ωτ )
A− − A+
A− + A+
gc ∼=
gc .
(18)
A+
A− ! gc ,
gc−stdp ∝ 1 −
which implies that gc−stdp is much smaller than gc when A+ ∼= A− . Particularly, gc−stdp
is extinguished when A+ ≥ A− .
In figure 3, we plot gc−stdp evaluated by numerical integration of equation (16) and
(19)
the approximation given by equation (18). We also plot gc−stdp obtained by numerical
simulations of our model (equations (3), (6), and (7)), in which gc−stdp is determined by
varying the initial synaptic weight g (0). The evaluation by equation (16) (solid line) is
in good agreement with gc−stdp obtained by numerical simulations of the model (circles)
for a broad range of A+/A− . As expected, the approximate estimation by equation (18)
(dotted line) also agrees with the numerical data (circles) when A+/A− is close to unity.
In conclusion, asymmetric STDP drastically enhances frequency synchrony.
Regarding symmetric STDP, for values of g such that ψ falls in the positive learning
window (refer to the dotted line in figure 1), g is strengthened to eventually exceed gc .
13
Therefore, frequency synchrony is facilitated. To what extent synchrony is promoted
depends on the width of the learning window.
When Ω < ω , there are two solutions ψ ∗ ∈ (−π , 0), one of which is stable. Then,
STDP elicits LTD. Even though ψ ∗ changes, the relation −π < ψ ∗ < 0 is preserved until
the pacemaker and the oscillator get disconnected. As a result, frequency synchrony
does not happen.
3.2 Two Oscillators
To examine how the connectivity between a pair of oscillators evolves in a large network,
we analyze the following toy model of two bidirectionally coupled oscillators:
φ1 = ω + ∆ω + g1 sin(φ2 − φ1),
φ2 = ω + g2 sin(φ1 − φ2),
(20)
where g1 , g2 ∈ [0, gmax ]. The network is depicted in figure 2(b). Now two oscillators
influence each other, which contrasts to the case of the pacemaker-oscillator network
examined in section 3.1. The term ∆ω represents the mismatch in natural frequen-
cies. Although the oscillators are identical in our original model (see equation (2)),
we introduce ∆ω because of the following reason. In oscillator networks with a pace-
maker, the oscillators are not completely phase synchronized. The oscillators directly
connected to the pacemaker are the first to fire after the pacemaker does. Then, other
oscillators adjacent to those connected to the pacemaker fire after some delay, and so
forth. Therefore, the oscillators closer to the pacemaker tend to have more advanced
phases, and the distribution of the phases is associated with the hierarchical organiza-
tion of the network. Imagine two oscillators coupled unidirectionally or bidirectionally
in a large network. We denote one that fires first and the other by oscillators 1 and 2
respectively. Precisely, the difference in the firing timing stems from complex effects of
coupling with other oscillators. For analytical tractability, here we replace such effects
14
by the frequency mismatch ∆ω , by which the difference in the firing timing can be
easily introduced. As shown in the following, for ∆ω > 0, oscillator 1 tends to fire
in advance of oscillator 2. Accordingly, we regard that oscillator 1 is closer to the
pacemaker than oscillator 2.
We analyze the model given by equation (20). By introducing ψ ≡ φ1 − φ2 , we
derive
ψ = ∆ω − (g1 + g2 ) sin ψ .
(21)
If
g1(0) + g2(0) ≥ ∆ω ,
two oscillators are locked with phase lag 0 < ψ ∗ < π/2, where
ψ ∗ = arcsin ∆ω
g1 + g2 ! .
If synaptic plasticity is absent, equation (22) gives the condition for frequency syn-
(23)
(22)
chrony.
In contrast to the network of one pacemaker and one oscillator analyzed in sec-
tion 3.1, equation (22) does not guarantee that frequency synchrony is maintained
throughout STDP. When equation (22) is satisfied, the synaptic dynamics are written
as
where
ωτ ! ,
g1 = −A− exp −
ψ ∗
g2 = A+ exp −
ωτ ! ,
ψ ∗
(24)
(25)
ω = ω +
g2∆ω
g1 + g2
is the frequency common to the two oscillators. Because A+ < A− , g1 + g2 decreases
with time. The oscillators desynchronize in frequency if g1 + g2 ≥ ∆ω is violated
via synaptic plasticity. For sufficiently small g1 (0) + g2(0), the two oscillators are
(26)
15
disconnected even from the beginning. In these cases, ψ slips due to the absence of
frequency synchrony.
The average frequencies of the two oscillators out of frequency synchrony are cal-
culated as
,
.
ω1 = ω +
ω2 = ω +
g2∆ω + g1q∆ω 2 − (g1 + g2)2
g1 + g2
g2∆ω − g2q∆ω 2 − (g1 + g2 )2
g1 + g2
The synaptic weights evolve according to
A+e−ψ/ ω1 τ
A−e−ψ/ ω2 τ
∆ω − (g1 + g2) sin ψ # dψ
0 "
1
2π Z π
∆ω + (g1 + g2) sin ψ −
A+
e−ψ/ ω1 τ "
∆ω − (g1 + g2) sin ψ # dψ ,
A−
1
2π Z π
∆ω + (g1 + g2) sin ψ −
0
A+e−ψ/ ω2 τ
A− e−ψ/ ω1 τ
0 "
∆ω + (g1 + g2) sin ψ # dψ
1
2π Z π
∆ω − (g1 + g2 ) sin ψ −
A+
e−ψ/ ω1 τ "
∆ω + (g1 + g2) sin ψ # dψ ,
A−
1
2π Z π
∆ω − (g1 + g2) sin ψ −
0
where we approximated e−ψ/ ω2 τ by e−ψ/ ω1 τ , as we did in section 3.1.
g2 =
g1 =
=
=
(27)
(28)
(29)
(30)
Since g1 < 0 is always satisfied, g1 eventually reaches 0; backward connectivity from
a downstream oscillator to an upstream oscillator is eliminated. Whether a ‘forward’
connectivity from the upstream oscillator to the downstream oscillator survives relies
on g2(t) where t is the time g1 reaches 0. If g2(t) is larger than gc−stdp obtained in
section 3.1, frequency synchrony will be eventually established. In this case, the final
oscillation frequency is equal to ω + ∆ω so that oscillator 2 is enslaved by oscillator 1.
If g2(t) < gc−stdp , the two oscillators are finally disconnected.
The critical value of g2 (0) above which frequency synchrony occurs is plotted in
figure 4 for different values of A+ /A− and g1(0). The critical g2(0) decreases with g1(0),
implying that a large g1(0) enhances frequency synchrony. Such backward connectivity
transiently serves to keep ψ small so that forward connectivity is enhanced. However,
only the synapse from the faster oscillator to the slower oscillator survives eventually.
16
The feedforward network is not created via symmetric STDP, by which g1 and
g2 evolve in the same direction. When initial mutual connectivity is strong enough,
synchrony is established so that the two synaptic weights are saturated (g1 = g2 =
gmax ). Then, based on equation (26), the common firing frequency is equal to ω +∆ω/2,
but not to the frequency of the faster oscillator (= ω + ∆ω ).
The effect of LTP-LTD balance is also shown in figure 4. When A+ is close to A− ,
critical g2(0) is lowered, and entrainment occurs easily. Even when LTP is rather weak
compared to LTD (A+/A− = 0.8, thinnest line), the critical g2 is much reduced from
gc .
4 Numerical Results for Large Networks
So far, we have analyzed small networks composed of two elements only. In this section,
we examine how frequency synchrony can be facilitated by STDP in larger networks.
In particular, we compare gc and gc−stdp and also investigate evolution of network
structure. To this end, we numerically simulate randomly connected n = 100 elements
(99 oscillators and one pacemaker) based on equations (3), (6), and (7).
4.1
Initial Setup
We generate a directed random network as follows. Starting from a set of n isolated
vertices, we add a directed edge that connects a randomly chosen pair of oscillators.
We forbid multiple directed edges between a pair of oscillators and self loops, i.e. edges
whose origin and destination are identical. This procedure is repeated n hki times. In
the following, we assume that hki = 10. In other words, each oscillator is presynaptic
to 10 other oscillators and postsynaptic to the same number of oscillators on average.
We define the distance li of oscillator i from the pacemaker by the smallest number
of directed edges necessary to reach from the pacemaker to oscillator i. For example,
the number of the oscillators at distance 1 is equal to those that receive direct synaptic
17
contacts from the pacemaker. Therefore, about hki oscillators have distance 1. Among
the other oscillators, those receiving an edge from an oscillator with distance 1 have
distance 2. The depth L of a network is defined as the distance averaged over all the
oscillators: L = Pn−1
i=1 li/(n − 1).
The initial phases φi (0 ≤ i ≤ n − 1) are taken independently from the uniform
density on [0, 2π). For all the synapses, we initially set gj i = g (0).
For a specific random network used in the following numerical simulations, we
obtained L = 3.22. For this network, we numerically found that frequency synchrony
happens without synaptic plasticity if g ≥ gc ∼= 100.7. In this case, the pacemaker first
fires in each cycle, and oscillators with smaller distances tend to fire with smaller lag
with respect to the pacemaker (Kori and Mikhailov, 2004).
4.2 Measured Quantities
We define the degree of frequency synchrony r ≡ r ([t1 , t2 ]) for a time interval [t1 , t2 ].
The mean frequency of each oscillator for this time interval is equal to
r =
(32)
.
.
(31)
ωi =
φi (t2 ) − φi (t1 )
t2 − t1
We note that ω0 = Ω. Then, the synchrony measure is defined by
n−1 Pn−1
1
i=1 ωi − ω
Ω − ω
When the oscillators are in frequency synchrony with the pacemaker, the mean fre-
quency of the oscillators Pn−1
i=1 ωi/(n − 1) is equal to the frequency of the pacemaker Ω,
and we have r = 1. If the pacemaker does not at all affect the other n − 1 oscillators,
the oscillators fire at their natural frequency ω , and we have r = 0. We divide the
total simulation time into consecutive bins of the width t2 − t1 = 100 for oscillator
simulations in sections 4.3 and 4.4.
More microscopically, we inspect possible formation of feedforward chains originat-
ing from the pacemaker. To quantify this, we track several order parameters derived
18
from synaptic weights. The first is the depth L extended to networks with heteroge-
neous synaptic weights in the following way. Let us consider a path from the pace-
maker to oscillator i. A path is equivalent to a chain of directed synapses: (j0 , j1 ) ∈ E ,
(j1 , j2 ) ∈ E , . . ., (jki−1 , jki ) ∈ E , where j0 = 0 and jki = i. The length of this path
is given by Pki−1
k=0 gmax /gjk jk+1 . The distance li of oscillator i from the pacemaker is
the shortest path length among all possible paths from the pacemaker to oscillator
i (Braunstein et al., 2003). This definition generalizes the prior definition for networks
with unit synaptic weights. The redefined distance is associated with how much a
downstream oscillator is influenced by the pacemaker. The depth of the network is
again defined by L = Pn−1
i=1 li/(n − 1) and measures effective proximity of the oscilla-
tors from the pacemaker. By definition, the generalized L is equal to or larger than
L of the unweighted network, with equality realized only when gj i = gmax for all the
synapses that appear in the shortest paths.
A synaptic connection (j, i) with lj < li (lj > li ) is a forward (backward) connection
in the meaning that it complies with the feedforward chain emanating from the pace-
maker. Accordingly, we define the amount of forward connection wf , that of backward
connection wb , and that of lateral connection wl by
,
,
.
(33)
gj i
Gf = Xli−lj >ǫ
n hki
gj i
Gb = Xli−lj <−ǫ
n hki
gj i
Gl = X−ǫ≤li−lj ≤ǫ
n hki
Summation is taken over the pairs of oscillators forming synapses ((j, i) ∈ E ). Note that
Gl quantifies the connection between oscillators whose distances from the pacemaker
are approximately equal. The number of synapses in the network (= n hki) gives
normalization, and thus 0 ≤ Gf , Gb , Gp ≤ gmax . The average synaptic weight is given
by 0 ≤ Gf + Gb + Gp ≤ gmax . The tolerance level is chosen to be sufficiently small:
(35)
(34)
19
ǫ = 0.05. We also define local quantities to evaluate feedforwardness. The average
weight of the synapses postsynaptic to the pacemaker is denoted by G0
f . This is equal
to the average of g0i over i with (0, i) ∈ E . This corresponds to g used in section 3.1.
Similarly, the average weight of the synapses presynaptic to the pacemaker is denoted
by G0
b . This is equal to the average of gi0 over i with (i, 0) ∈ E . We note that
f , G0
0 ≤ G0
b ≤ gmax .
4.3 Asymmetric Learning Window
We apply asymmetric STDP with LTD slightly stronger than LTP (Song et al., 2000;
Song and Abbott, 2001): A+ = 0.009 and A− = 0.01. By setting gmax = 15 < gc , ho-
mogeneous enhancement of all the synapses does not lead to synchrony. We determine
gc−stdp by running numerical simulations with various values of the initial synaptic
weight gj i = g (0).
Because A+ is close to A− , our results in section 3 predict the following.
• As shown in section 3.2, backward connection will be eliminated via the asym-
metric STDP so that Gb and G0
b decrease.
• The unidirectional connection between the pacemaker and the oscillator will be
easily established (section 3.1). As a result, a feedforward chain rooted at the
pacemaker is expected to form.
• gc−stdp is much smaller than gc .
Dynamics of synaptic-weight order parameters for g (0) = 0.7 are shown in fig-
ure 5(a, b). The average synaptic weight (dotted line in figure 5(a)) increases in the
initial stage because some synapses between the oscillators are potentiated. However,
its stationary value is much smaller than the maximal possible value (gmax = 15). There
is no selective potentiation of forward synapses (Gf , thick solid line in figure 5(a)) or
20
depression of backward synapses (Gb , thin solid line in figure 5(a)). The forward con-
nectivity from the pacemaker (G0
f , thick line in figure 5(b)) also degrades with time.
Eventually, the oscillators disconnect from the pacemaker, which is observed as indef-
initely growing L (uppermost line in figure 5(e)). Accordingly, frequency synchrony
between the pacemaker and the oscillators is not achieved; figure 5(f ) indicates that r
stays near 0.
By contrast, frequency synchrony without phase synchrony is established when
g (0) = 1.5, as supported by the rastergrams in figures 6(a) and 6(b) corresponding to
initial and final periods of simulations, respectively. More in detail, forward connec-
tivity Gf (thick solid line in figure 5(c)) and G0
f (thick line in figure 5(d)) grow toward
gmax to result in frequency synchrony at t ∼= 12500 (figure 5(f )), accompanied by a
decrease in L (lowermost line in figure 5(e)). Backward synapses directly pro jecting
to the pacemaker are pruned in an initial stage (G0
b ; thin line in figure 5(d)). It takes
longer time for Gb to decay (thin solid line in figure 5(c)). Although randomness in
the initial condition blurs the phase transition, we estimate gc−stdp ∼= 0.9 based on
figures 5(e) and (f ). Consistent with the results in section 3.1, gc−stdp is much smaller
than gc ∼= 100.7.
Frequency synchrony is made possible by combined effects of sufficiently large for-
ward weights and sufficiently small backward weights. It is not an immediate conse-
quence of increased average synaptic weights; achieving synchrony merely by homo-
geneously strong synapses necessitates Gf + Gb + Gp ≥ gc . Because of our choice of
gmax (< gc ), homogeneous LTP does not induce frequency synchrony. Elimination of
backward weights is essential for frequency synchrony. The final network structure re-
constructed from synapses with gj i > g (0) = 1.5 is shown in figure 5(g), with forward
and backward edges shown by thin and thick lines, respectively. Few backward edges
survive asymmetric STDP. The network is close to a feedforward network rooted at
the pacemaker, which enslaves the oscillators.
21
We remark that detailed behavior of the network order parameters varies according
to the value of ǫ and the definition of the distance li , which is inherently arbitrary for
weighted networks. However, the general tendency that forward synapses are poten-
tiated and backward synapses are depressed for g ≥ gc−stdp is observed irrespective of
these details. For a reasonably defined li , whether L increases or decreases (figure 5(e))
and whether feedforward networks form (figure 5(g)) are determined independently of
the definition of li .
We have performed additional numerical simulations in which every presynaptic
spike spends for τ /5 = T /30 before exciting the postsynaptic neurons. Because T =
50–200 ms, the corresponding synaptic delay is equal to a few milliseconds. The value
of gc−stdp hardly changes with this synaptic delay (results not shown).
To examine the effect of heterogeneity in oscillators, we pick the intrinsic frequency
of each oscillator from the gaussian distribution with mean ω = 8.1 and standard devi-
ation 0.1. Time courses of r are shown in figure 5(h). We estimate 1.2 < gc−stdp < 1.5,
implying the robustness of our results against heterogeneity. We remark that r con-
verges to a positive level when g < gc−stdp . This is because, even if the oscillators
disconnect from the pacemaker, some oscillators form feedforward networks of small
size in which fast oscillators entrain and speed up slow oscillators. If the heterogeneity
is even larger so that some oscillators are as fast as the pacemaker, frequency syn-
chrony seeding from the pacemaker would be difficult because the pacemaker and fast
oscillators compete in entraining slow oscillators.
4.4 Symmetric Learning Window
Now we examine symmetric STDP. We set gmax = 200 > gc ∼= 100.7 so that frequency
synchrony with small phase lags is achieved if gj i = gmax for all (j, i) ∈ E . For the
network same as that used in section 4.3, evolution of synaptic weights are summarized
in figures 7(a, b) and (c, d) for g (0) = 140 and g (0) = 150, respectively. Because
22
the oscillators share an identical natural frequency, the phases are fairly close among
them even with weak coupling. The synapses among these oscillators are potentiated.
Accordingly, for both values of g (0), the average synaptic weight initially increases
(dotted lines in figure 7(a, c)). Note that the forward weights (Gf , thick solid lines
in figure 7(a, c)) and the backward weights (Gb , thin solid lines in figure 7(a, c))
are equally potentiated. Accordingly, the distance between the oscillators is initially
shortened to result in a decrease in L (figure 7(e), t ≤ 2000).
When g (0) = 140, the average synaptic weight stops increasing at a value slightly
smaller than gmax = 200 (dotted line in figure 7(a)). This is because the synapses link-
ing the pacemaker to the oscillators have not been potentiated. Actually, G0
f (thick line
in figure 7(b)) and G0
b (thin line in figure 7(b)) decrease to eventually decouple the os-
cillators from the pacemaker. Consequently, L diverges (uppermost line in figure 7(e)),
and frequency synchrony is eventually lost (figure 7(f )).
f (thick line in figure 7(d)) and G0
When g (0) = 150, G0
b (thin line in figure 7(d)) as
well as Gf (thick solid line in figure 7(c)) and Gb (thin solid line in figure 7(c)) increase.
Consequently, L continues to decrease to reach the minimum possible value for which
gj i = gmax is achieved for most synapses (lowermost line in figure 7(e)). The synchrony
measure r stays near unity throughout (figure 7(f )). In fact, approximate phase syn-
chrony as well as frequency synchrony has been realized quickly. The synchrony arises
not owing to STDP but to sufficiently strong initial coupling.
Based on figures 7(e) and (f ), which show time courses of L and r for several values
of g (0), we estimate 145 < gc−stdp < 146. This value of gc−stdp is much larger than
the case of asymmetric STDP and comparable to gc ∼= 100.7, that is, the threshold for
frozen synapses.
The fact that symmetric STDP does not promote frequency synchrony man-
ifests the importance of the feedforwardness of networks.
In general,
forward
synapses promote frequency synchrony, whereas backward synapses hamper it
23
(Kori and Mikhailov, 2004). Symmetric STDP does not independently control for-
ward synaptic weights and backward synaptic weights. Consequently, it cannot get rid
of backward synapses without sacrificing forward synapses. Final network structure
is shown for g (0) = 150 in figure 7(g). In contrast to the case of asymmetric STDP
(figure 5(g)), many backward synapses (thick lines) remain. Under symmetric STDP,
frequency synchrony is due to strong mutual interaction but not to formation of a
feedforward network.
5 Networks of Pulse-coupled Spiking Neurons
To inspect whether the results obtained for coupled phase oscillators qualitatively ap-
ply to more general neuron models, we numerically simulate pulse-coupled spiking
neurons under STDP. We adopt a two-dimensional neuron model (Izhikevich, 2003;
Izhikevich et al., 2004). The subthreshold dynamics of the i-th neuron are described
by
vi = 0.04v 2
i + 5vi + 140 − ui − Isyn,i − Iext,i ,
ui = a(bvi − ui ),
(36)
(37)
where vi is the membrane potential (mV), ui denotes the recovery variable that evolves
slowly relative to vi , and the time unit is millisecond. The spiking mechanism is im-
plemented by resetting the dynamical variables to (vi , ui) = (c, d) as soon as vi exceeds
30 mV. We set a = 0.02, b = 0.2, c = −65, and d = 8, which are standard parameter
values for modeling pyramidal neurons (Izhikevich, 2003; Izhikevich et al., 2004).
The input Iext,i and Isyn,i are the external bias input and the synaptic input, re-
spectively. We set Iext,0 = 8.4 and Iext,i = 8 (1 ≤ i ≤ n − 1). The inherent firing rate
of the pacemaker (= 18.8 Hz) is about 5 % higher than that of the oscillators (= 17.9
Hz). In figure 8, example traces of the membrane potentials of the pacemaker (solid
line) and that of an oscillator (dashed line) are shown.
24
The synaptic input Isyn,i is composed of superposition of incident spikes from the
neurons presynaptic to the i-th neuron. A presynaptic spike of the j -th neuron ((j, i) ∈
E ) is assumed to change the postsynaptic membrane potential according to the time
course given by the alpha function:
t ≥ 0,
where t = 0 corresponds to the spike time. We set α = 1, so that the unit synaptic
gj is(t) = gj iα2 te−αt ,
(38)
input s(t) peaks at t = 1/α = 1 ms and then decays slowly. We set A+ = 0.09,
A− = 0.1, and τ = 10 ms.
The random network with n = 100 used in the following simulations are the same
as that used in section 4. We set the initial synaptic weight gj i = g (0) for all the
synapses. The initial values of vi and ui are independently chosen according to the
uniform distributions on [−75, −50] and [−8, −6], respectively. Under these conditions,
we track time courses of Gf , Gb , Gl , G0
f , G0
b , L defined in section 4.2, and r = r([t1 , t2 ])
r =
.
redefined based on spike counts:
n−1 Pn−1
1
i=1 (number of spikes from the i-th neuron)
(number of spikes from the pacemaker)
Frequency synchrony yields r ∼= 1. If the pacemaker and the oscillators fire indepen-
dently, r is the ratio of the single-neuron firing rate to the pacemaker firing rate. We
set t2 − t1 = 10000 ms.
With these parameter values, we first determined gc without STDP. We numerically
obtained gc ∼= 1.6 for the network of one pacemaker and one oscillator, and gc ∼= 38
for the random network. In the following simulations with asymmetric STDP, we set
(39)
gmax = 35 < gc so that uniform increases in gj i do not cause synchronization. Frequency
synchrony requires feedforward network structure.
For homogeneous initial synaptic weights g (0) = 5 and g (0) = 10, evolution of
synaptic weights via asymmetric STDP is shown in figures 9(a, b) and 9(c, d), re-
spectively. For g (0) = 5, Gf (thick solid line in figure 9(a)) and G0
f (thick line in
25
figure 9(b)) do not grow during the course of plasticity, similar to figures 5(a, b). The
oscillators disconnect from the pacemaker, and frequency synchrony is not realized
(figure 9(f )). For g (0) = 10, the forward connection from the pacemaker to the set of
oscillators is established (thick line in figure 9(d)), and backward connection is gradu-
ally removed (thin line in figure 9(d)), similar to figures 5(c, d). As a result, frequency
synchrony is reached (figure 9(f )). Based on figures 9(e) and 9(f ), which respectively
show L and r for different g (0), we estimate gc−stdp ∼= 7, which is much smaller than
gc ∼= 38. Note that g (0) = 7 is a marginal case, which yields a long transient before
frequency synchrony is reached. Frequency synchrony is not induced with, for exam-
ple, g (0) = 10 < gc if synapses are frozen. These numerical simulations confirm that
the results derived in the previous sections apply to networks of pulse-coupled spiking
neurons.
6 Discussion
We have shown that asymmetric STDP greatly reduces the threshold for frequency
synchronization of neural networks with a pacemaker. This reduction is efficient
particularly when LTP and LTD are nearly balanced, as assumed for stabilization
of synaptic weights in previous literature (Kempter et al., 1999; Song et al., 2000;
Song and Abbott, 2001; van Rossum et al., 2000). Our analytical results for two-
oscillator networks provide theoretical understanding of STDP-induced synchrony of
two-body networks with real neurons (Nowotny et al., 2003) and with Hodgkin-Huxley
neurons (Zhigulin et al., 2003). Our numerical results for large networks extend ear-
lier numerical simulations (Zhigulin and Rabinovich, 2004) and provide mechanisms of
synchrony.
More microscopically, we have shown that STDP guides formation of feedforward
networks originating from the pacemaker (figure 5(g)). By eliminating backward con-
nection, frequency synchrony is promoted in terms of required synaptic weights. Net-
26
works self-organize by asymmetric STDP so that upstream neurons entrain downstream
neurons. Even though engineered learning algorithms can promote formation of feed-
forward networks (Kori and Mikhailov, 2006), asymmetric STDP naturally achieves
this goal. Facilitation of frequency synchrony does not occur for symmetric STDP,
which cannot suppress backward synapses without sacrificing forward synapses.
In recurrent networks, synaptic delay may destabilize otherwise stable synchrony,
leading to oscillatory or chaotic population dynamics (e.g. Gerstner and van Hemmen,
1993; Gerstner, 2000; Timme et al., 2002). However, our numerical simulations suggest
that this is not the case in our system, which can be explained as follows. In the phase
oscillator model, the effect of delay can be replaced in a good approximation by the
phase shift in the coupling function (Kori and Kuramoto, 2001; Kori, 2003). Dynamics
of the oscillator system under consideration do not change qualitatively for a large class
of the coupling function (Kori and Mikhailov, 2006), which effectively includes the case
of synaptic delay. Although a synaptic delay enlarges the phase difference between
connected neurons, the oscillator dynamics in our model are thus robust against delay.
Another possible complication is that synaptic delay may change synaptic evolution
because it could interact with the learning window. However, since the delay simply
increases the phase difference between connected neurons, the causality of spike timing
does not change even with delay, as corroborated by our numerical experiments.
In terms of network structure, the feedforward structure is distinct from pruning
of synapses in a predefined unidirectional network with many presynaptic neurons
pro jecting to a single postsynaptic neuron (Kempter et al., 1999; Song et al., 2000;
Song and Abbott, 2001; van Rossum et al., 2000).
It also differs from multipar-
tite networks each part of which forms a cluster of synchronously firing neurons
(Horn et al., 2000; Levy et al., 2001).
In a sense, feedforward structure and hierar-
chy are straightforward consequences of asymmetric STDP (Song and Abbott, 2001),
which opts for causality. We stress that feedforward structure is naturally organized
27
when a network has a distinct pacemaker. The idea of growing feedforward structure
by unsupervised learning dates back to pioneering work by Bienenstock (1991, 1995),
which employed Hebbian plasticity. We have analyzed the network formation in detail
under asymmetric STDP. In real neural networks, there may be multiple pacemakers
as well as a huge number of follower neurons. It is straightforward to extend our results
to the case in which a collection of neurons in a network serves as a pacemaker. Pace-
maker neurons are relevant to, for example, regulation of respiration, internal clock,
and Parkisonian diseases (see section 1 for references). In these brain regions, pace-
maker neurons may recruit downstream neurons for frequency synchrony in order to,
for example, amplify rhythmic activity.
Formation of feedforward structure could occur even when predetermined pacemak-
ers are absent. In this case, neurons with relatively high natural frequencies may play
a role of pacemaker. Backward connection to these fast neurons, which would per-
turb their periodic firing, can be eventually eliminated by asymmetric STDP. Then,
the fast neurons can serve as distinct pacemakers. Regardless of the initial presence
of pacemakers, asymmetric STDP creates frequency synchrony, which can be called
feedforward synchrony. This mechanism of synchrony differs from that of synchrony
based on mutual coupling (Kuramoto, 1984).
Our results do not suggest that asymmetric STDP promotes phase synchrony,
namely, spike synchrony. This is in contrast to the finding that phase synchrony is
caused by asymmetric STDP (Karbowski and Ermentrout, 2002). In their work, fixed
inhibitory coupling as well as excitatory coupling with asymmetric STDP was assumed.
Perfectly balanced LTP and LTD, at least as the average, is a key condition for the
maintenance of bidirectional connectivity and phase synchrony. By contrast, we as-
sumed that LTD is stronger than LTP. This yields a considerable decrease in the
threshold for synchrony. However, this synchrony is frequency synchrony but not spike
synchrony.
28
With symmetric STDP, neurons whose spike times are close are likely to bind
together. Then, in addition to frequency synchrony, approximate phase synchrony
whose time resolution is specified by the width of the learning window can develop
(Seliger et al., 2002). In some situations, neurons divide into clusters in each of which
rough spike synchrony is maintained (Masuda and Aihara, 2004). Unlike asymmetric
STDP, symmetric STDP does not lessen the threshold for synchrony.
When feedforward frequency synchrony is achieved, neurons at different distances
from the pacemaker fire asynchronously. On top of that, phase synchrony can be ob-
served for neurons receiving the common signal. For example, the neurons directly
connected to the pacemaker are excited by the common drive from the pacemaker,
so that they are synchronized in phase. Likewise, neurons with the same distance
from the pacemaker tend to fire simultaneously.
Indeed, figure 6(b) and its mag-
nification in figure 6(c) indicate that clusters of phase-synchronized neurons can be
aligned according to the distance from the pacemaker (Kori and Mikhailov, 2004).
Even though spike time difference between neurons with different distances is usu-
ally small, the order of firing is fixed and reproducible. The pacemaker triggers a
volley of spikes, which travels down the hierarchy delineated by the distance. This
phenomenon is consistent with propagation of synfire volley through feedforward
neural networks in the excitable regime (Bienenstock, 1995; Diesmann et al., 1999;
Reyes, 2003; Vogels and Abbott, 2005). However, stably embedding synfire volley in
recurrent networks is usually difficult (Mehring et al., 2003). It needs, for example, se-
lective enhancement of forward synapses by 10-fold, which corresponds to large evoked
EPSPs of 8 mV (Vogels and Abbott, 2005). With asymmetric STDP, forward synapses
are enhanced. In addition, automatic elimination of backward synapses appreciably
lessens the forward synaptic strength (or the size of EPSP) needed for stable synfire
volley.
We have shown that the facilitation of frequency synchrony by STDP is robust
29
against some heterogeneity in the inherent firing frequency of the neurons. If synap-
tic delays or neurons are strongly heterogeneous, we would obtain more complex but
reproducible spatiotemporal spike patterns (Izhikevich et al., 2004; Izhikevich, 2006;
Lengyel et al., 2005).
Oscillatory neurons can model,
for example, temporal coding of place cells
(Mehta et al., 2002) and hippocampal associative memory (Lengyel et al., 2005). By
contrast, many neural circuits operate in the excitable regime,
in which neurons
are not spontaneously oscillatory.
Investigation of the excitable case is warranted
for future studies. However, we believe that the conclusion that asymmetric STDP
but not symmetric STDP induces feedforward synchrony generalizes to the ex-
citable case, as is consistent with previous numerical work (Song and Abbott, 2001;
Zhigulin and Rabinovich, 2004).
Acknowledgments
We thank Brent Doiron and Taro Toyoizumi for critical reading of the manuscript and
Tohru Ikeguchi and Tomoya Suzuki for discussion. Naoki Masuda thanks the Special
Postdoctoral Researchers Program of RIKEN. Hiroshi Kori thanks financial support
from the Humboldt foundation (Germany) and from 21st Century COE program “Non-
linearity via Singularity” in Department of Mathematics, Hokkaido University.
References
Abbott LF, Nelson SB (2000) Synaptic plasticity: taming the beast. Nat. Neurosci.
Supp. 3:1178–1183.
Bienenstock E (1991) Notes on the growth of a composition machine. Proceedings of
the First Interdisciplinary Workshop on Compositionality in Cognition and Neural
30
Networks. Abbaye de Royaumont, France (eds. Andler D, Bienenstock E, Laks B):
25–43.
Bienenstock E (1995) A model of neocortex. Network: Computation in Neural Systems
6:179–224.
Braunstein LA, Buldyrev SV, Cohen R, Havlin S, Stanley HE (2003) Optimal paths
in disordered complex networks. Phys. Rev. Lett. 91:168701.
Bell CC, Han VZ, Sugawara Y, Grant K (1997) Synaptic plasticity in a cerebellum-like
structure depends on temporal order. Nature 387:278–281.
Bi G, Poo M (1998) Synaptic modifications in cultured hippocampal neurons: de-
pendence on spike timing, synaptic strength, and postsynaptic cell type. J. Neurosci.
18(24):10464–10472.
Buzs´aki G, Draguhn A (2004) Neuronal oscillations in cortical networks. Science
304:1926–1929.
Dan Y, Poo M (1992) Hebbian depression of isolated neuromuscular synapses in vitro.
Science 256:1570–1573.
Diesmann M, Gewaltig M-O, Aertsen A (1999) Stable propagation of synchronous
spiking in cortical neural networks. Nature 402:529–533.
Froemke RC, Dan Y (2002) Spike-timing-dependent synaptic modification induced by
natural spike trains. Nature 416:433–438.
Gerstner W, van Hemmen JL (1993) Coherence and incoherence in a globally coupled
ensemble of pulse-emitting units. Phys. Rev. Lett. 71:312–315.
Gerstner W, Kempter R, van Hemmen JL, Wagner H (1996) A neuronal learning rule
for sub-millisecond temporal coding. Nature 383:76–78.
31
Gerstner W (2000) Population dynamics of spiking neurons:
fast transients, asyn-
chronous states, and locking. Neural Comput. 12:43–89.
Gerstner W, Kistler WM (2002) Spiking neuron models. Cambridge University Press,
Cambridge.
Glass L, Mackey MC (1988) From Clocks to Chaos – the Rhythms of Life. Princeton
University Press, Princeton.
Hansel D, Mato G, Meunier C (1993) Phase dynamics for weakly coupled Hodgkin-
Huxley neurons. Europhys. Lett. 23(5):367–372.
Hansel D, Mato G, Meunier C (1995) Synchrony in excitatory neural networks. Neural
Comput. 7:307–337.
Horn D, Levy N, Meilijson I, Ruppin E (2000) Distributed synchrony of spiking neu-
rons in a Hebbian cell assembly. In S. A. Solla, T. K. Leen, & K. -R. Muller (Eds.),
Advances in Neural Information Processing Systems 12:129–135. MIT Press, Cam-
bridge, MA.
Hutcheon B, Yarom Y (2000) Resonance, oscillation and the intrinsic frequency pref-
erences of neurons. Trends in Neurosci. 23(5):216–222.
Izhikevich EM (2003) Simple model of spiking neurons. IEEE Trans. Neur. Netw.
14(6):1569–1572.
Izhikevich EM, Gally JA, Edelman GM (2004) Spike-timing dynamics of neuronal
groups. Cereb. Cort. 14(8):933–944.
Izhikevich EM (2006) Polychronization: computation with spikes. Neural Comput.
18:245–282.
32
Jefferys JGR, Traub RD, Whittington MA (1996) Neuronal networks for induced ‘40
Hz’ rhythms. Trends in Neurosci. 19(5):202–208.
Karbowski J, Ermentrout GB (2002) Synchrony arising from a balanced synaptic
plasticity in a network of heterogeneous neural oscillators. Phys. Rev. E 65:031902.
Kempter R, Gerstner W, van Hemmen JL (1999) Hebbian learning and spiking neu-
rons. Phys. Rev. E 59:4498–4514.
Kori H, Kuramoto Y (2001). Slow switching in globally coupled oscillators: robustness
and occurrence through delayed coupling. Phys. Rev. E 63:046214.
Kori H (2003) Slow switching in a population of delayed pulse-coupled oscillators.
Phys. Rev. E 68:021919.
Kori H, Mikhailov AS (2004) Entrainment of randomly coupled oscillator networks
by a pacemaker. Phys. Rev. Lett. 93:254101.
Kori H, Mikhailov AS (2006) Strong effects of network architecture in the entrainment
of coupled oscillator systems. Phys. Rev. E 74:066115.
Kuramoto Y (1984). Chemical oscillations, waves, and turbulence. Springer-Verlag,
Berlin.
Kuramoto Y (1991) Collective synchronization of pulse-coupled oscillators and ex-
citable units. Physica D 50:15–30.
Lengyel M, Kwag J, Paulsen O, Dayan P (2005) Matching storage and recall: hip-
pocampal spike timing-dependent plasticity and phase response curves. Nat. Neurosci.
8:1677–1683.
Levy N, Horn D, Meilijson I, Ruppin E (2001) Distributed synchrony in a cell assembly
of spiking neurons. Neural Networks 14:815–824.
33
Markram H, Lubke J, Frotscher M, Sakmann B (1997) Regulation of synaptic efficacy
by coincidence of postsynaptic APs and EPSPs. Science 275:213–215.
Masuda N, Aihara K (2004). Self-organizing dual coding based on spike-time-
dependent plasticity. Neural Comput. 16:627–663.
Mehring C, Hehl U, Kubo M, Diesmann M, Aertsen A (2003) Activity dynamics
and propagation of synchronous spiking in locally connected random networks. Biol.
Cybern. 88:395–408.
Mehta MR, Lee AK, Wilson MA (2002) Role of experience and oscillations in trans-
forming a rate code into a temporal code. Nature 417:741–746.
Nishiyama M, Hong K, Mikoshiba K, Poo M, Kato K (2000) Calcium stores regulate
the polarity and input specificity of synaptic modification. Nature 408:584–588.
Nowotny T, Zhigulin VP, Selverston AI, Abarbanel HDI, Rabinovich MI (2003) En-
hancement of synchronization in a hybrid neural circuit by spike-timing dependent
plasticity. J. Neurosci. 23(30):9776–9785.
Pikovsky A, Rosenblum M, Kurths J (2001) Synchronization – A Universal Concept
in Nonlinear Sciences. Cambridge University Press, Cambridge, UK.
Plenz D, Kitai ST (1999) A basal ganglia pacemaker formed by the subthalamic
nucleus and external globus pallidus. Nature 400:677–682.
Ramirez JM, Tryba AK, Pena F (2004) Pacemaker neurons and neuronal networks:
an integrative view. Curr. Opinion in Neurobiol. 14:665–674.
Reyes AD (2003) Synchrony-dependent propagation of firing rate in iteratively con-
structed networks in vitro. Nature Neurosci. 6(6):593–599.
34
Ritz R, Sejnowski TJ (1997) Synchronous oscillatory activity in sensory systems: new
vistas on mechanisms. Curr. Opinion in Neurobiol. 7:536–546.
Seliger P, Young SC, Tsimring LS (2002) Plasticity and learning in a network of
coupled phase oscillators. Phys. Rev. E 65:041906.
Shouval HZ, Bear MF, Cooper LN (2002) A unified model of NMDA receptor-
dependent bidirectional synaptic plasticity. Proc. Natl. Acad. Sci. USA, 99(16):10831–
10836.
Singer W, Gray CM (1995) Visual feature integration and the temporal correlation
hypothesis. Ann. Rev. Neurosci. 18:555–586.
Song S, Miller KD, Abbott LF (2000) Competitive Hebbian learning through spike-
timing-dependent synaptic plasticity. Nat. Neurosci. 3(9):919–926.
Song S, Abbott LF (2001) Cortical development and remapping through spike timing-
dependent plasticity. Neuron 32:339–350.
Timme M, Wolf F, Geisel T (2002) Prevalence of unstable attractors in networks of
pulse-coupled oscillators. Phys. Rev. Lett. 89:154105.
van Rossum MCW, Bi GQ, Turrigiano GG (2000) Stable Hebbian learning from spike
timing-dependent plasticity. J. Neurosci. 20(23):8812–8821.
Vogels TP, Abbott LF (2005) Signal propagation and logic gating in networks of
integrate-and-fire neurons. J. Neurosci. 25(46):10786–10795.
Winfree AT (1980) The Geometry of Biological Time. Springer-Verlag, New York.
Woodin MA, Ganguly K, Poo M (2003) Coincident pre- and postsynaptic activity
modifies GABAergic synapses by postsynaptic changes in Cl− transporter activity.
Neuron 39:807–820.
35
Zhang LI, Tao HW, Holt CE, Harris WA, Poo M (1998) A critical window for coop-
eration and competition among developing retinotectal synapses. Nature 395:37–44.
Zhigulin VP, Rabinovich MI, Huerta R, Abarbanel HDI (2003) Robustness and en-
hancement of neural synchronization by activity-dependent coupling. Phys. Rev. E
67:021901.
Zhigulin VP, Rabinovich MI (2004) An important role of spike timing dependent
synaptic plasticity in the formation of synchronized neural ensembles. Neurocomputing
58–60:373–378.
36
P
T
L
f
o
t
n
u
o
m
a
0
-0.3
-0.15
0.15
0.3
0
tpost - tpre
Figure 1: Asymmetric (solid line) and symmetric (dashed line) learning windows of
STDP as a function of tpost − tpre , namely, the postsynaptic spike time relative to the
presynaptic spike time.
(a)
g
φ
0
Ω
pacemaker
φ
1
ω
oscillator
(b)
φ
1
ω
∆+ ω
oscillator
g
2
g 1
φ
2
ω
oscillator
Figure 2: Schematic diagrams showing (a) the network of one pacemaker and one
oscillator, and (b) the network of two oscillators.
Figure 3: gc−stdp for the network with one pacemaker and one oscillator. The evaluation
by equation (16) (solid line), that by equation (18) (dotted line), and gc−stdp determined
by numerical simulations of the coupled phase oscillators (circles) are compared. We
set gmax = 1.25.
0.25
0.2
0.15
0.1
0.05
2
g
l
a
c
i
t
i
r
c
0
0
0.2
0.6
0.8
0.4
g1
Figure 4: The critical g2(0) as a function of g1(0) for the network with two oscillators
(and no pacemaker). Three lines correspond to A+/A− = 0.96 (thickest line), 0.9, and
0.8 (thinnest line).
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
(a)
1.5
1
0.5
0
0
(c)
15
10
5
0
total
Gf Gb Gl
5000
10000 15000 20000
time
total
Gf Gb Gl
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
0
5000
10000 15000 20000
time
(b)
1.5
1
0.5
0
0
5000
10000 15000 20000
time
(d)
15
10
5
0
0
5000
10000 15000 20000
time
Figure 5: Results for 100 randomly coupled oscillators sub ject to asymmetric STDP.
Evolution of the synaptic-weight order parameters are shown for (a, b) g (0) = 0.7
and (c, d) g (0) = 1.5. In (a) and (c), Gf (thick solid lines), Gb (thin solid lines), Gl
(moderate solid lines), and the average weight (dotted lines) are shown. In (b) and (d),
f (thick lines) and G0
G0
b (thin lines) are indicated. Time courses of (e) L and (f ) r are
compared for g (0) = 0.7, 0.9, 1, 1.2, and 1.5. In (e), lower lines correspond to larger
g (0). (g) Final network structure for g (0) = 1.5. Only the synapses (j, i) ∈ E with
gj i > g (0) are presented. The pacemaker is labeled P . Forward edges and backward
edges are indicated by thin lines and thick lines, respectively. The network is drawn
by Pa jek (http://vlado.fmf.uni-lj.si/pub/networks/pa jek/ ). (h) Time courses of r for
some values of g (0) when the oscillators are heterogeneous.
(e)
50
40
30
20
10
0
L
(f)
1
0.75
g(0)=1.2,1.5
g(0)=1
r
0.5
g(0)=0.9
0.25
g(0)=0.7
0
0
5000
10000 15000 20000
time
5000 10000 15000 20000
time
(g)
P
(h)
1
0.75
r
0.5
0.25
0
g(0)=1.5
g(0)=1.0
g(0)=1.2
g(0)=0.8
5000 10000 15000 20000
time
(a)
100
75
50
25
0
0
x
e
d
n
i
r
o
t
a
l
l
i
c
s
o
4
2
time
(c)
100
75
50
25
0
x
e
d
n
i
r
o
t
a
l
l
i
c
s
o
(b)
100
75
50
25
0
x
e
d
n
i
r
o
t
a
l
l
i
c
s
o
li=4
3
19996
2
1
19998
time
20000
li=4
3
2
19999.2
19999.3
1
time
Figure 6: Rastergrams of the oscillators under asymmetric STDP in (a) initial and
(b) final cycles. We set g (0) = 1.5. The oscillators are aligned according to their
distances li from the pacemaker, which is calculated at time 0 in (a) and 19995 in (b).
After sufficient time, li is quantized, and the values of li are shown in (b). (c) is a
magnification of (b).
(a)
200
150
100
50
0
0
(c)
200
150
100
50
0
0
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
(b)
200
150
100
50
0
0
(d)
200
150
100
50
0
0
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
total
Gf
Gb
Gl
5000
10000 15000 20000
time
total
Gl
Gf, Gb
5000
10000 15000 20000
time
5000
10000 15000 20000
time
5000
10000 15000 20000
time
Figure 7: Results for 100 randomly coupled oscillators sub ject to symmetric STDP.
Evolution of the synaptic weights are shown for g (0) = 140 (a, b) and g (0) = 150 (c,
d). See the caption of figure 5 for legends. In (c), Gf (thick solid line) and Gb (thin
solid line) overlap almost completely. Time courses of (e) L and (f ) r are compared for
g (0) = 140, 145, 146, 148, and 150. In (e), lower lines correspond to larger g (0). (g)
Final network structure for g (0) = 150.
(e)
6
5
4
3
L
(f)
1
0.75
g(0)=146,148,150
r
0.5
g(0)=145
0.25
g(0)=140
0
5000
10000
time
15000
20000
0
0
5000 10000 15000 20000
time
(g)
P
)
)
V
V
m
m
(
(
v
v
-30
-30
-50
-50
-70
-70
49.5 49.6 49.7 49.8 49.9
49.5 49.6 49.7 49.8 49.9
time (sec)
time (sec)
50
50
Figure 8: Sample traces of v0 (solid line) and v1 (dashed line) when the spiking neurons
are uncoupled.
(b)
20
15
10
5
0
(d)
20
15
10
5
0
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
total
Gf
Gb
Gl
0
100
300
200
time (s)
400
500
total
Gf
Gb
Gl
0
100
200
300
time (s)
400
500
0
100
300
200
time (s)
400
500
0
100
200
300
time (s)
400
500
(a)
20
15
10
5
0
(c)
20
15
10
5
0
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
s
t
h
g
i
e
w
c
i
t
p
a
n
y
s
(e)
15
12
L
9
6
3
(f)
1
0.98
r
0.96
g(0)=10
g(0)=8
g(0)=7
g(0)=6
400
500
0
100
200
300
time (s)
400
500
0
100
200
300
time (s)
Figure 9: Results for 100 randomly coupled spiking neurons sub ject to asymmetric
STDP. Evolution of the synaptic weights are shown for (a, b) g (0) = 5 and (c, d)
g (0) = 10. See the caption of figure 5 for legends. Time courses of (e) L and (f ) r are
compared for g (0) = 5, 6, 7, 8, and 10. In (e), lower lines correspond to larger g (0).
|
1806.03047 | 1 | 1806 | 2018-06-08T09:40:32 | On sound-based interpretation of neonatal EEG | [
"q-bio.NC",
"cs.SD",
"eess.AS",
"stat.AP"
] | Significant training is required to visually interpret neonatal EEG signals. This study explores alternative sound-based methods for EEG interpretation which are designed to allow for intuitive and quick differentiation between healthy background activity and abnormal activity such as seizures. A novel method based on frequency and amplitude modulation (FM/AM) is presented. The algorithm is tuned to facilitate the audio domain perception of rhythmic activity which is specific to neonatal seizures. The method is compared with the previously developed phase vocoder algorithm for different time compressing factors. A survey is conducted amongst a cohort of non-EEG experts to quantitatively and qualitatively examine the performance of sound-based methods in comparison with the visual interpretation. It is shown that both sonification methods perform similarly well, with a smaller inter-observer variability in comparison with visual. A post-survey analysis of results is performed by examining the sensitivity of the ear to frequency evolution in audio. | q-bio.NC | q-bio | On sound-based interpretation of neonatal EEG
S. Gómez, M. O'Sullivan, E. Popovici
Electrical and Electronic Engineering,
University College Cork,
, Ireland
[email protected]
Abstract-Significant training is required to visually interpret
neonatal EEG signals. This study explores alternative sound-based
methods for EEG interpretation which are designed to allow for
intuitive and quick differentiation between healthy background
activity and abnormal activity such as seizures. A novel method
based on frequency and amplitude modulation (FM/AM) is
presented. The algorithm is tuned to facilitate the audio domain
perception of rhythmic activity which is specific to neonatal
seizures. The method is compared with the previously developed
phase vocoder algorithm for different time compressing factors. A
survey is conducted amongst a cohort of non-EEG experts to
quantitatively and qualitatively examine the performance of
sound-based methods
visual
interpretation. It is shown that both sonification methods perform
similarly well, with a smaller inter-observer variability in
comparison with visual. A post-survey analysis of results is
performed by examining the sensitivity of the ear to frequency
evolution in audio.
comparison with
the
in
Keywords- EEG sonification, neonatal seizure detection, phase
vocoder, frequency modulation, amplitude modulation.
I. INTRODUCTION
Neonatal seizures are the most common sign of acute
neonatal encephalopathy. Failure to detect such events and the
resulting lack of treatment can result in potentially life-
threatening outcomes. It is estimated that only 34% of neonatal
seizures present clinical signs, the remainder can only be
diagnosed using electroencephalography (EEG) monitoring [1].
Previous studies have shown EEG monitoring drastically
improves the percentage of correct diagnoses compared to
diagnosing seizures based on clinical manifestations alone [1].
Visual interpretation of EEG signals requires significant
training and this expertise is often available only in tertiary care
units. Even in such centres, it is not available on a 24-hour basis,
7 days a week. In order to ameliorate this situation, a simpler
form of EEG called amplitude integrated EEG (aEEG) is often
used where 1-2 channels of EEG are recorded and converted to
a compressed trend of EEG amplitude over time. However, the
use of aEEG amongst the neonatal population has several
limitations, waveform information is lost and its effectiveness in
seizure detection varies with experience and is poor when
compared to EEG [2]. Significant research has been conducted
in the area of objective detection of seizure events using artificial
intelligence [3-6]. These algorithms aim to provide clinicians
with a support in diagnosing abnormal EEG activity, and can
achieve accurate seizure detection, though no algorithm will
detect all seizures or perform without any false alarms.
However, objective methods need to be accompanied with
Research was supported in part by Wellcome Trust Seed Award
(12/RC/2272), SFI TIDA
(200704/Z/16/Z), SFI
INFANT Centre
(17/TIDA/504) and HRB (KEDS-2017-020).
978-1-5386-6046-1/18/$31.00 ©2018 IEEE
S. Mathieson, G. Boylan, A. Temko
Irish Centre for Fetal and Neonatal Translational Research,
University College Cork, Ireland
subjective methods such as visualisation of EEG traces to assure
that the clinician is engaged in the final decision making.
Methods for EEG sonification aim to facilitate EEG
interpretation in a quicker and easier way. In fact, the perception
of evolution in frequency over time and the presence of structure
(rhythm), which is characteristic of abnormal events such as
seizures [7], are more identifiable with hearing rather than with
visual aids. Sound-based interpretation aims to mitigate this
effect and release the visual sense for other tasks [8]. Several
techniques of EEG sonification have been proposed for various
applications in the area of adult EEG interpretation, in particular
for detection of epileptic discharges [10-13]. Neonatal EEG is
different to adult EEG; preterm EEG is different to full term
EEG, and even preterm EEG differs dramatically across
gestational ages. In previous work, the EEG signal has been
usually bottle-necked to just a few core features which are used
to drive the sonification process. The auditory representations
proposed in this work preserve the structure and completeness
of the original EEG signal, consider and analyse the ability to
discriminate patterns in the chosen clinical task (intelligibility)
and user preferences as a part of the auralisation process. The
analysis performed is relevant for the new proposed sonification
method as well as the phase vocoder based algorithm which was
previously developed for neonatal EEG interpretation [14, 15].
In this paper, the use of frequency and amplitude modulation
(FM/AM) in the space of neonatal EEG sonification is explored.
FM has been previously used in the space of adult EEG
sonification as in [16]. The algorithm implemented in this paper
differs in its additional use of AM and customised compression
techniques. The proposed algorithm is tuned to increase the
sensitivity of the human ear to the presence of rhythm. The phase
vocoder algorithm is further explored and its parameters are
fine-tuned in order to obtain the highest accuracy of correctly
diagnosed neonatal EEG seizure activity. The survey is
conducted and the analysis of results is performed along with the
subsequent examination of the ear sensitivity to changes in
frequency.
II. METHODS AND MATERIALS
A. Database
A database of 100 seizure and 100 non-seizure segments was
created for the purpose of this study, to be used in the survey.
In order to create this database of 200 segments, a larger
database which was previously used for seizure detection
algorithm development was utilized [4, 8]. This large database
consists of long unedited multichannel EEG recordings from 18
newborns totalling 816 hours of duration with 1389 seizures
Figure 1. Signal processing chart for PV and FM/AM. PV performs a spectra-to-spectra EEG-to-audio mapping, whereas FM/AM performs a waveform-to-
spectra EEG-to-audio mapping. Audio signals are time compressed by a factor of 10 in this example.
annotated by a neonatal neurophysiologist. The dataset
contained a wide variety of seizure types including both
electrographic-only and electro-clinical seizures of focal, multi-
focal and generalized types. Therefore, this dataset is truly
representative of the real-life situation in the NICU.
From the 1389 annotated seizures, only a small fraction was
annotated on a per-channel basis and rarely the whole seizure
from the beginning and the end had a per-channel annotation.
The seizure detection algorithm presented in [8] was first used
to create AI-based per-channel annotations for all seizures in the
dataset. Since the dataset was used to train the algorithm, its
performance in a patient-specific setting is very high [9]. On
seizure segments with overall seizure annotations, the maximum
probability of the seizure detection algorithm was used to select
a single channel in which the seizure had the highest support of
the algorithm. Only a single channel was selected in order to
minimize the chances of the error in this process, as seizures can
be focal, multi-focal and even migrate across channels.
From a pool of the per channel annotated seizures, the lowest
amplitude seizure events of 1-5 min duration were selected from
each patient to avoid seizures that were too short or too long,
seizures which were easily identifiable by amplitude alone and
to preserve the variety of patients, resulting in approximately
200 examples. These examples were further reviewed by an
experienced neurophysiologist to discard seizures which were
still too obvious or events which were incorrectly selected by the
algorithm. This resulted in 100 seizure examples which were
confirmed by the clinician.
Using the seizure annotations, from the same dataset of long
EEG recordings, the non-seizure segments were extracted to
have at least 5 min clearance from the nearest seizure.
Subsequently, 200 background EEG segments with the highest
amplitude and an algorithmic probability of seizure of at least
0.25 were selected, similarly to avoid any trivial segments,
segments which are easily identifiable by amplitude alone, and
to preserve equal representation of background activity from
each patient. These were then reduced manually to 100 by a
neonatal neurophysiologist, by confirming true non-seizure
segments and discarding evident high-amplitude artifacts. The
resulting 100 non-seizure segments were then randomly
cropped in order to have the same distribution of lengths for
seizure and non-seizure.
The resulting database was constructed therefore in such a
way that seizure identification could not be performed based on
amplitude or length and many non-seizure events had seizure-
like activity such as respiratory, ECG, sweating or movement
artifacts.
B. Neonatal EEG sonification algorithms
Seizures are partially defined by their slowly decreasing
dominant frequencies [17]. This evolution differentiates
seizures from artifacts such as respiratory or electrocardiogram
[18], which lack temporal frequency evolution.
The Phase Vocoder (PV) algorithm previously developed,
allows for the perception of such frequency changes over time
in the audio [14, 15]. The PV performs spectra to spectra
mapping by preserving a horizontal phase coherence [19] as
shown schematically in Fig. 1. The spectra of the EEG signal are
preserved in the audio signal obtained with PV.
The FM/AM algorithm presented in this study performs
waveform to spectra mapping as illustrated in Fig. 1. The
waveform amplitudes of the EEG signal become the spectra of
the audio signal. The presence of any structure in the signal
(repetitive waveforms) will be perceived as a rhythm in the
resultant audio.
Fig. 2. PV methodology: the EEG signal is processed by frames and reconstructed with coherent phases.
(1)
operate on a logarithmic scale, in decibels (dB). Thus, the input
the threshold is exceeded, the amplitude of the signal is reduced
the amplitude of the envelope is under the specified threshold
(cid:1831)(cid:3031)(cid:3003)(cid:3404)20(cid:3400)log (cid:1831)
distortion. For that purpose, the envelope (cid:4666)(cid:1831)(cid:4667) is calculated as the
convolution on the absolute value of (cid:1845) using an exponential
impulse response (cid:1860)(cid:3032)(cid:4666)(cid:1866)(cid:4667) with a decay time of 8s. Compressors
signal (cid:4666)(cid:1831)(cid:4667) is firstly converted to (cid:4666)(cid:1831)(cid:3031)(cid:3003)(cid:4667) using Eq. 1:
The envelope ((cid:1831)(cid:3031)(cid:3003)(cid:4667) is used to control the gain reduction. When
((cid:1846)(cid:3031)(cid:3003)), no gain reduction is applied to the signal. However, when
proportional to the compression ratio ((cid:1844)) as in Eq. 2:
0 (cid:1831)(cid:3031)(cid:3003)(cid:3407)(cid:1846)(cid:3031)(cid:3003)
(cid:3398)(cid:4672)1(cid:3398)(cid:2869)(cid:3019)(cid:4673)(cid:4666)(cid:1831)(cid:3031)(cid:3003)(cid:3398)(cid:1846)(cid:3031)(cid:3003)(cid:4667) (cid:1867)(cid:1872)(cid:1860)(cid:1857)(cid:1870)(cid:1875)(cid:1861)(cid:1871)(cid:1857)
For every 1 dB that the input signal exceeds (cid:1846)(cid:3031)(cid:3003), the amplitude
of the output will be reduced by an amount of (cid:4666)1(cid:3398)1/(cid:1844)(cid:4667). As
(cid:1844)(cid:3410)1, the gain reduction ((cid:1833)(cid:3031)(cid:3003)) is always ≤ 0. (cid:1833)(cid:3031)(cid:3003) is
subsequently converted back to the linear domain (cid:4666)(cid:1833)(cid:4667) by using
the reverse of Eq. 1. (cid:1833) is then used to control the gain of the
input (cid:1845) to obtain the output (cid:1843) as seen in Fig. 3.
In this algorithm, (cid:1846)(cid:3031)(cid:3003) of -20dB (5µV) was chosen, which is 1/10
of the expected full-scale amplitude. (cid:1844) was chosen to be 1.5,
The output, (cid:1843), is then amplified from ±50µV to ±1V, and hard-
meaning that a spike of 150µV would be reduced to 50µV,
keeping the output within the desired dynamic range.
(cid:1833)(cid:3031)(cid:3003)(cid:3404)(cid:4682)
limited as in Eq. 3 to ensure the signal is free from aliasing when
used as the modulator signal in the subsequent FM stage:
(2)
1
2(cid:3400)10(cid:2872)(cid:1843)(cid:3398)1
(cid:1843)(cid:3408)50(cid:2020)(cid:1848)
(cid:3398)50(cid:2020)(cid:1848)(cid:3407)(cid:1843)(cid:3407)50(cid:2020)(cid:1848)
(cid:1843)(cid:3407)(cid:3398)50(cid:2020)(cid:1848)
(cid:1839)(cid:4666)(cid:1843)(cid:4667)(cid:3404)(cid:3421)
The next subsections provide further details of both
algorithms.
1) Phase vocoder
PV is an analysis-synthesis method used for scaling
frequency while maintaining phase coherence and preserving
the spectral envelope of the original EEG signal [18]. The
developed PV algorithm converts the low frequency EEG signal
(0.5-8Hz) into the audible frequency domain (250-4000Hz). The
time length of the resulting output signal can be preserved to
match the input signal, or altered to be faster/slower. The effect
of time scaling on the discrimination of healthy versus non-
healthy EEG is explored in the survey. The algorithm can be
broken into four blocks.
Pre-processing: EEG signal is band-pass filtered between 0.5-
8Hz and down-sampled to 16Hz. These parameters were tuned
towards maximising the perception of seizure frequency
evolution and are different from those in [15].
Analysis: Short Time Fourier Transform (STFT) with a window
length/shift of 64s/16s was used. This results
the
decomposition of the EEG segment into magnitude and phase
for each frequency bin.
Processing: The magnitude of the signal is linearly interpolated
by a variable time compression factor and saved at an 8kHz
sampling rate. The phase is measured and unwrapped to track
the cumulative phase variation in order to preserve the phase
consistency across interpolated frames.
Synthesis: An inverse STFT and overlap-add is applied using the
window from the analysis stage, resulting in an 8kHz audio
signal.
in
used
synthesis,
commonly
and AM
2) Frequency & amplitude modulation
FM
in
telecommunication, is the process of varying a carrier wave
frequency and amplitude proportionally to a modulating wave.
In this work, the EEG is used as the modulating waveform. The
developed algorithm incorporates a number of digital signal
processing techniques as outlined in Fig. 3.
DC Removal: DC removal is achieved by applying a moving
average filter to the input signal. As the lowest frequency of
interest is 0.5Hz, a minimum window length of 2 s is required.
Pre-processing: EEG signal is band-pass filtered between 0.5-
7.5 Hz and down-sampled to 16Hz.
Compressor: EEG signals often contain amplitude spikes that
exceed the common amplitude range of ±50µV. In order to
avoid aliasing in the FM stage, compression is applied to reduce
the dynamic range of the input signal (cid:4666)(cid:1845)(cid:4667) with minimal
Fig. 3. FM/AM methodology: the EEG signal is filtered, compressed and modulated.
(3)
FM: The modulator signal (cid:4666)(cid:1839)(cid:4667) is firstly up-sampled (x 1000) to
[20], an exponential transform is applied to (cid:1839). If (cid:1839) is 0V, then
the value of the frequency ((cid:1832)) is 500Hz, if (cid:1839) is -1V, then (cid:1832) goes
down to 50Hz and if (cid:1839) is 1V, (cid:1832) goes up to 5kHz as in Eq. 4:
16kHz in order to achieve an audio bandwidth. As the human
hearing system perceives frequency on an exponential scale
F(cid:4666)(cid:1839)(cid:4667)(cid:3404)500(cid:3400)10(cid:3014)
(4)
The resulting output wave, (cid:1849), is a frequency-varying signal
AM: The envelope (cid:1831) is used to modulate the amplitude of (cid:1849).
[21], in which the amplitude variations of the EEG are mapped
to the frequency variations of the output signal, based on an
sine-wave FM synthesiser.
Thus, the amplitude of the EEG signal is also embedded in the
output audio.
3) Time compression
Both sonification algorithms can be used to manipulate the
time scale, so that the duration of the audio output can be
different to that of the input EEG. This is convenient for
generating faster sonification, allowing for the review of long
EEG signals in much shorter segments of audio. This
characteristic is also investigated for the discrimination and
perception of changes in the EEG morphology, which often
evolve too slowly to be perceived by the human ear in real time.
The time compression factor (cid:4666)(cid:1870)(cid:3410)1(cid:4667) defines the relationship
of (cid:1870)(cid:3404)10 will be used, (cid:1870)(cid:3404)1 results in direct 1:1 time scaling.
is decreased with respect to (cid:1870). In FM/AM, the time compression
is applied by simply reducing the value of the up-sampler by (cid:1870).
between the durations of the EEG and the resulting audio. For
instance, to scale 60 mins of EEG into 6 mins of audio, a value
In PV, time compression is applied in the processing stage when
interpolating the magnitude and phase. The interpolation factor
III. RESULTS & DISCUSSION
A. Survey design
In order to quantitatively assess the clinical performance of
the sonification algorithms, a survey was conducted amongst a
small cohort (N = 11) of non-EEG-expert participants. The
survey tests and compares the seizure identification rate using
seven tested scenarios – visual means (scenario 1), PV with a
time-compression factor of 1, 5, and 10 denoted by PVx1, PVx5,
and PVx10, scenarios 2 to 4; and FM/AM with a time-
compression factor of 1, 5 and 10 denoted by FMx1, FMx5 and
FMx10, respectively, scenarios 5 to 7.
Training data (3 seizures and 3 non-seizure segments) were
presented to each survey participant. The data were processed
according to the tested scenario for each method. During the
survey, 10 examples of neonatal EEG were randomly selected
from the created database for each scenario. The participant was
required to mark each example as seizure or non-seizure. Each
participant sequentially went through each of 7 tested scenarios.
The individual percentage scores of correct diagnoses were
calculated.
B. Sonification survey results
Fig. 4 shows the results of the survey as the mean and 95%
confidence intervals. It can be seen that both FM/AM and PV
sonification methods when speeded up by a factor of 10 perform
equally well, at 76% and 73% accuracy, respectively. These
methods slightly outperform visual interpretation, which obtains
69%. PV at a speed of 1, achieves the best real-time (1:1)
sonification result at 69%. The variance of visual interpretation
is
inconsistent accuracy across
participants.
largest denoting
the
its
Figure 4. Sonification survey results.
Fig. 5 shows the same results divided into 2 groups. Group
A (N = 3) includes the participants who performed below the
average correct rate for visual interpretation, and group B
includes the remainder of the cohort (i.e. above the average, N =
8). It can be seen that visual interpretation differs by more than
40% between the two groups. For sonification methods, the
difference is smaller than 15 % in the worst case. Moreover, the
same trends in the sonification methods are preserved across
both groups, with the increased speed generally resulting in the
increased performance of correct diagnoses. These results show
that audio interpretation is more consistent.
The participants were also asked about their preferred
method of EEG sonification. Three participants preferred PV
and eight preferred FM. Among them, two gave preference to
PV×10 and four chose FM×10. As the question was answered
without knowing the accuracy achieved with each method, this
Figure 5. Sonification survey results by groups. Plot (a) shows the group
with an accuracy below average on visual assesment. Plot (b) shows the
group with the accuracy above average.
result represents the aesthetic choice of the sonification which is
complementary to the accuracy-based assessment in Fig. 4.
C. Post-survey analysis of seizure frequency evolution
Previous studies have shown that during seizures, the
dominant frequencies in the EEG evolves over time, decreasing
in many cases [17]. However, that evolution can be so slow that
for some sonification methods the human hearing sense is
incapable of distinguishing a change in the frequency of the tone.
In order to generate a comparison between the evolution of the
seizure frequency and human hearing sensitivity of tone change,
a model of seizure frequency evolution and human sensitivity of
frequency change is required. A suitable figure of merit for
change in frequency is octaves per minute (oct/min), which is
the logarithmic measure of change in frequency over time. A
change of +1 oct/min or -1/oct/min implies that in 1 minute the
frequency doubles or is divided by 2, respectively. The relative
change of frequency over time can then be expressed as:
(5)
∆(cid:1858)(cid:4670)(cid:1867)(cid:1855)(cid:1872)(cid:4671)(cid:3404)log(cid:2870)(cid:3033)(cid:3118)(cid:3033)(cid:3117)
(cid:1871)(cid:1864)(cid:1867)(cid:1868)(cid:1857)(cid:4674)(cid:3042)(cid:3030)(cid:3047)(cid:3040)(cid:3036)(cid:3041)(cid:4675)(cid:3404)∆(cid:3033)(cid:4670)(cid:3042)(cid:3030)(cid:3047)(cid:4671)
∆(cid:3047)(cid:4670)(cid:3040)(cid:3036)(cid:3041)(cid:4671)
where ∆(cid:1858)=relative change in frequency in octaves, (cid:1858)(cid:2869), (cid:1858)(cid:2870)= start
& end frequencies, (cid:1871)(cid:1864)(cid:1867)(cid:1868)(cid:1857)= change in frequency over time.
(6)
The seizure and non-seizure examples were analysed with
respect to the evolution of the dominant frequencies in each EEG
epoch. Each EEG example is segmented into 2s epochs with
1 sample shift and the STFT is computed. The location of the
dominant frequency is found. Since the time difference between
epoch is 1 sample, the number of octaves between two
consecutive epochs is the slope in oct/sample, that can be
converted to oct/min as shown in Eq. 7:
(cid:1871)(cid:1864)(cid:1867)(cid:1868)(cid:1857)(cid:4674)(cid:3042)(cid:3030)(cid:3047)(cid:3040)(cid:3036)(cid:3041)(cid:4675)(cid:3404)(cid:1871)(cid:1864)(cid:1867)(cid:1868)(cid:1857)(cid:4674) (cid:3042)(cid:3030)(cid:3047)
(cid:3046)(cid:3028)(cid:3040)(cid:3043)(cid:3039)(cid:3032)(cid:4675)(cid:3400)(cid:2869)(cid:2874)(cid:3046)(cid:3028)(cid:3040)(cid:3043)(cid:3039)(cid:3032)(cid:3046)
(cid:2869)(cid:3046)
(cid:3400) (cid:2874)(cid:2868)(cid:3046)(cid:2869)(cid:3040)(cid:3036)(cid:3041)
(7)
Fig. 6 shows the distribution of frequency evolution rates for
the seizure and non-seizure examples in the database. It can be
seen that the peak of the distribution of the frequency changes
for seizures is at -0.373oct/min. This indicates that there is a
Figure 6. Estimated PDF of seizure and non-seizure frequency slope
distribution (red and blue respectively). Green area depicts the 95%
confidence interval where human hear are less accurate at detecting
frequency change. Best viewed in color.
negative slope associated with the seizure events, denoting a
decrease in frequency over time. For example, if the dominant
frequency of the seizure was 6Hz, then after 60s, on average, it
will decrease to 4.6Hz. In comparison, the peak of the
distribution for non-seizure is at -0.01oct/min, which implies
that there are no consistent changes in the frequency for
background EEG and seizure-like artifacts.
From Fig. 6 it can also be seen that in the range
between -1.69 oct/min and -0.13 oct/min, the seizure slope
distribution is above the non-seizure slope distribution and
seizures are separable from non-seizures. This indicates that the
measure can also be of use in automated seizure detection
algorithms.
D. Human sensitivity to frequency evolution
A test of human sensitivity to changes of continuous
frequency was performed to gain further insight in the results
from Fig. 4. A small survey was conducted with a cohort of 6
participants who repeated the test multiple times. In this survey,
the participants were asked to listen to a tone. The frequency of
the tone was randomly chosen in the range between 500Hz and
3500Hz. The frequency linearly varied over time, with a
randomly generated slope in the range between -0.5oct/min and
0.5oct/min. Participants were asked a binary question of whether
they perceived increase or decrease of a frequency. The results
were compiled for every slope. The power density function
(PDF) can be constructed by computing the ratio between errors
and trials.
The distribution of human errors in the sensitivity survey is
shown in Fig. 6. The distribution is centred around 0 and 95% of
human errors lie in the range between -0.389oct/min and
0.285oct/min. These results explain why speeding up the audio
resulted in a better discrimination between seizure and non-
seizure. For a factor of 10, the slope of the frequency evolution
for seizures is pushed outside the range of slopes in which the
human hearing system is incapable of distinguishing a change in
frequency. The perception of the evolution of the low-frequency
activity is facilitated with time compression.
IV. CONCLUSION
A non-expert EEG interpretation survey is presented and
provide promising results for the application of sonification in
neonatal seizure detection. Both methods of sonification
performed similarly well. Visual interpretation resulted in the
largest variance, denoting that non-expert visual detection of
seizures is inconsistent. It is observed that speeding up the EEG
in audio results in a higher accuracy. Analysing the frequency
evolution of seizures, it was shown that the peak value of the
frequency slope is -0.373oct/min. The human sensitivity to
changes
indicated a decreased
sensitivity to changes in frequency with a slope inside the range
of -0.389 to 0.285 oct/min. Thus, sonification methods with
larger time-scale factors will result in more accurate sound-
based seizure identification.
in continuous frequency
Future work will expand the results to include a larger cohort
and targeting clinical end-users. The parameters of sonification
algorithm will be tuned according to the objective results
obtained in the frequency evolution and human sensitivity to
frequency tests. The optimised sonification algorithm will be
implemented in the portable acquisition and interpretation
system presented [21], providing a viable medium to bring
sonification into clinical practice.
REFERENCES
[1]. D. Murray, G. Boylan, I. Ali, C. Ryan, B. Murphy and S. Connolly,
"Defining the gap between electrographic seizure burden, clinical
expression and staff recognition of neonatal seizures," Archives of
Disease in Chidhood: Fetal and Neonatal Edition, vol. 93, no. 3, pp. 187-
191, 2007.
[2]. A. Rakshasbhuvankar, S. Paul, L. Nagarajan, et al., "Amplitude-
integrated EEG for detection of neonatal seizures: a systematic review,"
Seizure, v. 33, pp. 90-98, 2015.
[3]. S. Mathieson, N. Stevenson, E. Low, W. Marnane, J. Rennie, A. Temko,
G. Lightbody and G. Boylan, "Validation of an automated seizure
detection algorithm for term neonates," Clinical Neurophysiology, vol.
127, no. 1, pp. 156-168, 2016.
[4]. E. Thomas, A. Temko, W. Marnane, G. Boylan, G. Lightbody,
"Discriminative and generative classification techniques applied to
automated neonatal seizure detection," IEEE Journal of Biomedical and
Health Informatics, v. 17, i. 2, pp. 297-304, 2013.
[5]. N. Stevenson, I. Korotchikova, A. Temko, G. Lightbody, W. Marnane and
G. Boylan, "An Automated System for Grading EEG Abnormality in
Term Neonates with Hypoxic-Ischaemic Encephalopathy," Annals of
Biomedical Engineering, vol. 41, no. 4, pp. 775-785, 2013.
[6]. AH. Ansari, P. Cherian, A. Caicedo Dorado, G. Naulaers, M. De Vos, S.
Van Huffel, "Neonatal Seizure Detection Using Deep Convolutional
Neural Networks," International Journal of Neural Systems, 2018.
[7]. S. Patrizi, G. Holmes, M. Orzalesi and F. Allemand, "Neonatal seizures:
characteristics of EEG ictal activity in preterm and fullterm infants,"
Brain and Development, vol. 25, no. 6, pp. 427-437, 2003.
[8]. A. Temko, W. Marnane, G. Boylan and G. Lightbody, "Clinical
implementation of a neonatal seizure detection algorithm," Decision
Support Systems, vol. 70, pp. 86-96, 2015.
[9]. A. Temko, A. K. Sarkar, G. B. Boylan, S. Mathieson, W. P. Marnane and
G. Lightbody, "Toward a Personalized Real-Time Diagnosis in Neonatal
Seizure Detection," in IEEE Journal of Translational Engineering in
Health and Medicine, vol. 5, pp. 1-14, 2017.
[10]. T. Väljamäe, T. Steffert, S. Holland, X. Marimon, R. Benitez, S. Mealla,
A. Oliveira and S. Jordà, "A review of real-time eeg sonification
research," in International Conference on Auditory Display (ICAD),
Łódź, 2013.
[11]. H. Khamis, A. Mohamed, S. Simpson and A. McEwan, "Detection of
temporal lobe seizures and identification of lateralisation from audified
EEG," Clinical Neurophysiology, vol. 123, no. 9, pp. 1714-1720, 2012.
[12]. Jeng-Wei Lin, Wei Chen, Chia-Ping Shen, Ming-Jang Chiu, Yi-Hui Kao,
Feipei Lai, Qibin Zhao, Andrzej Cichocki, "Visualization and sonification
of long-term epilepsy electroencephalogram monitoring," Journal of
Medical and Biological Engineering, pp 1–10. 2018.
[13]. T. Hermann, A. Hunt, J. Neuhoff (editors). The Sonification Handbook.
Logos Publishing House, Berlin, Germany, 2011.
[14]. A. Temko, G. Lightbody, L. Marnane and G. Boylan, "Real time
audification of neonatal electroencephalogram (EEG) signals". US Patent
US20170172523A1, 07 07 2014.
[15]. A. Temko, W. Marnane, G. Boylan, J. O'Toole and G. Lightbody,
"Neonatal EEG audification for seizure detection," in IEEE International
Conference of Engineering in Medicine and Biology Society (EMBC),
Chicago, 2014.
[16]. J. Parvizi, K. Gururangan, B. Razavi, C. Chafe, "Detectig silent seizures
by their sound," Epilepsia In Press, vol. 59, no. 4, pp. 877-884, 2018
[17]. A. Krystal, R. Prado and M. West, "New methods of time series analysis
of non-stationary EEG data:," Clinical Neurophysiology, vol. 110, no. 12,
pp. 2197-2206, 1999.
[18]. C. Hagmann, N. Robertson and D. Azzopardi, "Artifacts on
Electroencephalograms May Influence the Amplitude-Integrated EEG
Classification: A Qualitative Analysis in Neonatal Encephalopathy,"
Pediatrics, vol. 118, no. 6, pp. 2552-2554, 2006.
[19]. J. Flanagan and R. Golden, "Phase Vocoder," in IEEE The Bell System
Technical Journal, Minneapolis, 1966.
[20]. J. McDermott, M. Keebler, C. Micheyl and A. Oxenham, "Musical
intervals and relative pitch: Frequency resolution, not interval resolution,
is special," The Journal of the Acoustical Society of America, vol. 128, no.
4, p. 1943, 2010.
[21]. J. Chowning, "The synthesis of complex audio spectra by means of
frequency modulation," Computer Music Journal, vol. 1, no. 2, pp. 46-54,
1977.
[22]. M. O'Sullivan, S. Gomez, A. O'Shea, E. Salgado, K. Huillca, S.
Mathieson, G. Boylan, E. Popovici, A. Temko, "Neonatal EEG
interpretation and decision support framwork for mobile platforms", in
IEEE International Conference of Engineering in Medicine and Biology
Society (EMBC), Honolulu, 2018.
|
1702.06538 | 1 | 1702 | 2017-02-21T18:38:58 | Applications of Discrete Mathematics for Understanding Dynamics of Synapses and Networks in Neuroscience | [
"q-bio.NC"
] | Mathematical modeling has broad applications in neuroscience whether modeling the dynamics of a single synapse or an entire network of neurons. In Part I, we model vesicle replenishment and release at the photoreceptor synapse to better understand how visual information is processed. In Part II, we explore a simple model of neural networks with the goal of discovering how network structure shapes the behavior of the network.
To fully understand how visual information is processed requires an understanding of the way signals are transformed at the ribbon synapse of photoreceptor neurons. These synapses possess a ribbon-like structure capable of storing around 100 synaptic vesicles, allowing graded responses through the release of different numbers of vesicles in response to visual input. These responses depend critically on the ability of the ribbon to replenish itself as ribbon sites empty upon release. The rate of vesicle replenishment is thus an important factor in shaping neural coding in the retina. In collaboration with experimental neuroscientists we developed a mathematical model to describe the dynamics of vesicle release and replenishment at the ribbon synapse.
To learn more about how network architecture shapes the dynamics of the network, we study a specific type of threshold-linear network that is constructed from a simple directed graph. The network construction guarantees that differences in dynamics arise solely from differences in the connectivity of the underlying graph. By design, the activity of these networks is bounded and there are no stable fixed points. Computational experiments show that most of these networks yield limit cycles where the neurons fire in sequence. We devised an algorithm to predict the sequence of firing using the structure of the underlying graph. Using the algorithm we classify all the networks of this type on five or fewer nodes. | q-bio.NC | q-bio | APPLICATIONS OF DISCRETE MATHEMATICS FOR UNDERSTANDING
DYNAMICS OF SYNAPSES AND NETWORKS IN NEUROSCIENCE
7
1
0
2
b
e
F
1
2
]
.
C
N
o
i
b
-
q
[
1
v
8
3
5
6
0
.
2
0
7
1
:
v
i
X
r
a
by
Caitlyn M. Parmelee
A DISSERTATION
Presented to the Faculty of
The Graduate College at the University of Nebraska
In Partial Fulfilment of Requirements
For the Degree of Doctor of Philosophy
Major: Mathematics
Under the Supervision of Professor Carina Curto
Lincoln, Nebraska
August, 2016
APPLICATIONS OF DISCRETE MATHEMATICS FOR UNDERSTANDING
DYNAMICS OF SYNAPSES AND NETWORKS IN NEUROSCIENCE
Caitlyn M. Parmelee, Ph.D.
University of Nebraska, 2016
Adviser: Professor Carina Curto
Mathematical modeling has broad applications in neuroscience whether we are
modeling the dynamics of a single synapse or the dynamics of an entire network of
neurons. In Part I, we model vesicle replenishment and release at the photoreceptor
synapse to better understand how visual information is processed. In Part II, we
explore a simple model of neural networks with the goal of discovering how network
structure shapes the behavior of the network.
Vision plays an important role in how we interact with our environments. To fully
understand how visual information is processed requires an understanding of the way
signals are transformed at the very first synapse: the ribbon synapse of photoreceptor
neurons (rods and cones). These synapses possess a ribbon-like structure on which
approximately 100 synaptic vesicles can be stored, allowing graded responses through
the release of different numbers of vesicles in response to visual input. These responses
depend critically on the ability of the ribbon to replenish itself as ribbon sites empty
upon release. The rate of vesicle replenishment is thus an important factor in shaping
neural coding in the retina. In collaboration with experimental neuroscientists we
developed a mathematical model to describe the dynamics of vesicle release and
replenishment at the ribbon synapse.
To learn more about how network architecture shapes the dynamics of the network,
we study a specific type of threshold-linear network that is constructed from a simple
directed graph. These networks are particularly well suited for our study because
the network construction guarantees that differences in dynamics arise solely from
differences in the connectivity of the underlying graph. By design, the activity of these
networks is bounded and there are no stable fixed points. Computational experiments
show that most of these networks yield limit cycles where the neurons fire in sequence.
Can we predict the order in which the neurons fire? To this end, we devised an
algorithm to predict the sequence of firing using the structure of the underlying graph.
Using the algorithm we classify all the networks of this type on five or fewer nodes.
iv
COPYRIGHT
© 2016, Caitlyn M. Parmelee
v
DEDICATION
To my parents, to whom I owe everything.
And to Matt, to whom I owe this.
ACKNOWLEDGMENTS
vi
I would like to first thank my advisor, Dr. Carina Curto, for her support and
encouragement over the last few years. The amazing opportunities she has provided
me have been invaluable.
Another thank you goes out to the other members of my committee, Dr. Bo
Deng, Dr. Vladimir Itskov, Dr. Jamie Radcliffe, and Dr. Wallace Thoreson, for their
conversations and support. I would also like to thank our exceptional collaborators,
Dr. Wallace Thoreson, Dr. Matthew Van Hook, and Dr. Katherine Morrison for their
discussions and insights.
I would also like to thank the University of Nebraska -- Lincoln Mathematics De-
partment for providing the resources to pursue my research and develop my teaching.
There are so many people I am grateful to have met along my journey and owe
my deepest thanks: my labmates, for helping me navigate the world of mathematical
neuroscience, my fellow "first-years," for being an incredible support system and filling
my life with fun facts, and my officemates, for creating a safe space and making me
smile even on the most difficult days.
A special thanks goes to my undergraduate advisor, Dr. Matt Koetz, and the other
math faculty at Nazareth College for setting me on this journey and continuing to
believe in me every step of the way.
I could not have done this without my partner in crime, Nate Clayburn. He has
been a source of strength and love for the last five years. I am strong if you are strong.
Lastly, my eternal thanks to my family, especially my parents and sister, whose
love fills my life and to whom I owe so much.
GRANT INFORMATION
vii
This work was supported in part by the National Science Foundation grants DMS-
1516881, DMS-1225666/DMS-1537228 and P200A120068. This work was also sup-
ported in part by a sub-award from the NIH grant R01-EY010542-19.
Table of Contents
I Dynamics of Ribbon Synapses
1 Introduction to Part I
1.1 Synaptic transmission . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2 Photoreceptor neurons . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3 Synaptic ribbon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.4 Questions about release and replenishment at the synaptic ribbon . .
2 Model of vesicle release and replenishment
2.1 Dynamics of release and replenishment . . . . . . . . . . . . . . . . .
2.2 Measuring available pool size
. . . . . . . . . . . . . . . . . . . . . .
2.2.1 Pulse train experiments
. . . . . . . . . . . . . . . . . . . . .
2.2.2 An apparent paradox . . . . . . . . . . . . . . . . . . . . . . .
2.3 Using our model to predict pool size
. . . . . . . . . . . . . . . . . .
2.3.1 Derivation of pool size and release probability formulas . . . .
2.3.2 Estimating pool size and release probability from data . . . .
2.4 Generalization of release/replenishment model
. . . . . . . . . . . . .
2.4.1 Generalized pulse trains
. . . . . . . . . . . . . . . . . . . . .
2.4.2
Setting up and solving the recursion . . . . . . . . . . . . . .
2.4.3
Special cases of the generalized model . . . . . . . . . . . . . .
viii
1
2
4
5
6
8
11
12
15
15
16
17
18
20
21
21
22
24
2.4.4
Supporting lemmas . . . . . . . . . . . . . . . . . . . . . . . .
27
ix
3 Random walk model of vesicle replenishment
3.1 Replenishment timescale . . . . . . . . . . . . . . . . . . . . . . . . .
3.1.1 Paired pulse recordings . . . . . . . . . . . . . . . . . . . . . .
3.1.2 Derivation of the replenishment time constant . . . . . . . . .
3.1.3 Comparison of model predictions with data . . . . . . . . . . .
3.2 Role of calcium in replenishment
. . . . . . . . . . . . . . . . . . . .
3.2.1 Calcium affects vesicles . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Calcium affects the ribbon . . . . . . . . . . . . . . . . . . . .
3.2.3 Comparison to experimental results . . . . . . . . . . . . . . .
3.3 Other quantities of interest . . . . . . . . . . . . . . . . . . . . . . . .
3.3.1 Hit rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2 Expected waiting time . . . . . . . . . . . . . . . . . . . . . .
4 Computational model
4.1 Description of the computational model . . . . . . . . . . . . . . . . .
4.2 Effect of ribbon geometry on replenishment . . . . . . . . . . . . . . .
4.3 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
II Neural Sequences in Threshold-Linear Networks
5 Introduction to Part II
5.1 Threshold-linear networks
. . . . . . . . . . . . . . . . . . . . . . . .
5.2 Description of the CTLN model
. . . . . . . . . . . . . . . . . . . . .
5.3 Examples and behaviors
. . . . . . . . . . . . . . . . . . . . . . . . .
6 Sequence prediction algorithm
31
32
32
33
37
38
40
40
41
42
42
45
48
48
50
52
56
57
59
60
62
67
6.1 Description of the algorithm . . . . . . . . . . . . . . . . . . . . . . .
6.2 Performance of the algorithm . . . . . . . . . . . . . . . . . . . . . .
6.3 Extending the algorithm to oriented graphs
. . . . . . . . . . . . . .
6.4 An application: classification of oriented graphs on n ≤ 5 . . . . . . .
6.5 Dictionary of attractors for n ≤ 5 . . . . . . . . . . . . . . . . . . . .
6.6 Future work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
A Code for random walk model of replenishment
B Code for sequence prediction algorithm
C Catalogue of n ≤ 5 oriented graphs with no sinks
Bibliography
x
67
70
72
73
73
80
82
94
103
114
1
Part I
Dynamics of Ribbon Synapses
2
Chapter 1
Introduction to Part I
Vision plays an important role in how we interact with our environments. In fact, half
of our cerebral cortex is dedicated to processing the visual world [5]. Part I explores
how visual information is processed at the very first synapse of the visual pathway,
the photoreceptor synapse. We specifically look at the role of a structure called the
synaptic ribbon.
In this introductory section we will discuss some background information about
the structure and function of the visual system. Visual processing begins when light
enters the retina and is absorbed by photoreceptor neurons. Photoreceptor neurons
are the principal light-sensitive cells in the retina. When light enters the retina, it
passes through several layers of cells1 before being absorbed by the outer segments
of photoreceptors (see Figure 1.1). The absorption of light initiates the process of
phototransduction which ultimately triggers changes in membrane potential. These
signals are passed to a layer of bipolar cells2, and then to a layer of ganglion cells.
The ganglion cells send axons to the optic nerve and are the only source of outputs
1The other cells in the retina are relatively transparent, so when light passes through them there
is very little image distortion.[5]
2Photoreceptors also synapse onto horizontal cells, which modify the bipolar cells laterally.
from the retina.
3
Figure 1.1: Diagram of retinal pathway: (1) photoreceptor rods, (2) photoreceptors
cones, (3) horizontal cells, (4) bipolar cells, (5) amacrine cells, (6) retinal ganglion
cells. Adapted from [40].
To better understand vision, we wish to first understand how visual information
is processed at the photoreceptor synapse. To do this we need to set up some
necessary background information about the biology of these synapses. We will start
by discussing neurons and how they communicate with each other. Then we will look
more specifically at how photoreceptor neurons communicate. Finally we will describe
the synaptic ribbon, a specialized structure in photoreceptor synapses, and discuss
what is currently known about its role in the vesicle cycle and information processing.
Following the introduction, Chapters 2-4 discuss our work involving the ribbon
synapse. These results were obtained in collaboration with experimental neurosci-
entists in the Thoreson Lab at the University of Nebraska Medical Center. Our
contribution has been to develop theoretical models describing the dynamics of release
and replenishment in the ribbon synapse. Results from Sections 2.1-2.3 were published
in [36] and results from Sections 3.1-3.2 were published in [38].
4
1.1 Synaptic transmission
Neurons are cells involved in the transmission of information in the nervous system.
The neurons receive inputs from other neurons at the dendrites and once a threshold is
reached the neuron can send a signal, often in the form of an action potential, down its
axon to other cells. The pattern of action potentials codes the information being sent.
This information transfer between the two cells takes place at the synapse. Figure
1.2 shows a diagram of a conventional synapse. The information is passed between
neurons through the release of vesicles, which are small spheres made of membrane
and packed with neurotransmitters.
Figure 1.2: Diagram of a synapse: When a signal reaches the terminal, it triggers
vesicles to dock and fuse with the cell membrane and release neurotransmitters. These
neurotransmitters travel across the synaptic cleft and bind with the receptors on the
postsynaptic cell.
synaptic vesiclesaxonpostsynaptic receptorsneurotransmittersSYNAPSEsignal 5
When an action potential reaches the cell terminal, it triggers vesicle exocytosis.
Vesicle exocytosis is the process in which the vesicles dock and fuse with the cell
membrane and release their neurotransmitters into the synaptic cleft, the space
between the two cells. The area of the cell membrane where this occurs is referred
to as the "active zone." The neurotransmitters then bind with the receptors on the
postsynaptic cell, passing the information. For example, these receptors may open or
close ion channels or activate second messenger systems.
Vesicles are recycled through endocytosis, which is the process by which vesicles
are reformed using parts of the cell membrane and refilled with neurotransmitters.
These recycled vesicles then become part of the mobile vesicle pool inside the cell.
1.2 Photoreceptor neurons
We are particularly interested in studying synaptic transmission at photoreceptor
neurons. Photoreceptor neurons are the first cells of the visual system. Photoreceptors
are located in the retina and their function is to convert light into changes in membrane
potential. Light is absorbed by membranous disks, located on the outer segments,
containing photopigment. There are two main types of photoreceptors: rods and cones.
Rods are involved in night vision, motion detection, and peripheral vision and they are
dense everywhere but the center of the eye. Cones are located in the center of the retina
and are involved in color vision and detecting finer detail. Unless otherwise specified,
all the experimental data in Part I refers only to cone photoreceptors, specifically in
the aquatic tiger salamander.
In a conventional synapse the neuron responds to action potentials with discrete
vesicle events. Photoreceptor cells instead respond directly to the absorption of
photons by releasing vesicles constantly in darkness and slowing release as light
6
increases, i.e. the cell is depolarized in darkness and an increase in light causes the
cell to hyperpolarize. The graded responses given by photoreceptor cells allow for
a quicker processing of information as well as a larger range of responses [19]. This
graded release is facilitated by a structure called the synaptic ribbon, described in the
next section.
1.3 Synaptic ribbon
The synaptic ribbons present in cone photoreceptors are plate-like rectangular3 pro-
teinaceous structures anchored to the inside of the cell membrane close to the Ca2+
channels [33]. Daily and seasonal changes in the size, shape, number, and location of
synaptic ribbons can occur based on light conditions [39]. Vesicle release at the active
zone is controlled by the opening and closing of the calcium channels. In cones, less
than three channel openings are required to cause the fusion of a single vesicle, which
allows for precise timing of release to accurately reflect changes in light intensity [3].
The increase in intracellular Ca2+ also speeds the replenishment of vesicles, allowing
for sustained release [1].
Recall that in conventional synapses vesicles dock and fuse directly with the
cell membrane. The vesicles in ribbon synapses are instead first collected on the
synaptic ribbon. In the cone photoreceptors of the aquatic tiger salamander there are
approximately 11 rows of 5 vesicles stacked on each side of the ribbon, for a total of
110 vesicles [2]. The vesicles become tethered to the ribbon via tiny filaments and
then move along the ribbon towards the active zone. Not much is known about how
the vesicles move down the ribbon to the active zone, but recent research posits that
vesicles passively diffuse along the ribbon without an active transport mechanism
3Synaptic ribbons in different cells may have different shapes. For example, ribbons in the
auditory system can be spherical or ellipsoidal rather than rectangular ([32],[17]).
7
Figure 1.3: Electron micrographs of synaptic ribbons in rod terminals: The larger
arrows indicate the active zone at the bottom of each ribbon and the smaller arrows
indicate the hexagonally packed vesicles tethered to the ribbon. Adapted from [37].
[10]. Once the vesicles reach the bottom two rows of the ribbon they are considered
part of the rapidly releasable pool (RRP). Experiments have shown that the ribbon
may play a role in priming the vesicles for release [31]. As a result, the RRP can
be released almost immediately following the opening of calcium channels. Once the
RRP is depleted, additional vesicles from the reserve pool on the ribbon take their
place. Empty sites on the ribbon are refilled by the mobile vesicles in the cell terminal.
See Figure 1.4 for a cartoon of the vesicle cycle in a ribbon synapse.
There are many theories regarding the function of the synaptic ribbon. The ribbon
appears to support high rates of sustained vesicle release [33], but how the ribbon
achieves this is still an open question. One theory is that the ribbon acts as a "conveyor
belt" shuttling vesicles toward the active zone [23]. Another theory posits that it
serves to hold the vesicles in contact with each other to facilitate multivesicular release
via compound fusion [23]. Another theory asserts that the ribbon slows the delivery
of vesicles, regulating the timing of release [16]. Yet another proposes that the ribbon
functions to store the vesicles close to the active zone [45].
8
Figure 1.4: A cartoon depicting the vesicle cycle in the ribbon synapse [35].
1.4 Questions about release and replenishment at
the synaptic ribbon
With the goal of better understanding how visual information is processed at the
photoreceptor synapse in mind, we ask some questions about the role of the synaptic
ribbon in the vesicle cycle of this synapse.
The number of vesicles released is stimulus-dependent, with stronger stimuli
resulting in more vesicles released. What causes this stimulus-dependence in the vesicle
release? Does it depend solely on the probability of release or does it also depend
the number of vesicles currently available on the ribbon? To answer this question we
created a model of release and replenishment using experimentally measured quantities
9
to predict the unknown quantities of pool size and release probability. This allowed us
to independently predict the pool size and the release probability to determine which
changes with the stimulus strength. See Chapter 2.
There is an upper limit on the rate of sustained vesicle release. What is the
rate-limiting factor for release? Studies indicate that vesicle replenishment is the
rate-limiting step in sustained release [16], so we take a closer look at replenishment.
Vesicles move randomly in the cell terminal without a directed movement toward the
ribbon or active zone [24]. This may be due to the fact that ribbon synapses lack
synapsins, proteins that help bind vesicles to the actin cytoskeleton, allowing the
vesicles to diffuse freely [14]. Is this random motion of vesicles the rate-limiting step
for replenishment? To answer this question we designed a three-dimensional random
walk model of vesicle replenishment to calculate the replenishment timescale. We
conclude that the random motion is not rate-limiting for replenishment. See Section
3.1.
It is known that Ca2+ speeds the replenishment process [1]. By what mechanism
does Ca2+ speed replenishment? To answer this question we modified our random
walk model to test whether Ca2+ acted on the ribbon or on the vesicles. We conclude
that Ca2+ affects the probability of attachment at sites on the ribbon rather than
directly affecting vesicles. See Section 3.2.
To further study replenishment we ask two additional replenishment-related ques-
tions: (1) How many vesicles collide with the ribbon per second? and (2) How long
does it take for the ribbon to fill? We can use our random walk model of replenishment
to answer both. See Section 3.3.
Our random walk model does not take into account the geometry of the synaptic
ribbon, so we designed a computational model to test the effects of ribbon geometry.
The computational model indicates that ribbon geometry does play a role in replen-
ishment, so we also explore changes in local concentration near the ribbon and the
effect of attachment probability on replenishment in an effort to explain the effects of
geometry. The results appear in Section 4.
10
11
Chapter 2
Model of vesicle release and
replenishment
Photoreceptors respond to changes in light by releasing vesicles from the ribbon.
The amount of release depends on several key quantities: available pool size, release
probability, and quantal amplitude. The available pool size, N , is the number of
vesicles on the ribbon that are primed and ready for release. The release probability, P ,
is the probability that a vesicle on the ribbon will be released. The quantal amplitude,
Q, is the postsynaptic influence of a single vesicle. We experimentally measure the
response of photoreceptors to a given stimulus by measuring the postsynaptic currents1
(PSCs) evoked in the postsynaptic cells onto which the photoreceptor synapses. Since
the postsynaptic current is a linear sum of mini-EPSCs2 [7], we can then estimate the
number of vesicles released from the postsynaptic current using the quantal amplitude.
The amount of release depends on the stimulus, so which of N , P , and/or Q
contribute to this stimulus-dependence? Changes in quantal amplitude occur on a
1PSCs are generally measured in units of picoamps (pA).
2 Mini-EPSCs (mEPSCs) are the change in current resulting from a single vesicle releasing its
neurotransmitters.
12
longer timescale than our experiments [4, 15, 18, 43]. Thus Q cannot change quickly
enough to be the cause of the stimulus-dependent changes in postsynaptic response.
Stimulus-dependent changes in postsynaptic response are often due to Ca2+-dependent
changes in P [36], but it is also possible that stronger stimuli allow Ca2+ to spread
further up the ribbon, effectively increasing N . With these possibilities in mind, are
the stimulus-dependent changes then due only to changes in the release probability,
P , or are they a result of changes in N as well?
In this section we will discuss a paradox that arises when asking this question. We
then provide a model that estimates N and P independently, based on experimental
data, allowing us to resolve this paradox. We will also describe a generalization of the
model. The results in Sections 2.1-2.3 are published in [36]. The generalized model
results in Section 2.4 are an unpublished extension of this work.
2.1 Dynamics of release and replenishment
Vesicles on the ribbon are released when the photoreceptor is stimulated and the
amount released depends on the stimulus. As the vesicles are released, the empty sites
on the ribbon are replenished by vesicles freely diffusing in the cell terminal. To study
release and replenishment we model the available pool size during alternating periods
of release and replenishment. Figure 2.1 shows a cartoon of the model. The variable
A(t) tracks the number of vesicles on the ribbon at time t. When the stimulus is on,
A(t) decreases due to vesicle release and when the stimulus is off, A(t) increases due to
vesicle replenishment. When designing the model, we make several key assumptions.
We assume that no replenishment occurs during the periods of release. We also assume
identical stimuli for the basic model, but the generalized model allows for multiple
stimulus types in the same trial. By patching together the release and replenishment
13
Figure 2.1: Cartoon of release/replenishment model
dynamics, using the ending value of one period as the initial value of the next, we can
create a model of available pool size.
Release dynamics. The cumulative release at time t, c(t), is governed by the
differential equation:
dc
dt
=
ps(Ai − c)
τr
(2.1)
where ps is the stimulus-specific probability of release, Ai is the pool size at the
beginning of the ith pulse, and τr is the time constant of release. In salamander cones,
the cumulative release curve can be fit by a two term exponential, one of the release
time constants, τr, is around 5 ms and the other is too long to be accurately measured
in our experimental setup, so we omit it from the model. The timescale τr regulates
the release for strong stimuli (e.g. steps to -19 mV), and for weaker stimuli (e.g. steps
to -39 mV) the time constant is made effectively slower by the release factor ps. Hence
number of vesicles on ribbontimerelease replenishmentA(t)stimulus onstimulus o! for strong stimuli we have ps = 1. Solving the differential equation yields
c(t) = Ai(1 − e−pst/τr).
Thus the release during the ith pulse is given by
Ri = c(∆t) = PsAi
14
(2.2)
(2.3)
where Ps = 1−e−ps∆t/τr. Note that for 25 ms steps to -19 mV, P−19 = 1−e−5 ≈ 0.9933.
This is consistent with previous work showing that steps to -19 mV are strong enough
to stimulate the release of nearly the entire pool of vesicles [2].
Replenishment dynamics The cumulative replenishment at time t, a(t), is gov-
erned by the differential equation:
da
dt
=
n − a
τa
(2.4)
where n is number of sites unoccupied at the end of a pulse and τa is the time constant
of replenishment. In salamander cones, replenishment is modeled with a two term
exponential with time constants τfast = 815 ms and τslow = 13 s. Since the experiments
occur on a much faster timescale than τslow, we ignore τslow in our model. Solving the
differential equation yields
a(t) = n(1 − e−t/τa).
Thus the amount of replenishment after the ith pulse is given by
a(T ) = n(1 − β)
(2.5)
(2.6)
15
where β = e−T /τa. Recall that we have chosen to omit the slow replenishment time
constant, so after the initial release we assume only the fast-replenishing sites have
time to fill during our replenishment period. Depending on the stimulus, we have a
different fraction, fs, of vesicles that are subject to the fast time constant. For steps
to -19 mV, f−19 = 0.76 and for steps to -39 mV, f−39 = 0.55. Thus the number of
available sites at the end of the ith pulse is n = fsAs − Ai(1 − Ps), where As is the
maximum pool size for stimulus s.
2.2 Measuring available pool size
2.2.1 Pulse train experiments
To study the release and replenishment dynamics we consider a pulse train experiment.
In this setup the cone is voltage-clamped near resting membrane potential and a
steady train of pulses, i.e. voltage jumps, is applied to the presynaptic cell. Each pulse
has duration ∆t ms, and the time between pulses is T ms. During a given pulse the
voltage jumps up to a chosen voltage step (e.g. a step to -19 mV) and between steps
the voltage returns to -79 mV (see Figure 2.2). The postsynaptic currents (PSCs)
are measured in the postsynaptic horizontal cells. Since postsynaptic response is the
result of a linear sum of independent quantal release events (mEPSCs) [7], we can
use the postsynaptic measurements to estimate the number of vesicles released from
the cones3. Once the release and replenishment reach an equilibrium where release is
limited by replenishment, we can measure the limiting release. Our goal is to design
a model that can predict maximum pool size and release probability using the first
release and limiting release values measured during such an experiment.
3The mean amplitude of an mEPSC in the salamander retina, i.e. how much a single vesicle
contributes, is around 6.5 ± 1.6 pA [7].
16
Figure 2.2: Pulse trains for voltage jumps to -19 mV,-29 mV, -39 mV, and -49 mV.
Note that the stronger pulses have larger first release peaks [36].
2.2.2 An apparent paradox
Previous work uses a method of back-extrapolation to estimate the maximum pool
size, A [25]. This method considers the cumulative release curve and fits a line to
the steady state response that occurs when release is limited by replenishment. Back-
extrapolating to the time 0 gives an estimate of A. This method predicts that the
maximum pool size is significantly smaller for weaker stimuli (see Figure 2.3). The
amplitude of the releasable pool predicted by back-extrapolation is 80 pA for -39 mV,
105 pA for -29 mV, and 132 pA for -19 mV [36].
One of the pitfalls of the back-extrapolation method is that it assumes that the
replenishment rate is constant. As we saw in Section 2.1, the replenishment rate is
certainly not constant in salamander cones. This causes the method to underestimate
17
Figure 2.3: Back-extrapolation method for -19 mV, -29 mV, and -39 mV pulses: Panel
A shows a plot of cumulative amplitude in pA and Panel B show a plot of cumulative
charge in fC [36]. To predict maximum pool size we back-extrapolate from the steady
state to t = 0.
the maximum available pool size since replenishment is faster when the ribbon has
more available space. The method is close for the stronger stimuli because the pulse
train stabilizes to the limiting release right away when exposed to a strong stimulus.
It also does not take into account the fast and slow replenishing sites. After the first
pulse, slow sites don't have time to fill between pulses, so the limiting release reflects
only the replenishment of the fast sites.
2.3 Using our model to predict pool size
In our model, we let both the maximum pool size As and the release probability Ps
vary independently to determine which causes the voltage-dependent changes in vesicle
release. We want to estimate both As and Ps in terms of the measured first release
(R1)s and limiting release Rs for each stimulus s.
18
2.3.1 Derivation of pool size and release probability
formulas
Let Ai be the pool size at the beginning of the ith pulse, ci(t) be the cumulative
release t milliseconds into the ith pulse, and ai(t) be the cumulative replenishment t
seconds into the ith pulse. Then we can compute Ai by taking the pool size at the
beginning of the previous pulse, subtracting the release during that pulse, and adding
the amount replenished before the ith pulse, i.e. Ai = Ai−1 − ci−1(∆t) + ai−1(T ). Note
Figure 2.4: Schematic of the pulse train setup [36]: A(t) keeps track of pool size at
time t. Stimulus pulses of duration ∆t cause vesicle release and between pulses we
have replenishment periods of duration T . Note that the pool size decreases during the
pulses due to vesicle release and increases in between the pulses due to replenishment.
The maximum possible pool size is denoted by A.
that ci(∆t) = AiPs and ai(T ) = (fsAs − Ai−1(1 − Ps))(1 − β). Thus
Ai = bAi−1 + c
where b = β(1 − Ps) and c = fsAs(1 − β).
Solving the recursion we get
Ai = A1bi−1 + c
1 − bi−1
1 − b
.
19
(2.7)
(2.8)
The details of solving the recursion appear in Lemma 2.1. Taking the limit as i → ∞
gives
(cid:18)
A1bi−1 + c
(cid:19)
1 − bi−1
1 − b
A∞ = lim
i→∞ Ai = lim
i→∞
=
c
1 − b
=
fsAs(1 − β)
1 − β(1 − Ps)
,
(2.9)
which represents the pool size at the beginning of each pulse during the steady state.
Since Ri = PsAi is the amount released during the ith pulse, then the limiting release
is given by
Solving for As yields
Rs = PsA∞ =
PsfsAs(1 − β)
1 − β + βPs
.
As =
Rs(1 − β + βPs)
Psfs(1 − β)
=
Rs
fs
(cid:18) 1
Ps
(cid:19)
.
+
β
1 − β
(2.10)
(2.11)
Also, note that (R1)s = A1Ps = AsPs. Solving for Ps and substituting into our
equation for As gives
As =
(cid:19)
(cid:18) β
1 − β
Rs(R1)s
(fs(R1)s − Rs)
.
(2.12)
Using the fact that (R1)s = AsPs, we can also find an expression for Ps,
20
(cid:18)1 − β
(cid:19) fs(R1)s − Rs
.
(2.13)
Ps =
β
Rs
Equations 2.12 and 2.13 give formulas for independently estimating the maximum
pool size and release probability for each stimulus given the first release and limiting
release.
2.3.2 Estimating pool size and release probability from data
During pulse trains with steps to -19 mV and -39 mV we measure the amplitude of the
first pulse as well as the limiting release. Limiting release is estimated by measuring
the cumulative increase in amplitude 1 -- 2 seconds into the pulse train [36]. With
f−19 = 0.76, f−39 = 0.55, ∆t = 25ms, τa = 815ms, and τr = 5ms, we can estimate the
pool size for the two stimulus types. The results are recorded in Table 2.1. Note that
although the amplitude of the first pulse varies significantly with stimulus strength,
the predicted pool sizes are roughly the same. Using our formula to predict the release
probabilities in the 5 mM EGTA cases, we see P−19 ≈ 1 for the strong stimuli and
P−39 ≈ 0.5 for the weak stimuli. This supports the hypothesis that changes in release
probability alone cause voltage-dependent changes in release. Note that the first
release during strong pulses is nearly identical to the estimated pool size, consistent
with a release probability of 1. Also recall that back-extrapolation predicted a pool
size of 80 pA for a stimulus of -39 mV while our model estimates that the pool size is
closer to 131 pA.
Additional experiments were done with a weaker Ca2+ buffer of 0.05mM EGTA
compared to 5 mM EGTA. These results were similar to those with 5 mM EGTA
(see Table 2.1). The slightly smaller first responses in the experiments with 0.05 mM
21
stimulus
EGTA
T
-19 mV
50 ms
5 mM
-39 mV
5 mM
50 ms
-19 mV
5 mM 125 ms
-39 mV
5 mM 125 ms
-19 mV 0.05 mM 50 ms
-39 mV 0.05 mM 50 ms
PSC amplitude
(first pulse, R1)
128.2±10.9 pA
70.9±7.4 pA
135.5± 15.8 pA
71.3± 12.8 pA
91.1± 18.2 pA
38.5± 9.5 pA
Predicted pool size, As
Ratio
(PSC amplitude)
A−39/A−19
131.3 pA
131.2 pA
136.9 pA
131.2 pA
110.9 pA
113.6 pA
1.0
0.96
1.02
Table 2.1: Pool size predictions for several experimental conditions [36].
EGTA are likely due to the smaller number of ribbon contacts per postsynaptic HC
(an average of 2.79 and 2.95 ribbon contacts for experiments with 5 mM EGTA versus
an average of 1.98 ribbon contacts for experiments with 0.05 mM EGTA)[36]. This
provides additional evidence that the increased spread of Ca2+ does not increase the
available pool size.
2.4 Generalization of release/replenishment
model
In this section, we generalize the model from Section 2.3. We originally assumed that
all pulses were of equal strength and duration. In the generalized model, both pulses
and replenishment periods can vary in duration and pulses can also vary in stimulus
strength. The results in this section make use of several lemmas and a proposition
whose statements and proofs appear in Section 2.4.4.
2.4.1 Generalized pulse trains
We consider a generalized pulse train with n periods of release and replenishment. The
length of the (cid:96)th period is ∆t(cid:96). We do not necessarily alternate between release and
22
Figure 2.5: Generalized model setup: Ai,(cid:96) denotes available pool size at the beginning
of the (cid:96)th period of the ith cycle and ∆t(cid:96) is the duration of the (cid:96)th period.
replenishment. Once we have cycled through all n periods we started at the beginning
and repeat the periods in the same order. We denote the available pool size at the
beginning of the (cid:96)th period of the ith cycle by Ai,(cid:96) and assume the ribbon is full at
t = 0, so A = A1,1 is the maximum available pool size. See Figure 2.5. Let ai,(cid:96)(t) be
the total change in pool size t seconds into period (cid:96) of cycle i with ai,(cid:96)(0) = 0. If index
j denotes a release period then the dynamics is governed by
dai,j
dt
= − 1
τj
(Ai,j + ai,j).
If index k denotes a replenishment period then the dynamics is governed by
dai,k
dt
=
1
τk
(A − (Ai,k + ai,k)).
2.4.2 Setting up and solving the recursion
Note that this setup gives
Ai,(cid:96) = Ai,(cid:96)−1 + ai,(cid:96)−1(∆t(cid:96)−1) for 1 < (cid:96) ≤ n and i ≥ 1
Ai,1 = Ai−1,n + ai−1,n(∆tn)
for i > 1
A1,1 = A
(2.14)
(2.15)
(2.16)
A1,1A1,2A1,l-1A1,lA1,l+1...A1,n-1A1,nA2,1A2,2A2,l-1A2,lA2,l+1......A2,n-1A2,n...∆t1∆tn-1∆tl∆tl-1∆tn∆t1... 23
Then we get
Ai,1 = Ai−1,1 +
n(cid:88)
(cid:96)=1
ai−1,(cid:96)(∆t(cid:96)).
(2.17)
In release periods ai,j(t) = −Ai,j(1 − e−t/τj ) and in replenishment periods ai,k(t) =
(A − Ai,k)(1 − e−t/τk). Let α(cid:96) = e−∆t(cid:96)/τ(cid:96) so we can rewrite the terms in the sum using
ai,(cid:96)(∆t(cid:96)) = (θ((cid:96))A − Ai,(cid:96))(1 − α(cid:96))
(2.18)
where
θ((cid:96)) =
Ai−1,1:
1 if (cid:96) indexes a replenishment period
(cid:33)
(cid:32) n(cid:89)
n(cid:89)
0 if (cid:96) indexes a release period
(cid:32)(cid:88)
(cid:33)
(1 − α(cid:96))
αr
r=(cid:96)+1
Ai,1 = Ai−1,1
α(cid:96)
+ A
(cid:96)=1
(cid:96)∈L
Using Equations 2.14-2.16 and simplifying we can write Equation 2.17 in terms of
(2.19)
for i ≥ 2 where α(cid:96) = e−∆t(cid:96)/τ(cid:96), and L = {(cid:96) : θ((cid:96)) = 1}. See the proof of Proposition
2.4 in Section 2.4.4 for the details. Now we have a recurrence in Ai,1, which we can
solve using ordinary generating functions. Note that the recurrence is of the form
Ai = bAi−1 + c where b =(cid:81)n
(cid:96)=1 α(cid:96) and c = A(cid:80)
generating functions gives a solution of
r=(cid:96)+1 αr. Then ordinary
(cid:96)∈L(1 − α(cid:96))(cid:81)n
(cid:18) bi−1 − 1
(cid:19)
b − 1
.
(cid:96)∈L(1 − α(cid:96))(cid:81)n
(cid:80)
1 −(cid:81)n
(cid:96)=1 α(cid:96)
r=(cid:96)+1 αr
.
Ai,1 = Abi−1 + c
Note that since b =(cid:81)n
(cid:96)=1 α(cid:96) < 1, then
A∞,1 = lim
i→∞ Ai,1 =
c
1 − b
= A
Thus we have a limit cycle.
For (cid:96) s.t. θ((cid:96)) = 0, we have that the total release during that period is given by
24
Ri,(cid:96) = Ai,(cid:96)(1 − α(cid:96)).
Lemma 2.2 gives a closed formula for Ai,(cid:96) in terms of Ai,1, so
Ri,(cid:96) = Ai,1(1 − α(cid:96))
αr + A
θ(r)(1 − αr)(1 − α(cid:96))
(cid:96)−1(cid:89)
(cid:96)−1(cid:88)
r=1
r=1
(cid:96)−1(cid:89)
s=r+1
αs.
2.4.3 Special cases of the generalized model
Release only. Consider the case where each period is a release period. Then L = ∅.
So for the recursion we have
Ai,1 =
which has solution
α(cid:96)
(cid:32) n(cid:89)
(cid:32) n(cid:89)
(cid:96)=1
(cid:33)
Ai−1,1,
(cid:33)i−1
α(cid:96)
.
Ai,1 = A
When we take the limit we get
(cid:96)=1
A∞,1 = lim
i→∞ Ai,1 = A · 0 = 0
as expected.
Replenishment only. Consider the case where each period is a replenishment
period. Then L = {1, . . . , n}. So for the recursion we have
n(cid:88)
(1 − α(cid:96))
n(cid:89)
(cid:32) n(cid:89)
(cid:33)
(cid:32)
1 − n(cid:89)
(cid:33)
Ai,1 =
α(cid:96)
Ai−1,1 + A
αr =
α(cid:96)
Ai−1,1 + A
α(cid:96)
,
(cid:96)=1
(cid:96)=1
r=(cid:96)+1
(cid:96)=1
(cid:96)=1
(cid:32) n(cid:89)
(cid:33)
25
which has solution
(cid:33)i−1
α(cid:96)
+ A
(cid:32) n(cid:89)
(cid:96)=1
1 −
(cid:32) n(cid:89)
(cid:96)=1
(cid:33)i−1 .
α(cid:96)
Ai,1 = A
When we take the limit we get
A∞,1 = lim
i→∞ Ai,1 = A
as expected.
Alternating release and replenishment periods. Consider the case where we
have alternating periods of release and replenishment, starting with release. Then
L = {2, 4, . . . , n}. So for the recursion we have
(cid:32) n(cid:89)
(cid:33)
(cid:88)
(1− α(cid:96))
n(cid:89)
(cid:32) n(cid:89)
(cid:33)
n(cid:88)
n(cid:89)
(−1)(cid:96)
αr,
Ai,1 =
α(cid:96)
Ai−1,1 + A
=
α(cid:96)
Ai−1,1 + A
(cid:96)=1
(cid:96) even
r=(cid:96)+1
(cid:96)=1
(cid:96)=1
r=(cid:96)+1
which has solution
(cid:32) n(cid:89)
(cid:96)=1
(cid:33)i−1
(cid:32)
α(cid:96)
+
A
n(cid:88)
(cid:96)=1
(−1)(cid:96)
(cid:33)(cid:32)
1 − ((cid:81)n
1 −(cid:81)n
(cid:96)=1 α(cid:96))i−1
(cid:96)=1 α(cid:96)
(cid:33)
.
αr
n(cid:89)
r=(cid:96)+1
Ai,1 = A
When we take the limit we get
A∞,1 = lim
i→∞ Ai,1 =
A(cid:80)n
(cid:96)=1(−1)(cid:96)(cid:81)n
1 −(cid:81)n
(cid:96)=1 α(cid:96)
r=(cid:96)+1 αr
.
We can calculate the total release during release period (cid:96) by
(cid:96)−1(cid:89)
(cid:96)−1(cid:88)
(cid:96)−1(cid:89)
αs.
Ri,(cid:96) = Ai,1(1 − α(cid:96))
αr + A(1 − α(cid:96))
(−1)r
r=1
r=1
s=r+1
Alternating release and replenishment, original model.
In the case of our
experiment we have α1 = α, α2 = β, L = {2}, and n = 2. So for the recursion we have
26
Ai,1 = Ai−1,1αβ + A(1 − β),
which has solution
Ai,1 = A(αβ)i−1 + A(1 − β)
1 − (αβ)i−1
1 − αβ
.
Taking the limit we get
A∞,1 = lim
i→∞ Ai,1 =
A(1 − β)
1 − αβ
.
Note that Ri,1 = Ai,1(1 − α) for all i. So
Ri,1 = A(αβ)i−1(1 − α) + A(1 − β)(1 − α)
1 − (αβ)i−1
1 − αβ
.
Thus R = limi→∞ Ri,1 =
2.3.1 (without the f ).
A(1 − α)(1 − β)
1 − αβ
, which matches our prediction from Section
Two different release periods, but identical replenishment. Here we have
alternating periods of release and replenishment with two different release periods,
starting with release, so L = {2, 4}. We also assume that all of the replenishment
periods have the same dynamics. Let β := α2 = α4. So for the recursion we have
Ai,1 = β2α1α3Ai−1,1 + A(1 − β)(1 + α3β),
27
which has solution
Ai,1 = A(β2α1α3)i−1 + A(1 − β)(1 + α3β)
1 − (β2α1α3)i−1
1 − β2α1α3
.
Taking the limit we get
A∞,1 = lim
i→∞ Ai,1 =
A(1 − β)(1 + βα3)
1 − β2α1α3
.
2.4.4 Supporting lemmas
In this section, we give the proof of the recursion in Equation 2.17. We also give
proofs of the supporting lemmas used to prove Proposition 2.4 (Equation 2.19).
Lemma 2.1. The solution to the recursion Ai = bAi−1 + c is Ai = A1bi−1 + c
1 − bi−1
1 − b
.
In the following proof we use a standard generating function technique for solving
recursions found in [42].
Proof. To solve using generating functions we first multiply the recursion by xi and
sum over i ≥ 2 to get:
(cid:88)
i≥2
(cid:88)
i≥2
Aixi = b
Ai−1xi + c
(cid:88)
i≥2
xi.
(cid:18) 1
1 − x
(cid:19)
.
− 1 − x
Let A(x) =(cid:80)
i≥1 Aixi. Then
A(x) − A1x = bxA(x) + c
Solving for A(x) yields:
28
bixi
A(x) = A1
x
1 − bx
(cid:32)
(cid:88)
i≥1
=
Thus, Ai = A1bi−1 + c
(cid:88)
xi(cid:88)
bixi + cx2(cid:88)
(cid:19)
i≥0
xi.
x2
(cid:33)
(cid:18)
A1bi−1 + c
k=0
bk
i≥1
+ c
xi =
= A1x
(1 − x)(1 − bx)
(cid:88)
i≥0
1 − bi−1
1 − b
i≥0
A1bi−1 + c
i−2(cid:88)
(cid:19)
(cid:18)1 − bi−1
1 if (cid:96) indexes a replenishment period
(cid:96)−1(cid:88)
0 if (cid:96) indexes a release period
(cid:96)−1(cid:89)
(cid:96)−1(cid:89)
1 − b
.
θ(r)(1 − αr)
αs.
.
r=1
r=1
s=r+1
Recall α(cid:96) = e−∆t(cid:96)/τ(cid:96) and θ((cid:96)) =
Lemma 2.2. Ai,(cid:96) = Ai,1
αr + A
Proof. We induct on (cid:96). Note that for (cid:96) = 1 we have
0(cid:89)
0(cid:88)
0(cid:89)
Ai,1
αr + A
θ(r)(1 − αr)
αs = Ai,1.
r=1
r=1
s=r+1
So the result holds for (cid:96) = 1. Let (cid:96) > 1 and assume that the result holds for all
smaller (cid:96). Then
Ai,(cid:96) = Ai,(cid:96)−1 + ai,(cid:96)−1(∆t(cid:96)−1) by Equation 2.14
= Ai−1,(cid:96) + (θ((cid:96) − 1)A − Ai,(cid:96)−1)(1 − α(cid:96)−1) by Equation 2.18
= Ai,(cid:96)−1α(cid:96)−1 + θ((cid:96) − 1)A(1 − α(cid:96)−1)
θ(r)(1 − αr)
αr + A
(cid:33)
(cid:32)
Ai,1
αs
=
α(cid:96)−1 + θ((cid:96) − 1)A(1 − α(cid:96)−1)
(cid:96)−2(cid:89)
(cid:96)−1(cid:89)
r=1
(cid:96)−2(cid:88)
(cid:96)−2(cid:88)
r=1
by the induction hypothesis
(cid:96)−2(cid:89)
(cid:96)−1(cid:89)
s=r+1
= Ai,1
αr + A
θ(r)(1 − αr)
αs + θ((cid:96) − 1)A(1 − α(cid:96)−1)
r=1
r=1
s=r+1
(cid:96)−1(cid:89)
r=1
αr + A
(cid:96)−1(cid:88)
r=1
= Ai,1
θ(r)(1 − αr)
(cid:96)−1(cid:89)
s=r+1
αs.
n(cid:88)
(cid:96)=1
(1 − α(cid:96))
(cid:96)−1(cid:88)
r=1
θ(r)(1 − αr)
(cid:96)−1(cid:89)
s=r+1
αs =
Lemma 2.3.
Proof.
n(cid:88)
(cid:96)=1
(1 − α(cid:96))
(cid:96)−1(cid:88)
r=1
θ(r)(1 − αr)
(cid:96)−1(cid:89)
s=r+1
αs = θ(1)(1 − α1)
r=(cid:96)+1
(cid:32)
(cid:96)=1
s=2
r=2
θ((cid:96))(1 − α(cid:96))
(1 − αr)
n−1(cid:88)
(cid:32) n(cid:88)
(cid:32) n(cid:88)
n(cid:88)
(cid:32) r−1(cid:89)
n(cid:88)
(cid:32)
1 − n(cid:89)
r=(cid:96)+1
r=(cid:96)+1
αs
1 − n(cid:89)
(cid:33)
(cid:33)
r−1(cid:89)
r−1(cid:89)
r−1(cid:89)
αs − r(cid:89)
(cid:33)
s=(cid:96)+1
αs
(1 − αr)
s=(cid:96)+1
αr
s=(cid:96)+1
+θ(2)(1 − α2)
αs
+··· + θ(n − 1)(1 − αn−1)(1 − αn)
(1 − αr)
r=3
s=3
(cid:96)=1
n(cid:88)
n(cid:88)
n(cid:88)
(cid:96)=1
=
=
=
θ((cid:96))(1 − α(cid:96))
θ((cid:96))(1 − α(cid:96))
θ((cid:96))(1 − α(cid:96))
29
(cid:33)
αr
(cid:33)
αs
Proposition 2.4. Ai,1 = Ai−1,1
(cid:96)=1
(cid:33)
α(cid:96)
+ A
(cid:32) n(cid:89)
(cid:96)=1
(cid:32)(cid:88)
(cid:96)∈L
(1 − α(cid:96))
r=(cid:96)+1
(cid:33)
n(cid:89)
r=(cid:96)+1
αr
Proof. We know that
Ai,1 = Ai−1,1 +
n(cid:88)
(cid:96)=1
ai−1,(cid:96)(∆t(cid:96)) = Ai−1,1 +
n(cid:88)
(cid:96)=1
(θ((cid:96))A − Ai−1,(cid:96))(1 − α(cid:96)).
Then we have
Ai,1 = Ai−1,1 +
n(cid:88)
(cid:96)=1
θ((cid:96))A(1 − α(cid:96))
(cid:32)
− n(cid:88)
(cid:96)=1
(cid:96)−1(cid:89)
r=1
αr + A
(cid:96)−1(cid:88)
r=1
(cid:96)−1(cid:89)
(cid:33)
αs
30
(1 − α(cid:96)) by Lemma 2.2
s=r+1
= Ai−1,1 +
θ((cid:96))A(1 − α(cid:96)) − Ai−1,1
αr(1 − α(cid:96))
θ(r)(1 − αr)
(cid:96)−1(cid:89)
r=1
n(cid:88)
(cid:96)−1(cid:89)
(cid:96)=1
Ai−1,1
n(cid:88)
(cid:96)=1
(cid:96)=1
n(cid:88)
n(cid:89)
(cid:32) n(cid:89)
(cid:32) n(cid:89)
(cid:96)=1
(cid:96)=1
(cid:96)=1
r=1
(cid:96)−1(cid:88)
n(cid:88)
(cid:33)
(cid:33)
(cid:96)=1
by Lemma 2.3
= Ai−1,1
α(cid:96)
+ A
= Ai−1,1
α(cid:96)
+ A
+A
(1 − α(cid:96))
θ(r)(1 − αr)
= Ai−1,1
α(cid:96) +
θ((cid:96))A(1 − α(cid:96)) − A
θ((cid:96))(1 − α(cid:96))
(cid:33)
(cid:32)
1 − n(cid:89)
(cid:33)(cid:33)
(cid:32)
1 − n(cid:89)
r=(cid:96)+1
αr
αr
r=(cid:96)+1
αs
(cid:96)=1
s=r+1
n−1(cid:88)
(1 − α(cid:96)) −(cid:88)
n(cid:89)
(1 − α(cid:96))
(cid:96)∈L
αr
r=(cid:96)+1
(cid:32)(cid:88)
(cid:32)(cid:88)
(cid:96)∈L
(cid:96)∈L
(1 − α(cid:96))
(cid:33)
31
Chapter 3
Random walk model of vesicle
replenishment
Previous work indicates that vesicle replenishment is the rate-limiting step in sustained
vesicle release [16], so in this chapter we take a closer look at the replenishment process.
What part of the replenishment process limits release? In Section 3.1, we design a
simple random walk model to theoretically predict the time constant of replenishment,
τa, initially discussed in Section 2.1. Using the model we can determine which
parameters affect replenishment. We discover that τa relies on four fundamental
parameters: vesicle diffusion, vesicle concentration, vesicle size, and the probability of
attachment to the ribbon. The model predicts an exponential replenishment curve
with time constant
τa =
1
Dρδs
where D is the vesicle diffusion coefficient, ρ is the vesicle concentration, δ is the
diameter of a single vesicle, and s is the attachment probability. The nature of vesicle
movement within the synapse leads us to ask if the random diffusion of vesicles is
32
rate-limiting. We compare experimental data with our model results and conclude
that diffusion is not, in fact, rate-limiting.
Further exploring replenishment in Section 3.2, we introduce two variations of
the original model to investigate the role of Ca2+ on replenishment. The results in
Sections 3.1 and 3.2 are published in [38].
We can also use the model to calculate several other quantities of interest: how
many vesicles hit the ribbon per second (hit rate) and how long it takes to fill up the
ribbon (expected waiting time). The derivations in Section 3.3 are unpublished.
3.1 Replenishment timescale
In this section we wish to answer the question: Is vesicle diffusion a rate-limiting
step for replenishment? We first discuss how to measure the replenishment curve
experimentally. Then since vesicles move randomly in the cell terminal without a
directed movement toward the ribbon or active zone [24], we create a random walk
model of vesicle movement and replenishment.
3.1.1 Paired pulse recordings
To experimentally measure the replenishment of vesicles onto the synaptic ribbon, we
use paired pulse recordings. First, when the ribbon is full, a large voltage jump, or
pulse, is applied to the cell (similar to the pulse trains in Section 2.2.1) and vesicle
release is measured. The size of the voltage jump is such that all vesicles from the
ribbon are released. After t seconds a second pulse of the same amplitude is applied
and vesicle release is again measured (see Figure 3.1). This is repeated for multiple
values of t to approximate the replenishment curve. For each inter-pulse interval, t,
we plot the ratio R2/R1 where R1 is the release on the first pulse and R2 is the release
33
on the second pulse (see Figure 3.1). The ratio, R2/R1, can be thought of as the
percentage of vesicles replenished. In the case of salamader cones, the replenishment
curve can be fit with a double exponential with time constants τfast=815 ms (76%)
and τslow=13 s [38].
Figure 3.1: Panel A shows an example of two paired pulse recordings with interpulse
intervals of 500ms and 2s. Note that the longer interpulse interval gives more time for
the ribbon to replenish and hence the the second pulse is larger in the 2s trial. Panel
B shows the replenishment curve. The horizontal axis gives the interpulse interval
and the vertical axis gives the ratio of the two responses. This ratio can be thought of
as the percentage replenished. Adapted from [38].
In the next section our goal is to predict the time constant of replenishment theoretically
using a random walk model.
3.1.2 Derivation of the replenishment time constant
To answer this question, we developed a three-dimensional random walk model. We
modeled the vesicle motion in the synapse by spherical vesicles undergoing random
walks on a rectangular lattice of spacing δ. During each time step, ∆t, every vesicle
moves to an adjacent lattice site in each dimension. We update each of the three
dimensions simultaneously, resulting in a diagonal move overall. The (macroscopic)
AB 34
diffusion coefficient,
D =
δ2
2∆t
,
relates δ and ∆t in the (microscopic) random walk model, so these quantities cannot
be chosen independently [6]. Moreover, we would like to assume that each lattice
site can be occupied by at most one vesicle, and that the occupation probabilities for
distinct lattice sites are independent. These two assumptions can only be satisfied if
we choose δ to be equal to the vesicle diameter.
Figure 3.2: Random walk model of ribbon replenishment: the vesicles undergo a
random walk on a rectangular lattice of spacing δ [38].
We use p to denote the probability that a given lattice site (or tethering site) on
the ribbon will become occupied in a given time step. If we assume the vesicles are
distributed randomly and uniformly within the cell, the probability of a given lattice
site being occupied is independent from one time step to the next. For a lattice site
far from the ribbon, this probability is simply given by the vesicle density per lattice
site, ρδ3, where ρ is the overall density of vesicles inside the cell. Since the ribbon
sites can only be accessed from one side, for ribbon sites we have collision probability
35
p =
1
2
ρδ3.
Let s be the attachment probability, the probability that a vesicle that comes into
contact with the ribbon will "stick."1 Then sp is the probability of a vesicle actually
sticking to a ribbon site in a given time step. Thus the probability of having to wait
at least t seconds before a ribbon site is "permanently" occupied is:
P (t) = (1 − sp)t/∆t,
with t/∆t giving the total number of time steps that have elapsed in t seconds. Note
that 1 − sp is the probability that a given lattice site on the ribbon is not occupied
permanently in a given time step.
Now we make a crucial approximation for P (t), which is valid for sp << 1.2 The
approximation stems from the fact that ln(1 + x) ≈ x for x<< 1. To use it, we
first take the natural log of the P (t) equation, and then plug in ∆t = δ2/2D and
p = ρδ3/2:
ln P (t) =
t
∆t
ln(1 − sp) ≈ t
∆t
t = −2Dρδ3s
(−sp)
t = −(Dρδs)t.
= −2Dps
δ2
2δ2
Exponentiating both sides we obtain
P (t) ≈ e−t/τa, where
τa =
1
Dρδs
.
1i.e. become tethered to the ribbon until release, not drifting away at a future time step.
2In fact, the approximation is still quite good up to values of sp ∼ 0.1.
36
Solving for P (t) without making the approximation we get
P (t) = e−t/τexact, where
τexact =
−δ2
2D ln(1 − 1
2ρδ3s)
.
Next, observe that the expected number of ribbon sites that are filled at time t,
assuming all sites are empty at t = 0, is given by
a(t) =
m
(cid:18) n
(cid:19)
m
m=1
(1 − P (t))mP (t)n−m = n(1 − P (t)).
n(cid:88)
(cid:1)xmyn−m = (x + y)n, by the Binomial Theorem. Differentiating
n(cid:88)
(cid:18) n
xm−1yn−m = n(x + y)n−1.
(cid:19)
m
Recall that(cid:80)n
(cid:0) n
m=0
m
with respect to x:
The second equality is obtained using a familiar variant of the Binomial Theorem.
Now, letting x = 1 − P (t) and y = P (t), we obtain
m
m=1
(cid:18) n
(cid:19)
m
m
n(cid:88)
m=1
(1 − P (t))m−1P (t)n−m = n.
Finally, multiplying both sides by 1− P (t) we obtain the desired result. Note that each
term in the sum corresponds to the probability that exactly m sites are "permanently"
occupied at time t, weighted by m. On the other hand, given that each of the n
ribbon sites has an occupation probability of 1 − P (t) at time t, it is intuitive that
the expected number of occupied sites at this time is a(t) = n(1 − P (t)).
Using the approximate expression for P (t) we derived above, we obtain, a(t), the
expected number of vesicles on the ribbon at time t, in terms of our fundamental
constants:
a(t) = n(1 − e−t/τa), where
τa =
1
Dρδs
.
(3.1)
37
3.1.3 Comparison of model predictions with data
We wish to determine whether diffusion is rate-limiting for replenishment and thus
sustained release. We answer this question by comparing the experimentally measured
time constant with the model predictions. Recall that in Section 3.1.1, we fit an
exponential replenishment curve with two time constants to the data with τfast=815
ms (76%) and τslow=13 s.
.
Table 3.1 shows the experimental values for all of the fundamental constants for
salamander cones. Since we are interested in knowing whether diffusion is rate-limiting,
we use our model to calculate the fastest possible timescale of vesicle replenishment
due to vesicle diffusion. To do this we set the attachment probability s equal to 1.
Hence if all vesicles that collide with the ribbon due to diffusion attach to it with
probability s = 1 then the predicted time constant is
τa =
1
(.11)(2210)(.045)
seconds = 91 ms.
Thus the model predicts that the fastest replenishment time constant for salamander
cone ribbons is 91 ms, which is about an order of magnitude faster than the experi-
mentally measured τfast of 815 ms. This suggests that other factors beyond the rate of
vesicle collisions with the ribbon, such as an attachment probability s < 1, time of
descent down the ribbon, and/or vesicle priming must play a role in slowing down the
rate of vesicle accretion.
Since our theoretical model does not take into account ribbon geometry aside from
assuming that the sites are only accessible from one side, it is reasonable to use this
model to predict the replenishment time constant for other ribbon and conventional
synapses provided their vesicles also exhibit random motion. The vesicles in rod
38
constant
meaning
n
D
ρ
δ
s
max no. of vesicles on the ribbon
vesicle diffusion coefficient
(mobile) vesicle density
vesicle diameter
attachment probability
measured value
110 vesicles [2]
0.11 µm2/s [24]
2210 vesicles/µm3 [37]
45 nm = 0.045 µm[28]
0 < s ≤ 1
Table 3.1: Experimentally measured parameters for model of replenishment in sala-
mader cones.
bipolar cells, goldfish bipolar cells, and hippocampal cells all appear to exhibit random
motion [10, 14, 29]. Rod bipolar cells and goldfish bipolar cells both contain ribbons
[10, 14], but hippocampal cells do not [29]. Table 3.2 gives the parameters for rod
bipolar cells, goldfish bipolar cells, and hippocampal cells. In these cells, note that
our model predicts a replenishment time constant that is slower than the measured
replenishment time constant, indicating that the motion of vesicles may be rate-limiting
for replenishment in these synapses.
Diffusion coefficient, D
Rod Bipolar Cells Goldfish Bipolar Cells
0.015 µm2/s [14]
0.015 µm2/s [14]
Hippocampal Cells
0.0042 µm2/s [29]
Vesicle diameter, δ
38 nm [10]
30 nm [22]
38 nm [13]
Vesicle Concentration, ρ
1933 v/µm3 [10]
445 v/µm3 [14, 24]
270-465 v/µm3 [12, 26]
Measured τa
Predicted τa
400 ms [30]
908 ms
4 s [20]
5 s
7 s[34, 41, 9]
13-23 s
Table 3.2: Model predictions for the fastest possible timescale of replenishment for
other synapses based on experimentally measured D, δ, and ρ.
3.2 Role of calcium in replenishment
In this section we use variations on our model to test two different mechanisms by which
calcium (Ca2+) and calmodulin (CaM), a calcium-binding messenger protein, might
govern the attachment probability, s. It is known that Ca2+ speeds replenishment
39
[1]. However, the mechanism by which this occurs is unknown. Data suggests that
Ca2+/CaM do not accelerate vesicles from the top of the ribbon to the release sites,
nor do they increase the fusion rate at the membrane [38]. Increased intracellular Ca2+
does not affect the mobility of vesicles in the terminal [24, 14]. Hence D and ρ would
not be affected by calcium. It appears that vesicle size (quantal amplitude) is also
not affected by calcium [38], so we posit that Ca2+/CaM increases the attachment
probability, s. In this section, we will use two variations of the model to test two
hypotheses regarding the role of Ca2+/CaM. The first variant, which we call Model 1,
assumes that Ca2+/CaM acts as a switch making some vesicles more "sticky" than
others. In the second variant, Model 2, we assume Ca2+/CaM again acts as a switch,
but this time on the ribbon tethering sites, making some ribbon sites more "sticky"
than others but leaving the vesicles unchanged. Perhaps surprisingly, these two models
produce qualitatively different results. This may enable us to distinguish between the
two possible functions of Ca2+/CaM embedded into each model, by comparing the
model predictions to experimental observations.
Figure 3.3: Calcium Hypotheses: Model 0 is the initial setup from Section 3.1.2, Model
1 assumes changes in s occur at the vesicles, and Model 2 assumes changes in s occur
at the ribbon sites [38].
δ}δ}Model 0Model 1Model 2 40
3.2.1 Calcium affects vesicles
For Model 1, suppose we have two populations of vesicles, A and B. Vesicles in
population A have higher attachment probability when they collide with the ribbon,
given by attachment probability sA. Vesicles in population B have a lower attachment
probability, given by attachment probability sB. Then, 0 < sB ≤ sA ≤ 1. Let f be the
fraction of vesicles in population A, with 1 − f the fraction in population B. Since the
ribbon sites are identical, the probability of a vesicle collision resulting in attachment
is simply given by the weighted average of these attachment probabilities:
s = f sA + (1 − f )sB.
The rest of the model remains unchanged. In particular, we still have
a(t) = n(1 − e−t/τa), where
τa =
1
Dρδs
,
and s is the "average" attachment probability computed above.
If the effect of
Ca2+/CaM is to change the fraction f of vesicles in the stickier population, then this
effect will manifest itself as a change in the vesicle accretion timescale, τa. Inhibition
of Ca2+/CaM should cause a decrease in f , and hence an increase in τa. Note that
this model does not predict the existence of a second timescale, even though there are
two populations of vesicles.
3.2.2 Calcium affects the ribbon
For Model 2, suppose all vesicles are identical, but we have two populations, A and B,
of tethering sites on the ribbon. The ribbon sites in population A are more sticky,
modeled by a higher attachment probability sA, while the ribbon sites in population
41
B are less sticky, with sB < sA. We let nA and nB denote the number of sites in each
population, with n = nA + nB. If f is the fraction of ribbon sites in population A,
then nA = f n and nB = (1 − f )n.
Since attachment probabilities are different for different ribbon sites, we must use
different expressions for P (t): PA(t) = (1 − sAp)t/∆t for the sites in population A,
while PB(t) = (1 − sBp)t/∆t for population B. The result is that expected number of
vesicles on the ribbon at time t is given by the sum of two terms:
a(t) = nA(1 − e−t/τA) + nB(1 − e−t/τB ), where τA =
1
DρδsA
, and τB =
1
DρδsB
.
Note that since sA > sB, the population A timescale is faster, τA < τB. The presence of
two timescales makes this model qualitatively different from Model 1. This difference
is also seen in thinking about the effect of Ca2+/CaM in this model. If Ca2+/CaM
changes the fraction of ribbon sites f that belong to the stickier population, then this
will manifest itself as a change in the amplitudes nA and nB for each component of
a(t). Inhibition of Ca2+/CaM should cause a decrease in f , and hence a decrease
in nA and an increase in nB. This model predicts no Ca2+/CaM effect on the time
constants, in contrast to Model 1.
3.2.3 Comparison to experimental results
Recall that paired pulse experiments showed that the replenishment curve can be fit
with a double exponential. Table 3.3 shows the results of experimentally decreasing
Ca2+/CaM using BAPTA, nifedipine, Calmidazolium, and MLCK [38]. When the
fast timescale is constrained to match the control, we see that the fit is comparable to
the unconstrained case, but the percentage of fast-replenishing sites is much lower.
Thus, inhibiting Ca2+/CaM in these experiments caused slight changes in the fast
42
timescale, but more substantial changes to the amplitude of the fast component. This
is consistent with the predictions seen in Model 2, where we have two time constants
and inhibition of Ca2+/CaM causes the amplitude of the fast component to decrease.
We conclude that Ca2+/CaM more likely acts on ribbon sites rather than vesicles. It
is possible that vesicles are affected by Ca2+/CaM as well, but changes in the fast
timescale were not consistent across trials.
3.3 Other quantities of interest
3.3.1 Hit rate
In this section we compute the hit rate, i.e. the number of vesicles coming in contact
with the ribbon per second. One way to compute the hit rate is to do so macroscopically
by first calculating the flux near the ribbon and then multiplying by the surface area
of the ribbon. We consider the concentration of vesicles to be zero on the ribbon.
The concentration of mobile vesicles not attached to the ribbon is ρ as before. Thus
over the distance δ (one lattice step from a ribbon site to a nonribbon site) we have
vesicles
µm2 · s
[6]. The total surface area of the ribbon is nδ2 µm2. Thus computing the hit rate
a change in concentration from ρ to 0 giving us a flux of J = D
ρ − 0
=
Dρ
δ
δ
using this method yields H =
Dρ
δ
nδ2 = Dρδn =
n
τa
vesicles/s where τa =
1
Dρδ
is the
replenishment time constant from our previous calculations.
We can also microscopically compute the hit rate. First, find the expected number
of sites filled in a single time step. We know that the probability of a ribbon site
becoming occupied in the next time step is
1
2
ρδ3 and there are n sites on the ribbon.
1
2
ρδ3n. We
Thus the expected number of sites filled in a single time step is given by
know that each time step ∆t is
δ2
2D
seconds. Thus the formula for hit rate in our
43
K
C
L
M
M
µ
0
2
-
K
C
L
M
M
µ
0
2
.
m
l
a
C
M
µ
0
2
e
n
i
i
p
d
e
f
i
n
M
µ
3
A
T
P
A
B
M
m
1
A
T
G
E
M
m
5
r
e
t
e
m
a
r
a
P
7
1
9
.
0
9
.
3
2
8
.
3
1
0
.
1
b
1
5
9
.
0
3
.
5
2
1
.
4
1
0
.
1
7
1
l
o
r
t
n
o
c
1
5
9
.
0
1
.
8
6
6
.
7
1
8
9
.
0
6
8
9
.
1
1
.
3
3
9
.
5
2
9
9
.
0
a
6
1
8
.
0
5
.
7
1
4
.
7
1
9
9
.
0
4
9
3
.
1
1
.
7
3
6
.
7
1
9
9
.
0
a
6
1
8
.
0
9
.
6
2
2
.
2
1
9
9
.
0
5
1
.
2
2
.
6
5
8
.
8
2
9
9
.
0
a
6
1
8
.
0
1
.
2
2
6
5
.
9
7
9
.
0
7
)
l
o
r
t
n
o
c
(
6
1
8
.
0
7
.
5
7
9
.
2
1
7
9
.
0
0
1
d
e
n
i
a
r
t
s
n
o
c
n
U
)
s
(
t
s
a
f
τ
t
s
a
f
%
)
s
(
w
o
l
s
τ
2
R
d
e
n
i
a
r
t
s
n
o
C
)
s
(
t
s
a
f
τ
t
s
a
f
%
)
s
(
w
o
l
s
τ
2
R
n
.
n
o
i
t
c
n
u
f
l
a
i
t
n
e
n
o
p
x
e
e
l
b
u
o
d
a
h
t
i
w
t
fi
s
i
t
n
a
t
s
n
o
c
e
m
i
t
t
n
e
m
h
s
i
n
e
l
p
e
r
e
h
T
:
M
a
C
/
+
2
a
C
f
o
s
t
c
e
ff
E
:
3
.
3
e
l
b
a
T
,
s
e
s
a
c
l
o
r
t
n
o
c
o
w
t
e
h
t
m
o
r
f
e
d
i
s
A
.
s
e
s
a
c
d
e
n
i
a
r
t
s
n
o
c
d
n
a
d
e
n
i
a
r
t
s
n
o
c
n
u
e
h
t
n
i
r
a
l
i
m
i
s
e
r
a
s
t
fi
e
h
t
t
a
h
t
e
t
o
N
-
u
l
l
e
c
a
r
t
n
i
f
o
d
a
e
r
p
s
e
h
t
s
t
c
i
r
t
s
e
r
A
T
P
A
B
.
y
a
w
e
m
o
s
n
i
m
u
i
c
l
a
c
s
e
s
a
e
r
c
e
d
n
o
i
t
i
d
n
o
c
l
a
t
n
e
m
i
r
e
p
x
e
r
e
h
t
o
h
c
a
e
.
s
r
o
t
i
b
i
h
n
i
n
i
l
u
d
o
m
l
a
c
e
r
a
K
C
L
M
d
n
a
)
.
m
l
a
C
(
m
u
i
l
o
z
a
d
m
l
a
C
i
d
n
a
,
x
u
fl
n
i
+
2
a
C
s
e
c
u
d
e
r
e
n
i
i
p
d
e
f
i
n
,
+
2
a
C
r
a
l
l
o
r
t
n
o
c
-
K
C
L
M
r
o
f
t
s
a
f
τ
o
t
d
e
n
i
a
r
t
s
n
o
c
s
t
i
F
b
.
)
s
6
1
8
.
0
(
n
o
i
t
i
d
n
o
c
l
o
r
t
n
o
c
A
T
G
E
M
m
5
r
o
f
t
s
a
f
τ
o
t
d
e
n
i
a
r
t
s
n
o
c
s
t
i
F
a
.
]
8
3
[
m
o
r
f
d
e
t
p
a
d
A
.
)
s
1
5
9
.
0
(
n
o
i
t
i
d
n
o
c
44
model is
1
2
ρδ3n vesicles
1 time step
H =
· 1 time step
seconds
δ2
2D
= Dρδn vesicles/s,
in agreement with the flux-based calculations above.
Lastly, we can again confirm this result by approximating the hit rate from
, which is the formula for the rate of accumulation of vesicles onto the ribbon
da
dt
from previous calculations. The hit rate computed above ignores the decrease in
available surface area due to vesicles already on the ribbon and thus corresponds to
the accumulation rate only for small values of t. Recall that a(t) = n(1 − e−t/τa) is
the expected number of vesicles on the ribbon at time t where τa =
1
Dρδs
. Here we
assume s = 1 since we are just finding how many vesicles come in contact with the
ribbon per second and ignoring how many stick. For t = 0 the rate of accumulation
also corresponds to the hit rate since the form of a(t) assumes that the ribbon is empty
e−t/τa, so at t = 0 we have that the rate of accumulation is
=
da
dt
n
τa
at t = 0. Then
n
τa
= Dρδn vesicles/s. So again the hit rate is given by H = Dρδn vesicles/s.
In summary, when the ribbon is empty, the hit rate is
H =
n
τa
= Dρδn vesicles/s,
but as the ribbon becomes filled the hit rate decreases as
H(t) =
e−t/τa.
n
τa
Note that H = H(0) and
(cid:90) ∞
0
H(t)dt =
(cid:90) ∞
0
n
τa
e−t/τadt =
n
τa
(cid:12)(cid:12)(cid:12)(cid:12)∞
0
(−τae−t/τa)
=
n
τa
τa = n.
45
3.3.2 Expected waiting time
To calculate the expected waiting time, Twait, to fill all n lattice sites on the ribbon,
we first consider an individual lattice site. Let P (t) be the probability that we wait
at least t seconds to fill the given lattice site. Then 1 − P (t) is the probability that
the given lattice site fills before t seconds have passed and r(t) = (1 − P (t))n is
the probability that all n sites have filled before t seconds have passed. Hence the
probability we wait exactly t seconds is r(cid:48)(t)dt.
Thus the expected waiting time Twait is given by
(cid:90) ∞
Twait = E[t] =
tr(cid:48)(t)dt = τaHn
0
where τa is the vesicle accretion timescale and Hn =
This result is proven in Lemma 3.1.
n(cid:88)
k=1
1
k
tr(cid:48)(t)dt =
=
=
=
n
τa
n
τa
n
τa
n
τa
= τa
te−t/τa(1 − e−t/τa)n−1dt
0
0
te−t/τa
(cid:90) ∞
(cid:90) ∞
n−1(cid:88)
(cid:18)n − 1
(cid:19)
n−1(cid:88)
(cid:18)n − 1
(cid:19)
n−1(cid:88)
(cid:18)n − 1
(cid:19)
n−1(cid:88)
k=0
k=0
k
k
k=0
n
k
k=0
(cid:18)n − 1
(cid:19)
(cid:90) ∞
k
(−1)k
(−1)k
0
τ 2
a
(k + 1)2
(−1)k
1
(k + 1)2 .
(−1)ke−kt/τadt
te−(k+1)t/τadt
Lemma 3.1.
nth Harmonic number.
0
tr(cid:48)(t)dt = τaHn where r(t) = (1 − e−t/τa)n, and Hn =
Proof. We have r(cid:48)(t)dt = n
e−t/τa(1 − e−t/τa)n−1dt, so
τa
is the nth harmonic number.
n(cid:88)
k=1
1
k
is the
(cid:90) ∞
(cid:90) ∞
0
(cid:19)
(cid:18)n − 1
(cid:19)
k + 1
k
46
n
(cid:18) n
(cid:18)n − 1
(cid:19)
n−1(cid:88)
(cid:18)n
(cid:19)
n(cid:88)
(cid:90) 1
n−1(cid:88)
k=1
k=0
k
k
dx =
0
k=0
Using the identity (k + 1)
= n
and reindexing, we can rewrite
τa
(−1)k
1
(k + 1)2 = τa
(−1)k−1 1
k
.
(cid:18)n
(cid:19)
k
n(cid:88)
k=1
Now, we claim that
(cid:90) 1
0
1 − xn
1 − x
(−1)k−1 1
k
xkdx =
= Hn. First note that
(cid:90) 1
n−1(cid:88)
k=0
0
n(cid:88)
k=1
xk−1
k
xk =
(cid:12)(cid:12)(cid:12)(cid:12)1
0
n(cid:88)
k=1
1
k
=
= Hn.
Then, letting u = 1 − x, we have,
(cid:90) 1
0
1 − xn
1 − x
dx = −
1
u
=
k=0
du
k
u
1 − (1 − u)n
(cid:90) 0
(cid:0)n
(cid:1)(−1)kuk
1 −(cid:80)n
(cid:90) 1
1 −(cid:0)n
(cid:1)(−1)0u0 −(cid:80)n
(cid:90) 1
(cid:19)
(cid:18)n
(cid:90) 1
= − n(cid:88)
(cid:34)
(cid:35)
(cid:12)(cid:12)(cid:12)(cid:12)1
(cid:19)
(cid:18)n
n(cid:88)
(−1)k−1
(−1)k
uk
k
k=1
=
=
0
0
k
k
0
0
k=1
=
0
u
uk−1du
du
(cid:0)n
(cid:1)(−1)kuk
k=1
k
du
(cid:18)n
k
(cid:19)
(−1)k−1 1
k
.
n(cid:88)
k=1
(cid:18)n
k
n(cid:88)
k=1
(cid:19)
(−1)k−1 1
k
Thus,
(cid:90) ∞
0
tr(cid:48)(t)dt = τaHn.
= Hn and therefore
Recall that we predicted τa = 91 ms and H110 is approximately 5.2882, so the expected
waiting time Twait is 481 ms. Note that this calculation is useful to set the duration for
computer simulations in the computational model of replenishment (see Chapter 4).
Expected waiting time in Model 1 Recall that in Model 1, we have two popu-
lations of vesicles where sA and sB are the attachment probabilities for population
47
A and population B, respectively. Then the expected waiting time for Model 1 is
where s = f sA + (1 − f )sB and f is the fraction of vesicles in
Twait = τaHn =
Hn
Dρδs
population A.
Expected waiting time in Model 2 Recall that in Model 2, the stickiness occurs
in the ribbon sites instead. We have nA ribbon sites with attachment probability sA
and nB sites with attachment probability sB. We know that the probability of having
to wait at least t seconds before a ribbon site in population A is occupied is given by
PA(t) = e−t/τA where τA = 1
and the probability of having to wait at least t seconds
before a ribbon site in population B is occupied is given by PB(t) = e−t/τB where
. Then r(t) = (1 − PA(t))nA(1 − PB(t))nB = (1 − e−t/τA)nA(1 − e−t/τB )nB
τB = 1
DρδsA
DρδsB
is the probability that all n = nA + nB sites have filled before t seconds have passed.
Hence the probability we wait exactly t seconds is r(cid:48)(t)dt.
Now, we have that the expected waiting time for the ribbon to fill is
(cid:90) ∞
(cid:90) ∞
0
0
Twait =
=
(cid:18)
tr(cid:48)(t)dt
t(1 − e−t/τA)nA(1 − e−t/τB )nB−1e−t/τB
+t(1 − e−t/τB )nB (1 − e−t/τA)nA−1e−t/τA
(cid:18)nA
(cid:19)(cid:18)nB − 1
(cid:19)
nA(cid:88)
nB−1(cid:88)
(cid:19)(cid:18)nA − 1
(cid:18)nB
(cid:19)
nA−1(cid:88)
nB(cid:88)
j=0
i=0
j
i
k=0
l=0
k
l
=
nB
sB
+
nA
sA
nB
τB
(cid:19)
dt.
nA
τA
(−1)i+j
1
(j + τB
i + 1)2
τA
(−1)k+l
1
(l + τA
k + 1)2
τB
48
Chapter 4
Computational model
The random walk model of vesicle replenishment described in Chapter 3 does not take
into account the geometry of the ribbon. What effect does ribbon geometry have on
replenishment? In this chapter we discuss a computational model of replenishment
including ribbon geometry that was designed to complement the theoretical model.
This model is currently unpublished. In Section 4.1 we describe the setup of the
model. Then in Section 4.2, we compare the results of the computational model
with the results of the theoretical model to determine the role geometry plays in
replenishment. The Matlab code for the computational model of replenishment can
be found in Appendix A.
4.1 Description of the computational model
The cell space is modeled by a 3-dimensional array with entries in {0, 1} where 1s
indicate locations of vesicles within the cell. The array is randomly generated with
a given concentration of 1s computed from the vesicle concentration ρ. The total
number of 1s denoted by N . The matrix S is a N × 3 matrix where the ith row records
49
the current position of the ith vesicle within the array. In each time step the matrix
is updated by adding a random N × 3 matrix with entries in {−1, 1}. This ensures
that each vesicle moves one lattice space per dimension in each time step, a diagonal
move overall. A set of n coordinates, where n is the maximum number of vesicles that
fit on the ribbon, are designated as "ribbon sites." The coordinates of these sites are
recorded in the matrix SiteMat and during each time step the coordinates of all N
vesicles are checked against SiteMat to determine how many vesicles are occupying
ribbon sites. Then with probability s, the attachment probability discussed in Section
3.1.2, a vesicle occupying a ribbon site becomes permanently stuck and does not
update in subsequent time steps. This is done by zeroing out the corresponding row
in the update matrix. At each time step we record how many vesicles are permanently
stuck to the ribbon. This gives us the computational replenishment curve, which we
can compare with our theoretical prediction.
Figure 4.1 shows the arrangement of ribbon sites in the rectangular ribbon case
and the "nonribbon" case. The placement of sites in the nonribbon case allows us to
study the effects of ribbon geometry. For cases with the rectangular ribbon, we also
make sure that the ribbon is solid by not allowing any updates that would represent a
vesicle passing through the ribbon. This is achieved by returning any vesicles that pass
through the ribbon in the current time step to their original position before moving
to the next time step.
Note that because of the way the vesicles update it is possible for vesicles to occupy
the same lattice site during the same time step. For small concentrations (around 300
v/µm3), less than 1% of the vesicles are occupying the same site as another vesicle
and for larger concentrations (around 2300 v/µm3), less than 10% of the vesicles are
occupying the same site as another vesicle.
Since the theoretical model does not take into account geometry, the computational
50
Figure 4.1: Ribbon and nonribbon attachment sites: The left panel shows the ribbon
sites arranged in a flat rectangular plate based off the structure seen in cone photore-
ceptors. The right panel shows the sites spread out in the cell space to act as a control
when studying the effects of ribbon geometry.
model and the theoretical model should be close in the nonribbon case1. Figure 4.2
shows the comparison between the two models in the nonribbon case for several
different concentrations and attachment probabilities. Note that the models closely
match across a wide range of parameters.
4.2 Effect of ribbon geometry on replenishment
Since the theoretical model does not take into account the geometry of the ribbon
we use our computational model to approximate the replenishment curve in the case
where we have a rectangular ribbon attached to the edge of the cell space.
Figure 4.3 indicates that the geometry of the ribbon does in fact play a role
in replenishment. The ribbon sites in this case have a rectangular shape based
on the ribbons in cone photoreceptors and the trials are run for varying vesicle
1Recall that in the analytical model we have a factor of 1/2 that represents the fact that the ribbon
sites are only accessible from one side. When using the analytical model to predict replenishment in
nonribbon cases, we leave out the factor of 1/2 since these sites are accessible from all sides.
102030405051015202530510152025303540455010203040505101520253051015202530354045501020304050510152025305101520253035404550NONRIBBON SITES1020304050510152025305101520253035404550RIBBON SITES 51
concentrations and attachment probabilities. For low attachment probability and high
vesicle concentration, the computational model trial average shows faster replenishment
than predicted by the analytical model. For high attachment probability and low vesicle
concentration, the computational model trial average shows slower replenishment than
predicted by the analytical model.
Recall that cone photoreceptor synapses have a high vesicle concentration and
based on our random walk model of replenishment in Chapter 3 are also likely have a
low attachment probability. Studying the case of high vesicle concentration and low
attachment probability in our computational model, we note that the computational
trial average is faster than the theoretical model prediction. This indicates that having
a synaptic ribbon for this parameter regime actually speeds replenishment compared
to having no ribbon where the vesicles dock directly with the cell membrane. This
may provide evidence for why photoreceptor cones contain ribbons, but exactly how
the ribbon accelerates replenishment in this case is still unclear.
In the case of high attachment probability and low concentration, we hypothesize
that once the ribbon starts to fill up, the local concentration near the ribbon decreases
causing the ribbon to fill slower than predicted. To test this we calculate the concen-
tration of vesicles close to the ribbon and far away from the ribbon. Figure 4.4 shows
the results of this calculation. The concentration near the ribbon drops steeply as
the ribbon fills up and the concentration further away stays relatively constant. This
drop in local concentration is most pronounced in the high s/low ρ cases. This may
account for the slower replenishment we see in these cases.
Since the theoretical model incorporates the factor of 1/2 indicating that the sites
are only accessible from one side, but not specific ribbon geometry, the theoretical
prediction gives a reasonable approximation for time constant of vesicles reaching the
cell membrane in a terminal without a ribbon. Figure 4.3 indicates that having a
52
ribbon actually may slow replenishment in synapses with low vesicles concentration.
This suggests that having a ribbon would not be advantageous in synapses with low
vesicle concentration and random motion of vesicles. This is consistent with the case
of hippocampal synapses which have a low vesicle concentration and random motion,
but do not contain ribbons [29].
4.3 Future work
Local concentration. The computational model discussed in Chapter 4 revealed
that the local concentration near the ribbon drops sharply near the ribbon as the
ribbon fills up. This contradicts our assumption that the vesicle concentration is
constant. To improve our random walk model, we would like to find a formula to
describe the change in concentration as the ribbon fills up.
Movement on ribbon and vesicle fusion. The random walk model does not
take into account the movement of vesicles along the ribbon. As more becomes known
about this process, we would like to incorporate this step into the model. This model
also does not take into account vesicle release. Adding these features will allow us to
explore more questions regarding the function of the ribbon.
53
Figure 4.2: Comparison to the theoretical model: The computational model results
for the nonribbon case (see Figure 4.2) are averaged over 100 trials and the gray
area represents one standard deviation from the mean. We show trials for a low (300
vesicles/ µm3) and a high concentration (2300 vesicles/ µm3) as well as three different
attachment probabilities (0.1, 0.5, and 1). Note that the theoretical and computational
models appear to closely match across a variety of parameters, as expected in the
nonribbon case.
vesicle concentrationattachment proability00.511.522.533.544.50102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Analytical ModelComputational Trial Average00.10.20.30.40.50.60.70.80.90102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.5 Analytical ModelComputational Trial Average00.050.10.150.20.250.30.350.40.450102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Analytical ModelComputational Trial Average051015202530350102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Analytical ModelComputational Trial Average012345670102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.5 Analytical ModelComputational Trial Average00.511.522.533.50102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Analytical ModelComputational Trial Average 54
Figure 4.3: Effect of ribbon geometry on replenishment: The computational model
results for the rectangular ribbon case (see Figure 4.2) are averaged over 100 trials
and the gray area represents one standard deviation from the mean. We show trials
for a low (300 vesicles/ µm3) and a high concentration (2300 vesicles/ µm3) as
well as three different attachment probabilities (0.1, 0.5, and 1). Note that the
computational model shows the greatest deviation from the theoretical prediction in
the low concentration/high attachment probability case.
vesicle concentrationattachment probability012345670102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.5 Analytical ModelComputational Trial Average051015202530350102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Analytical ModelComputational Trial Average00.511.522.533.50102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Analytical ModelComputational Trial Average00.511.522.533.544.50102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Analytical ModelComputational Trial Average00.10.20.30.40.50.60.70.80.90102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.5 Analytical ModelComputational Trial Average00.050.10.150.20.250.30.350.40.450102030405060708090100110Time (sec)Number of Vesicles Accruedρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Analytical ModelComputational Trial Average 55
Figure 4.4: Local concentration: The above plots show the difference between the
local concentration near the ribbon (in red) and away from the ribbon (in blue) for
two different overall vesicle concentrations (300 vesicles/µm3 and 2300 vesicles/µm3)
and two different attachment probabilities (s = 0.1 and 1) in the case where the
ribbon sites are arranged in a rectangular plate. Notice the sharp drop in the local
concentration near the ribbon when the ribbon first begins to fill. The percentage
drop is largest for the low concentration/high attachment probability case.
0123050100150200250300350Time (sec)Concentrationρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Concentration near ribbonConcentration away from ribbon00.10.20.30.405001000150020002500Time (sec)Concentrationρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 1 Concentration near ribbonConcentration away from ribbon05101520253035050100150200250300350Time (sec)Concentrationρ = 300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Concentration near ribbonConcentration away from ribbon00.511.522.533.544.505001000150020002500Time (sec)Concentrationρ = 2300 vesicles/µm3, D = 0.11 µm2/sec, s = 0.1 Concentration near ribbonConcentration away from ribbon 56
Part II
Neural Sequences in
Threshold-Linear Networks
57
Chapter 5
Introduction to Part II
Part II focuses on neural networks and the interplay between network connectivity
and neural activity. In particular, we are interested in studying how network structure
shapes the behavior of the network.
To do this, we study the dynamics of a combinatorial family of competitive
threshold-linear networks constructed from simple directed graphs (the CTLN model)
as defined in [21]. This family of networks is particularly well suited for our study
because the network construction guarantees that differences in dynamics arise solely
from differences in the connectivity of the underlying graph. This allows us to focus
on the properties of the graphs themselves when trying to predict the behavior of
the corresponding network. This robust family of dynamical systems exhibits several
different nonlinear behaviors including limit cycles, quasiperiodic attractors, and chaos.
Figure 5.1 shows an example of a network that exhibits multiple behaviors depending
on the choice of initial conditions.
In this part, we begin by giving some background about competitive threshold-
linear graphs and the CTLN model. We then use the CTLN model to study how
the graph structure affects the resulting dynamics. Computational experiments show
58
that most CTLN networks yield limit cycles. We present an algorithm that uses
the structure of the underlying graph to predict the sequence of firing of neurons
in the limit cycle. Our algorithm predicts the sequence correctly for most small
graphs, but sometimes fails for certain classes of larger graphs. To gain further insight
into how the structure of the underlying graph shapes the dynamics, we classify the
behavior we see for small networks (n ≤ 5 nodes) arising from oriented graphs. Both
of these results work towards the larger goal of better understanding high-dimensional
nonlinear dynamics.
Figure 5.1: An example on n = 8 nodes having several different behaviors based on
initial conditions: Panel A shows the graph. Panel B shows one of two stable fixed
points, Panel C shows a limit cycle, and Panel D shows a chaotic attractor. The
traces of activity are color-coded to match the colors of the nodes in the graph. The
plots on the far right show random two-dimensional projections of the 8-dimensional
trajectories corresponding to the limit cycle and the chaotic attractor. Adapted from
[21].
1623458701020304050timefiring ratefiring ratetimelimit cycleprojection015020025030050100chaosfiring ratetime0102030405060708090100fixed pointACBD 59
5.1 Threshold-linear networks
The CTLN model is a specific type of threshold-linear network. Neuroscientists use
threshold-linear networks to model recurrent neural networks [27]. These networks
are thought to be involved in perception and memory processes [11]. Development of
the mathematical theory behind threshold-linear networks is ongoing [44, 11, 21].
Definition 5.1. A threshold-linear network on n neurons is defined by the following
system of differential equations:
= −xi +
dxi
dt
(cid:34) n(cid:88)
j=1
(cid:35)
Wijxj + θ
,
+
i ∈ [n].
(5.1)
where xi is the firing rate of the ith neuron, W is the matrix of connection strengths,
θ ∈ R is the external drive to the network, and [y]+ = max{0, y} is the threshold
nonlinearity.
In our neural network context, the −xi represents the leak term and guarantees
that the activity of neuron i will die out in the absence of other inputs. Inside the
nonlinearity we have a sum of inputs from all other neurons weighted by the connection
strengths. In inhibitory networks (Wij ≤ 0), the parameter θ must be positive in order
for the nonlinear term to be nonzero.
We study the behavior of threshold-linear networks of n neurons as they are one
of the simplest examples of a nonlinear system of ordinary differential equations. In
particular, we are interested in studying the dynamics of a competitive threshold-linear
network defined from a simple directed graph.
Definition 5.2. A competitive threshold-linear network is governed by Equation 5.1
with the added restriction that Wij ≤ 0 and Wii = 0 for all i, j = 1, . . . , n and θ > 0.
In the next section we will describe the CTLN model, which is a particular type of
competitive threshold-linear network.
60
5.2 Description of the CTLN model
The Combinatorial Threshold-Linear Network model (CTLN model) was first intro-
duced by Curto et al. in [8] and further explored by Morrison et al. in [21]. This model
was designed as a way to study high-dimensional nonlinear dynamics without using
a linear approximation [21]. Linear models are limited as tools for approximation
as they do not demonstrate complex behaviors such as limit cycles, multistability,
and chaos. The nonlinearity in the CTLN model captures the full range of nonlinear
behaviors, but is still simple enough that it is possible to develop a corresponding
mathematical theory. In this chapter we describe the CTLN model and necessary
background.
Definition 5.3. The Combinatorial Threshold-Linear Network (CTLN) model refers
to the competitive threshold linear network constructed from a simple directed graph
with only two values for the inhibitory connection strengths. For any δ > 0 and
0 < ε < 1, the n × n connectivity matrix W is given by
Wij =
0
−1 + ε
−1 − δ
if i = j
if i ← j in G
if i (cid:56) j in G
(5.2)
where i ← j represents a directed edge from node j to node i in the graph G and
i (cid:56) j means that no such edge exists in G [21].
The CTLN networks are therefore fully inhibitory, with each node acting on its
61
neighbors by quieting their activity. A biological motivation for the model is shown in
Figure 5.2. Inhibitory interneurons (gray circles) inhibit all neighboring excitatory
pyramidal cells (colored triangles) equally [8]. Connections between excitatory neurons
therefore have two strengths. A directed edge represents an overall connection strength
of −1 − ε, i.e. inhibition has been weakened by an excitatory connection. Lack of
an edge represents an overall connection strength of −1 − δ. The following theorem
from [21] gives constraints on the graph and the relationship between δ and ε that
guarantees bounded activity, but disallows stable fixed points.
Figure 5.2: Diagram of excitatory and inhibitory connections: The left panel shows
inhibitory interneurons (gray circles) and excitatory pyramidal cells (colored triangles).
Arrows indicate connections between neurons. The right panel shows just the excitatory
neurons and their connections. Adapted from [8].
Theorem 5.4. [21] Let G be an oriented1 graph with no sinks (i.e. every vertex has
outdegree at least 1), and consider the associated CTLN model with W = W (G, ε, δ).
If ε <
δ
1 + δ
, then the network has bounded activity and no stable fixed points.
By forbidding stable fixed points, Theorem 5.4 guarantees that the activity of the
network is either oscillatory or chaotic. Computational experiments show that most of
the small networks for which Theorem 5.4 holds exhibit limit cycles where the neurons
1An oriented graph is a directed graph with no bi-directional connections.
AB 62
often appear to fire in sequence. The construction of the CTLN model guarantees
that any differences in dynamics arise solely from differences in the underlying graph.
Our goal is to use the structure of the graphs to predict the resulting sequences. In
the following sections we will assume θ = 1, ε = 0.25, δ = 0.5, and we will focus on
oriented graphs with no sinks. Theorem 5.4 holds in these cases. The next section
shows some examples of these networks and their behaviors.
5.3 Examples and behaviors
In this section, we will explore examples of CTLN networks with a small number of
nodes. We will start with the simplest example of a network satisfying the conditions
of Theorem 5.4, a three-cycle as seen in Figure 5.3. Since a directed edge represents a
less inhibited connection, the activity of such networks often follows the direction of
the arrows, though not always. Note that in this case the dynamics are a limit cycle
where the peak firing of the three nodes happens in the same order as the three-cycle
in the graph.
This three-cycle is the only graph on n = 3 nodes that meets the criteria of
Theorem 5.4, i.e. is an oriented graph with no sinks. On n = 4 nodes there are seven
such graphs and on n = 5 there are 152 such graphs. See Appendix C for the full
catalogue of oriented graphs with no sinks on n ≤ 5 nodes. The number of graphs
explodes when looking at oriented graphs with no sinks on n > 5 nodes.
Figure 5.4 shows several examples of networks and their dynamics on n = 5 nodes.
Panels A, B, and C show networks with limit cycles and the Panel D shows a chaotic
attractor. Note that some of the nodes have different peak firing rates. In many cases
on n = 5 nodes we see three nodes with a relatively high peak firing rate and the
remaining two nodes have a much smaller firing rate as seen in Figure 5.4 Panel A. We
63
Figure 5.3: Example on n = 3 nodes: In the top left we show the only oriented graph
on n = 3 vertices with no sinks. The top right panel shows the dynamics of this
network: a limit cycle where the nodes fire in the order 123. The bottom left shows the
transposed adjacency matrix (which is used for the CTLN model construction). The
bottom right shows the matrix of connection strengths constructed using Equation
5.2. Adapted from [21].
sometimes see synchronous firing of nodes, where the nodes fire at exactly the same
rate, often resulting from a graph automorphism, as in Panel C. Chaotic attractors
occur in networks as small as n = 5 nodes. See Figure 5.4 Panel D for an example.
Note that it is possible for a network to have multiple limit cycles or chaotic attractors.
For example, the network in Panel D has four different chaotic attractors, only one of
which is shown.
Could we have predicted these dynamics by looking at the graphs? The limit
cycles in panels A and B, each follow a cycle in the corresponding graph. This is
common in smaller graphs, but for larger n we have seen examples where this is not
the case. Also, what happens when there is more than one 5-cycle in the graph? Note
that the graph in Panel B has two 5-cycles, 12534 and 15423, but the network has
only one limit cycle. Why does the network preferentially choose one 5-cycle over
the other? Figure 5.5 shows an example on n = 7 nodes. Note that the limit cycle
Dfiring ratetime123E051015202530 64
shown in the activity trace corresponds to a 6-cycle in the graph. Why does node 2
stop firing? It receives input from three other nodes while nodes 3 and 5 only receive
two inputs each. Additionally, there are multiple 7-cycles in the graph. Why doesn't
the limit cycle correspond to one of these 7-cycles? These are all questions that our
algorithm must address. The algorithm must be able to discern between limit cycles
and chaos, predict the firing sequence for limit cycles, and must also identify which
nodes stop firing, if any.
65
Figure 5.4: Examples on n = 5 nodes: Panels A1-D1 show some examples of oriented
graphs on n = 5 nodes. Panels A2-D2 show the dynamics of the networks corresponding
to the graphs. Panel A2 shows a typical limit cycle. Note that nodes 4 and 5 fire at a
much lower rate than nodes 1, 2, and 3. Panel B1 shows an example of a balanced
subgraph. Note that each node has indegree 2 and outdegree 2. Panels C1-C2 show
an example with synchronous firing: nodes 1,4, and 5 fire at the same rate. This
is caused by the graph automorphism in C1. Panels D1-D2 show an example of a
network with a chaotic attractor. Note that we show a longer trial in D2 to show the
chaotic behavior.
timeactivity1234512345timeactivity12345020406080100120140160timefiring rateactivity01020304050607080timefiring rate01020304050607080timefiring rateactivity of nodes 1, 4, and 5 is synchronoustimeD1B1C101020304050607080timefiring rate12345timeactivityA1A2B2C2D2 66
Figure 5.5: Example on n = 7 nodes: The limit cycle has a sequence of 634517 as
indicated above the dynamics. Note that node 2 (red) stops firing. Adapted from [21].
1234567timefiring rate01020405060301673452total pop activityAB 67
Chapter 6
Sequence prediction algorithm
In this section, we will discuss an algorithm we designed to predict the neural sequence
from the graph for CTLN networks. The basic premise of this algorithm comes from
the idea of removing the "weakest" node and looking at the dynamics of the remaining
network. Once we know how the smaller network behaves we work to figure out a way
to tell where the deleted node fits in the sequence. We start with a description of the
algorithm, discuss some examples, and then state some conjectures about when the
algorithm is successful. We will look first at the case of tournaments1 without sinks
and then examine oriented graphs without sinks.
6.1 Description of the algorithm
The algorithm has two separate phases. In the deconstruction phase, we will first
deconstruct the graph by deleting one vertex at a time, keeping track of the order
of deletion. Then in the reconstruction phase, we start with a base sequence based
on our deconstruction and rebuild the neural sequence by adding back in the deleted
1A tournament is a complete simple directed graph, where complete means that there is an edge
between each pair of vertices.
68
Figure 6.1: Example of algorithm on n = 5
vertices in reverse order. Let G be a tournament on n vertices with no sinks.
Deconstruction phase. At each step of the algorithm we delete one of the vertices
of G with smallest indegree such that the resulting reduced tournament has no sinks.
We continue to delete vertices until we can no longer do so. This occurs when the
resulting tournament is a three-cycle (see Proposition 6.1). We refer to this three-cycle
as the core cycle. At each step we record the current tournament and the vertex we
deleted. See Figure 6.1 for an example. Note that the choice of vertex to delete is not
necessarily unique, so it is possible for the algorithm to output multiple sequences.
123452345245vertex to delete: 1vertex to delete: 3vertex to delete: noneSTEP 1:STEP 2:STEP 3:DeconstructionReconstructionSTEP 1:STEP 2:STEP 3:sequence: 254sequence: 2354sequence: 23154X 69
Reconstruction phase. To reconstruct the sequence, we start with the three-cycle
from the last step of the deconstruction phase. Recording the vertices in order of the
three-cycle we insert the other vertices into this sequence in reverse order of deletion.
Proceeding backwards through the list of deleted vertices, we add each vertex back
into the sequence following the vertex that feeds into it in the graph from the preceding
step in the deconstruction. If more than one vertex feeds into the vertex to be added,
we look at the subgraph induced by these possibilities and if one of these possibilities
is a sink in the induced subgraph, we place the vertex to be added after the sink in
the sequence. See Figure 6.1 for an example. Note that it is possible for the algorithm
to fail if there are two or more edges feeding into the node we are adding back in.
Node death. When reconstructing the sequence, there are rules to predict the
death of a node. If the vertex we delete at a given step has indegree zero then we do
not add that node back in during reconstruction. If the vertex we delete at a given
step has indegree one with the one edge coming from a vertex not in the core cycle
then we do not add that node back in during reconstruction.
Final sequence. To predict the final sequence, we first consider the full list of
possibilities. If there are two possibilities that are identical except one is missing a
node, then we choose the shorter sequence. If the possibilities are different but the
same length, some neurons might fire synchronously. We predict the synchronous
firing of a subset of neurons if that subset appears in the same cyclic order in each of
the possibilities but with a different starting point. The neurons not in the subset
appear in the same order in all possibilities. For example, we would predict that 2, 3,
and 4 fire synchronously if the algorithm output sequences 12345, 13425, and 14235.
70
Implementation of the algorithm in Matlab. We have developed a Matlab
code to automate the prediction algorithm. The code can be found in Appendix B.
Proposition 6.1. For a tournament G with no sinks, the deconstruction phase of the
algorithm will terminate if and only if the graph corresponding to the current step is a
three-cycle.
Proof. Let G be a tournament on n vertices with no sinks. Assume you reach a step
in the algorithm where there are currently m nodes remaining and deleting any vertex
results in an illegal graph. Then each node must have at least one incoming edge from
a vertex with outdegree exactly 1. This implies that every vertex has outdegree exactly
1. Thus the graph has a total outdegree of m, i.e. we have m total edges. The number
of edges can also be given by(cid:0)m
2
(cid:1) since it is a tournament, so m =(cid:0)m
(cid:1). Solving for
2
m gives m = 3. Since we have outdegree 1 at every node, the current graph is a
3-cycle. Also note that if the graph at the current step is a three-cycle, then deleting
any vertex, will result in a graph with two vertices and a directed edge between them.
Thus one of the vertices is a sink, so the deconstruction phase terminates at at step
n − 3.
6.2 Performance of the algorithm
Proposition 6.2. For ε = 0.5, δ = 0.25, and θ = 1, the algorithm correctly predicts
the neural sequence for tournaments without sinks on n ≤ 5 nodes.
The proof of this proposition is done by checking each case computationally. We
also note that for oriented graphs, the algorithm appears to predict which neurons
will have high firing rates. For small examples, we often see three neurons with higher
firing rates than the remaining neurons. These three high-firing neurons appear to
71
correspond to the neurons in the core cycle predicted by the algorithm. Further
exploration of tournaments lacking sinks on n > 5 nodes indicates that there are
networks where the algorithm does not correctly predict the behavior, often in the
form of spurious predictions or incorrectly predicting neuron death. Analysis of these
graphs indicates that the networks for which the algorithm fails appear to have the
common property of having a balanced subgraph on n ≥ 5 nodes or are an outerneuron
construction.
Definition 6.3. A balanced subgraph is a complete induced subgraph of an oriented
graph G where all nodes have the same outdegree. Note that for a balanced subgraph
of size m, where m is odd, the outdegree of each vertex is m−1
2 . There are no balanced
subgraphs of even size. See Panel B in Figure 5.4 for an example of a balanced graph
on 5 vertices. Each vertex has indegree 2 and outdegree 2.
Definition 6.4. The outerneuron construction is the process of taking a simple
directed graph on n vertices and adding two vertices to the graph: one vertex with
edges directed to all vertices in the original graph (a pseudo-source) and one vertex
who receives directed edges from all the vertices in the original graph (a pseudo-sink).
We then add a directed edge from the pseudo-sink to the pseudo-source to guarantee
that the new graph on n + 2 vertices has no sinks.
We have looked at all tournaments having no sinks on up to n = 7 vertices.
Proposition 6.2 gives that the algorithm works for the 11 such graphs on n ≤ 5
vertices. On n = 6 nodes there are 44 graphs and on n = 7 nodes there are 400
such graphs. The algorithm fails for only two of the n = 6 graphs, one with a
spurious prediction and one with a spurious deletion. Both of these graphs have a
balanced subgraph on 5 vertices. For n = 7, if we ignore graphs with an outerneuron
construction and graphs with balanced subgraphs with n ≥ 5 vertices (153 total), the
algorithm only fails for 5 out of the remaining 247 graphs. Additionally, of the 153
graphs having an outerneuron construction or a balanced subgraph with n ≥ 5, 62 of
these are still correctly predicted by the algorithm.
72
6.3 Extending the algorithm to oriented graphs
We have also explored the success of the algorithm on oriented graphs without sinks
on n ≤ 5 nodes. Recall that unlike tournaments, oriented graphs do not require
an edge between every pair of vertices. As a result we are not guaranteed that the
algorithm will terminate at a three-cycle for oriented graphs. In fact, we have seen
examples on n = 5 where the algorithm terminates in a 4- or 5-cycle. A comprehensive
list of oriented graphs without sinks on n ≤ 5 nodes appears in Appendix C. The
algorithm correctly predicts the behavior in all but 6 of the 160 total networks on
n ≤ 5 nodes. Using the algorithm on the Graph #147, 148, 149, 152, and 158 predicts
the correct sequence, but also produces a spurious prediction. For example, the
network corresponding to Graph #147 has a limit cycle with sequence 12(45)3 where
the parentheses indicate that neurons 4 and 5 fire synchronously. The algorithm makes
three predictions: 12453, 12543, and 12534. The first two predictions result in a correct
final sequence of 12(45)3, but 12534 is a spurious prediction. Using the algorithm on
Graph #153, a node is deleted that does not die, which keeps the algorithm from
predicting the synchronous activity.
One weakness that arises when using the algorithm on oriented graphs rather than
tournaments is that it is possible to disconnect the graph during deconstruction. Since
the reconstruction rules will not necessarily make sense in this case, additional rules
will be necessary for the deconstruction of oriented graphs to avoid breaking the graph
into multiple components. We also still sometimes make incorrect predictions in the
case of graphs with the outerneuron construction and/or a balanced subgraph. The
next section shows a comprehensive study of all the oriented graphs on n ≤ 5 vertices
in order to investigate ways to adjust the algorithm for oriented graphs.
73
6.4 An application: classification of oriented
graphs on n ≤ 5
To investigate why the algorithm fails in some cases, we perform an exhaustive study
and classification of oriented graphs on n ≤ 5 nodes. See Appendix C for a complete
list of graphs and their classification. If we look at the graphs for which the algorithm
failed, we see that they all fall into a category where there are two or more different
n = 4 subgraphs possible in the first step in the algorithm. Using this classification
we hope to start refining the algorithm to work in more generality.
In the next section we also sort the graphs by dynamics. Note that the graphs
in the same entry in the dictionary often also appear in the same category of the
classification in Appendix C.
6.5 Dictionary of attractors for n ≤ 5
In this chapter we create a dictionary of graph behaviors. We sort the graphs into
groups based on the dynamics of the corresponding network. Each entry in the
dictionary corresponds to a specific limit cycle or chaotic attractor. AT denotes
"attractor type." For each entry, we show a representative graph for that particular
attractor type along with its dynamics. To make the dynamics plots we use ε = 0.5,
δ = 0.25, and θ = 1 as before. Initial conditions used are listed above the dynamics
plots. We use an asterisk (*) to indicate which initial condition we used to create the
74
plot. The sequence listed is for the representative graph only. We underline low-firing
neurons and synchronous neurons are in parentheses. All graphs listed in that entry
exhibit the same behavior (up to permutation) as the representative graph for some
initial condition. If a graph has more than one attractor we annotate the graph
number with ic1,
ic2, etc. The labelling of the graphs corresponds to the numbering
in the catalogue found in Appendix C. The compilation of the dictionary was carried
out in collaboration with Katherine Morrison.
75
Basic Dictionary of Attractors AT-1 (limit cycle) Rep. graph 1, seq 123. All graphs: 1, 2, 3, 4, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38. AT-2 (limit cycle) Rep. graph 6, seq 1234. All graphs: 6, 7, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 117, 118, 119, 120, 121, 122, 123, 124, 125, 126, 127, 128, 129, 130. AT-3 (limit cycle) Rep. graph 140, seq 123(45). All graphs: 140, 141, 142, 147, 148. AT-4 (limit cycle) Rep. graph 82, seq 15234. All graphs: 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 136, 137, 138, 139, 149. 13210011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94753derivative of total activitytotal pop activity134210011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94367derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94422derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.9402derivative of total activitytotal pop activity 76
AT-5 (limit cycle) Rep. graph 5, seq 1234. All graphs: 5, 43, 44, 45, 46, 47. AT-6 (limit cycle) Rep. graph 48, seq 12345. All graphs: 48, 49, 50, 51. AT-7 (limit cycle) Rep. graph 152, seq 125345. All graphs: 152 AT-8 (limit cycle) Rep. graph 160, seq 12345. All graphs: 160. 134210011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94803derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94484derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.94159derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.9475derivative of total activitytotal pop activity 77
AT-9 (limit cycle) Rep. graph 159, seq 12534. All graphs: 159. AT-10 (limit cycle) Rep. graph 111, seq 15243. All graphs: 111_ic1, 111_ic2*, 131_ic1, 133_ic1, 133_ic2. AT-11 (limit cycle) Rep. graph 39, seq 1235. All graphs: 39, 40, 41, 42, 114_ic1, 114_ic2, 115_ic2, 131_ic2. AT-12 (limit cycle) Rep. graph 153, seq (145)23. All graphs: 153, 154. 1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.92592derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [0 1 1 0 1]−0.0500.050.10.80.850.90.951mean pop activity = 0.92007derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.91048derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.91455derivative of total activitytotal pop activity 78
AT-13 (limit cycle) Rep. graph 116, seq (15)234. All graphs: 116 AT-14 (limit cycle) Rep. graph 132, seq 12(35)4. All graphs: 132 AT-15 (limit cycle) Rep. graph 8, seq 12431243. All graphs: 8_ic1*, 8_ic2, 100_ic1, 100_ic2, 101_ic1, 101_ic2, 102_ic1, 102_ic2, 103_ic1, 103_ic2, 104_ic1, 104_ic2, 105_ic1, 105_ic2, 106_ic1, 106_ic2, 107_ic1, 107_ic2, 108_ic1, 108_ic2, 109_ic1, 109_ic2, 110_ic1, 110_ic2, 143_ic1, 144_ic1, 145_ic2, 146_ic2, 155_ic2, 157_ic2, 158_ic2. AT-16 (limit cycle) Rep. graph 112, seq 1254312543. All graphs: 112_ic1, 112_ic2*, 113_ic1, 113_ic2, 115_ic1, 134_ic1, 134_ic2, 135_ic1, 135_ic2, 143_ic2, 144_ic2, 145_ic1, 146_ic1, 155_ic1, 156_ic1, 156_ic2, 157_ic1, 158_ic1. 1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.9173derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 1 0 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.91563derivative of total activitytotal pop activity134210011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [1 0 1 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.92477derivative of total activitytotal pop activity1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [0 1 1 0 1]−0.0500.050.10.80.850.90.951mean pop activity = 0.92692derivative of total activitytotal pop activity 79
AT-17 (quasiperiodic) Rep. graph 151, seq 12312(45)312(45)312(45)3. All graphs: 151_ic1*, 151_ic2, 151_ic3. AT-18 (chaotic) Rep. graph 150. All graphs: 150_ic1, 150_ic2*, 150_ic3, 150_ic4. 1234510011012013014015016017018019020000.10.20.30.40.50.60.70.80.91timefiring rateX0 = [0 1 1 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.91932derivative of total activitytotal pop activity1234510012014016018020022024026028030000.20.40.60.81timefiring rateX0 = [0 1 1 0 0]−0.0500.050.10.80.850.90.951mean pop activity = 0.919derivative of total activitytotal pop activity 80
6.6 Future work
Balanced subgraphs and outerneuron constructions. For the subset of graphs
whose sequences cannot be predicted by our current algorithm, we need to adjust the
current algorithm or design a new algorithm to handle these cases.
Refining the algorithm for oriented graphs. Using the classification of oriented
graphs on n ≤ 5 nodes, we can look for commonalities in the graphs where the algorithm
fails and adjust the algorithm accordingly.
Exploring larger networks. After classifying the networks and behaviors for
oriented graphs on n ≤ 5 one of the next steps is to take a closer look at graphs with
n > 5.
Proving conjectures about algorithm. We have many conjectures about the
prediction algorithm, so we would like to look for proofs or counterexamples.
Conjecture 6.5. In an oriented graph with no sinks on n = 5 nodes, two nodes fire
synchronously if the algorithm predicts two sequences that are identical except for with
the two nodes switched and neither node is part of the core cycle.
Conjecture 6.6. During the reconstruction phase, if node u has indegree zero in the
corresponding subgraph, then this node dies.
Conjecture 6.7. During the reconstruction phase, if node u, with indegree 1 in the
corresponding subgraph, is added to the sequence following a node that was not part of
the base sequence then u dies, unless u is part of a balanced subgraph.
Conjecture 6.8. If the node we are placing back in the sequence has indegree > 1,
then consider the subgraph induced by the vertices which contribute to the indegree. If
the induced subgraph has a sink then place the vertex after the sink in the sequence.
81
82
Appendix A
Code for random walk model of
replenishment
This appendix gives the Matlab code for the random walk model of replenishment.
There are six functions total:
• Trials for plot
• R evolve
• create SiteMat
• R update
• N evolve
• N update
The main function Trials for plot takes ribbon type (1 for ribbon and 0 for nonribbon),
vesicle diameter (in µm), diffusion coefficient (in µm2/s), attachment probabilities for
both populations (as in Models 1 and 2), the fraction in population A, the vesicle
83
concentration (in vesicles/µm3), and the number of trials and outputs a cell array
where entry {i, j, 1} is a vector of the number of vesicles on the ribbon at each time
step in the computational model and entry {i, j, 2} is a vector of the number of vesicles
on the ribbon at each time step as predicted by the theoretical random walk model
for the ith concentration value and the jth trial. Plotting these two vectors against
time gives us the replenishment curves in both cases.
function [ Trials ] = T r i a l s _ f o r _ p l o t ( ribbon , delta , Diff , s_A ,
s_B ,f , concen , nu mt r ia l s )
% %%%%% initial state %%%%%%%%
W = 50; % width of matrix ( pick an even number )
H = 50; % height of matrix
D = 31; % depth of matrix
n = 110; % number of ribbon sites , must be a multiple of 10
f = 1; % fraction of fast - r e p l e n i s h i n g sites 0.757 for -10
mV and 0.54 or -30 mV
s = s_A * f + s_B *(1 - f ) ; % average a t t a c h m e n t p r o b a b i l i t y
dim_vec = [H ,W , D ];
t i m e v e c t o r = floor ((( harmonic ( n ) *2) /( delta ^3* s ) ) ./( concen
.* fracmob ) ) ; % c o r r e s p o n d i n g length of time for each
conc .
c on ce n ma x = length ( concen ) ; % number of c o n c e n t r a t i o n
values
Trials = cell ( concenmax , numtrials ,2) ; % collect data
% %%%%% start trials % % % % % % % % % % %
for co nc e ni d x = 1: co nc e nm ax
84
rho = concen ( c o nc en i dx ) ; % vesicle density per
m i c r o m e t e r ^3
den = rho * delta ^3; % density of occupied lattice sites
T = t i m e v e c t o r ( c on ce n id x ) ; % number of time steps
for trial = 1: n um tr i al s
% %%%%%% creates initial state matrix %%%%%%%
for i = 1: D
Mat (: ,: , i ) = rand (H , W ) >(1 - den ) ;
end
m = nnz ( Mat ) ; % number of vesicles
[r , c , l ] = ind2sub ( size ( Mat ) , find ( Mat == 1) ) ;
S _i ni t ia l = [r ,c , l ]; % matrix whose rows are the
c o o r d i n a t e s of the vesicles
% %%%%% chooses update function %%%%%%%
if ribbon == 1
e v o l v e _ f u n = @( S_ i ni t ia l ) R_evolve ( S_initial , dim_vec ,T
, s_A , s_B ,f ,n , shape ) ;
e v o l v e _ f u n = @( S_ i ni t ia l ) N_evolve ( S_initial , dim_vec ,T
, s_A , s_B ,f , n ) ;
else
end
tic
[ S_array , vec ] = e v o l v e _ f u n ( S_ i ni ti a l ) ; % update matrix T
times
toc
85
if ribbon ==1
t a u _ a p p r o x = @( s ) 1/( Diff * rho * delta * s ) ;
% approx time
constant for ribbon
else
t a u _ a p p r o x = @( s ) 1/(2* Diff * rho * delta * s ) ;
% approx time
constant for n on r ib bo n
end
end
x = 1: T -1;
x1 = [0 x ].*( delta ^2/(2* Diff ) ) ;
y = zeros (1 , T ) ;
z = zeros (1 , T ) ;
for i =1: T
y ( i ) = length ( vec { i }) ; % vector of vesicles on ribbon
at each time step
z ( i ) = ( n * f ) *(1 - exp ( - x1 ( i ) / t a u _ a p p r o x ( s_A ) ) ) +( n *(1 - f ) )
*(1 - exp ( - x1 ( i ) / t a u _ a p p r o x ( s_B ) ) ) ; % vector of
t h e o r e t i c a l l y p r ed ic t ed number of vesicles at each
time step
end
Trials { concenidx , trial ,1} = y ;
Trials { concenidx , trial ,2} = z ;
figure ; % plot t h e o r e t i c a l p r e d i c t i o n vs . c o m p u t a t i o n a l
trial
plot ( x1 ,y , ' -k ')
86
hold on ;
plot ( x1 ,z , ' -b ')
hold off ;
end
end
end
function [ S_array , vec ] = R_evolve (S , dim_vec , Tsteps , s_A , s_B
,f , n )
% evolve function for ribbon case
SiteMat = c r e a t e _ S i t e M a t ( dim_vec , n /10) ; % creates matrix
of ribbon sites
W = dim_vec (1) ;
D = dim_vec (3) ;
perm = randperm ( n ) ;
% p e r m u t a t i o n of num of vesicles to
randomly de t er m in e p o p u l a t i o n s A and B
frac = floor ( f * n ) ;
% number in p o p u l a t i o n A
indices = [];
for k = 1: length ( S (: ,1) )
if S (k ,1) >( W /2) -1 && S (k ,1) <( W /2) +1 && S (k ,2) <10 && S (
k ,3) >(( D +1) /2) -3 && S (k ,3) <(( D +1) /2) +3
indices = [ indices , k ];
end
end
n e w i n d i c e s = setdiff ([1: size (S ,1) ] , indices ) ;
87
S = S ( newindices ,:) ;
R ib bo n Ma t = ones ( size ( S ) ) ; % matrix of ones with rows
c o r r e s p o n d i n g to vesicles on ribbon zeroed out
F il le d Ma t = [];
[ rowdim , coldim ] = size ( S ) ;
S_array = zeros ( rowdim , coldim , Tsteps ) ;
S_array (: ,: ,1) = S ;
vec = cell (1 , Tsteps ) ;
vec {1} = [];
for t = 2: Tsteps
index = find ( ismember (S , SiteMat , ' rows ') ) ;
ves_idx = [ vec {t -1}];
index = setdiff ( index , ves_idx ) ;
for j = 1: length ( index )
v e s i c l e _ i d x = find ( Ri b bo nM a t (: ,1) ==0) ;
if isempty ( find ( ismember ( S ( vesicle_idx ,:) ,S ( index (
j ) ,:) ,' rows ') ,1) ) == 1
if isempty ( find ( ismember ( SiteMat ( perm (1: frac )
,:) ,S ( index ( j ) ,:) ,' rows ') ) ) == 0
if rand (1) <= s_A
R ib bo n Ma t ( index ( j ) ,:) = 0;
F il le d Ma t = [ F il le d Ma t ; S ( index ( j ) ,:) ];
end
else
if rand (1) <= s_B
88
R ib bo n Ma t ( index ( j ) ,:) = 0;
F il le d Ma t = [ F il le d Ma t ; S ( index ( j ) ,:) ];
end
end
end
end
vec { t } = find ( R i bb on M at (: ,1) == 0) ;
S = R_update (S , RibbonMat , dim_vec , FilledMat , shape , n ) ;
S_array (: ,: , t ) = S ;
end
end
function SiteMat = c r e a t e _ S i t e M a t ( dim_vec , ht )
% creates matrix of the indices of the ribbon
W = dim_vec (1) ;
D = dim_vec (3) ;
SiteMat = [];
for ii = [ W /2 -1 , W /2+1]
for jj = 1: ht ;
for kk = ( D +1) /2 -2:( D +1) /2+2;
SiteMat = [ SiteMat ;[ ii , jj , kk ]];
end
end
end
function [ S_new ] = R_update ( currentS , RibbonMat , dim_vec ,
89
FilledMat , n )
% update function for the ribbon case
[ rowdim , coldim ]= size ( currentS ) ;
changeS = randi ([0 ,1] , size ( currentS ) ) ;
changeS =2* changeS -1;
D = dim_vec (3) ;
I l l e g a l _ M a t = F il l ed Ma t ;
for ii = 1: n /10
for jj = ( D +1) /2 -2:( D +1) /2+2
I l l e g a l _ M a t = [ I l l e g a l _ M a t ;[ dim_vec (1) /2 , ii , jj
]];
end
end
t e s t _ c h a n g e S = changeS .* Ri bb o nM at ;
test_S = currentS + changeS ;
for ii = 1: length ( dim_vec )
test_S (: , ii ) = min ( max ( test_S (: , ii ) ,1) , dim_vec ( ii ) ) ;
end
c h a n g e _ e n t r i e s = find ( ismember ( test_S , Illegal_Mat , ' rows ') )
;
changeS ( change_entries ,:) = 0;
changeS = changeS .* R ib bo n Ma t ;
preS = currentS + changeS ;
for jj = 1: length ( dim_vec )
S_new (: , jj ) = min ( max ( preS (: , jj ) ,1) , dim_vec ( jj ) ) ;
90
end
end
function [ S_array , vec ] = N_evolve (S , dim_vec , Tsteps , s_A , s_B
,f , n )
% evolve function for the no n ri bb o n case
SiteMat = [[4 ,7 ,6];[4 ,16 ,6];[4 ,25 ,6];[4 ,34 ,6];[4 ,43 ,6];
[15 ,7 ,6];[15 ,16 ,6];[15 ,25 ,6];[15 ,34 ,6];[15 ,43 ,6];
[26 ,7 ,6];[26 ,16 ,6];[26 ,25 ,6];[26 ,34 ,6];[26 ,43 ,6];
[37 ,7 ,6];[37 ,16 ,6];[37 ,25 ,6];[37 ,34 ,6];[37 ,43 ,6];
[4 ,7 ,10];[4 ,16 ,10];[4 ,25 ,10];[4 ,34 ,10];[4 ,43 ,10];
[15 ,7 ,10];[15 ,16 ,10];[15 ,25 ,10];[15 ,34 ,10];[15 ,43 ,10];
[26 ,7 ,10];[26 ,16 ,10];[26 ,25 ,10];[26 ,34 ,10];[26 ,43 ,10];
[37 ,7 ,10];[37 ,16 ,10];[37 ,25 ,10];[37 ,34 ,10];[37 ,43 ,10];
[4 ,7 ,14];[4 ,16 ,14];[4 ,25 ,14];[4 ,34 ,14];[4 ,43 ,14];
[15 ,7 ,14];[15 ,16 ,14];[15 ,25 ,14];[15 ,34 ,14];[15 ,43 ,14];
[26 ,7 ,14];[26 ,16 ,14];[26 ,25 ,14];[26 ,34 ,14];[26 ,43 ,14];
[37 ,7 ,14];[37 ,16 ,14];[37 ,25 ,14];[37 ,34 ,14];[37 ,43 ,14];
[4 ,7 ,18];[4 ,16 ,18];[4 ,25 ,18];[4 ,34 ,18];[4 ,43 ,18];
[15 ,7 ,18];[15 ,16 ,18];[15 ,25 ,18];[15 ,34 ,18];[15 ,43 ,18];
[26 ,7 ,18];[26 ,16 ,18];[26 ,25 ,18];[26 ,34 ,18];[26 ,43 ,18];
[37 ,7 ,18];[37 ,16 ,18];[37 ,25 ,18];[37 ,34 ,18];[37 ,43 ,18];
[15 ,7 ,22];[15 ,16 ,22];[15 ,25 ,22];[15 ,34 ,22];[15 ,43 ,22];
[26 ,7 ,22];[26 ,16 ,22];[26 ,25 ,22];[26 ,34 ,22];[26 ,43 ,22];
91
[37 ,7 ,22];[37 ,16 ,22];[37 ,25 ,22];[37 ,34 ,22];[37 ,43 ,22];
[15 ,7 ,26];[15 ,16 ,26];[15 ,25 ,26];[15 ,34 ,26];[15 ,43 ,26];
[26 ,7 ,26];[26 ,16 ,26];[26 ,25 ,26];[26 ,34 ,26];[26 ,43 ,26];
[37 ,7 ,26];[37 ,16 ,26];[37 ,25 ,26];[37 ,34 ,26];[37 ,43 ,26]];
W = dim_vec (1) ;
D = dim_vec (3) ;
perm = randperm ( n ) ;
% p e r m u t a t i o n of num of vesicles to
randomly de t er m in e p o p u l a t i o n s A and B
frac = floor ( f * n ) ;
% number in p o p u l a t i o n A
indices = [];
for k = 1: length ( S (: ,1) )
if S (k ,1) >( W /2) -1 && S (k ,1) <( W /2) +1 && S (k ,2) <10 && S (
k ,3) >(( D +1) /2) -3 && S (k ,3) <(( D +1) /2) +3
indices = [ indices , k ];
end
end
n e w i n d i c e s = setdiff ([1: size (S ,1) ] , indices ) ;
S = S ( newindices ,:) ;
R ib bo n Ma t = ones ( size ( S ) ) ;
F il le d Ma t = [];
[ rowdim , coldim ] = size ( S ) ;
S_array = zeros ( rowdim , coldim , Tsteps ) ;
S_array (: ,: ,1) = S ;
vec = cell (1 , Tsteps ) ;
vec {1} = [];
92
for t = 2: Tsteps
index = find ( ismember (S , SiteMat , ' rows ') ) ;
ves_idx = [ vec {t -1}];
index = setdiff ( index , ves_idx ) ;
for j = 1: length ( index )
v e s i c l e _ i d x = find ( Ri b bo nM a t (: ,1) ==0) ;
if isempty ( find ( ismember ( S ( vesicle_idx ,:) ,S ( index (
j ) ,:) ,' rows ') ,1) ) == 1
if isempty ( find ( ismember ( SiteMat ( perm (1: frac )
,:) ,S ( index ( j ) ,:) ,' rows ') ) ) == 0
if rand (1) <= s_A
R ib bo n Ma t ( index ( j ) ,:) = 0;
F il le d Ma t = [ F il le d Ma t ; S ( index ( j ) ,:) ];
end
else
if rand (1) <= s_B
R ib bo n Ma t ( index ( j ) ,:) = 0;
F il le d Ma t = [ F il le d Ma t ; S ( index ( j ) ,:) ];
end
end
end
end
vec { t } = find ( R i bb on M at (: ,1) == 0) ;
S = N_update (S , RibbonMat , dim_vec ) ;
S_array (: ,: , t ) = S ;
end
end
93
function [ S_new ] = N_update ( currentS , RibbonMat , dim_vec )
% update function for the n on r ib bo n case
[ rowdim , coldim ]= size ( currentS ) ;
changeS = randi ([0 ,1] , size ( currentS ) ) ;
changeS =2* changeS -1;
changeS = changeS .* R ib bo n Ma t ;
preS = currentS + changeS ;
S_new = zeros ( length ( currentS ) ,3) ;
for i =1: length ( dim_vec )
S_new (: , i ) = min ( max ( preS (: , i ) ,1) , dim_vec ( i ) ) ;
end
end
94
Appendix B
Code for sequence prediction
algorithm
In this appendix, we present the Matlab code for the implementation of the neural
sequence algorithm. There are three functions total:
• run Algorithm
• DeconstructGraph
• ReconstructCycle
The main function, run Algorithm takes the transposed adjacency matrix of a graph
as an input and outputs the list of predicted limit cycles for the corresponding network.
This function calls two subfunctions for the two phases of the algorithm: Deconstruct-
Graph and ReconstructCycle, also included here. This particular implementation of
the algorithm was done in collaboration with Katherine Morrison.
function [ E x p e c t e d L i m i t C y l e s ] = r u n _ A l g o r i t h m ( sA )
95
% This function takes a t r a n s p o s e d a d ja c en cy matrix ( sA
matrix ) as an input and outputs a list of expected
limit cycles .
D e l e t e d N o d e s =[ ];
% This should always be i n i t i a l i z e d as empty
D e l e t e d N o d e s L i s t =[ ];
% This should always be i n i t i a l i z e d as empty
C o r e C y c l e s L i s t =[ ];
% This should always be i n i t i a l i z e d as empty
[ DeletedNodesList , C o r e C y c l e s L i s t ] = D e c o n s t r u c t G r a p h ( sA ,
DeletedNodes , DeletedNodesList , C o r e C y c l e s L i s t ) ;
n = size ( sA ,2) ;
E x p e c t e d L i m i t C y c l e s = zeros ( size ( CoreCyclesList ,1) ,2* n ) ;
% We will r e c o n s t r u c t a full cycle for every core cycle
and we will have every unique core cycle listed as well
since we expect unstable fixed points at those core
cycles .
We will allow the E x p e c t e d L i m i t C y c l e s to have
length up to 2* n because of the po te n ti al for period
doubling -- h op e fu ll y we won ' t have any cycles longer
than this , but if we do , we ' ll get an error here
U n s t a b l e F i x e d P o i n t s = unique ( CoreCyclesList , ' rows ') ;
% Insert all unique core cycles at the end of our
E x p e c t e d L i m i t C y c l e s list
for i =1: size ( CoreCyclesList ,1)
D e l e t e d N o d e s = D e l e t e d N o d e s L i s t (i ,:) ;
96
C or eC y cl e = C o r e C y c l e s L i s t (i ,:) ;
F ul lC y cl e = R e c o n s t r u c t C y c l e ( sA , DeletedNodes , C or eC y cl e ) ;
E x p e c t e d L i m i t C y c l e s (i , 1: length ( F u ll C yc le ) ) = F u ll Cy c le ;
% Insert each r e c o n s t r u c t e d full cycle into a long row
padded with zeros
end
E x p e c t e d L i m i t C y c l e s
end
function [ DeletedNodesList , C o r e C y c l e s L i s t ] =
D e c o n s t r u c t G r a p h ( sA , DeletedNodes , DeletedNodesList ,
C o r e C y c l e s L i s t )
% sA is the full ad ja c en cy matrix
% D e l e t e d N o d e s is a row vector of the nodes deleted thus
far along a single path to to one core cycle
% D e l e t e d N o d e s L i s t is a matrix that will have rows of
length n -3 added to it once a full row vector of
D e l e t e d N o d e s has been c o mp le t ed ( this row vector will
be padded with zeros if the size of the core cycle is
greater than 3
% C o r e C y c l e s L i s t is a matrix that will have rows of length
n added to it as each core cycle is found ( the core
cycle will be padded with zeros to make it length n )
% This function r e c u r s i v e l y calls itself ( in the manner of
a depth - first search ) until there are no a d d i t i o n a l
97
nodes that can be deleted and yield a valid graph .
At
this point , the vector of deleted nodes is padded with
zeros and added as a row vector to D e l e t e d N o d e s L i s t .
The c o r r e s p o n d i n g core cycle is computed , padded with
zeros , and added as a row vector to C o r e C y c l e s L i s t
if nargin <2
D e l e t e d N o d e s =[ ];
end
if nargin <3
D e l e t e d N o d e s L i s t =[ ];
end
if nargin <4
C o r e C y c l e s L i s t =[ ];
end
N o d e s T o D e l e t e = F i n d D e l e t e N o d e s ( sA , D e l e t e d N o d e s ) ;
for i =1: length ( N o d e s T o D e l e t e )
[ DeletedNodesList , C o r e C y c l e s L i s t ] = D e c o n s t r u c t G r a p h (
sA ,[ DeletedNodes , N o d e s T o D e l e t e ( i ) ] ,
DeletedNodesList , C o r e C y c l e s L i s t ) ;
% Keep r e c u r s i v e l y calling the function , adding the
next node to delete to the end of the vector
D e l e t e d N o d e s ( in the function call but not outside
it so that we can loop through and call the
function multiple times with di f fe re n t nodes to
98
delete added on the end )
end
n = size ( sA ,2) ;
if isempty ( N o d e s T o D e l e t e )
% In this case , we ' ve gotten down to a core cycle
after an empty loop
D el et e Ro w = zeros (1 , n -3) ;
D el et e Ro w (1: length ( D e l e t e d N o d e s ) ) = D e l e t e d N o d e s ;
% Fill in the first entries of D e le te R ow so that the
r em ai n in g entries are all the padded zeros
D e l e t e d N o d e s L i s t =[ D e l e t e d N o d e s L i s t ; D e le te R ow ];
CycleRow = zeros (1 , n ) ;
[ sAsubmat , labels ]= M a k e S u b m a t ( sA , D e l e t e d N o d e s ) ;
C o r e C y c l e I n d i c e s (1) =1;
% Always start the cycle at an index of 1 ( which will
have a label c o r r e s p o n d i n g to the lowest re ma i ni ng
node )
for j =1: length ( labels ) -1
N ex tI n de x = find ( sAsubmat (: , C o r e C y c l e I n d i c e s ( j ) ) ==1)
;
% This finds the unique node in the core cycle
that C o r e C y c l e s I n d i c e s ( j ) feeds into
C o r e C y c l e I n d i c e s ( j +1) = Ne x tI nd e x ;
end
CycleRow (1: length ( labels ) ) = labels ( C o r e C y c l e I n d i c e s ) ;
99
% Fill in the first entries of CycleRow so that the
r em ai n in g entries are all the padded zeros
C o r e C y c l e s L i s t =[ C o r e C y c l e s L i s t ; CycleRow ];
end
function Fu l lC y cl e = R e c o n s t r u c t C y c l e ( sA , DeletedNodes ,
C or eC y cl e )
% sA is the full ad ja c en cy matrix
% D e l e t e d N o d e s is a row vector of all the nodes to
delete
in the graph to get down to the core cycle
% Co re C yc le is a row vector of the nodes in the core cycle
in order a c co rd i ng to which node feeds into which
% This function r e c o n s t r u c t s the full cycle by
a p p r o p r i a t e l y r e i n s e r t i n g the deleted nodes in reverse
order of when they were deleted .
D e l e t e d N o d e s = D e l e t e d N o d e s ( D e l e t e d N o d e s ~=0) ;
% This removes any padded zeros from the end
C or eC y cl e = C or e Cy cl e ( C or e Cy cl e ~=0) ;
% This removes any padded zeros from the end
cycle = C o re Cy c le ;
for i = length ( D e l e t e d N o d e s ) : -1:1
% run through the D e l e t e d N o d e s in reverse order
N o d e T o I n s e r t = D e l e t e d N o d e s ( i ) ;
[ sAsubmat , labels ]= M a k e S u b m a t ( sA , D e l e t e d N o d e s (1: i -1) )
;
100
% This r e c o n s t r u c t s the subgraph when the nodes before
the current node have been deleted -- when we ' re
on the last node , i . e . i =1 , D e l e t e d N o d e s (1: i -1)
will be empty and we ' ll just get back the original
sA and labels =1: n
I d x T o I n s e r t = find ( labels == N o d e T o I n s e r t ) ;
% This finds the index in the s u bm a tr ix c o r r e s p o n d i n g
to the node to be inserted
I d x I n t o N o d e = find ( sAsubmat ( IdxToInsert , :) ==1) ;
% This finds all the indices of nodes that feed into
the node to be inserted ( i . e . the l o ca ti o ns of 1 s
in the row c o r r e s p o n d i n g to the node to be inserted
)
if length ( I d x I n t o N o d e ) >=3
N o d e s I n S u b M a t = sAsubmat ( IdxIntoNode , I d x I n t o N o d e ) ;
SinkNode = find ( sum ( NodesInSubMat ,1) ==0) ;
if length ( SinkNode ) == 1
I d x I n t o N o d e = SinkNode ;
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e ) ) ;
else
disp ([ ' Error -- vertex ' num2str ( N o d e T o I n s e r t ) '
has 3 or more inputs in the subgraph for the
delete sequence ' mat2str ( D e l e t e d N o d e s ) '.
Check this graph by hand to update al g or it h m ' ])
101
;
return
% added
end
% At this point , we kill the function via ' return '
because the al go r it h m doesn ' t know how to
handle this s i tu a ti on .
This should be updated
once we ' ve seen graphs that have this feature
and de t er mi n e a he u ri st i c for how to handle
this s i tu at i on
elseif length ( I d x I n t o N o d e ) ==2 && sAsubmat ( I d x I n t o N o d e
(1) , I d x I n t o N o d e (2) ) ==0 && sAsubmat ( I d x I n t o N o d e (2) ,
I d x I n t o N o d e (1) ) ==0
% In this case there are 2 nodes that feed into
the N o d e T o I n s e r t and they are not adjacent to
each other , so the al g or i th m says that the new
node should be inserted into the sequence after
both incoming vertices
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e (1) ) ) ;
% This g u a r a n t e e s that we are only placing nodes
back into the sequence if they directly follow
a node from the core cycle
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e (2) ) ) ;
elseif length ( I d x I n t o N o d e ) ==2
102
% One of the nodes to insert feeds into the other .
The one that is fed into should have the new
node inserted after it
if sAsubmat ( I d x I n t o N o d e (1) , I d x I n t o N o d e (2) ) ==1
% Then node 2 feeds into node 1 , so the new
node should be inserted after node 1
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e (1) ) ) ;
else
end
% Then node 1 feeds into node 2 , so the new
node should be inserted after node 2
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e (2) ) ) ;
elseif length ( I d x I n t o N o d e ) ==1
if isempty ( in te r se c t ( CoreCycle , labels ( I d x I n t o N o d e )
) ) == 0
cycle = I n s e r t N o d e ( cycle , NodeToInsert , labels (
I d x I n t o N o d e ) ) ;
end
end
end
F ul lC y cl e = cycle ;
103
Appendix C
Catalogue of n ≤ 5 oriented graphs
with no sinks
This appendix includes a list of all the oriented graphs on n ≤ 5 vertices with no sinks.
The graphs on n = 5 vertices are sorted by which n = 4 subgraph(s) they reduce to in
the first step of the algorithm from Chapter 6. We group the n = 4 graphs into four
classes: #2, 3, and 4 reduce to a three-cycle, #5 is a 4-cycle, #6 and 7 have three
strong neurons and one weak, and #8 has two distinct limit cycles. For graphs whose
networks exhibit more than one behavior, like #8, we list the initial conditions for
each attractor type. Table C.1 gives an index for the classification indicating which
category each of the oriented graphs on n = 5 nodes without sinks fall into.
Reduces to
Graph indices Attractor type
104
2/3/4
5
6/7
8
2 or 2
2 or 5
5 or 5
9 -- 38
39 -- 42
43 -- 47
48 -- 51
52 -- 81
82 -- 99
100 -- 110
111
112 -- 113
114
115
116
2/3/4 or 6/7
117 -- 130
5 or 6
6/7 or 6/7
2/3/4 or 8
6/7 or 8
8 or 8
3, 3, or 3
4, 6, or 7
7, 7, or 7
2, 6, or 8
6, 6, or 8
3, 7, or 8
2, 8, or 8
7, 7, 7, 7, or 7
none
131
132
133
134 -- 135
136 -- 139
140 -- 142
143 -- 146
147 -- 148
149
150
151
152
153
154
155
156
157
158
159
160
AT-1
AT-11
AT-5
AT-6
AT-2
AT-4
AT-15
AT-10
AT-16
AT-11
AT-11, AT-16
AT-13
AT-2
AT-10, AT-11
AT-14
AT-10
AT-16
AT-4
AT-3
AT-15, AT-16
AT-3
AT-4
AT-18
AT-17
AT-7
AT-12
AT-12
AT-15, AT-16
AT-16
AT-15, AT-16
AT-15, AT-16
AT-9
AT-8
Table C.1: Index for classification of oriented graphs with no sinks on n = 5 nodes:
The left column gives the indices of the possible n = 4 subgraph(s) that appear after
step one of the algorithm, the middle column shows the indices of the graphs that
have this reduction (based on the classification indexing), and the right column shows
which attractors these graphs have (based on the dictionary). Subgraphs 2, 3, and 4
are grouped together since they have the same behavior and subgraphs 6 and 7 are
similarly grouped.
105
Canonical Labelling of Oriented Graphs on n≤5 nodes143251321342134213421342687134213421342ic1 = [1 0 1 0]ic2 = [1 0 0 1] 106
12345123451234512345123459101234512345123451234512345123451234512345123451234512345131823123451234512345123451234512345123451234512345123451234512345123451234512112833381415161719202122242526272930313234353637First step in algorithm reduces to #2,3,4 107
1234512345123451234539404142First step in algorithm reduces to #512345123451234512345123454346454447485150491234512345123451234512345525554535612345123451234512345First step in algorithm reduces to #6,7 108
123451234512345123451234557605958616265646366123451234512345123451234567706968711234512345123451234512345727574737612345123451234512345123457780797881123451234512345123451234582858483861234512345123451234512345 109
12345123451234587908988911234512345123451234592959493961234512345123451234512345979998123451234512345First step in algorithm reduces to #81001031021011041234512345123451051081071061091234512345123451234512345ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1] 110
123451234512345110113112111123451234512345114116115First step in algorithm reduces to #2 or #2First step in algorithm reduces to #2 or #5First step in algorithm reduces to #5 or #5First step in algorithm reduces to #2/3/4 or #6/71171201191181211234512345123451234512345123451221251241231261234512345123451234512345ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 1 0 0 0]ic2 = [0 0 1 0 1]ic1 = [1 1 0 0 0]ic2 = [0 0 1 0 1] 111
12345123451234512713012912813112345123451234512345First step in algorithm reduces to #5 or #6132First step in algorithm reduces to #6/7 or #6/7133136135134137123451234512345123451381411401391421234512345123451234512345ic1 = [1 1 0 0 0]ic2 = [0 0 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [1 0 1 0 1] 112
1234512345143146145144First step in algorithm reduces to #2/3/4 or #812345123451234512345147149148First step in algorithm reduces to #6/7 or #81234512345150151First step in algorithm reduces to #8 or #812345ic1 = [1 0 0 1 1]ic2 = [1 1 0 0.1 0]ic1 = [1 0 0 1 1]ic2 = [1 1 0 0.1 0]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [1 0 0 1 1]ic2 = [0 1 1 0 1]ic1 = [0 0 1 0 1]ic2 = [0 1 1 0 0]ic3 = [0 1 0 1 0]ic4 = [0 0 0 1 1]ic1 = [0 1 1 0 0]ic2 = [0 1 0 1 0]ic3 = [0 1 0 0 1] 113
First step in algorithm reduces to #3 or #3 or #5152153154123451234512345First step in algorithm reduces to #4 or #6 or #7First step in algorithm reduces to #7 or #7 or #7First step in algorithm reduces to #2 or #6 or #8155156157First step in algorithm reduces to #6 or #6 or #8First step in algorithm reduces to #3 or #7 or #8123451234512345First step in algorithm reduces to #2 or #8 or #8158159160First step in algorithm reduces to #7or #7 or #7 or #7 or #7No reduction123451234512345ic1 = [1 0 0 1 1]ic2 = [1 0 1 0 1]ic1 = [1 0 0 1 1]ic2 = [1 0 1 0 1]ic1 = [1 0 0 1 1]ic2 = [1 0 1 0 1] 114
Bibliography
[1] Norbert Babai, Theordore M. Bartoletti, and Wallace B. Thoreson. Calcium
regulates vesicle replenishment at the cone ribbon synapse. The Journal of
Neuroscience, 30(47):15866 -- 15877, 2010. 1.3, 1.4, 3.2
[2] Theodore M. Bartoletti, Norbert Babai, and Wallace B. Thoreson. Vesicle
pool size at the salamander cone ribbon synapse. Journal of Neurophysiology,
103(1):419 -- 423, 2010. 1.3, 2.1, 3.1.3
[3] Theodore M. Bartoletti, Skyler L. Jackman, Norbert Babai, Aaron J. Mercer,
Richard H. Kramer, and Wallace B. Thoreson. Release from the cone ribbon
synapse under bright light conditions can be controlled by the opening of only a
few Ca2+ channels. Journal of Neurophysiology, 106:2922 -- 2935, 2011. 1.3
[4] Theodore M. Bartoletti and Wallace B. Thoreson. Quantal amplitude at the cone
ribbon synapse can be adjusted by changes in cytosolic glutamate. Molecular
Vision, 17:920 -- 931, 2011. 2
[5] Mark F. Bear, Barry W. Connors, and Michael A. Paradiso. Neuroscience:
Exploring the Brain. Lippincott Williams and Wilkins, 3rd edition, 2007. 1, 1
[6] Howard Berg. Random Walks in Biology. Princeton University Press, revised
edition, 1993. 3.1.2, 3.3.1
115
[7] Lucia Cadetti, Daniel Tranchina, and Wallace B. Thoreson. A comparison of
release kinetics and glutamate receptor properties in shaping rod-cone differences
in EPSC kinetics in the salamander retina. The Journal of Physiology, 569(3):773 --
788, 2005. 2, 2.2.1, 3
[8] Carina Curto and Katherine Morrison. Pattern completion in symmetric thresold-
linear networks. arXiv:1512.00897 (To appear in Neural Computation), 2016. 5.2,
5.2, 5.2
[9] Elizabeth Garcia-Perez and John F. Wesseling. Augmentation controls the
fast rebound from depression at excitatory hippocampal synapses. Journal of
Neurophysiology, 99:1770 -- 1786, 2008. 3.1.3
[10] Cole W. Graydon, Jun Zhang, Alioscka A. Sousa, Richard D. Leapman, and
Jeffrey S. Diamond. Passive diffusion as a mechanism underlying ribbon synapse
vesicle release and resupply. The Journal of Neuroscience, 34(27):8948 -- 8962,
2014. 1.3, 3.1.3
[11] Richard H.R. Hahnloser, H. Sebastian Seung, and Jean-Jacques Slotine. Permitted
and forbidden sets in symmetric threshold-linear networks. Neural Computation,
15:621 -- 638, 2003. 5.1
[12] Kristen M. Harris and John K. Stevens. Dendritic spines of CA1 pyramidal
cells in the rat hippocampus: Serial electron microscopy with reference to their
biophysical characteristics. The Journal of Neuroscience, 9(8):2982 -- 2997, 1989.
3.1.3
[13] Kristen M. Harris and Peter Sultan. Variation in the number, location and size
of synaptic vesicles provides an anatomical basis for the nonuniform probability
116
of release at hippocampal CA1 synapses. Neuropharmacology, 34(11):1387 -- 1395,
1995. 3.1.3
[14] Matthew Holt, Anne Cooke, Andreas Neef, and Leo Lagnado. High mobility of
vesicles supports continuous exocytosis at a ribbon synapse. Current Biology,
14:173 -- 183, 2004. 1.4, 3.1.3, 3.2
[15] Taro Ishikawa, Yoshinori Sahara, and Tomoyuki Takahashi. A single packet
of transmitter does not saturate postsynaptic glutamate receptors. Neuron,
34:613 -- 621, 2002. 2
[16] Skyler L. Jackman, Sue-Yeon Choi, Wallace B. Thoreson, Katalin Rabl, and
Richard H. Kramer. Role of the synaptic ribbon in transmitting the cone light
response. Nature Neuroscience, 12(3):303 -- 310, 2009. 1.3, 1.4, 3
[17] Stuart L. Johnson, Andrew Forge, Marlies Knipper, Stefan Munker, and Walter
Marcotti. Tonotopic variation in the calcium dependence of neurotransmitter
release and vesicle pool replenishment at mammalian auditory ribbon synapses.
The Journal of Neuroscience, 28(30):7670 -- 7678, 2008. 3
[18] Shanker Karunanithi, Leo Marin, Kar Wong, and Harold L. Atwood. Quantal
size and variation determined by vesicle size in normal and mutant drosophila
glutamatergic synapses. The Journal of Neuroscience, 22(23):10267 -- 10276, 2002.
2
[19] Gary Matthews and Paul Fuchs. The diverse roles of ribbon synapses in sensory
neurotransmission. Nature, 11(12):812 -- 822, 2010. 1.2
117
[20] Steven Mennerick and Gary Matthews. Ultrafast exocytosis elicited by calcium
current in synaptic terminals of retinal bipolar neurons. Neuron, 17:1241 -- 1249,
1996. 3.1.3
[21] Katherine Morrison, Anda Degeratu, Vladimir Itskov, and Carina Curto. Diversity
of emergent dynamics in competitive threshold-linear networks: a preliminary
report. arXiv:1605.04463, 2016. 5, 5.1, 5.1, 5.2, 5.3, 5.2, 5.4, 5.3, 5.5
[22] Christophe Paillart, Jian Li, Gary Matthews, and Peter Sterling. Endocytosis and
vesicle recycling at a ribbon synapse. The Journal of Neuroscience, 23(10):4092 --
4099, 2003. 3.1.3
[23] Thomas D. Parsons and Peter Sterling. Synaptic ribbon: Conveyor belt or safety
belt? Neuron, 37:379 -- 382, 2003. 1.3
[24] Ruth Rea, Jian Li, Ajay Dharia, Edwin S. Levitan, Peter Sterling, and Richard H.
Kramer. Streamlined synaptic vesicle cycle in cone photoreceptor terminals.
Neuron, 41:755 -- 766, 2004. 1.4, 3.1, 3.1.3, 3.1.3, 3.2
[25] Takeshi Sakaba, Ralf Schneggenburger, and Erwin Neher. Estimation of quantal
parameters at the calyx of Held synapse. Neuroscience Research, 44:343 -- 356,
2002. 2.2.2
[26] Thomas Schikorski and Charles F. Stevens. Quantitative ultrastructural analysis
of hippocampal excitatory synapses. The Journal of Neuroscience, 17(15):5858 --
5867, 1997. 3.1.3
[27] H. Sebastian Seung and Rafael Yuste. Principles of Neural Science, chapter
Appendix E: Neural Networks, pages 1581 -- 1600. McGraw Hill Education/Medical,
5th edition, 2013. 5.1
118
[28] Zejuan Sheng, Sue-Yeon Choi, Ajay Dharia, Jian Li, Peter Sterling, and Richard H.
Kramer. Synaptic Ca2+ in darkness is lower in rods than cones, causing slower
tonic release of vesicles. The Journal of Neuroscience, 27(19):5033 -- 5042, May
2007. 3.1.3
[29] Matthew Shtrahman, Chuck Yeung, David W. Nauen, Guo qiang Bi, and Xiao
lun Wu. Probing vesicle dynamics in single hippocampal synapses. Biophysical
Journal, 89:3615 -- 3627, 2005. 3.1.3, 4.2
[30] Joshua H. Singer and Jeffrey S. Diamond. Vesicle depletion and synaptic depres-
sion at a mammalian ribbon synapse. Journal of Neurophysiology, 95:3191 -- 3198,
2006. 3.1.3
[31] Josefin Snellman, Bhupesh Mehta, Norbert Babai, Theordore M. Bartoletti,
Wendy Akmentin, Adam Francis, Gary Matthews, Wallace B. Thoreson, and
David Zenisek. Acute destruction of the synaptic ribbon reveals a role for the
ribbon in vesicle priming. Nature Neuroscience, 14(9):1135 -- 1143, 2011. 1.3
[32] Hanna M. Sobkowicz, Jerzy E. Rose, Grayson E. Scott, and Susan M. Slapnick.
Ribbon synapses in the developing intact and cultured organ of Corti in the
mouse. The Journal of Neuroscience, 2(7):942 -- 957, 1982. 3
[33] Peter Sterling and Gary Matthews. Structure and function of ribbon synapses.
TRENDS in Neurosciences, 28(1):20 -- 29, 2005. 1.3, 1.3
[34] Charles F. Stevens and John F. Wesseling. Activity-dependent modulation of the
rate at which synaptic vesicles become available to undergo exocytosis. Neuron,
21:415 -- 424, 1998. 3.1.3
119
[35] Wallace B. Thoreson. Diagrams of the synaptic ribbon. private communication,
2014. 1.4
[36] Wallace B. Thoreson, Matthew J. Van Hook, Caitlyn M. Parmelee, and Carina
Curto. Vesicle pools at the cone ribbon synapse: changes in release probability are
solely responsible for voltage-dependent changes in release. Synapse, 70(1):1 -- 14,
2015. 1, 2, 2.2, 2.2.2, 2.3, 2.4, 2.3.2, 2.1, 2.3.2
[37] Wallace B. Thoreson, Katalin Rabl, Ellen Townes-Anderson, and Ruth Heidel-
berger. A highly Ca2+-sensitive pool of vesicles contributes to linearity at the
rod photreceptor ribbon synapse. Neuron, 42:595 -- 605, May 2004. 1.3, 3.1.3
[38] Matthew J. Van Hook, Caitlyn M. Parmelee, Minghui Chen, Karlene M. Cork,
Carina Curto, and Wallace B. Thoreson. Calmodulin enhances ribbon replen-
ishment and shapes filtering of synaptic transmission by cone photoreceptors.
Journal of General Physiology, 144(5):357 -- 78, 2014. 1, 3, 3.1.1, 3.1, 3.2, 3.2, 3.3,
3.2.3, 3.3
[39] Lutz Vollrath and Isabella Spiwoks-Becker. Plasticity of retinal ribbon synapses.
Microscopy Research and Technique, 35(6):472 -- 487, 1996. 1.3
[40] Heinz Wassle. Parallel processing in the mammalian retina. Nature Reviews:
Neuroscience, 5:747 -- 757, 2004. 1.1
[41] John F. Wesseling and Donald C. Lo. Limit on the role of activity in controlling
the release-ready supply of synaptic vesicles. The Journal of Neuroscience,
22(22):9708 -- 9720, 2002. 3.1.3
[42] Herbert S. Wilf. generatingfunctionology. Academic Press, 2nd edition, 1994.
2.4.4
120
[43] Nathan R. Wilson, Jiansheng Kang, Emily V. Hueske, Tony Leung, Helene
Varoqui, Jonathan G. Murnick, Jeffrey D. Erickson, and Guosong Liu. Presynaptic
regulation of quantal size by the vesicular glutamate transporter VGLUT1. The
Journal of Neuroscience, 25(26):6221 -- 6234, 2005. 2
[44] Xiaohui Xie, Richard H.R. Hahnloser, and H. Sebastian Seung. Selectively
grouping neurons in recurrent networks of lateral inhibition. Neural Computation,
14:2627 -- 2646, 2002. 5.1
[45] D. Zenisek, J. A. Steyer, and W. Almers. Transport, capture, and exocytosis of
single synaptic vesicles at active zones. Nature, 406:849 -- 854, 2000. 1.3
|
1803.08541 | 1 | 1803 | 2018-03-22T18:46:21 | A spectrum of routing strategies for brain networks | [
"q-bio.NC"
] | Communication of signals among nodes in a complex network poses fundamental problems of efficiency and cost. Routing of messages along shortest paths requires global information about the topology, while spreading by diffusion, which operates according to local topological features, is informationally "cheap" but inefficient. We introduce a stochastic model for network communication that combines varying amounts of local and global information about the network topology. The model generates a continuous spectrum of dynamics that converge onto shortest-path and random-walk (diffusion) communication processes at the limiting extremes. We implement the model on two cohorts of human connectome networks and investigate the effects of varying amounts of local and global information on the network's communication cost. We identify routing strategies that approach a (highly efficient) shortest-path communication process with a relatively small amount of global information. Moreover, we show that the cost of routing messages from and to hub nodes varies as a function of the amount of global information driving the system's dynamics. Finally, we implement the model to identify individual subject differences from a communication dynamics point of view. The present framework departs from the classical shortest paths vs. diffusion dichotomy, suggesting instead that brain networks may exhibit different types of communication dynamics depending on varying functional demands and the availability of resources. | q-bio.NC | q-bio | A spectrum of routing strategies for brain networks
Andrea Avena-Koenigsberger1*, Xiaoran Yan2, Artemy Kolchinsky3, Martijn van den Heuvel 4, Patric
Hagmann5,6, Olaf Sporns1,2
1. Department of Psychological and Brain Sciences, Indiana University, Bloomington, IN, USA
2. IU Network Institute, Indiana University, Bloomington, IN, USA
3. Santa Fe Institute, Santa Fe, NM, USA
4. Brain Center Rudolf Magnus, University Medical Center Utrecht, University Utrecht, The Netherlands
5. Department of Radiology, Centre Hospitalier Universitaire Vaudois (CHUV) and University of Lausanne
(UNIL), Lausanne, Switzerland
6. Signal Processing Laboratory 5 (LTS5), Ecole Polytechnique Federale de Lausanne (EPFL), Lausanne,
Switzerland
* Corresponding Author
Andrea Avena-Koenigsberger
[email protected]
Keywords. Connectome, Communication cost, Communication efficiency, Biased Random Walks
1
Abstract
Communication of signals among nodes in a complex network poses fundamental problems of
efficiency and cost. Routing of messages along shortest paths requires global information about the
topology, while spreading by diffusion, which operates according to local topological features, is
informationally "cheap" but inefficient. We introduce a stochastic model for network communication
that combines varying amounts of local and global information about the network topology. The model
generates a continuous spectrum of dynamics that converge onto shortest-path and random-walk
(diffusion) communication processes at the limiting extremes. We implement the model on two cohorts
of human connectome networks and investigate the effects of varying amounts of local and global
information on the network's communication cost. We identify routing strategies that approach a
(highly efficient) shortest-path communication process with a relatively small amount of global
information. Moreover, we show that the cost of routing messages from and to hub nodes varies as a
function of the amount of global information driving the system's dynamics. Finally, we implement the
model to identify individual subject differences from a communication dynamics point of view. The
present framework departs from the classical shortest paths vs. diffusion dichotomy, suggesting instead
that brain networks may exhibit different types of communication dynamics depending on varying
functional demands and the availability of resources.
2
Introduction
The function of many real world complex networks is to relay information within and between their
constituent elements. Efficient communication, i.e. the passing of information at high speed and high
reliability at low cost to the system, is essential to the functioning of systems in many domains, ranging
from technological to social and biological applications. For example, communication is central to the
operation of brain networks, as it is necessary for information integration and for distributed neural
computation [1]. However, the mechanisms that enable information to flow efficiently among large
numbers of distributed elements interacting through a complex topology remain mostly unexplained.
Previous work on optimal routing in networks highlighted the importance of small-world topologies for
promoting short communication pathways at low wiring cost [2,3]. Indeed, information transfer that
takes place through topologically shortest paths is both fast and direct, and reduces a message's
vulnerability to errors and attack [4]. Yet, such a communication model also has disadvantages: it
discounts the vast majority of a network's structural connections [5,6], it is prone to bottlenecks and
congestion [7-9], and it lacks robustness to edge failures [10]. Most importantly, a system's ability to
route along shortest paths relies on all of the system's elements having information about the global
topology of the network [11,12]. Therefore, an explicit analysis of the costs and benefits of efficient
communication should take into account the cost associated with global information, in addition to
better-known costs such as wiring and energy consumption [1,13-17]. We refer to the cost of the
information necessary for signal routing as the informational cost.
A drastically different picture emerges if we discard the premise that the system's elements are capable
of accessing information about the global topology of the network. Under this scenario, signals are
dispersed according to a random walk or diffusion process [18-20], driven only by local topological
properties. While diffusion has no associated cost of storing global topological information,
communication is inefficient if measured in terms of the time needed for a signal to arrive at a specific
destination. This results in increased vulnerability to signal corruption and slower integration of
information as signals are broadcast and spread indiscriminately across the network.
While shortest paths and diffusion have been extensively studied in the context of network
communication, they merely represent the extremes of a spectrum of communication processes that
3
deserve greater attention. As an example, for some types of network topologies, a preferential choice
policy where messages are preferentially routed towards high degree nodes [21, 22] decreases search
times significantly compared to random walks, yet the informational cost is small since nodes only need
to "know" the degree of their neighbors. Brain networks are a case in point: on average, shortest paths
tend to follow a low-to-high and then high-to-low degree sequence [23] and closeness centrality
sequence [24], suggesting that efficient routing patterns in brain networks could be driven by a mixture
of degree and closeness preferential choice policies. Preferential policies are often modeled as biased
random walks [25], where the motion of a random walker located at a given node is biased according to
an attribute (e.g. degree) associated with the neighboring nodes. It has been shown that biased random
walks can generate relatively efficient communication processes (high speed, low cost) and are able to
account for navigation rules that are observed in real world systems [26-29], offering alternative
interpretations of node centralities and community structures [30].
Here, we focus on a specific family of biased random walks, governed by routing strategies generated by
a stochastic model that combines local and global information about the network topology. This
framework allows us to explore a continuous spectrum of dynamics that converge onto shortest-path
communication processes at one extreme, and random-walk (diffusion) communication processes at the
other extreme. We apply this framework to investigate communication cost from a dynamic point of
view in large-scale brain networks. We suggest that brain networks may exhibit different types of
communication dynamics depending on varying functional demands and the availability of resources.
A continuous spectrum of routing strategies combining local and global information.
We model messages or signals transferred from a source brain region to a target brain region as random
walkers traversing a brain network, where network nodes and edges represent small cortical parcels that
are connected by bundles of axons. We consider the dynamics of such random walkers
(signals/messages) on the network, where walkers must reach an a priori specified target node t.
Formally, let X be a random variable indicating the current node of the walker, Y the random variable
indicating the node to which the walker will move in the next time step, and T the random variable
indicating the target node where the walk will terminate (we assume that all nodes can be reached from
all nodes in finite time). For all t, we denote the transition probabilities at X=i as:
4
Pt
ij = Pr(Y = jX = i,T = t) where ∑j Pt
ij = 1, and Pt
ij = 0 when there is no connection between nodes i
and j. Finally, the walk ends when i = t, in which case Pt
ij = 1 for j=t and 0 for all other j. Formally, the
network dynamics for each separate target t form a Markov chain with state t as an absorbing state (see
Methods). The set of transition probabilities for all t express the routing strategy that governs the
dynamics of walkers (signals) navigating the network.
We specify transition probabilities at every node using a family of dynamical processes that combine
local and global information about the network's topology. To this end, we define the dynamics of the
system by controlling the effect of global information using the following stochastic model:
where
is a normalization factor. Transition probabilities are governed
by two sources of information:
• a local source of information dij denoting the length of the edge connecting i and j (
if
and only if a connection between i and j exists).
• a global source of information dij + gjt, denoting the minimum distance from node i to target t
though node j. This is the sum of the distance between node i and neighbor node j (dij), plus the
distance from node j to the target node t through the shortest path (gjt - note that this term has no
dependence on i).
The parameter controls the extent to which transition probabilities are shaped by global information.
Most importantly, gradually changes the dynamics on the network from an unbiased random walk
towards a shortest-path routing strategy (see Fig 1):
• When =0, a walker's motion is driven only by local information. Transition probabilities are
simply given by
and do not depend on the target node
(nonetheless, the walk still terminates when it eventually reaches the target node t). In the case of
brain networks, where edge-weights express connection strengths or node proximities in the
interval (0,1) (this can always be enforced through a unique linear normalization function [32];
See Methods), we apply the proximity-to-distance function
and map all edge-
5
tiijjtijZdgdtTiXjYP1)))((exp(),(jijjtijtidgdZ)))((exp(ijdtiijZdtTiXjYP1)exp(),(dij=-log(wij) weights onto edge-distances. The resulting dynamics,
, where si
= ∑j wij is the strength of node i, defines an unbiased random walk on the network where walkers
favor transitions through edges with shorter connection distances (i.e. closer proximities). We
refer to the unbiased random walk as the reference navigation strategy, Pref, as it represents a
null model of navigation that would naturally take place on the network if no bias is introduced.
• When → ∞ global information governs the model and transition probabilities converge to Pt
ij =
1 if the edge
lies on the (unique) shortest path between i and t (degenerate shortest paths,
i.e. more than one shortest path from i to t, are unlikely in weighted networks, but see Methods
for the case where degenerative shortest paths exist) and Pt
ij = 0
otherwise. Hence, statistics
computed on such walks will correspond to a "shortest-path" routing strategy - in particular
average walk length will be equal to shortest path length.
Fig 1. A spectrum of routing strategies. The parameter controls the extent to which routing strategies
(transition probabilities) are reshaped by global information. Toy networks in the top row illustrate how
transition probabilities, represented by orange arrows, are reshaped as the parameter increases. At each
node, the orange arrows are proportional to the probability of a walker moving to a neighboring node via
that edge. Blue arrows on the toy networks in the bottom row illustrate a possible walk followed by a
6
iijswtTiXjY1),(P0{i,j} random walker (signal) going from node 1 to node t, while operating according to the routing strategy
represented by the orange arrows. When =0, transition probabilities at each node are proportional to the
strength of its connections. Random walkers operating under this routing strategy (the reference
navigation strategy, Pref) diffuse through the network until they eventually arrive at the target node.
When → ∞, transition probabilities at each node route walkers through the shortest path to the target
node; a walker starting at node 1 will follow the shortest path to node t, as illustrated by the blue arrows.
In the middle of the spectrum, walker's dynamics are influenced by global information but still driven
partially by local topological properties. Notice that only transition probabilities vary with while the
underlying network structure remains invariant.
It is worth noting that the model acts on the routing strategies by changing the transition probabilities at
each node, but the topology and weight structure of the network remain unchanged (see Fig 1).
The cost of reshaping the system's dynamics.
We are interested in characterizing the communication cost of the dynamics generated by our model as
we gradually increase . Here, we focus on two aspects of the cost associated with a communication
process. First, we consider a transmission cost, which is the cost associated with messages being
transmitted from one node to another. Second, we consider an informational cost, which is the cost
associated with using global information to reshape the system's dynamics and thus route messages
efficiently.
We consider a walker navigating the network and acting according to the routing strategies P(Y X,T).
Let 𝑐𝜆
𝑡𝑟𝑎𝑛𝑠(i,t) = ∑j P( Y= j X=i,T = t) dij be the immediate transmission cost at node i for a walker
going to node t with routing strategy P( Y X=i,T = t). The immediate transmission cost quantifies the
cost associated with X=i partaking in the communication process by relaying the message to one of its
neighbors, and in this setting it is equivalent to the expected distance that a walker at node i has to travel
to move to a neighbor of i. Let nt
(i,k) be the mean number of times node k is visited by a walker
starting at a source node X0=i and acting according to a routing strategy P(Y X=i,T = t). We define the
transmission cost of a walk starting at source node X0=i and terminating at the target node t as the sum
of the immediate transmission costs accumulated at each visited node along a walk, that is 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(i,t) =
∑k nt
(i,k) 𝑐𝜆
𝑡𝑟𝑎𝑛𝑠(k,t). Thus, a walk's transmission cost is equivalent to the mean walk length between
nodes i and t, under the routing strategy defined by λ. Noting that the transmission cost is not a
symmetric measure, (i.e. 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(i,t) may not be the same as 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(t,i), except for when → ∞), we can
7
define the average transmission cost of a node acting as a source as 𝐶⃗
𝑡𝑟𝑎𝑛𝑠(𝑖) = 1
𝜆
𝑁
∑ 𝐶𝜆
𝑡
𝑡𝑟𝑎𝑛𝑠(𝑖,𝑡)
, and the
average transmission cost of a node acting as a target as 𝐶⃖
𝑡𝑟𝑎𝑛𝑠(𝑡) = 1
𝜆
𝑁
∑ 𝐶𝜆
𝑖
𝑡𝑟𝑎𝑛𝑠(𝑖,𝑡)
. These measures
quantify the source and target closeness centrality of each node under a routing strategy: 𝐶⃗
𝑡𝑟𝑎𝑛𝑠(𝑖)
𝜆
quantifies the average walk length from a node i to any other target node in the network, whereas
𝐶⃖
𝑡𝑟𝑎𝑛𝑠(𝑡) quantifies the average walk length from any source node to the target node t.
𝜆
To quantify the informational cost associated with routing messages to a target node t under the routing
strategy P(Y X=i,T = t), we define 𝑐𝜆
𝑖𝑛𝑓𝑜
(i,t) = KL(P(Y X=i,T = t)Pref(Y X=i,T = t)) as the
informational cost at node X=i, measuring the Kullback-Leibler divergence between the routing strategy
P(Yi,t) and the reference routing strategy Pref (Yi,t). The informational cost quantifies the effect of the
bias due to global information by measuring how much reshaping of the system's dynamics has taken
place at node X=i [33]. Then, the informational cost of routing a message from a starting at node X=i to
a target node t is the weighted average informational cost across all nodes in the network, weighted by
the frequency with which each node is visited along the walk: 𝐶𝜆
𝑖𝑛𝑓𝑜
(i,t) = ∑ (
𝑘
𝑡 (𝑖,𝑘)
𝑛𝜆
𝑘
∑ 𝑛𝜆
𝑡 (𝑖,𝑘)
𝑡𝑟𝑎𝑛𝑠(𝑘, 𝑡))
𝑐𝜆
.
𝑖𝑛𝑓𝑜(𝑖) =
Finally, we can define the average informational cost of a node acting as a source as 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜(𝑖,𝑡)
∑ 𝐶𝜆
𝑡
𝑖𝑛𝑓𝑜(𝑡) = 1
, and the average informational cost of a node acting as a target as 𝐶⃖
𝜆
𝑁
𝑖𝑛𝑓𝑜(𝑖,𝑡)
.
∑ 𝐶𝜆
𝑖
1
𝑁
In the following sections we will study the communication costs of routing strategies generated by our
stochastic model applied to the structural brain connectivity matrices of two cohorts of healthy subjects.
In the main text, we focus on 173 unrelated subjects from the Human Connectome Project (HCP) dataset
[34,35]. The Supplementary Fig S7 - S11 show results from the replication dataset (LAU), composed of
40 healthy subjects (see Methods). We first analyze cost measures at the global, nodal, and pairwise
level, measured and averaged across all subjects (within each cohort). In the last section we demonstrate
the utility of this approach for identifying individual subject differences from a communication
dynamics point of view, hence departing from the classical routing vs. diffusion dichotomy [12,36].
Brain networks are more efficient within an intermediate region of the communication spectrum.
By construction, the transmission and informational cost have a competing relationship (or trade-off)
such that as we increase in the stochastic model, the mean walk lengths (𝐶𝜆
𝑡𝑟𝑎𝑛𝑠) of messages acting
8
according to P become shorter while the bias effect due to global information (𝐶𝜆
𝑖𝑛𝑓𝑜
) increases. This
trade-off is shown in Fig 2a where averages of 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
across all {i,t} pairs are plotted as a
function of . It can be seen that 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠, measuring the average walk length, approaches a shortest path-
length regime at around = 1 (ln() = 0 in Fig 2), suggesting that in this regime messages can be
efficiently routed at a low informational cost.
𝑡𝑟𝑎𝑛𝑠 (red) and 𝐶𝜆
𝑖𝑛𝑓𝑜
𝑖𝑛𝑓𝑜‖ (blue) across all node pairs. These curves are computed by normalizing 𝐶𝜆
Fig 2: A spectrum of communication processes. (a) Averages of 𝐶𝜆
node pairs, as a function of . Solid red and blue lines correspond to the median across all subjects,
whereas the shaded red and blue regions denote the 95th percentile. (b) Averages of ‖𝐶𝜆
‖𝐶𝜆
respect to the same cost measures computed on ensembles of 500 randomized networks (per subject).
𝑖𝑛𝑓𝑜‖ that are smaller than 1,
Shaded red and blue areas indicate sections of the curves ‖𝐶𝜆
respectively, indicating the regions in the spectrum where the communication cost of empirical networks
is smaller than the cost computed on the randomized ensembles. The dashed vertical lines are placed at
the minimum and maximum of ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ (2 and 3, respectively), and at two points near the extremes of
𝑡𝑟𝑎𝑛𝑠‖ and ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ (red) and
with
𝑖𝑛𝑓𝑜
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
(blue) across all
9
the spectrum ((1 and 4). (c) pairwise values of 𝐶𝜆
𝐶𝜆
𝑖𝑛𝑓𝑜
(i,t) for all node pairs. In all panels, 1=e-4.49, 2=e-1.64, 3=e0.37 and 4=e1.79.
𝑡𝑟𝑎𝑛𝑠(i,t) for all node pairs. (d) pairwise values of
Next, we consider an ensemble of random networks and compare average transmission and
informational costs incurred in empirical brain networks and in randomized ensembles of networks. All
randomized networks preserve node degree, node strength (evaluated with respect to the proximity edge-
weights), and the network's weight distribution (see Methods). We generate routing strategies P for all
randomized networks and normalize the cost measures 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
of each subject's empirical
brain network with respect to the average cost measures computed across the corresponding randomized
counterparts. Fig 2b shows normalized cost measures ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ = 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(𝑒𝑚𝑝)/𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(𝑟𝑎𝑛𝑑) (red
line) and ‖𝐶𝜆
𝑖𝑛𝑓𝑜‖ = 𝐶𝜆
𝑖𝑛𝑓𝑜(𝑒𝑚𝑝)/𝐶𝜆
𝑖𝑛𝑓𝑜(𝑟𝑎𝑛𝑑) (blue line) as a function of . In accordance with prior
work (37-39), we find that average walk lengths are shorter for random networks (i.e. ‖𝐶𝜆
the extremes of the spectrum, representing the unbiased random walk (Pref ) and shortest path regimes.
𝑡𝑟𝑎𝑛𝑠‖ > 1) at
Interestingly, our analysis reveals an interval of values (shaded region in Fig 2b) for which empirical
networks exhibit shorter walk-lengths compared to the randomized counterparts (i.e. ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ < 1).
Moreover, the informational cost behaves similarly, although the regions ‖𝐶𝜆
𝑖𝑛𝑓𝑜‖ < 1 and ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ < 1
barely overlap. Overall, these results show that the randomized counterparts of empirical brain networks
are more efficient only at the extremes of the communication spectrum.
Fig 2c and 2d show pairwise 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
(median across subjects) computed for routing strategies
generated with 1=e-4.49, 2=e-1.64, 3=e0.37 and 4=,e1.79 (see dashed vertical lines in Fig 2a and 2b).
These values of correspond to two points located near the extremes of the communication spectrum,
and two points located at the minimum and maximum of the curve ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖, where the empirical
networks are most and least efficient compared to their randomized counterparts. As evidenced by the
column-like patterns in the matrices corresponding to 1 and 2, the dynamics of messages navigating
the network are strongly determined by the local connectivity of the target node when the global
information bias is small. As the bias increases and routing strategies depart from the reference strategy
Pref, the dynamics of messages are less dependent on the target node only. Finally, as walk-lengths
converge towards shortest-path, the transmission cost becomes symmetric, i.e., 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(i,t) = 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(t,i).
10
Source vs. target communication cost
We now analyze cost measures at the nodal level. Fig 3a and 3b show scatter plots of the average source
and target transmission costs (𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠, respectively), and the average source and target
𝜆
𝜆
𝑖𝑛𝑓𝑜
informational costs (𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
, respectively) associated to all nodes (median across all subjects)
for the same values of highlighted in Fig 2. Nodes are colored according to their membership in
functional intrinsic connectivity networks (ICNs; see Methods), highlighting a tendency of some ICNs
to contain an overabundance of costly sources and/or targets, while other ICNs' cost varies as a function
of . For example, nodes belonging to the somatomotor network (SM, colored green) tend to exhibit a
𝑖𝑛𝑓𝑜
high 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and low 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 for < e0.37, while they also exhibit a consistent low 𝐶⃗
𝜆
𝜆
𝜆
; nodes belonging
𝑖𝑛𝑓𝑜
to the visual network (VIS, colored red) exhibit high 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
for >e-4.49 , while 𝐶⃗
𝑡𝑟𝑎𝑛𝑠and
𝜆
𝐶⃖
𝑡𝑟𝑎𝑛𝑠 vary as a function of .
𝜆
11
𝑖𝑛𝑓𝑜
) and target (𝐶⃖
𝜆
) during communication processes taking place under routing strategies
Fig 3: Nodal average transmission costs for four increasingly biased routing strategies. (a) Scatter plots
show the transmission cost associated to each node when it acts as source (𝐶⃗
𝑡𝑟𝑎𝑛𝑠) and target (𝐶⃖
𝑡𝑟𝑎𝑛𝑠)
𝜆
𝜆
during communication processes taking place under routing strategies generated with the values 1, 2,
3 and 4. (b) Scatter plots show the transmission cost associated to each node when it acts as source
𝑖𝑛𝑓𝑜
(𝐶⃗
𝜆
generated with the values 1, 2, 3 and 4. Markers in the scatter plots in (a) and (b), representing each
node, are colored according to the node's membership in the 7 intrinsic connectivity networks (ICN)
defined by Yeo et al. (2011) [61]: Visual (VIS), Somatomotor (SM), Dorsal Attention (DA), Ventral
Attention (VA), Limbic (LIM), Frontal Parietal (FP), and Default Mode Network (DMN). The size of
the markers is proportional to node's strength. (c) Correlations between node strength and 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 (red),
𝜆
𝑖𝑛𝑓𝑜
𝐶⃖
𝑡𝑟𝑎𝑛𝑠(orange), 𝐶⃗
(blue) as a function of . Solid lines show median correlation
𝜆
𝜆
across all subjects, shaded areas surrounding the lines show 95th percentile. Shaded colored areas
between the vertical dashed lines indicate regions where the correlations were not significant (p >
𝑖𝑛𝑓𝑜
0.001). (d) Correlation between 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠(red), and 𝐶⃗
𝜆
𝜆
𝜆
Solid lines show medians across all subjects and shaded areas surrounding solid lines show the 95th
𝑖𝑛𝑓𝑜
(green) and 𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
(blue), as a function of .
12
percentile. Shaded areas between the vertical dashed lines indicate areas where correlation values were
not significant (p > 0.001). In all panels, 1=e-4.49, 2=e-1.64, 3=e0.37 and 4=e1.79.
In order to assess to what extent high or low nodal costs are driven by the network's overall topology, as
opposed to nodal degree or strength distribution, we standardize nodal costs with respect to the
corresponding nodal cost distributions measured on the randomized network ensembles. Significantly
high or low standardized nodal cost measures are indicative of global connectivity patterns that are
encountered only in empirical brain networks. Supplementary Fig S1 - S4 show thresholded z-scores (α
= 0.01) for all nodal cost measures as a function of lambda. As expected, near the extremes of the
spectrum ( = 0 and > 1), most nodes exhibit significantly higher costs, compared to the randomized
networks, however, significantly low cost regions are found in the middle of the spectrum. Prominent
𝑖𝑛𝑓𝑜
low 𝐶⃗
𝜆
regions include the right and left hemisphere superior frontal and caudal middle frontal,
𝑖𝑛𝑓𝑜
precentral, paracentral and postcentral regions; low 𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
regions overlap with the low 𝐶⃗
𝜆
regions,
but also include the right and left posterior cingulate, the supramarginal gyrus, the superior parietal
cortex, the precuneus, and the inferior parietal cortex. Prominent low 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 regions are mainly located
𝜆
in the frontal cortex (frontal pole, medial orbital frontal and rostral middle frontal regions), right and left
superior parietal regions, the right and left precuneus, and the left cuneus. No significantly low 𝐶⃖
𝑡𝑟𝑎𝑛𝑠
𝜆
regions were identified.
Our analyses also reveal a varying relationship (as a function of ) between the nodal cost measures and
node strength (see Fig 3c). At the extremes of the spectrum, transmission cost is strongly driven by node
degree. When =0, the correlation between node strength and 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 is r = 0.55 and r = -
𝜆
𝜆
0.61, respectively (p < 0.001), indicating that high degree nodes (hubs) are costly sources but low cost
targets with respect to transmission cost. In other words, when the global information bias is low (or
zero), messages can be routed at a low transmission cost from any brain region to a hub; conversely,
routing a message from a hub to any brain region incurs a high transmission cost. At the other end of the
spectrum (i.e. for large values of ), hub nodes are low cost sources and targets with respect to
transmission cost (r = -0.53, p<0.001 ; note that the orange and red lines in Fig 3c converge). However,
in the middle of the spectrum, the average correlation between node strength and 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 is close to zero,
𝜆
whereas the correlation between node strength and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 remains significant (r ≈ -0.5 , p<0.001)
𝜆
13
throughout the entire spectrum. The relationship between source and target costs also varies as a
𝑖𝑛𝑓𝑜
function of (see Fig 3d). For low values of , both 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠, and 𝐶⃗
𝜆
𝜆
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
are
negatively correlated. In other words, nodes that are costly sources are efficient targets, and nodes that
are costly targets are efficient sources. However, the correlations undergo a sign flip as increases and
𝑖𝑛𝑓𝑜
𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠, and 𝐶⃗
𝜆
𝜆
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
become positively correlated. Note that the correlation
between 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 converges to 1 as these two measures are identical at the shortest-path
𝜆
𝜆
extreme (the symmetry between 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 at the shortest path extreme will hold for any
𝜆
𝜆
undirected network).
A node's propensity to be a costly transmission/informational source or target is projected onto the
cortical surface in Fig 4, where we show the difference between a node's source and target costs for
1=e-4.49, 2=e-1.64, 3=e0.37 and 4=,e1.79 (same values highlighted in Fig 2 and Fig 3) . Cortical regions
that are costly sources (compared to the cost of being a target) are colored red whereas regions that are
costly targets (compared to the cost of being a source) are colored blue. To assess whether a region's
propensity to be a costlier source or target is significant or not, given the node degrees and strengths, we
𝑖𝑛𝑓𝑜
standardize the empirical cost differences (i.e. 𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃗
𝜆
𝜆
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
) with respect to the
distribution of cost differences computed on the ensembles of randomized networks, and test whether
the empirically measured source and target cost difference is significantly larger than the difference
measured in the randomized ensembles. Our results indicate that the right and left hemisphere superior
parietal regions, the precuneus, and the fusiform gyri are significantly costlier sources, in terms of
transmission cost, for > e-0.8. Regions such as the right insula and rostral middle frontal cortex, right
and left superior frontal cortex, and the precentral gyri are significantly costlier targets in terms of
transmission cost, for > e-0.8. In terms of informational cost, we find that the precentral gyri, paracentral
lobule, right lateral-occipital cortex and left lingual gyrus are costlier sources, whereas the right
posterior cingulate and supramarginal gyrus, left precuneus, right and left frontal pole, superior parietal
cortex, and inferior parietal lobules are significantly costlier targets, for < e-1.1. All z-scored cost
differences as a function of are shown in Supplementary Fig S5 and S6.
14
Fig 4. A brain region's propensity to be a costly source or target. Cortical surfaces show the difference
between a node's source and target transmission costs. (a) 𝐶⃗
𝑡𝑟𝑎𝑛𝑠 − 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 for routing strategies
𝜆
𝜆
𝑖𝑛𝑓𝑜 − 𝐶⃖
𝑖𝑛𝑓𝑜
generated with the values 1, 2, 3 and 4. (b) 𝐶⃗
𝜆
𝜆
values 1, 2, 3 and 4. Red colored areas on the cortical surfaces correspond to nodes whose source
transmission/informational cost is higher than their target transmission/informational cost. Blue colored
areas correspond to nodes whose target transmission/informational cost is higher than their source
transmission/informational cost. In all panels, 1=e-4.49, 2=e-1.64, 3=e0.37 and 4=e1.79.
for routing strategies generated with the
Routing strategies for privileged nodes
In this section we will explore a different scenario where, in the interest of economizing on
informational cost, we allow only a subset of privileged nodes to have access to global information. We
consider increasingly larger size sets of r privileged nodes that are able to reshape their routing strategies
according to the influence of global information. Privileged nodes are selected according to different
node centrality rankings. Given a centrality-based ranking of nodes, we generate routing strategies for
the r-highest ranked (privileged) nodes according to the stochastic model, where is an attribute that
only applies to the set of privileged nodes; all non- privileged nodes' routing strategies remain unbiased
𝑡𝑟𝑎𝑛𝑠 and
and are equal to Pref(X). The left and middle panel of Fig 5 show network average values of 𝐶𝜆
𝑖𝑛𝑓𝑜
𝐶𝜆
(median across all subjects) as a function of for varying fractions of privileged nodes that are
selected according to various centrality-based rankings. The black dotted lines show 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
respectively, for the case in which all nodes' routing strategies are biased.
,
15
𝑡𝑟𝑎𝑛𝑠 (left panel) and 𝐶𝜆
𝑖𝑛𝑓𝑜
(middle panel) as a function of (node
Fig 5. Network average values of 𝐶𝜆
medians across all subjects) for 22, 55, 110, and 165 privileged nodes (corresponding to 10%, 25%, 50%
and 75% of the network's nodes) that are selected according to betweenness centrality ranking (yellow
line), strength ranking (purple line), shortest-path-based closeness centrality (green line), and random-
walk-based closeness centrality (blue line). For comparison purposes, we also show cost measures for
randomly sampled nodes (red line represents average across 500 samples). The dotted lines show 𝐶𝜆
and 𝐶𝜆
privileged nodes). Right panel shows node stretch distributions for the different sets of privileged nodes
and centrality rankings. Black markers indicate the median of the distributions.
, respectively, for the case in which all nodes' routing strategies are biased (i.e. 100%
𝑡𝑟𝑎𝑛𝑠
𝑖𝑛𝑓𝑜
This approach reveals three interesting properties about the routing capacity of the brain. First, the
composition of the set of privileged nodes matters, as evidenced by the differences in 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
that are obtained as the set size and composition is varied. Second, for a fixed number of privileged
𝑖𝑛𝑓𝑜
nodes, the more the system economizes on informational cost, the more it expends on transmission cost.
16
For example, routing strategies where we select privileged nodes according to betweenness centrality
ranking yield smaller 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠and larger 𝐶𝜆
𝑖𝑛𝑓𝑜
throughout the entire spectrum, compared to other
centrality-based privileged node selections. Conversely, routing strategies where we select privileged
𝑡𝑟𝑎𝑛𝑠, but least
nodes according to a random walk centrality ranking are the most costly in terms of 𝐶𝜆
costly in terms of 𝐶𝜆
𝑖𝑛𝑓𝑜
. Third, a small number of privileged nodes can achieve a 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 that is nearly as
small as the 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 achieved for shortest paths, but at a significantly smaller informational cost
compared to what is needed to route messages through shortest-paths. Following [24] we define the
stretch of a walk as the difference between the walk length and the shortest path length (both measured
in terms of the number of edges/steps). Then, the stretch of a source node is the average stretch with
respect to all target nodes in the network. For all sets of privileged nodes, we compute the stretch
between the walks generated with =e3.2 (i.e. the walks with the minimum 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠) and the shortest
paths. Node stretch distributions (medians across all subjects) are shown in the right-side panel of Fig 5.
For 25% privileged nodes, the median stretch for the BC ranking is 5.8; for the strength ranking it is 7;
for the SP closeness ranking it is 7.1; for the random ranking it is 13.5; for the RW closeness it is 46.5.
For 50% privileged nodes, the median stretch for the BC ranking is 4.2; for the strength ranking it is 4.7;
for the SP closeness ranking it is 4.9; for the random ranking it is 6.4; for the RW closeness it is 12.7.
Overall, these results indicate that efficient routing patterns can emerge even when less than half of the
nodes are capable of routing information.
A communication cost trade-off within subjects
Our approach allows us to study the variability of communication cost measures across subjects. We
first examine whether subjects who exhibit higher values of 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 at = 0 (that is, longer walk lengths
for the unbiased random walk) will maintain a high 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 throughout the entire spectrum. Fig 6a shows
correlations between all subject's 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠across all values of . These correlations show that subjects
who exhibit higher values of 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 at < e-3.1 are also subjects with the highest 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 at >1, but the
relationship is inverted in the middle of the spectrum.
17
𝑡𝑟𝑎𝑛𝑠 across
𝑖𝑛𝑓𝑜
𝑖𝑛𝑓𝑜
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
), respectively. Notice how some subject's 𝐶𝜆
curves after normalization with respect to the max(𝐶𝜆
Fig 6. Communication cost trade-off within subjects. (a) Correlations between all subject's 𝐶𝜆
all values of . Positive correlations are colored in red, negative correlations are colored in blue. (b)
Eight subject's 𝐶𝜆
max(𝐶𝜆
some subject's 𝐶𝜆
normalized 𝐶𝜆
𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 curves decay faster than others, and how
curves grow faster than others. (c) Scatter plot of the computed areas under the
curves, sowing a trade-off between the decay of 𝐶𝜆
) is r = -0.74, p < 0.001).
𝑡𝑟𝑎𝑛𝑠) and A(𝐶𝜆
𝑖𝑛𝑓𝑜
(the correlation between A(𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
𝑡𝑟𝑎𝑛𝑠) and
𝑡𝑟𝑎𝑛𝑠 and the growth of
𝑖𝑛𝑓𝑜
Finally, we investigate if there are differences in how individual subject's brain networks take advantage
of the global information bias. We address this question by measuring the area under each subject's
𝑡𝑟𝑎𝑛𝑠 curve and 𝐶𝜆
𝐶𝜆
𝑖𝑛𝑓𝑜
curve. Moreover, since we are interested in capturing the rate of decay and
growth of subject's 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
curves, we first normalize each subject's 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 curve with
respect to 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 at = 0 (that is, the average length of unbiased random walks), and we normalize each
subject's 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝐶𝜆
𝑖𝑛𝑓𝑜
𝑖𝑛𝑓𝑜
curve with respect to 𝐶𝜆
𝑖𝑛𝑓𝑜
at = e3.2 (that is, the max value of 𝐶𝜆
𝑖𝑛𝑓𝑜
). The normalized
curves of 8 subjects are shown in Fig 6b, illustrating curves that decay/grow faster
with , which we can capture by measuring the area under the curve. Fig 6c shows a scatter plot of the
areas under the normalized 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
curves of all subjects, exhibiting a strong negative
18
correlation between the normalized areas under 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
(r = -0.74 , p<0.001). This strong
relationship indicates that there is a trade-off between a brain network's ability to take advantage of
global information to route messages in a fast manner, and the amount of informational cost required to
achieve optimally fast routing. How this trade-off is negotiated varies across individual subjects.
Discussion
The efficiency of communication in real world networks is not only determined by the speed with which
messages are relayed, but the informational cost associated with selecting efficient routes is equally
important. Here we introduce a stochastic model that generates routing strategies on a network by
controlling the effect of global information over the actions of random walkers. We characterize the
trade-offs between the cost of reshaping the system's dynamics (𝐶𝜆
𝑖𝑛𝑓𝑜
) and the cost of relaying
messages through the network (𝐶𝜆
𝑡𝑟𝑎𝑛𝑠), and characterize these costs at a global, nodal and subject-wise
level. Our results show that biased random walks can rapidly approach a shortest-path communication
regime when afforded gradual small increases in the bias on global information.
Several properties inherent in this framework have important implications for the study of
communication processes in brain networks. First, routing patterns are derived from a dynamical point
of view, and not from a purely topological analysis of the system, allowing us to make use of well-
established theoretical results about linear processes and biased random walks [25-27]. Second, routing
patterns generated by these processes take place through multiple paths, combining structural robustness
with tolerance to noisy interference. Third, routing strategies at each node are dynamic and allow us to
model systems whose function demands flexible routing patterns that switch depending on intrinsic (e.g.
amount of global information available) and/or extrinsic factors. Finally, building on the concept of
dynamic routing patters, the notion of dynamic measures of centrality emerge naturally as a means to
quantify the varying importance of nodes and edges under different underlying dynamics [30]. Here we
𝑖𝑛𝑓𝑜
have proposed the nodal cost measures 𝐶⃗
𝑡𝑟𝑎𝑛𝑠, 𝐶⃖
𝑡𝑟𝑎𝑛𝑠, 𝐶⃗
𝜆
𝜆
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
as dynamic source and target
closeness centrality measures, but we note that additional centrality measures can be evaluated over the
underlying flow graphs [26], that is, the weighted networks where the patterns of flow generated by our
stochastic model are embedded.
19
The concept of communication dynamics has become increasingly important in the context of brain
networks [40,41]. Here, we address some of the assumptions behind two widely used brain
communication models, namely routing and diffusion models. On the one side, communication that
takes place through shortest paths assumes that neural elements are able to identify the optimal path and
route a signal/message through such path; however, the mechanisms by which signals are routed and the
informational cost associated with routing them are rarely discussed. On the other side, communication
that takes place through (unbiased) random walks assumes that signals are able to "bounce between
nodes" for long periods of time. Yet, such a scheme raises issues about signal integrity and strength as
well as metabolic cost. In our analysis, communication cost is not measured as a structural property of
the network [17,23,36]. While wiring cost affects brain communication by means of being an important
driver of brain geometry and network topology [1,17,15], it should be noted that wiring cost is a static
property of the network (within relatively short time-scales) that is invariant under any communication
process taking place on the network. In contrast, our framework approaches communication cost by
considering two different cost components that are measured from the dynamics generated by a specific
communication model. First, we consider the transmission cost which we interpret as a proxy for the
metabolic cost of transmitting neural signals from one neural node to another. It has been estimated that
about 50% of the brain's energy is used to drive signals across axons and synapses [1], suggesting that
energy consumption is a strong incentive to minimize the length of communication pathways in neural
systems. A natural derivation from the transmission cost measure results from its reciprocal (or inverse),
thus extending and generalizing the global (or routing) efficiency [37] and diffusion efficiency [12,36]
measures for shortest path and diffusion-based communication, respectively. Second, we consider the
cost of reshaping the patterns of information flow (informational cost) that allow a signal to be
efficiently routed towards a specific brain region. We conceptualize this cost as associated with
modulatory processes that take place at the mesoscale or microscale, where signal traffic may be
regulated as two neuronal population's firing rates change in order to synchronize and thus communicate
[42], or as a process that emerges on top of the collective oscillatory dynamics of neural elements [43].
Hence, the communication cost measures proposed here are dynamic as they vary with the patterns of
information flow generated by the communication process taking place on the network. Moreover, these
cost measures intrinsically capture the informational cost associated with traversing high-degree nodes,
that is, those comprising the brain's rich club. Indeed, it has been proposed that rich-club nodes facilitate
integration of information within the network at the expense of a high wiring cost [23]; nonetheless,
20
hubs are only advantageous for communication if signals can be routed through them, which implies
high information cost [44]. Interestingly, a strong relationship between node degree, and the
directionality with which signals are preferentially transferred through the network structure has been
found in analytical, computational and empirical studies [45, 46], where it has been noted that high
degree nodes' oscillatory activity lags in phase whereas low degree nodes' activity leads. These findings
suggest that hub nodes tend to be directional targets, while low degree nodes act like sources [46]. Here
we find similar routing patterns at the low-information end of the spectrum, where hub nodes tend to be
low cost targets, while low degree nodes are efficient as sources. Building on these findings, our results
also suggest that there is a regime within the spectrum where empirical networks are more efficient than
their randomized counterparts; within this regime, the frontal cortex has an overabundance of high
informational and high transmission cost source nodes; conversely, the parietal cortex exhibits an
overabundance of high informational and high transmission cost target nodes.
Our results also contrast with well-established notions about the efficiency of random topologies
[37,38], demonstrating that the randomized counterparts of empirical brain networks are only more
efficient at the extremes of the communication spectrum. These findings have implications in terms of
the use of randomized networks as a point of reference to normalize graph-efficiency measures [12,47].
Furthermore, our findings support the idea that the brain's topology does not only optimize a trade-off
between wiring cost and efficient communication [15], but informational cost, the ability to access
multiple pathways, and flexible routing patterns are additional important factors driving the network's
organization.
Our findings regarding the selection of privileged nodes that have access to global information show that
some nodes are poised to take advantage of global information more efficiently than others; in brain
networks, efficient routing patterns can be achieved by allowing as few as 25% of the highest
betweenness or strength centrality nodes to reshape their routing strategies according to a bias on global
information. These results offer a new perspective on the role of highly central nodes in facilitating the
co-existence of functional integration and segregation between and within neural sub-systems: densely
connected clusters of nodes (network communities) tend to "trap" random walkers [48] which promotes
their segregation, while a few well-connected privileged nodes are specialized to direct the sharing and
exchange of information between clusters. Hence, the privileged nodes framework presented here may
21
provide some insight about the underlying communication processes allowing the exchange of
information between sub-systems [49,50].
Some limitations are worth mentioning. First, for this study, our application of the stochastic model is
limited by restricting to be a global attribute for all nodes, or for a set of privileged nodes; nonetheless,
it is feasible (although computationally intensive) and perhaps more realistic to define as a
continuously varying nodal property, (i). Second, the stochastic model considers a scenario where
communication between all nodes and a given target is equally salient. In systems such as the brain,
where different sub-systems are associated with specific cognitive tasks, it is unlikely that all node pairs
require the ability to efficiently exchange information with all other nodes. In this sense, the cost
measures computed here may serve as an upper bound for the actual communication cost. Third, linear
dynamics may not be appropriate for systems that exhibit highly complex non-linear dynamics. Indeed,
the brain is highly complex, topologically and dynamically. Yet, its complexity allows us to study it at
different scales [51]. While it is clear that both structure and dynamics must be considered
simultaneously to achieve a more comprehensive description of the system, it is still unclear how
communication dynamics manifest at the various scales at which we are able to capture brain structure
and dynamics. Hence, there is no evidence to discard linear dynamics as good approximation of the
routing patterns taking place on large-scale brain networks.
Taken together, our work establishes a theoretical framework to study the efficiency of a broad range of
communication processes on complex networks. While we have focused on a particular class of biased
random walks where biases depend on the topological distance to target nodes, we note that biases may
also depend on other aspects of the global topology or the embedding of a network in physical space
[14,28]. Overall, this framework can be used to study any real world network that employs
communication or navigation processes in its operation. It may be used, for instance, to infer pathways
through which information is preferentially transferred, or, when such pathways are known, to infer the
search and navigation strategies that allow accessing these pathways. In the context of brain networks,
this theoretical framework may prove useful to identify efficient communication strategies that balance
different aspects of the cost associated with neural communication.
22
Materials and Methods
Data sets.
LAU. Informed written consent in accordance with the Institutional guidelines (protocol approved by the Ethics
Committee of Clinical Research of the Faculty of Biology and Medicine, University of Lausanne, Switzerland)
was obtained for all subjects. Forty healthy subjects (16 females; 25.3 ± 4.9 years old) underwent an MRI session
on a 3T Siemens Trio scanner with a 32-channel head coil. Magnetization prepared rapid acquisition with gradient
echo (MPRAGE) sequence was 1-mm in-plane resolution and 1.2-mm slice thickness. DSI sequence included 128
diffusion weighted volumes + 1 reference b0 volume, maximum b value 8000 s/mm2, and 2.2 × 2.2 × 3.0 mm
voxel size. EPI sequence was 3.3-mm in-plane resolution and 3.3-mm slice thickness with TR 1920 ms. DSI and
MPRAGE data were processed using the Connectome Mapper Toolkit [52]. Each participant's gray and white
matter compartments were segmented from the MPRAGE volume. The grey matter volume was subdivided into
68 cortical and 15 subcortical anatomical regions, according to the Desikan-Killiany atlas, defining 83 anatomical
regions. These regions were hierarchically subdivided to obtain five parcellations, corresponding to five different
scales [53]. The present study uses a parcellation comprising 233 regions of interest (ROI). Whole brain
deterministic streamline tractography was performed on reconstructed DSI data, initiating 32 streamline
propagations (seeds) per diffusion direction, per white matter voxel [54]. Within each voxel, seeds were randomly
placed and for each seed, a fiber streamline was grown in two opposite directions with a 1mm fixed step. Fibers
were stopped if a change in direction was greater than 60 degrees/mm. The process was complete when both ends
of the fiber left the white matter mask. For each individual subject, connection weights between pairs of ROI are
quantified as a fiber density [55]. Thus, the connection weight between the pair of brain regions {u,v} captures the
average number of streamlines per unit surface between u and v, corrected by the average length of the
streamlines connecting such brain regions. The aim of these corrections is to control for the variability in cortical
region size and the linear bias toward longer streamlines introduced by the tractography algorithm. Fiber densities
were used to construct individual subject structural connectivity matrices. Each structural connectivity matrix is
then modeled as the adjacency matrix A={aij}of a graph G = {V,G} with nodes V = {v1,...,vn} representing ROIs,
and weighted, undirected edges E = {e1,...,em} representing anatomical connections with their fiber densities.
HCP. High-resolution diffusion-weighted (DWI) data from the Human Connectome Project [34] including 173
subjects (Q3 release; males and females mixed, age 22–35 years; imaging parameters: voxel size 1.25 mm
isotropic, TR/TE 5520/89.5 ms, 90 diffusion directions with diffusion weighting 1000, 2000, or 3000 s/mm2) was
used to reconstruct macroscale human connectomes for each subject. DWI data processing included the
following: (1) eddy current and susceptibility distortion correction, (2) reconstruction of the voxelwise diffusion
profile using generalized q-sampling imaging, and (3) whole-brain streamline tractography (see ref 56 for details).
23
Cortical segmentation and parcellation was performed on the basis of a high-resolution T1-weighted image (voxel
size: 0.7 mm isotropic) using FreeSurfer [57], automatically parcellating the complete cortical sheet into 219
distinct regions using a subdivision of the Desikan-Killiany atlas. White matter pathways were reconstructed
using generalized Q-sampling imaging (GQI), and streamline tractography [56]. A streamline was started in each
white matter voxel, following the most matching diffusion direction from voxel to voxel until a streamline
reached the gray matter, exited the brain tissue, made a turn of >45 degrees or reached a voxel with a low
fractional anisotropy (<0.1). For each individual subject, a 219 x 219 weighted connectivity matrix was
constructed by taking the strength of reconstructed region-to-region connections as the number of tractography
streamlines between i and j, and dividing by the average cortical surface area of both regions [53].
Defining topological distances for human structural connectivity networks. The edge weights of human brain
structural connectivity networks are normally defined in terms of proximity measures such as the number of
streamlines or fiber densities. These proximity edge-weights are often interpreted as a measure of information
flow or traffic capacity that can travel through a connection (a notion that is analogous to the concept of
bandwidth in telecommunication networks). Hence, the proximity between two brain regions is determined by the
sequence of edges that maximize the traffic or flow capacity. In order to define topological distances on human
brain structural connectivity networks, a proximity-to-distance mapping must be applied over the set of edge-
weights, such that large edge-weights (large edge-proximities) are mapped onto small edge-distances, and small
edge-weights are mapped onto large edge-distances. The proximity-to-distance mapping can be defined in various
ways. Following previous work [6,44], in this study we use the mapping dij = log(1/wij), where wij are edge-
proximities (i.e. fiber densities) and dij are the resulting edge-distances. This mapping yields edge-distances with a
log-normal distribution, which is consistent with evidence showing log-normal distributions of synaptic strengths
between cortical cells [58] and cortico-cortical projections [59]. Finally, in order to implement this mapping, we
first normalize all edge-weights, to ensure that wij are bounded in the interval [0,1]. As shown previously [32],
there is a unique linear function that can normalize any weighted graph onto the unit interval without affecting
network properties:
𝑤𝑖𝑗 =
(1 − 2𝜖)𝑤𝑖𝑗 + (2𝜖 − 1) ∙ 𝑀𝐼𝑁(𝑤𝑖𝑗)
𝑀𝐴𝑋(𝑤𝑖𝑗) − 𝑀𝐼𝑁(𝑤𝑖𝑗)
+ 𝜖
Here we use 𝜖=MIN(wij), in order to obtain normalized edge-weights in the interval (0,1) which allows us to apply
the proximity-to-distance map dij = log(1/𝑤𝑖𝑗).
24
Computation of 𝒏𝝀
𝒕 . Let M = {S, P } be a Markov chain composed by a set of N states S={1,2,..., iN} that
correspond element by element to the set of nodes of a graph G with N nodes and E edges; P is the matrix of
transition probabilities characterizing the probability of transitioning from one state to another. Then, P(i,j) ≠ 0 if
and only if an edge exists between nodes i and j in graph G.
Let X be a random variable indicating the current state of the chain, or equivalently, the current node where the
walker is located; Y is the random variable indicating the node to which the walker will move in the next time
step, and T is the random variable indicating the target node where the walk will terminate (we assume that M is
an irreducible chain). For a given value of , and an specified target T = t, let P be the NxN matrix of transition
probabilities where elements of P are defined as
where
is a normalization factor, dij is the distance from i to j and gjt is the
geodesic distance from j to the target node t.
We make M an absorbing chain and t an absorbing state by setting all transition probabilities P(Y=jX=t,T=t) = 0
for j ≠ t and P(Y=jX=t,T=t) = 1 for j = t, and define Qt
as the N-1xN-1 matrix of transition probabilities from
non-absorbing to non-absorbing states. Then, nt
= (I- Qt
)-1 is the fundamental matrix for the absorbing chain [31],
and the elements nt
(i,j) denote the amount of time that the chain spends in the j-th non-absorbing state when the
chain is initialized in the i-th non-absorbing state. In other words, if we take P to represent the transition
probabilities for a (biased) random walker on graph G, and going from a source node i to a target node t, then
nt
(i,j) represents the number of times that the random walker starting at node i visits node j before it reaches node
t.
Transition probabilities for degenerate paths. Let π1 and π2 be any two paths going from node i to node t
through edges {i,j}, and {i,k}, respectively. The ratio between the transition probabilities Pt
ij and Pt
ik is:
𝑡
𝑃𝑖𝑗
𝑡 =
𝑃𝑖𝑘
exp (−𝜆(𝑑𝑖𝑗 + 𝑔𝑗𝑡)) exp (−𝑑𝑖𝑗)
exp(−𝜆(𝑑𝑖𝑘 + 𝑔𝑘𝑡)) exp (−𝑑𝑖𝑘)
Assume that the length of π1 and π2 is equal, so dij+gjt = dik+gkt. Then we can write:
𝑡
𝑃𝑖𝑗
𝑡 =
𝑃𝑖𝑘
exp (−𝑑𝑖𝑗)
exp (−𝑑𝑖𝑘)
25
tiijjtijZdgdtTiXjYP1)))((exp(),(jijjtijtidgdZ)))((exp( Now, let S indicate the set of edges leaving from node i along which there is a shortest path from node i to node t.
Since all edges in S lie on shortest paths, for any pair of edges {𝑖, 𝑗}, {𝑖, 𝑘} ∈ 𝑆, it must be that dij+gjt = dik+gkt.
Then, when λ→∞, we can write
exp (−𝑑𝑖𝑗)
𝑖𝑓 {𝑖, 𝑗} ∈ 𝑆
𝑡 = {
𝑃𝑖𝑗
{𝑖,𝑗′}∈𝑆
exp (−𝑑𝑖𝑗′)
∑
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
If the network is unweighted, then all dij = const. In that case, all edges in S will have a uniform transition
probability from node i.
Note that in the λ→∞ case, only transitions along shortest paths will be allowed. This means that the random walk
path lengths will be equal to shortest path lengths.
Randomized Networks. For each subject, we created a population of 500 randomized brain networks, with
preserved degree strength sequence, and preserved weight distribution, following the procedure described in [60].
The empirical networks were first binarized and then randomized by swapping pairs of connections as described
in [12], thus preserving the binary degree of each node. In order to approximate the strength sequence of the
empirical structural connectivity matrices, we used a simulated annealing algorithm that minimizes the cost
function C = ∑isi - ri, where si is the strength of node i in the empirical network and ri is the strength in the
randomized network. The cost function is minimized by randomly permuting weight assignments across edges
and probabilistically accepting the permutations that reduced the energy as the temperature parameter of the
algorithm is decreased. The annealing schedule consisted of 123 iterations and a starting temperature of t0=100,
which was scaled by 0.125 after each iteration. The result of this procedure was an average final energy of
C=0.2797±0.04, which indicates that the average strength discrepancy per node was between 0.0011 - 0.0014.
Intrinsic Connectivity Networks. We mapped the Desikan Killiany anatomical parcels used to construct
individual subject structural connectivity networks, onto the seven intrinsic connectivity networks (ICN) defined
by Yeo et al. (2011) [61]. This parcellation was derived by using a clustering algorithm to partition the cerebral
cortex of 1000 healthy subjects into networks of functionally coupled regions. The clustering procedure resulted
in the definition of seven clusters comprising systems previously described in the literature including the visual
(VIS) and somatomotor (SM) regions, dorsal (DA) and ventral (VA) attention networks, frontoparietal control
(FP), limbic (LIM) and default mode network (DMN). The mapping between the Desikan-Killiany anatomical
parcels and the seven ICNs from the ICN parcellation was obtained by extracting the vertices of the brain surface
corresponding to each anatomical region in the Desikan-Killiany atlas, and then evaluating the mode of the
vertices' assignment in the ICN parcellation
26
References
1. Laughlin, S.B., & Sejnowski, T.J. Communication in neuronal networks. Science 301, 1870-1874 (2003)
2. Hu, Y., Wang, Y., Li, D., Havlin, S., & Di, Z. Possible origin of efficient navigation in small worlds. Physical Review
Letters,106(10), 108701 (2011)
3. Kleinberg, J. M. Navigation in a small world. Nature, 406(6798), 845-845 (2000)
4. Crucitti, P., Latora, V., Marchiori, M., & Rapisarda, A. Efficiency of scale-free networks: error and attack tolerance.
Physica A: Statistical Mechanics and its Applications, 320, 622-642 (2003)
5. Kalavri, V., Simas, T., & Logothetis, D. The shortest path is not always a straight line: leveraging semi-metricity in graph
analysis. Proceedings of the VLDB Endowment, 9(9), 672-683 (2016)
6. Avena-Koenigsberger, A., Mišić, B., Hawkins, R. X., Griffa, A., Hagmann, P., Goñi, J., & Sporns, O. Path ensembles and
a tradeoff between communication efficiency and resilience in the human connectome.Brain Structure and Function, 222(1),
603-618 (2017)
7. Yan, G., Zhou, T., Hu, B., Fu, Z. Q., & Wang, B. H. Efficient routing on complex networks.Physical Review E, 73(4),
046108 (2006)
8. Yu, H., Kim, P. M., Sprecher, E., Trifonov, V., & Gerstein, M. The importance of bottlenecks in protein networks:
correlation with gene essentiality and expression dynamics. PLoS Comput Biol, 3(4), e59 (2007)
9. Tombu M, Asplund C, Dux P, Godwin D, Martin J, et al. A unified attentional bottleneck in the human brain. Proc Natl
Acad Sci USA 108: 13426–13431 (2011)
10. Dodds, P. S., Watts, D. J., & Sabel, C. F. Information exchange and the robustness of organizational
networks. Proceedings of the National Academy of Sciences,100(21), 12516-12521 (2003)
11. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., & Hwang, D. U. Complex networks: Structure and dynamics. Physics
reports, 424(4), 175-308 (2006)
12. Goñi, J., Avena-Koenigsberger, A., de Mendizabal, N. V., van den Heuvel, M. P., Betzel, R. F., & Sporns, O. Exploring
the morphospace of communication efficiency in complex networks. PLoS One, 8(3), e58070 (2013)
13. Chen, B. L., Hall, D. H., & Chklovskii, D. B. Wiring optimization can relate neuronal structure and function. Proceedings
of the National Academy of Sciences of the United States of America, 103(12), 4723-4728 (2006)
14. Barthélemy, M. Spatial networks. Physics Reports, 499(1), 1-101 (2011)
15. Bullmore, E., & Sporns, O. The economy of brain network organization. Nature Reviews Neuroscience, 13(5), 336-349
(2012)
16. Clune, J., Mouret, J. B., & Lipson, H. The evolutionary origins of modularity. In Proc. R. Soc. B (Vol. 280, No. 1755, p.
20122863). The Royal Society (2013)
17. R. Betzel, R. F., Avena-Koenigsberger, A., Goñi, J., He, Y., De Reus, M. A., Griffa, A., ... & Van Den Heuvel, M.
Generative models of the human connectome. Neuroimage, 124, 1054-1064 (2016)
18. Noh, J. D., & Rieger, H. Random walks on complex networks. .Physical review letters, 92(11), 118701 (2004)
19. Yang, S. J. Exploring complex networks by walking on them. Physical Review E, 71(1), 016107 (2005)
27
20. Tejedor, V., Bénichou, O., & Voituriez, R. Global mean first-passage times of random walks on complex
networks. Physical Review E, 80(6), 065104 (2009)
21. Adamic, L. A., Lukose, R. M., Puniyani, A. R., & Huberman, B. A. Search in power-law networks. Physical review
E, 64(4), 046135 (2001)
22. Yin, C. Y., Wang, B. H., Wang, W. X., Zhou, T., & Yang, H. J. Efficient routing on scale-free networks based on local
information. Physics letters A,351(4), 220-224 (2006)
23. van den Heuvel, M. P., Kahn, R. S., Goñi, J., & Sporns, O. High-cost, high-capacity backbone for global brain
communication. Proceedings of the National Academy of Sciences, 109(28), 11372-11377 (2012)
24. Csoma, A., Kőrösi, A., Rétvári, G., Heszberger, Z., Bíró, J., Slíz, M., ... & Gulyás, A. Routes Obey Hierarchy in Complex
Networks. Scientific Reports (2017)
25. Fronczak, A., & Fronczak, P. Biased random walks in complex networks: The role of local navigation rules. Physical
Review E, 80(1), 016107 (2009)
26. Lambiotte, R., Sinatra, R., Delvenne, J. C., Evans, T. S., Barahona, M., & Latora, V. Flow graphs: Interweaving
dynamics and structure.Physical Review E, 84(1), 017102 (2011)
27. Zlatić, V., Gabrielli, A., & Caldarelli, G. Topologically biased random walk and community finding in
networks. Physical Review E, 82(6), 066109 (2010)
28. Boguna M, Krioukov D, Claffy KC. Navigability of complex network. Nat Phys 5: 74–80 (2009)
29. Simsek O, Jensen D. Navigating networks by using homophily and degree. Proc Natl Acad Sci USA 105:12758–12762
(2008)
30. Yan X, Teng S, Lerman K, Ghosh R.Capturing the interplay of dynamics and networks through parameterizations of
Laplacian operators.PeerJ Computer Science2:e57 (2016)
31. Grinstead, C. M., & Snell, J. L. Introduction to Probability 2nd edn (Providence, RI: American Mathematical Society)
(2003)
32. Simas, T., & Rocha, L. M. Distance closures on complex networks.Network Science, 3(02), 227-268 (2015)
33. Todorov, E. Efficient computation of optimal actions. Proceedings of the national academy of sciences, 106(28), 11478-
11483 (2009)
34. Van Essen DC, Smith SM, Barch DM, Behrens TE, Yacoub E, Ugurbil K, WU-Minn HCP Consortium. The WU-Minn
Human Connectome Project: an overview. Neuroimage 80:6279 (2013)
35. M.F. Glasser, et al.The minimal preprocessing pipelines for the human connectome project NeuroImage, 80, 105-124
(2013)
36. Avena-Koenigsberger, A., Goñi, J., Betzel, R. F., van den Heuvel, M. P., Griffa, A., Hagmann, P., ... & Sporns, O. Using
Pareto optimality to explore the topology and dynamics of the human connectome. Phil. Trans. R. Soc. B, 369(1653),
20130530 (2014)
37. Latora, V., & Marchiori, M. Efficient behavior of small-world networks.Physical review letters, 87(19), 198701 (2001)
38. Humphries, M. D., & Gurney, K. Network 'small-world-ness': a quantitative method for determining canonical network
equivalence. PloS one, 3(4), e0002051 (2008)
28
39. Mišić B, Sporns O, McIntosh AR. Communication Efficiency and Congestion of Signal Traffic in Large-Scale Brain
Networks. PLoS Comput Biol 10(1): e1003427 (2014)
40. Avena-Koenigsberger A., Mišić B, Sporns O. Communication Dynamics. Nature Review Neuroscience 19, 17-33 (2018)
41. Roland, P. E., Hilgetag, C. C., & Deco, G. Cortico-cortical communication dynamics. Frontiers in systems
neuroscience, 8 (2014)
42. Fries, P. Rhythms for cognition: communication through coherence. Neuron, 88(1), 220-235 (2015)
43. Palmigiano, A., Geisel, T., Wolf, F., & Battaglia, D. Flexible information routing by transient synchrony. Nature
Neuroscience (2017)
44. Goñi, J., van den Heuvel, M. P., Avena-Koenigsberger, A., de Mendizabal, N. V., Betzel, R. F., Griffa, A., ... & Sporns,
O. Resting-brain functional connectivity predicted by analytic measures of network communication.Proceedings of the
National Academy of Sciences, 111(2), 833-838 (2014)
45. Stam, C. J., & van Straaten, E. C. Go with the flow: use of a directed phase lag index (dPLI) to characterize patterns of
phase relations in a large-scale model of brain dynamics. Neuroimage, 62(3), 1415-1428 (2012)
46. Moon, J. Y., Lee, U., Blain-Moraes, S., & Mashour, G. A. General relationship of global topology, local dynamics, and
directionality in large-scale brain networks. PLoS computational biology, 11(4), e1004225 (2015)
47. Van Wijk, B. C., Stam, C. J., & Daffertshofer, A. Comparing brain networks of different size and connectivity density
using graph theory. PloS one, 5(10), e13701 (2010)
48. Delvenne, J. C., Schaub, M. T., Yaliraki, S. N., & Barahona, M. The stability of a graph partition: A dynamics-based
framework for community detection. In Dynamics On and Of Complex Networks, Volume 2 (pp. 221-242). Springer New
York (2013)
49. Bertolero, M. A., Yeo, B. T. T., & D'Esposito, M. The Diverse Club: The Integrative Core of Complex Networks. Nature
Communications 8; 1277 (2017)
50. Bertolero, M. A., Yeo, B. T., & D'Esposito, M. The modular and integrative functional architecture of the human
brain. Proceedings of the National Academy of Sciences, 112(49), E6798-E6807. (2015)
51. Betzel, R. F., & Bassett, D. S. Multi-scale brain networks. Neuroimage (2016)
52. Daducci, A., Gerhard, S., Griffa, A., Lemkaddem, A., Cammoun, L., Gigandet, X., ... & Thiran, J. P. The connectome
mapper: an open-source processing pipeline to map connectomes with MRI. PloS one, 7(12), e48121 (2012)
53. Cammoun, L., Gigandet, X., Meskaldji, D., Thiran, J. P., Sporns, O., Do, K. Q., ... & Hagmann, P. Mapping the human
connectome at multiple scales with diffusion spectrum MRI. Journal of neuroscience methods,203(2), 386-397 (2012)
54. Wedeen, V. J., Wang, R. P., Schmahmann, J. D., Benner, T., Tseng, W. Y. I., Dai, G., ... & de Crespigny, A. J. Diffusion
spectrum magnetic resonance imaging (DSI) tractography of crossing fibers. Neuroimage, 41(4), 1267-1277 (2008)
55. Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Honey, C. J., Wedeen, V. J., & Sporns, O. Mapping the structural
core of human cerebral cortex. PLoS Biol, 6(7), e159 (2008)
56. de Reus MA, van den Heuvel MP. Simulated rich club lesioning in brain networks: a scaffold for communication and
integration? Front Hum Neurosci 8:647, pmid:25191259 (2014)
57. Fischl B, et al. Automatically parcellating the human cerebral cortex. Cereb Cortex 14:11–
22, doi:10.1093/cercor/bhg087, pmid:14654453 (2004)
29
58. Buzsa´ki G, Mizuseki K. The log-dynamic brain: how skewed distributions affect network operations. Nat Rev Neurosci
15(4):264–278 (2014)
59. Markov NT, Ercsey-Ravasz MM, Gomes AR, Lamy C, Magrou L, Vezoli J, Sallet J. A weighted and directed interareal
connectivity matrix for macaque cerebral cortex. Cereb Cortex. bhs270 (2012)
60. Mišić, B., Betzel, R. F., Nematzadeh, A., Goñi, J., Griffa, A., Hagmann, P., ... & Sporns, O. Cooperative and competitive
spreading dynamics on the human connectome. Neuron, 86(6), 1518-1529 (2015)
61. Yeo BT, et al. The organization of the human cerebral cortex estimated by intrinsic functional connectivity. J
Neurophysiol 106(3):1125–1165 (2011)
Acknowledgments. The authors thank Alessandra Griffa for providing the processed DTI data as well as
computational tools for visualization. We also thank Filip Miscevic and Bratislav Mišić for fruitful discussions.
O.S. was supported by the National Institutes of Health (R01-AT009036). P.H. was supported by the Leenaards
Foundation.
Conflict of Interest.
The authors have no competing interests.
Author Contributions. A.A-K and A.K. designed the experiments; A.A-K. performed the computational analysis;
A.A-K, A.K, X.Y and O.S. Analyzed the results; M.van den H. and P.H. provided the data; A.A-K., X.Y., A.K.
and O.S. Wrote and prepared the manuscript; all authors reviewed and edited the manuscript.
30
SUPPLEMENTAL INFORMATION
31
Fig S1. z-scored source transmission costs as a function of . For each subject's structural connectivity
𝑡𝑟𝑎𝑛𝑠) of every node was standarized with respect to the corresponding
network, the source transmission cost (𝐶⃗
𝜆
𝑡𝑟𝑎𝑛𝑠 measured on an ensemble of 500 randomized networks. z-scored values were then
distribution of 𝐶⃗
𝜆
thresholded according to a type I error α = 0.01. Red regions on the spectrum indicate nodes with significantly
𝑡𝑟𝑎𝑛𝑠, whereas blue regions on the spectrum indicate nodes with significantly low 𝐶⃗
𝑡𝑟𝑎𝑛𝑠.
high 𝐶⃗
𝜆
𝜆
32
33
Fig S2. z-scored target transmission costs as a function of . For each subject's structural connectivity
𝑡𝑟𝑎𝑛𝑠) of every node was standarized with respect to the corresponding
network, the target transmission cost (𝐶⃖
𝜆
𝑡𝑟𝑎𝑛𝑠 measured on an ensemble of 500 randomized networks. z-scored values were then
distribution of 𝐶⃖
𝜆
thresholded according to a type I error α = 0.01. Red regions on the spectrum indicate nodes with significantly
𝑡𝑟𝑎𝑛𝑠, whereas blue regions on the spectrum indicate nodes with significantly low 𝐶⃖
𝑡𝑟𝑎𝑛𝑠.
high 𝐶⃖
𝜆
𝜆
34
35
Fig S3. z-scored source informational costs as a function of . For each subject's structural connectivity
𝑖𝑛𝑓𝑜
network, the source informational cost (𝐶⃗
𝜆
) of every node was standardized with respect to the corresponding
𝑖𝑛𝑓𝑜
distribution of 𝐶⃗
𝜆
measured on an ensemble of 500 randomized networks. z-scored values were then
thresholded according to a type I error α = 0.01. Red regions on the spectrum indicate nodes with significantly
𝑖𝑛𝑓𝑜
high 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
, whereas blue regions on the spectrum indicate nodes with significantly low 𝐶⃗
𝜆
.
36
37
Fig S4. z-scored target informational costs as a function of . For each subject's structural connectivity
𝑖𝑛𝑓𝑜
network, the target informational cost (𝐶⃖
𝜆
) of every node was standarized with respect to the corresponding
𝑖𝑛𝑓𝑜
distribution of 𝐶⃖
𝜆
measured on an ensemble of 500 randomized networks. z-scored values were then
thresholded according to a type I error α = 0.01. Red regions on the spectrum indicate nodes with significantly
𝑖𝑛𝑓𝑜
high 𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
, whereas blue regions on the spectrum indicate nodes with significantly low 𝐶⃖
𝜆
.
38
39
Fig S5. z-scored difference between source and target transmission costs as a function of . For each
𝑡𝑟𝑎𝑛𝑠-
subject's structural connectivity network, the difference between the source and target transmission cost (𝐶⃗
𝜆
𝑡𝑟𝑎𝑛𝑠) of every node was standardized with respect to the corresponding distribution of 𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠 measured
𝐶⃖
𝜆
𝜆
𝜆
on an ensemble of 500 randomized networks. A one-tailed test (α = 0.01) was performed to test the hypothesis
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠> null(𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠) if 𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠>0, or 𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠< null (𝐶⃗
𝑡𝑟𝑎𝑛𝑠-
that the difference 𝐶⃗
𝜆
𝜆
𝜆
𝜆
𝜆
𝜆
𝜆
𝜆
𝜆
𝑡𝑟𝑎𝑛𝑠) if 𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠<0, where null(𝐶⃗
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠) is the mean difference obtained from the ensemble of
𝐶⃖
𝜆
𝜆
𝜆
𝜆
𝜆
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠,
randomized networks. Red regions on the spectrum indicate nodes with significantly high (>0) 𝐶⃗
𝜆
𝜆
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠.
whereas blue regions on the spectrum indicate nodes with significantly small (<0) 𝐶⃗
𝜆
𝜆
40
41
Fig S6. z-scored difference between source and target informational costs as a function of . For each
𝑖𝑛𝑓𝑜
subject's structural connectivity network, the difference between the source and target transmission cost (𝐶⃗
𝜆
-
𝑖𝑛𝑓𝑜
𝐶⃖
𝜆
𝑡𝑟𝑎𝑛𝑠-𝐶⃖
𝑡𝑟𝑎𝑛𝑠 measured
) of every node was standardized with respect to the corresponding distribution of 𝐶⃗
𝜆
𝜆
on an ensemble of 500 randomized networks. A one-tailed test (α = 0.01) was performed to test the hypothesis
𝑖𝑛𝑓𝑜
that the difference 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
> null(𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
) if 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
>0, or 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
< null(𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
) if 𝐶⃗
𝜆
-
𝑖𝑛𝑓𝑜
𝐶⃖
𝜆
𝑖𝑛𝑓𝑜
<0, where null(𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
) is the mean difference obtained from the ensemble of randomized networks.
𝑖𝑛𝑓𝑜
Red regions on the spectrum indicate nodes with significantly high (>0) 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
, whereas blue regions on the
𝑖𝑛𝑓𝑜
spectrum indicate nodes with significantly small (<0) 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
-𝐶⃖
𝜆
.
42
Replication data set (LAU).
Fig S7. A spectrum of communication processes. (a) Averages of 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 (red) and 𝐶𝜆
𝑖𝑛𝑓𝑜
(blue) across all node
pairs, as a function of . Solid red and blue lines correspond to the median across all subjects, whereas the shaded
red and blue regions denote the 95th percentile. Vertical dashed lines correspond to the values 1=e-4.19, 2=e-2.16,
𝑡𝑟𝑎𝑛𝑠‖ (red) and ‖𝐶𝜆
𝑖𝑛𝑓𝑜‖ (blue) across all node pairs. These curves are
3=e-042 and 4=,e1.31. (b) Averages of ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
computed by normalizing 𝐶𝜆
𝑖𝑛𝑓𝑜
with respect to the same cost measures computed on ensembles of
500 randomized networks (per subject). Shaded red and blue areas indicate sections of the curves ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ and
43
𝑖𝑛𝑓𝑜‖ that are smaller than 1, respectively, indicating the regions in the spectrum where the communication
‖𝐶𝜆
cost of empirical networks is smaller than the cost computed on the randomized ensembles. The dashed vertical
lines are placed at the minimum and maximum of ‖𝐶𝜆
𝑡𝑟𝑎𝑛𝑠‖ (2 and 3, respectively), and at two points near the
extremes of the spectrum ((1 and 4). (c) pairwise values of 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠(i,t) for all node pairs. (d) pairwise values of
(i,t) for all node pairs. For the panels 1=e-4.19, 2=e-2.16, 3=e-042 and 4=e1.31.
𝑖𝑛𝑓𝑜
𝐶𝜆
Fig S8. Nodal average transmission costs for four increasingly biased routing strategies. (a) Scatter plots
𝑡𝑟𝑎𝑛𝑠) and target (𝐶⃖
𝑡𝑟𝑎𝑛𝑠) during
show the transmission cost associated to each node when it acts as source (𝐶⃗
𝜆
𝜆
communication processes taking place under routing strategies generated with the values 1=e-4.19, 2=e-2.16, 3=e-
𝑖𝑛𝑓𝑜
042 and 4=e1.31. (b) Scatter plots show the transmission cost associated to each node when it acts as source (𝐶⃗
𝜆
)
44
𝑖𝑛𝑓𝑜
and target (𝐶⃖
𝜆
) during communication processes taking place under routing strategies generated with the
values 1, 2, 3 and 4. Markers in the scatter plots in (a) and (b), representing each node, are colored according
to the node's functional role according to the 7 intrinsic connectivity networks (ICN) defined by Yeo et al. (2011)
(61): Visual (VIS), Somatomotor (SM), Dorsal Attention (DA), Ventral Attention (VA), Limbic (LIM), Frontal
Parietal (FP), and Default Mode Network (DMN) (see Figure SI1 showing ICNs projected on a cortical surface).
𝑡𝑟𝑎𝑛𝑠
The size of the markers is proportional to node's strength. (c) Correlations between node strength and 𝐶⃗
𝜆
𝑖𝑛𝑓𝑜
𝑡𝑟𝑎𝑛𝑠(orange), 𝐶⃗
(red), 𝐶⃖
𝜆
𝜆
𝑖𝑛𝑓𝑜
(green) and 𝐶⃖
𝜆
(blue) as a function of . Solid lines show median correlation
across all subjects, shaded areas surrounding the lines show 95th percentile. Shaded colored areas between the
vertical dashed lines indicate regions where the correlations were not significant (p > 0.001). (d) Correlation
𝑖𝑛𝑓𝑜
𝑡𝑟𝑎𝑛𝑠 and 𝐶⃖
𝑡𝑟𝑎𝑛𝑠(red), and 𝐶⃗
between 𝐶⃗
𝜆
𝜆
𝜆
𝑖𝑛𝑓𝑜
and 𝐶⃖
𝜆
(blue), as a function of . Solid lines show medians across
all subjects and shaded areas surrounding solid lines show the 95th percentile. Shaded areas between the vertical
dashed lines indicate areas where correlation values were not significant (p > 0.001).
Fig S9. A brain region's propensity to be a costly source or target. Cortical surfaces show the difference
𝑡𝑟𝑎𝑛𝑠 − 𝐶⃖
𝑡𝑟𝑎𝑛𝑠 for routing strategies generated with
between a node's source and target transmission costs. (a) 𝐶⃗
𝜆
𝜆
45
𝑖𝑛𝑓𝑜 − 𝐶⃖
𝑖𝑛𝑓𝑜
the values 1, 2, 3 and 4. (b) 𝐶⃗
𝜆
𝜆
for routing strategies generated with the values 1, 2, 3 and 4.
Red colored areas on the cortical surfaces correspond to nodes whose source transmission/informational cost is
higher than their target transmission/informational cost. Blue colored areas correspond to nodes whose target
transmission/informational cost is higher than their source transmission/informational cost. For all panels, 1=e-
4.19, 2=e-2.16, 3=e-042 and 4=e1.31.
Fig S10. Network average values of 𝑪𝝀
𝒕𝒓𝒂𝒏𝒔 (left panel) and 𝑪𝝀
𝒊𝒏𝒇𝒐
(middle panel) as a function of (node
medians across all subjects) for 22, 55, 110, and 165 privileged nodes (corresponding to 10%, 25%, 50% and 75%
or the network's nodes) that are selected according to betweenness centrality ranking (yellow line), strength
46
ranking (purple line), shortest-path-based closeness centrality (green line), and random-walk-based closeness
centrality (blue line). For comparison purposes, we also show cost measures for randomly sampled nodes (red
line). The dotted lines show 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
, respectively, for the case in which all nodes' routing strategies are
biased (i.e. 100% privileged nodes). Right panel shows node stretch distributions for the different sets of
privileged nodes and centrality rankings. Black markers indicate the median of the distributions.
Fig S11. Communication cost trade-off within subjects. (a) Correlations between all subject's 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 across all
values of . Positive correlations are colored in red, negative correlations are colored in blue. (b) Scatter plot of
𝑡𝑟𝑎𝑛𝑠 and 𝐶𝜆
𝑖𝑛𝑓𝑜
curves, showing a trade-off between the decay of
) is r = -0.6, p < 0.001).
47
the computed areas under the normalized 𝐶𝜆
𝑡𝑟𝑎𝑛𝑠 and the growth of 𝐶𝜆
𝐶𝜆
𝑖𝑛𝑓𝑜
(correlation between A(𝐶𝜆
𝑡𝑟𝑎𝑛𝑠) and A(𝐶𝜆
𝑖𝑛𝑓𝑜
|
1210.6789 | 1 | 1210 | 2012-10-25T10:40:05 | Associative memory of phase-coded spatiotemporal patterns in leaky Integrate and Fire networks | [
"q-bio.NC"
] | We study the collective dynamics of a Leaky Integrate and Fire network in which precise relative phase relationship of spikes among neurons are stored, as attractors of the dynamics, and selectively replayed at differentctime scales. Using an STDP-based learning process, we store in the connectivity several phase-coded spike patterns, and we find that, depending on the excitability of the network, different working regimes are possible, with transient or persistent replay activity induced by a brief signal. We introduce an order parameter to evaluate the similarity between stored and recalled phase-coded pattern, and measure the storage capacity. Modulation of spiking thresholds during replay changes the frequency of the collective oscillation or the number of spikes per cycle, keeping preserved the phases relationship. This allows a coding scheme in which phase, rate and frequency are dissociable. Robustness with respect to noise and heterogeneity of neurons parameters is studied, showing that, since dynamics is a retrieval process, neurons preserve stablecprecise phase relationship among units, keeping a unique frequency of oscillation, even in noisy conditions and with heterogeneity of internal parameters of the units. | q-bio.NC | q-bio |
Associative memory of phase-coded spatiotemporal patterns in leaky Integrate and
Fire networks
Silvia Scarpetta1,2 , Ferdinando Giacco3
1 Dept. of Physics "E. R. Caianiello", University of Salerno, 84084 Fisciano (SA), IT
3 Dep. of Environmental Sciences, Second University of Naples, 81100 Caserta, Italy
2INFN Unita' di Napoli Gruppo coll. di Salerno, IT
(Dated: Received: date / Accepted: date)
We study the collective dynamics of a Leaky Integrate and Fire network in which precise relative
phase relationship of spikes among neurons are stored, as attractors of the dynamics, and selectively
replayed at different time scales. Using an STDP-based learning process, we store in the connec-
tivity several phase-coded spike patterns, and we find that, depending on the excitability of the
network, different working regimes are possible, with transient or persistent replay activity induced
by a brief signal. We introduce an order parameter to evaluate the similarity between stored and
recalled phase-coded pattern, and measure the storage capacity. Modulation of spiking thresholds
during replay changes the frequency of the collective oscillation or the number of spikes per cycle,
keeping preserved the phases relationship. This allows a coding scheme in which phase, rate and
frequency are dissociable. Robustness with respect to noise and heterogeneity of neurons parameters
is studied, showing that, since dynamics is a retrieval process, neurons preserve stable precise phase
relationship among units, keeping a unique frequency of oscillation, even in noisy conditions and
with heterogeneity of internal parameters of the units.
I.
INTRO
spikes relatively to the ongoing oscillation.
It has been hypothesized that, in many areas of the
brain, having different brain functionality, repeatable
precise spatiotemporal patterns of spikes play a crucial
role in coding and storage of information. Temporally
structured replay of spatiotemporal patterns have been
observed to occur during sleep, both in the cortex and
hippocampus [1 -- 4], and it has been hypothesized that
this replay may subserve memory consolidation. The se-
quential reactivation of hippocampal place cells, corre-
sponding to previously experienced behavioral trajecto-
ries, has been observed also in the awake state (awake
replay) [6 -- 9], namely during periods of relative immobil-
ity. Awake replay may reflect trajectories through either
the current environment or previously, spatially remote,
visited environments. A possible interpretation is that
spatiotemporal patterns, stored in the plastic synaptic
connections of hippocampus, are retrieved when a cue
activates the emergence of a stored pattern, allowing
these patterns to be replayed and then consolidated in
distributed circuits beyond the hippocampus [9]. Cross-
correlogram analysis revealed that in prefrontal cortex
the time scale of reactivation of firing patterns during
post-behavioral sleep was compressed five- to eightfold
relative to waking state [3, 5], a similar compression ef-
fect may also be seen in primary visual cortex[2]. Inter-
nally generated spatiotemporal patterns have also been
observed in the rat hippocampus during the delay pe-
riod of a memory task, showing that the emergence of
consistent pattern of activity may be a way to maintain
important information during a delay in a task [10].
Among repeating patterns of spikes a central role is
played by phase-coded patterns [11 -- 13, 15], i.e. patterns
with precise relative phases of the spikes of neurons par-
ticipating to a collective oscillation, or precise phases of
First experimental evidence of the importance of spike
phases in neural coding was observed in experiments
on theta phase precession in rat's place cells [17, 19],
showing that spike phase is correlated with rat's posi-
tion. Recently, the functional role of oscillations in the
hippocampal-entorinal cortex circuit for path-integration
has been deeply investigated [3, 16 -- 19, 28], showing that
place cells and grid cells form a map in which precise
phase relationship among units plays a central role. In
particular it has been shown[21, 23, 24] that both spatial
tuning and phase-precession properties of place cells can
arise when one has interference among oscillatory cells
with precise phase relationship and velocity-modulated
frequency.
Further evidence of phase coding comes from the ex-
periments on spike-phase coding of natural stimuli in au-
ditory and visual primary cortex [12, 13], and from ex-
periments on short-term memory of multiple objects in
prefrontal cortices of monkeys [11].
These experimental works support the hypothesis that
collective oscillations may underlie a phase dependent
neural coding and an associative memory behavior which
is able to recognize the phase coded patterns.
The importance of precise timing relationships among
neurons, which may carry information to be stored, is
supported also by the evidence that precise timing of few
milliseconds is able to change the sign of synaptic plas-
ticity. The dependence of synaptic modification on the
precise timing and order of pre- and post-synaptic spiking
has been demonstrated in a variety of neural circuits of
different species. Many experiments show that a synapse
can be potentiated or depressed depending on the precise
relative timing of the pre- and post-synaptic spikes. This
timing dependence of magnitude and sign of plasticity,
observed in several types of cortical [39, 40, 47] and hip-
pocampal [41 -- 43, 47] neurons, is usually termed Spike
Timing Dependent Plasticity (STDP).
The role of STDP has been investigated both in super-
vised learning framework [63], in unsupervised framework
in which repeating patterns are detected by downstream
neurons [14], cortical development [64],generation of se-
quences [65, 66] and polychronous activity [67], and in an
associative memory framework with binary units [68, 69].
However, this is the first time that this learning rule has
been used to make a IF network to work as associative
memory for phase-coded patterns of spike, each of which
becomes a dynamic attractor of the network. Notably,
in a phase coded pattern not only the order of activation
matters, but the precise spike timing intervals between
units.
We therefore present a possibility to build a circuit
with stable phase relationships between the spikes of a
population of IF neurons, in a robust way with respect
to noise and changes of frequency. The first important
result of the paper is the measurement of the storage
capacity of the model, i.e. the maximum number of dis-
tinct spatiotemporal patterns that can be stored and se-
lectively retrieved, since it has never been computed in a
spiking model for spatiotemporal patterns.
Several classic papers (see [51] and references therein)
have focused on storage capacity of binary model with
static binary patterns [37], and much efforts have been
done to use more biophysical models and patterns [29 --
35, 38, 50, 56, 57, 68], but, up to our knowledge, without
any calculation of the storage capacity of spatiotemporal
patterns in IF spiking models. Notably, by introducing
an order-parameter which measures the overlap between
phase coded spike trains, we are able quantitatively
measure of the overlap between the stored pattern and
the replay activity, and to compute the storage capacity
as a function of the model parameters.
Another important result is the study of the different
regimes observed by changing the excitability parameters
of the network. In particular, we find that near the re-
gion of the parameter space where the network tends to
become unresponsive and silent there is a regime in which
the network responds selectively to cue presentation with
a short transient replay of the phase-coded pattern. Dif-
ferently, in the region of higher excitability, the patterns
are replayed persistently and selectively, and eventually
with more then one spike per cycle.
The paper is organized as follows: Section II introduces
the Leaky-Integrate-and-Fire (IF) neuronal model; Sec-
tion III describes the STDP learning rule used to design
the connections; in Section IV we study the emergence
of collective dynamics and introduce an order parameter
to measure the overlap between the collective dynamics
and the stored phase coded patterns; Section V reports
on the storage capacity of the network, i.e. the maximum
number of patterns that can be stored and selectively re-
trieved in the network; the parameter space and the dif-
ferent working regions are also investigated in Section V;
2
in Section VI we study the robustness of the retrieval dy-
namics wrt noise and heterogeneity; Section VII reports
on the implication of this model in the framework of oscil-
latory interference model of path-integration; summary
and discussion are outlined in Section VIII.
II. THE MODEL
We consider a recurrent neural network with N (N − 1)
possible connections Jij, where N is the number of neural
units. The connections Jij are designed during the learn-
ing mode, when the connections change their efficacy ac-
cording to a learning rule inspired to the STDP. After
the learning stage, the connections values are frozen, and
the collective dynamics is studied. This distinction in two
stages, plastic connection in the learning mode and frozen
connections in the dynamics mode, is a useful framework
to simplify the analysis. It also finds some neurophys-
iological motivations in the effects of neuromodulators,
such as dopamine and acetylcholine [58, 59], which regu-
late excitability and plasticity.
The single neuron model is a Leaky Integrate-and-Fire
(IF) [55]. This simple choice, with few parameters for
each neuron, is suitable to study the emergence of collec-
tive dynamics and the diverse regimes of the dynamics,
instead of focusing on the complexity of the neuronal
internal structure. We use the Spike Response Model
(SRM) formulation [55, 56] of the IF model, which al-
lows us to use an event-driven programming and makes
the numerical simulations faster with respect to a differ-
ential equation formulation.
In this picture, the postsynaptic membrane potential
is given by:
hi(t) = Xj
Jij Xtj >ti
ǫ(t − tj),
(1)
where Jij are the synaptic connections, ǫ(t) describes the
response kernel to incoming spikes on neuron i, and the
sum over tj runs over all presynaptic firing times follow-
ing the last spike of neuron i. Namely, each presynaptic
spike j, with arrival time tj, is supposed to add to the
membrane potential a postsynaptic potential of the form
Jijǫ(t − tj), where
ǫ(t−tj) = K (cid:20)exp(cid:18)−
t − tj
τm (cid:19) − exp(cid:18)−
t − tj
τs (cid:19)(cid:21) Θ(t−tj)
(2)
where τm is the membrane time constant (here 10 ms),
τs is the synapse time constant (here 5 ms), Θ is the
Heaviside step function, and K is a multiplicative con-
stant chosen so that the maximum value of the kernel is
1. The sign of the synaptic connection Jij sets the sign of
the postsynaptic potential's change, so there's inhibition
for negative Jij and excitation for positive Jij .
When the membrane potential hi(t) exceeds the spik-
th, a spike is scheduled, and the membrane
ing threshold θi
potential is reset to the resting value zero. We use the
same threshold θth for all the units, except in sec VI
where different values θi
th are used and the robustness
w.r.t. the heterogeneity is studied. Clearly the spiking
threshold θth of the neurons is related to the excitabil-
ity of the network, an increase of the value of θth is also
equivalent to a decrease of K, the size of the unitary post-
synaptic potential, or, equivalently to a global decrease
in the scaling factor of synaptic connections Jij.
Numerical simulations of this dynamics are performed for
a network with P stored patterns, where connections Jij
are determined via a learning rule described in the next
paragraph. We found that a few number of spikes, given
a in proper time order, are able to selectively induce the
emergence of a persistent collective spatiotemporal pat-
tern, which replays one of the stored pattern (see sec IV).
III. DESIGNING THE CONNECTIONS OF THE
NETWORK
In a learning model previously introduced in [49 -- 51],
the average change in the connection Jij, occurring in
the time interval [−tlearn, 0] due to periodic spike trains
of period T, with tlearn >> T , was formulated as follows:
δJij =
T
tlearn
0
Z
dt
0
Z
−tlearn
−tlearn
dt′ xi(t)A(t − t′)xj (t′) (3)
where T /tlearn is a normalization factor, xj (t) is the ac-
tivity of the pre-synaptic neuron at time t, and xi(t) the
activity of the post-synaptic one. It means that the prob-
ability for unit i to have a spike in the interval (t, t + ∆t)
is proportional to xi(t)∆t in the limit ∆t → 0. The learn-
ing window A(τ = t − t′) is the measure of the strength
of synaptic change when a time delay τ occurs between
pre and post-synaptic activity. To model the experimen-
tal results of STDP in hippocampal neurons, the learn-
ing window A(τ ) should be an asymmetric function of
τ , mainly positive (LTP) for τ > 0 and mainly negative
(LTD) for τ < 0.
Equation (3) holds for activity pattern x(t) which rep-
resents instantaneous firing rate, and is suitable to use
in analog rate models [49 -- 53] and spin network models
[36, 68]. Differently, here, being interested in spiking
neurons, the patterns to be stored are defined as precise
periodic sequence of spikes, i.e. spike-phase coded pat-
terns. Namely, activity of the neuron j is a spike train
at times tµ
j ,
xµ
j (t) = Xn
δ(t − (tµ
j + nT µ)),
(4)
where tµ
j + nT µ is the set of spikes times of unit j in the
pattern µ with period T µ, and frequency νµ = 1/T µ.
Therefore, following Eq.(3), the change in the connec-
tions Jij due to the learning of the pattern µ when the
time duration of the learning process tlearn is longer then
a single period T µ, is simply given by
3
J µ
ij =
∞
Xn=−∞
A(tµ
j − tµ
i + nT µ).
(5)
The window A(τ ), shown in Fig. 1, is given by
A(τ ) = (cid:26) ape−τ /Tp − aDe−ητ /Tp
apeητ /TD − aDeτ /TD
for
for
τ > 0
τ < 0,
(6)
with the same parameters used in [54] to fit the exper-
imental data of [41], namely ap = γ [1/Tp + η/TD]−1,
aD = γ [η/Tp + 1/TD]−1, with Tp = 10.2 ms, TD = 28.6
ms, η = 4, γ = 0.42. This function satisfies the balance
−∞ A(τ )dτ = 0. Notably, when A(τ ) is used
in eq. 5 to learn phase-coded patterns with uniformly
condition R ∞
distributed phases, then the property R A(τ )dt = 0 as-
citation (1/N )Pi,Jij >0 Jij and the summed inhibition
(1/N )Pi,Jij <0 Jij are equal in the thermodynamic limit,
and therefore it assures a balance between excitation and
inhibition.
sures that in the connection matrix the summed ex-
Writing Eq. (3-5), implicitly we have assumed that,
with periodic phase-coded spike trains used to induce
plasticity, the effects of all separate spike pairs sum lin-
early, each weighted by the same STDP window reported
in Fig. 1. Timing-dependent learning curves as the one
reported in Fig. 1 are indeed typically measured by giv-
ing an order of 100 pairs of spikes repeatly, with fixed
phase relationship, and fixed frequency in a proper range.
However, in different situations, for instance if the fre-
quency is too low or to high [47], or in case of few spike
pairs [27], the timing dependence of plasticity is not well
described by the bidirectional window used here, and a
more detailed model is needed to account for integration
of spike pairs when arbitrary trains are used (see [44, 45]
and references therein).
FIG. 1: a) Plot of the learning window A(τ ) used in the
learning rule (see Eqs. (3), (5), (6)) to model STDP effects.
The parameters of the function A(τ ) (Eq. (6)) are determined
by fitting the experimental data reported in [41].
The spikes patterns used in this work are periodic spa-
tiotemporal sequences, made up of one spike per cycle
and each of which has a phase φµ
j randomly chosen from
a uniform distribution in [0, 2π). In each pattern, infor-
mation is coded in the precise time delay between spikes
of unit i and unit j, which corresponds to a precise phase
relationship among units i and j. A spatiotemporal pat-
tern represented in this way is often called phase coded
pattern. Pattern's information is coded in the spiking
phases which, in turn, shape the synaptic connectivity
responsible of the emerging dynamics and the memory
formation.
The set of timing of spikes of unit j can be defined as
j + nT µ = (φµ
tµ
j )/(2πνµ) + n/νµ, where νµ is the oscil-
lation frequency of the neurons. Thus, each pattern µ
is represented through the frequency νµ and the specific
phases of spike φµ
j of the neurons j = 1, .., N . The change
in the connection Jij provided by the learning of pattern
µ is given by
A(tµ
j −tµ
J µ
ij = Xn
i +nT µ) = Xn
φµ
2πνµ +n/νµ).
i
(7)
When multiple phase coded patterns are stored, the
learned connections are simply the sum of the contri-
butions from individual patterns, namely
φµ
j
2πνµ −
A(
Jij =
P
Xµ=1
J µ
ij .
(8)
Note that ring-like topology with strong unidirectional
connections is formed only in the case P=1, when a sin-
gle pattern is stored. When multiple patterns are stored
in the same connectivity, with phases of one pattern un-
correlated with the others, bidirectional connections are
possible, and the more the stored patterns, the less the
ring-like is the connectivity. Even in the cases when the
connectivity is not ring-like the network is still able to
retrieve each of the P stored patterns in a proper range
of threshold values (see storage capacity in Sec V).
IV. EMERGING OF COLLECTIVE PATTERNS
IN THE NEURAL DYNAMICS OF THE
NETWORK
We study a recurrent network with N leaky Integrate
and Fire units, with connections fixed to the values cal-
culated in Eqs. (7,8) for different values of P. The results
show that, within a well specified range of parameters,
our IF network is able to work as an associative memory
for spike-phase patterns.
In order to check if the network is able to retrieve se-
lectively each of the stored patterns, we give an initial
signal, made up of M ≪ N spikes, taken from the stored
pattern µ, and we check if this initial short cue is able
to selectively trigger a collective sustained activity that
is the replay of the same stored pattern µ, i.e. checking
if the sustained activity has spikes aligned to the phases
φµ
i of pattern µ.
4
i = Tstimφµ
An example of successful selective retrieval process is
shown in Fig. 2 where, depending on the partial cue
presented to the network, a different collective activity
emerges with the phases of the firing neurons which re-
semble one or another of the stored patterns.
In this work the cue is a stimulation with M spikes, with
M = N/10, at times tµ
i , 0 < i < M , with
Tstim = 50ms. In the example shown in Fig. 2 the short
stimulation (which lasts less then 5 ms, shown in pink
in all the figures) has the effect to selectively trigger the
sustained replay of pattern µ.
Note that the retrieval dynamics has the same phase re-
lationship among units than the stored pattern, but the
replay may happen on a time scale different from the scale
used to store the pattern, and the collective spontaneous
dynamics is a time compressed (or dilated) replay of the
stored pattern. Indeed, the period of the collective pe-
riodic pattern which emerges during retrieval stage may
be different then the period of the periodic pattern used
in the learning stage. In the example of Fig. 2 the time
scale of the retrieval dynamics (Fig. 2c,d) is faster then
the time scale used to learn the patterns (Fig. 2a,b). In
the following we will study the factors affecting the time
scale during retrieval, given the time scale of the pattern
used during learning.
Clearly, regions of the parameter space in which the net-
work is unable to retrieve selectively the patterns also
exist. In these regions the retrieval dynamics may cor-
respond to a mixture of patterns or a spurious state, i.e.
a state which is not correlated with any of the stored
patterns because the number of stored patterns exceeded
the storage capacity of the network. As discussed below,
the storage capacity, defined as the maximum number of
encoded and successfully retrieved patterns, depends on
the frequency used during the learning stage (which af-
fects connectivity), and on the spiking threshold of the
units (which affects excitability and network dynamics).
Example of failure are shown in Fig.3. In Fig. 3b the
network has too low excitability and the response is not
persistent, while in Fig. 3a the emerging dynamics is not
correlated with any of the stored patterns.
To measure quantitatively the success of the retrieval, in
analogy with the Hopfield model, we introduce an order
parameter, which estimates the overlap between the net-
work collective activity during the spontaneous dynamics
and the stored phase-coded pattern. This quantity is 1
when the phases φj of neurons j coincides with the stored
phases φµ
j , and is close to zero when the phases are un-
correlated with the stored ones. Therefore, we consider
the following dot product mµ(t) =< ξ(t)ξµ > where ξµ
is the vector having components eiφµ
j , namely:
e−i2πt∗
j /T ∗
eiφµ
(9)
j(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where t∗
is the spike timing of neuron j during the
j
spontaneous dynamics, and T ∗ is an estimation of the
mµ(t) =
1
N Xj=1,...,N
t−T ∗<t∗
j
<t
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
5
a)
b)
a)
b)
FIG. 3: Example of neural response in two case of failure
of retrieval. A spurious state emerge in (a), while a short
transient response emerges in the case shown in (b). N =
3000 and ν µ = 3Hz as in previous picture, while the values
θth and P are θth = 10, P = 5 in (a) and θth = 95, P = 5 in
(b). The dynamics emerging after a short train of M = N/10
spikes with phases equal to the stored pattern (pattern shown
in Fig 2a), is not a self-sustained retrieval of the pattern. For
clarity, the raster plot of only 50 (randomly chosen) units are
1 of
shown, sorted according to increasing value of phase φi
the stored pattern.
invariant by a simple change in time scale. This is a
suitable choice especially when the replay of a spatio
temporal pattern has to be detected independently from
the compression of the time scale. Note that if we have
a spike train that is not periodic, we cannot define the
period, however we can define the order parameter (9)
looking at the time-window T ∗ which maximize the
order parameter. This can be useful in the case when
one looks for a short replay hidden in a not-periodic
spike train, such in many experimental situations.
The value of mµ(t) after a transient converges to a stable
value which is close to one when pattern µ is retrieved
(for example in Fig. 2c at large times mµ=1 = 1, and
mµ=2 = 0.01) while mµ(t) is of order ≃ 1/√N for all µ
in the case of failure of retrieval. Two further cases of
in Fig. 3a mµ(t) after the transient
failure can occur:
has values in the range 0.01 − 0.02 for all µ because the
emerging dynamics is a spurious state not correlated
with any of the stored phase-patterns, while mµ(t) is
zero in Fig. 3b since the network becomes silent.
In the following, the storage capacity of the model is ana-
lyzed considering the maximum number of patterns that
the network is able to store and selectively recall. In par-
ticular we investigate the role of two model parameters:
the frequency of the stored patterns νµ, and the spik-
ing threshold θth affecting the excitability of the network.
V. STORAGE CAPACITY
Numerical
the
IF network with
N = 3000 neurons were performed by systematically
simulations of
c)
d)
FIG. 2: Examples of selective successful retrieval (c,d) of two
stored patterns (a,b). The raster plot of 50 units (randomly
chosen) are shown sorted on the vertical axis according to in-
creasing values of phase φ1
i of the first stored pattern µ = 1.
The network has N = 3000 IF neurons, Θth = 70 and connec-
tions given by Eq. (7,8) with P = 5 stored patterns at ν µ = 3
Hz. Two of the stored patterns used during the learning mode
are shown in a,b. The dynamics emerging after a short train
of M = N/10 spikes with phases similar to the pattern shown
in a,b, is shown in c,d respectively. The dynamics of the net-
work, after a transient, is periodic of period T . The spikes
which belong to the trigger are shown in pink in (c,d), the
other different colors represent the value of ti/T mod 4, where
ti is the time of the spike of the unit i during the emerging
spontaneous dynamics. Figure c shows that when the network
dynamic is stimulated by a partial cue of pattern µ = 1, the
neurons oscillate with phase alignments resembling pattern
µ = 1, but at different frequency. Otherwise, in d, when the
partial cue is taken from pattern µ = 2, the neurons phase
relationships, even if periodic, are uncorrelated with pattern
µ = 1, and recall the phase of pattern µ = 2.
period of the collective spontaneous periodic dynamics.
The overlap in Eq.
(9) is equal to 1 when the phase-
coded pattern is perfectly retrieved (i.e. same sequence
and phase relationships among spikes, even though on
a different time scale), while is of order ≃ 1/√N when
phases of spikes are uncorrelated to the stored phases.
The order parameter mµ allow us to measure the
network storage capacity in the space of parameters θth
and νµ.
Note that the value of mµ(t) between two periodic spike
trains measures the similarity in the sequence of spiking
neurons and in the phase lag between the spikes, being
changing the value of the spiking threshold θth, the
connections Jij , and for different number of patterns
P and frequency νµ. Here we propose results for a
unique value of the spiking threshold θth for all neurons,
however the behavior is also robust with respect to a
variability in the threshold values among neurons, as
reported in the next Section.
Network storage capacity is defined as αc = Pmax/N ,
where N is the number of neurons and Pmax is the
maximum number of patterns that can be stored and
successfully retrieved with an overlap mµ larger than a
certain value, which measures the degree of similarity.
Given that in our simulations the overlap mµ(t) at
large times has mostly two possible values, close to one
(success) or close to 1/√N (failure), we fixed the desired
similarity value to 0.5, since the whole storage capacity
analysis is very robust with respect to this parameter
(since the transition between low values and high values
of mµ(t) as a function of P is sharp).
Patterns with random phases were extracted and used to
define the network connections Jij with the rule Eq. (8).
After the stimulation with a short train of M = N/10
spikes taken at times ti
from the first pattern, the
dynamics is simulated and the overlap defined in Eq.
(9) with µ = 1 is evaluated at large times. If the overlap
mµ=1(t), averaged over 50 runs, is greater than 0.5 at
time t > ¯t (where ¯t = 600 ms for all the simulations),
then we consider the retrieval successful for that pattern.
The maximum value of P, for which the network is able
to successfully replay each of the stored patterns, defines
the storage capacity of the network.
The storage capacity as a function of the spiking thresh-
old θth and storing frequency νµ is reported in Fig. 4a,
where Pmax is shown in a color-coded legend. The largest
capacity is achieved when the frequency of the stored
patterns during learning is νµ ≃ 8 Hz and the spiking
threshold of the units during retrieval is θth ≃ 130,
which provides a capacity αmax = Pmax/N = 0.016.
In Fig.
4b we show the storage capacity Pmax as a
function of the frequency νµ once fixed the value of the
threshold θth to the optimal value, corresponding to
highest capacity for each frequency.
The optimal storing frequencies and threshold values
depend on the time constants of the model, such as the
τs, τm of the IF units and the temporal shape of the
learning kernel A(τ ), whereas different shapes of A(τ )
may subserve to different storing frequency ranges.
In
this work τs, τm and A(τ ) are set to the values described
in Sec.
II, and the emergent collective dynamics is
studied as a function of the other network parameters.
Indeed, Fig. 4b shows that for the learning kernel A(τ )
used here, there is peak in the storage capacity around 8
Hz, in the range 2Hz-20Hz. Figure 4a also proves that,
for each stored frequency, a large interval of spiking
threshold values θth exists for which the network is
still able to work properly as associative memory for
phase-coded patterns.
The associative memory properties as a function of
d
l
o
h
s
e
r
h
t
g
n
i
r
i
F
250
200
150
100
50
a)
Frequency of the stored patterns (Hz)
30
10
20
6
40
40
30
20
10
0
x
a
m
P
55
50
45
40
35
30
25
20
15
10
0
b)
10
40
Frequency of the stored patterns (Hz)
20
30
FIG. 4: (a) Storage capacity in a network of N = 3000 units,
as a function of the spiking threshold θth and oscillation fre-
quency ν µ of stored patterns. The maximum number of pat-
terns successful retrievable Pmax is shown in color-coded leg-
end, the value grows from Pmax = 0 (dark blue) to Pmax = 50
(strong red). (b) The storage capacity Pmax as a function of
the frequency of stored patterns, once fixed the threshold θth
to the optimal value for each frequency.
the spiking threshold are reported in Fig.
5, when
the oscillation frequency of the patterns stored during
learning is νµ = 3Hz. The region marked in green in
Fig. 5 corresponds to cases in which the retrieval is
successful and the cue is able to selectively activate
the self-sustained replay of the stored pattern (with an
order parameter mµ larger than 0.5). When spiking
threshold changes in the range 10 < θth < 90 the
storage capacity changes between Pmax/N = 1/3000
and Pmax/N = 29/3000.
Outside the green region the number of patterns exceeds
the storage capacity and the retrieval
fails. There
are two possible reasons for this behavior. At low
threshold, when the number P of patterns exceeds the
storage capacity Pmax, the network responds with a
self-sustained activity that is not correlated with any of
the stored patterns, i.e. a spurious state. In this regime,
marked with red color in Fig. 5a, the order parameter
mµ(t) is of order 1/√N for all the stored patterns (see
also raster plot in Fig.3a). On the contrary, in the high
θth regime, the network tends to become silent and
unresponsive.
Indeed, in the region marked with blue
color, the network responds to the initial cue stimulation
with a short transient and then became silent. In this
case the value of mµ(t) is zero because there is no
self-sustained activity at time t > ¯t, meaning that the
stored attractors become unstable when θth is too high
Capacity of the network for patterns of 3 Hz frequency
30
25
20
15
10
5
s
n
r
e
t
t
a
p
d
e
r
o
t
s
f
o
r
e
b
m
u
N
a)
20
40
1
0,8
0,6
0,4
0,2
0
80
100
)
n
>
S
(
P
t
o
t
b)
60
80
Firig Threshold
100
120
140
n = 0
n = N/2
n = N
120
140
Firing Threshold
160
180
200
FIG. 5: a) Storage capacity at ν µ = 3 Hz: the region of suc-
cessful retrieval as a function of spiking threshold and number
of patterns is marked in green. The region with persistent ac-
tivity not correlated with any of the stored pattern is marked
in red (spurious states), and the region in which the network
responds with only a short transient and then becomes silent
is marked in blue (see examples in fig 3). b) The probability
that the size Stot of the network response, measured as the
number of the spikes that follow the cue stimulation, is larger
than n, with n = 0, N/2, N , is shown as a function of spiking
threshold θth, in a network with ν µ = 3Hz and P = 1. As
always in this paper the number of units is N = 3000. The
figure shows that near θcrit
th ≃ 90 there's a transition from a
region of persistent replay to a region of silence.
th = 90 .
(see raster plot in Fig.3b).
For values of the threshold greater than θcrit
th , indepen-
dently from P, the network activity is never persistent,
as reported in Fig. 5a where θcrit
At thresholds close to this critical value the network
responds with a transient activity that is a short replay
of the stored pattern, but not a persistent replay. The
size Stot of the network response, measured as the
number of spikes that follow the cue stimulation,
is
reported in Fig. 5b as a function of θth, for a network
with νµ = 3 Hz, N = 3000 and P = 1.
In the following we investigate the replay activity in
the region with successful retrieval. We focused on the
dependence of the frequency of collective oscillations
during replay on the model parameters. Fig. 6a shows
the collective frequency of replay as a function of the
frequency νµ of the patterns stored in the learning stage
with N=3000. The red dots in the figure refer to the
frequencies of oscillations observed during retrieval at
the optimal spiking threshold (where the maximum
storage capacity occurs), while the bar indicates the
available range of
replay, accessible
through a change in the spiking threshold. Important to
frequencies of
)
z
H
(
y
c
n
e
u
q
e
r
f
n
o
i
t
a
l
l
i
c
s
o
t
u
p
t
u
O
30
20
10
0
0
a)
5
10
15
20
25
30
35
Frequency of the stored patterns (Hz)
7
Patt. Freq. 2 Hz
Patt. Freq. 3 Hz
Patt. Freq. 4 Hz
Patt. Freq. 10 Hz
Patt. Freq. 20 Hz
Patt. Freq. 30 Hz
35
30
25
20
15
10
5
)
z
H
(
y
c
n
e
u
q
e
r
f
n
o
i
t
a
l
l
i
c
s
o
t
u
p
t
u
O
40
b)
0
0
50
100
150
Firing threshold
200
e
l
c
y
c
r
e
p
e
k
i
p
s
f
o
r
e
b
m
u
N
18
16
14
12
10
8
6
4
2
0
c)
Patt. Freq. 2 Hz
Patt. Freq. 3 Hz
Patt. Freq. 4 Hz
Patt. Freq. 10 Hz
Patt. Freq. 20 Hz
Patt. Freq. 30 Hz
0
20
40
Firing threshold
60
80
100
d)
FIG. 6:
(a) Frequency of the collective dynamics during
replay as a function of the frequency of stored patterns in the
network with N=3000 units. Dots refer to replay frequency
observed at optimal spiking threshold. The bars refers to
the range of frequency available through changes in spiking
threshold. (b) Frequencies of the collective dynamics during
replay as a function of the spiking threshold and for different
stored frequencies (see colors legend). Pattern is replayed on
a time scale which becomes faster if we decrease threshold θth
for the most of the frequencies ν µ. The dependence is much
stronger for ν µ ≤ 4 Hz. N=3000. (c) The number of spikes
per cycle as a function of the spiking threshold θth in networks
with different frequency ν µ of stored pattern. (d) Frequency
of the dynamics during replay as a function of ratio between
spiking threshold and network size N. Red dots are results for
a network with N=10000 units, while blue squares are results
for a network with N=1000 units. Size of the symbols refers
to the stored frequency, small symbols (on the top of the
picture) correspond to stored frequency ν µ = 10Hz, medium
size symbols correspond to stored frequency ν µ = 3Hz, and
large symbols (bottom) correspond to stored frequency ν µ =
1Hz.
note is that, in most of the cases, the frequency of the
stored pattern and the collective replay frequency do
not coincide, since the pattern is replayed compressed
(or dilated) in time, on a time scale dependent on the
network parameters. We observe that for the chosen
parameters τm, τs of the network, and the given shape
of A(τ ), the replay occurs on a compressed time scale
for all stored patterns of frequency lower then 25 Hz,
while the two time scale coincide when νµ ≃ 25 Hz. The
dependence of the frequency of the collective oscillations
on the spiking threshold is shown in Fig.
6b. This
dependence is weak for stored frequencies higher than
10 Hz. Besides, for low stored frequencies (1-4 Hz) the
frequency of the replay is very sensitive to the threshold
value, changing from 6 Hz at high spiking threshold to
30 Hz at low threshold.
We also investigate the frequency's dependence on
network size N .
In fig. 6.d red dots are results for a
network with N = 10000 units, while blue squares are
results for a network with N = 1000 units.
If what
counts is only the time lag between the single units
consecutive in the sequence, then one expects that result
with 1Hz stored at N = 10000 would be similar to 10Hz
stored at N = 1000, and this is not the case. We see
that when we use a storage frequency equal to 10 Hz
(which corresponds to different time lag between cells
depending on N), then the oscillation frequency during
replay is around 30Hz in both networks (both N = 1000,
and N = 10000), while, on the other hand, if we have a
storage frequency equal to 1 Hz the oscillation frequency
during replay may span a large range (5Hz -25Hz) in
both networks. Fig. 6.d also shows that frequency of
replay depends on the ratio between spiking threshold
and network size N, and that the high sensitivity on
spiking threshold value holds, when stored frequency is
low (1-4 Hz), also at different values of the network size.
This open the possibility to govern the oscillation
frequency of the collective replay activity via neuromod-
ulators which change the excitability and therefore the
spiking threshold of the neurons. Since in our model
(see Eq. 1,2) a change of the threshold is equivalent to
a change in the scale factor of all synaptic connections,
a similar effect might be achieved also by simply driving
the cells more due to increased synaptic input. Impor-
tantly, the sensitivity of collective oscillation frequency
on spiking threshold is not a sensitivity of the single unit
but of the collective behavior, since, as discussed in Sec.
VI, if we change the spiking threshold of few units the
collective rhythm is still unique for the whole population.
The replay frequency depends on the average threshold
among units, but all the units have the same oscillation
frequency during replay.
Moreover, for networks with νµ ≥ 10 Hz, whose replay
frequency does not considerably change with spiking
threshold, the replay dynamics is still affected by the
spiking threshold.
Indeed, in this case, the number of
spikes per cycle increases with lowering of the spiking
threshold. An example is reported in Fig. 7. The raster
plots show the same pattern replayed in three networks
having different values of the spiking threshold θth: a
burst of activity takes place within each cycle, with
phases aligned with the pattern, with a number of spikes
per cycle dependent on the value of θth. This behavior
is summarized in Fig. 6c where the number of spikes
per cycle is reported as a function of spiking threshold,
at different values of stored frequencies. Therefore, by
lowering the spiking threshold the replay activity occurs
with more than one spike per cycle, or on a faster time
scale (see Fig. 6b,c).
The behavior of the output oscillation frequency
suggests that a parameters region exists where the net-
work always responds with one spike per cycle. In this
region an increase of the excitability produces a growth
of the frequency of oscillation up to a plateau value.
Differently,
for higher excitability the frequency does
8
a)
b)
c)
FIG. 7: Modulation of spiking threshold changes the number
of spikes per cycle, keeping preserved the phase relationship
among units. Recall of the pattern µ = 1 for networks of
N = 3000 units, ν µ = 20 Hz and different values of spiking
threshold θth = 80, 65, 40 is shown respectively in a,b,c. De-
pending on the value of the spiking threshold θth, the phase-
coded pattern is replayed with a different number of spikes per
cycle. Spike of the cue stimulation are shown in pink, while
the response of the network in black. For clarity, the raster
plot of only 50 (randomly chosen) units are shown, sorted ac-
1 of the stored pattern.
cording to increasing value of phase φi
not increase while the number of spikes per cycle grows.
This means that the different frequencies, in addition
to the information coded in the phase relationship, can
code other information in relationship with the level
of spiking threshold: at high frequency the threshold
changes the number of spikes per cycle, while at low
frequency the threshold changes the frequency of the
collective oscillations during the replay.
This open the possibility to have a coding scheme in
which, while the phases encode pattern's information,
a change in frequency or a change in rate in each cycle
represents the strength and saliency of the retrieval or
it may encode another variable [60]. The recall of the
same phase-coded pattern with a different number of
spikes per cycle is particularly interesting at the light of
recent observations of Huxter et al. [20] in hippocampal
place cells, showing the occurrence of the same phases
with different rates. The authors prove that the phase
of firing and firing rate are dissociable and can represent
two independent variables, e.g.
location
within the place field and its speed of movement through
the field.
Notably, the recall of the same phase coded pattern with
the animal
different frequencies of oscillation is also relevant and
accords well with the need to have stable precise phase
relationship among cells with frequency of oscillation
modulated by parameters such as the speed of the
animal [23, 24].
The value of the frequency of collective activity during
the replay clearly is related not only to the threshold and
the stored frequency, but also to the shape of the learning
window A(τ ) and on the two characteristic times of the
model τs, τm. A systematic study of the dependence of
the replay time scale on the shape of STDP and the char-
acteristic times of the neuron model has not yet done in
a spiking model, however a dependence on the asymme-
try of A(t) has been analytically found in a simple model
with analog neurons and a single characteristic time [49].
VI. EFFECTS OF NOISE AND ROBUSTNESS
OF COLLECTIVE OSCILLATION FREQUENCY
AND PHASE RELATIONSHIPS
1 , ..., φµ
N with φµ
N , with Sµ
While in the Hopfield model the patterns are static,
and information is coded in a binary pattern Sµ =
Sµ
1 , . . . , Sµ
i ∈ {±1}, here, in this study, the
patterns are time dependent, and information is coded
in the phase pattern φµ = φµ
i ∈ [0, 2π],
where the value φµ
i /(2πνµ) represents the time shift of
the spike of unit i with respect to the collective rhythm,
i.e the time delay among units. However, as for the Hop-
field model, the patterns stored in the network are at-
tractors of the dynamics, when P do not exceeds storage
capacity, and the dynamics during the retrieval is robust
with respect to noise. We firstly check robustness w.r.t.
input noise, i.e when a Poissonian noise ηi(t) is added to
the postsynaptic potential hi(t) given in Eq.1 . The total
postsynaptic potential of each neuron i is then given by
hi(t) = ηi(t) +Xj
Jij Xtj >ti
ǫ(t − tj)
where ηi(t) is modelled as
ηi(t) = Jnoise Xtnoise>ti
ǫ(t − tnoise).
(10)
(11)
The times tnoise are randomly extracted for each neuron i,
and Jnoise are random strengths, extracted independently
for each neuron i and time tnoise. The intervals between
times tnoise are extracted from a Poissonian distribution
P (δt) ∝ e−δt/(N τnoise), while the strength Jnoise is ex-
tracted from a Gaussian distribution with mean ¯Jnoise
and standard deviation σ(Jnoise).
The network dynamics during the retrieval of a pattern
in presence of noise is shown in Fig. 8 with different levels
of noise (τnoise = 10ms, ¯J = 0 and σ(Jnoise) = 0, 10, 20, 30
in a,b,c,d respectively). Results show that when the noise
is not able to move the dynamics out of the basin of at-
traction, the errors do not sum up, and the phase rela-
tionship is preserved over time (see Fig. 8a,b,c). If the
9
FIG. 8:
(a,b,c,d) Robustness wrt noise. Raster plots show
that, when the pattern retrieval is triggered, network's spikes
continue to have phase alignments resembling the pattern
even in presence of noise. Errors do not sum up until the
system is in the basin of attraction of the phase-coded pat-
tern, as in a,b,c. Different levels of noise are used in a,b,c,d
(σ(Jnoise) = 0, 10, 20, 30 respectively), and pattern is trig-
gered with M = N/10 as in previous cases. Only in d the
level of noise is too high and the system goes out of the basin
of attraction. (e,f) For comparison, the dynamics, when the
retrieval is not triggered (M = 0), is shown in subplot e,f in
presence of the same noise used in c,d. Figure e shows that
the noise used in c usually affects strongly the dynamics of
the network, however if the collective oscillation is retrieved
the system is robust wrt noise. Thresholds in all figures are
θth = 80, N = 3000, and synaptic connections Jij are build
learning P = 2 phase-patterns at ωµ = 3Hz.
input noise is very high, as in the example of Fig. 8d,
the dynamics moves out of the basin of attraction.
In order to see the effects of input noise level used in Fig
8cd, we report in Fig. 8ef the network dynamics when
the pattern retrieval is not initiated (M = 0). In partic-
ular, Fig. 8e shows that the noise level used in Fig. 8c is
strong enough to generate spontaneous random activity
in absence of the initial triggering, but is not sufficient
to destroy the attractive dynamics during a successful
retrieval. As in the Hopfield model, errors do not sum
up and the dynamics spontaneously goes back to the re-
trieved phase-coded pattern for all the perturbations that
leave the system inside the basin of attraction.
Lastly, the robustness of retrieval w.r.t. heterogeneity
of the spiking thresholds is investigated. This analysis
can be carried out by using a different value θi
th of spiking
threshold for each neuron i:
θi
th = (1 + zζi)θth
(12)
where ζi is a random number extracted from a uniform
distribution in [−1, 1], and z is the degree of heterogene-
ity. Even with high degree of heterogeneity, the emer-
gence of the retrieval collective dynamics forces all neu-
rons to have exactly the same frequency of oscillation
and to keep a precise phase relationship, in a very robust
manner. Figure 9a,b shows the dynamics with threshold
heterogeneity z = 0.2, 0.5 respectively, while the remain-
ing parameters are set to the values of Fig. 8a.
The above analysis shows that a unique collective fre-
quency emerges, which is the frequency corresponding
to the mean value of the θi
th. This is also evident by
comparing Fig. 9a,b with Fig. 8a. At z = 0.5 it can
be seen an additional small phase shift proportional to
the value of θ of each neuron, but collective activity is
preserved. Clearly if z is too high and threshold values
are distributed out of the region of successful retrieval
the network is unable to retrieve the pattern and failure
happens, as already discussed in Sec. V.
a)
b)
FIG. 9: Robustness wrt heterogeneity of spiking threshold
values. Raster plots show that, when the pattern retrieval is
triggered, units participate to the network collective oscilla-
tion, showing all the same frequency and phase alignments
resembling the pattern, even in presence of threshold values
heterogeneity among units. Spiking thresholds of neuron i
are distributed according to θi
th = θth(1 + zζ) with average
θth = 80 and z = 0.2, 0.5 in (a,b) respectively. All other pa-
rameters are as in Fig.8a ( N=3000, M=N/10, and synaptic
connections Jij are build learning P = 2 phase-patterns at
ωµ = 3Hz).
VII. RELATIONSHIP OF THIS MODEL WITH
THEORIES OF PATH-INTEGRATION
Recently,
system and
path-integration
the
hippocampal-entorinal cortex circuit have been deeply
investigated [3, 16, 17, 19, 21, 23, 28], showing that
place cells and grid cells form a map in which precise
phase relationship among units play a central role to
generate the spatial tuning. A number of models of
the spatial firing properties of place and grid cells were
offered. Generally two main categories are distinguished:
models which focus on continuous attractor mechanisms
and models which use interference between oscillators
at dynamically modulated frequencies; however, deeper
computational principles may exist
that unify the
different cases of neural integration [26].
In oscillatory
interference models [19, 21, 23, 24, 62] (see also [25] for
a review) the total synaptic input to a neuron (such
as a grid cell or a place cell) is a weighted sum of
10
the activities of n oscillatory inputs, whose oscillation
frequency is modulated by the rat velocity and head
direction. Grid and place cells, according to these mod-
els, derive their temporal and spatial properties simply
by detecting synchrony among such velocity-modulated
oscillatory inputs. The oscillatory interference theory is
one possible hypothesized mechanism of path integration
and it has not been conclusively accepted or rejected. It
is supported by recent studies suggesting that the pre-
dicted velocity-modulated oscillators exist as theta cells
(interneurons found throughout the septo-hippocampal
circuit) whose inter-burst frequency shows a cosine
modulation by running direction and a linear increase
with running velocity [24].
Such velocity-modulated oscillatory input was hypoth-
esized to come from single oscillatory neurons [19], or
networks such as subcortical "ring attractors" generat-
ing velocity-modulated theta oscillations[23, 24, 62].
The properties of modulation and stability of
fre-
quency, and stability of multiple phase relationships,
make our circuit a possible mechanism to build the
velocity-modulated oscillators of the oscillatory interfer-
ence theory. Indeed our circuit has a collective oscilla-
tion frequency, which depends on the frequency stored
in the connectivity matrix, and that can be modulated
by changing the parameters such as θth. Each neuron in
the circuit has a phase determined by its position in that
network, i.e. determined by the phase φµ
i of the stored
phase-pattern. If the parameter θth is modulated by an-
imal speed, then the collective oscillation frequency of
the circuit is modulated by the animal's speed, while the
neurons preserve stable phase relationship among them.
The other persistent-firing models [23, 24, 62] of the os-
cillators needed by the oscillatory interference theory suf-
fer from problems related to robustness, as those en-
countered by the single-cell oscillatory models [70], due
to the variability in the frequency of persisting spik-
ing [25]. Indeed the oscillatory interference models im-
pose strict constraints upon the dynamical properties of
the velocity-modulated oscillatory inputs, which have to
preserve robust velocity-modulated frequency and stable
phase relationships among them on relevant time scales
in a manner robust to noise (many seconds, or dozens of
theta cycle periods) [23, 25, 61, 70].
Notably, in our model, since connections Jij among units
in the circuit are fixed by the learning rule (7), dynamics
is a retrieval process and neurons preserve stable precise
phase relationship among units and stable frequency even
in noisy conditions, at least when the dynamics is in the
basin of attraction of that phase relationship.
Even in the more recent spiking models [24, 62] of the
oscillatory interference principle, in which many prob-
lems related to noise are solved, the heterogeneity of pa-
rameters of the cells which participate to the ring oscil-
lator is not taken into account.
Here we show that the circuit level interactions among
units make the oscillation frequency and the phase-
relationship of the system robust even with respect to
heterogeneity of the spiking thresholds of the units (see
Sec. VI) .
This robustness, due to the proposed coupling which
forces all the units of the circuit to have exactly the same
period of oscillation and to have precisely the same phase
relationships of stored pattern, may be useful in all cases
of sequence coding.
Moreover, our circuit can be easily programmed to
cycle in different phase orders, by storing more than
one phase-pattern as attractors. The circuit is a robust
phase-shuffling ring oscillator, since the network has the
capability of shuffling the order in which its neurons fire,
by storing a variety of different phase-patterns within the
connectivity. If the oscillators predicted by the interfer-
ence theory can generate more than one phase sequence,
as in the model presented here, then this could provide a
potential mechanism to explain the phenomenon of hip-
pocampus remapping[77, 78]. One of the more interesting
discovery of place cells behavior is indeed the remapping
of the place cell representation of space in response to
a changes in sensory or cognitive inputs, i.e. place cells
change their firing properties (place cells can appear dis-
appear or move to other unpredictable locations). This
change may be abrupt and similar to the switch from one
attractor to another[77]. If the place cell will fire at a spe-
cific place where its inputs become synchronized[3, 23 --
25], by recalling a different phase-coded pattern among
the ones stored in our circuit, it will change the phases
of theta cells that are the inputs of the place cell, and it
will change the specific 'place' where the inputs become
Finally we note that even thou our model is not a con-
tinuous attractor, it shares many similarity with such a
class of models. Our model is a circuit with many dis-
tinct attractors, one for each phase-relationship stored
in the network, and the number of different attractor
states is set by the maximal storage capacity studied
here. Furthermore each attractor is a phase-coded pat-
tern, replayed with a collective resonant frequency that
can be modulated by changing for example the spiking
threshold of the units.
Even thou during exploration the activity of place
cells may be explained by the superposition of velocity-
controlled oscillators inputs, the recurrent connections
inside the place cells network may have anyway a rele-
vant role.
During sleep, in absence of external input, the role of
recurrent connections increases, probably due to an in-
crease of excitability via neuromodulators or other mech-
anisms, and the spontaneous activity of the network show
temporarily short replay of stored patterns, probably ini-
tiated simply by noise.
So the pattern activated repeatly during experience,
is stored in the connectivity, and then activated during
sleep when the network is near a critic point and noise is
able to initiate short replay sequences.
Replay of phase-coded patterns of neural activity
during sleep has been observed in hippocampus and
neocortex[5]
11
Notably in our model the time scale of reactivation
is different from the time scale of storing, depends on
the collective frequency which emerges from the connec-
tivity, and may be accelerated or slowed down changing
parameters such as spiking thresholds. Therefore, our
model might be also relevant for replay in prefrontal cor-
tex (PFC) or other cortical areas in which replay is ac-
celerated with respect to awake activity.
In the hippocampus, spikes representing adjacent place
fields occur in rapid succession within a single theta cycle
during behavior. Therefore, relative to this withintheta
cycle rate, reactivation during sleep in hippocampus is
not accelerated. However, reactivation in rat PFC is
clearly compressed five to eight times relative to the wak-
ing state[3, 5]. Indeed, while in hippocampus one may
think that the coding sequence is the within-theta cy-
cle, in prefrontal cortex it is clearly seen that the cross-
correlation among cells during sleep replay is time com-
pressed compared with the cross-correlation during wak-
ing state. The playback speed declines over time as does
the strength of the replay, which is consistent for exam-
ple with a simple increases of spiking threshold in our
model.
VIII. DISCUSSION
We studied the temporal dynamics, including the stor-
age and replay properties, in a network of spiking in-
tegrate and fire neurons, whose learning mechanism is
based on the Spike-Timing Dependent Plasticity. The
temporal patterns we consider are periodic spike-timing
sequences, whose features are encoded in the relative
phase shifts between neurons.
The importance of oscillations and precise temporal
patterns has been pointed out in many brain structure,
such as cortex [73], cerebellum [71, 72], or olfactory sys-
tem [74]. The proposed associative memory approach,
with selective replay of stored sequence, can be a method
for recognize an item, by activating the same memorized
pattern in response to a similar input. Another possibil-
ity is to have a way to transfer a memorized item to an-
other area of the brain, such as for memory consolidation
during sleep. During sleep, indeed, few spikes with the
right phase relationship may initiate the retrieval of one
of the patterns stored in the network and this reactivation
may be useful for memory consolidation. The stored pat-
tern is an attractor of the network dynamics, that is the
dynamics spontaneously goes back to the retrieved phase-
coded pattern for all the perturbations which leave the
system within the basin of attraction. Therefore phase
errors do not sum up, and the phase relationships may
be transferred and kept stable over long time scales.
The time scale of the pattern during retrieval, i.e. the pe-
riod of oscillation 1/ν, depends on (1) the time constants
of the single neuron τm and τs, (2) the spiking thresh-
old θth of the neurons, and (3) the connectivity, through
the STDP learning shape A(τ ) and the time scale of the
pattern during learning mode 1/νµ. Different areas of
the brain may have different shape of STDP to subserve
different oscillation frequencies and different functional
role. Here we fix the shape of A(τ ) to the one observed
in hippocampal cultures (see Fig. 1 and [41]) and fo-
cus on the dependence on the spiking threshold θth of
the neurons. The spiking threshold can modulate the
frequency of the collective oscillation, leaving unaffected
the phase relationships among the units. This opens a
possible way to govern the frequency of collective oscil-
lation via neuromodulators, and to encode information
(such as velocity of the animal) in the frequency of the
oscillations, in addition to the information encoded in
the phase relationships. Notably, in a particular range
of frequencies, the spiking threshold does not affect the
frequency of oscillation but changes the number of spike
per cycle during the retrieval dynamics. This means that
information can be encoded via the number of spikes per
cycle, independently from the information coded in the
phase relationship among units, in agreement with the
observations of independent rate and phase coding in
hippocampus [20].
Important to note, the phase rela-
tionships and the frequency of the collective oscillation
are both robust with respect to noise and to heterogene-
ity of the spiking threshold of the units.
A systematic study of the retrieval capacity of the net-
work is proposed as a function of two parameters of the
model: the frequency of the input pattern and the spiking
threshold. The storage capacity, evaluated as Pmax/N ,
is always lower than the storage capacity of the Hop-
field model. However, the information content of a single
pattern in our dynamical model with N units is higher
than the information content of a pattern in the Hopfield
model with N-units. Indeed, an Hopfield pattern is a set
of N binary values while our phase-coded pattern is a set
of N real number φµ
The role of STDP in the formation of sequences has
been recently investigated in [65, 66]. These studies have
shown how it's possible to form long and complex se-
quences , but they did not concern themselves on how it's
possible to learn and store not only the order of activa-
tion in a sequence, but the precise relative times between
spikes in a closed sequence, i.e. a phase-coded pattern.
In our model not only the order of activation is preserved,
but also the precise phase relationship among units. The
tendency to synchronization of units is avoided in our
model, without need to introduce delays or adaptation,
due to the balance between excitation and inhibition that
is in the connectivity of large networks when the phase
coded pattern with random phases is learned using the
rule in eqs. (5-8). In our rule all the connections, both
positive and negative, scale with the time of presentation
of patterns, keeping always a balance. Indeed, since (1)
the stored phase are uniformly distributed in [0, 2π) and
j ∈ [0, 2π].
then the connectivity matrix in eq. (7) has the prop-
(2) the learning window has the property R A(τ )dt = 0,
erty that the summed excitation (1/N )Pi,Jij >0 Jij and
12
the summed inhibition (1/N )Pi,Jij <0 Jij are equal in
the thermodynamic limit (indeed they are of order unity,
while their difference is of order 1/√N ).
Under this conditions, we investigate how multiple
phase-coded patterns can be learned and selectively re-
trieved in the same network, as a function of time scale
of patterns and network parameters.
The task of storing and recalling phase-coded memo-
ries has been also investigated in [75] in the framework
of probabilistic inference. While we study the effects of
couplings given by Eqn.(5) in a network of IF neurons,
the paper [75] studies this problem from a normative the-
ory of autoassociative memory, in which real variable xi
of neuron i represents the neuron i spike timing with re-
spect to a reference point of an ongoing field potential,
and the interaction H(xi, xj) among units is mediated by
the derivative of the synaptic plasticity rule used to store
memories.
The model proposed here is a mechanism which com-
bines oscillatory and attractor dynamics, which may be
useful in many models of path-integration, as pointed
out in sec.
(VII). Our learning model offers a IF cir-
cuit able to keep robust phase-relationship among cells
participating to a collective oscillation, with a modulated
collective frequency, robust with respect to noise and het-
erogeneities. Notably the frequency of the collective os-
cillation in our circuit is not sensible to the single value
of the threshold of each unit, but to the average value of
the threshold of all units, since all units participate to a
single collective oscillating pattern which is an attractor
of the dynamics.
Recently there is renewed interest in reverberatory
activity[81] and in cortical spontaneous activity[79, 80]
whose spatiotemporal structure seems to reflect the un-
derlying connectivity, which in turn may be the result of
the past experience stored in the connectivity.
Similarity between spontaneous and evoked cortical ac-
tivities has been shown to increase with age [83], and
with repetitive presentation of the stimulus [82]. Inter-
estingly, in our IF model, in order to induce spontaneous
patterns of activity reminiscent of those stored during
learning stage, few spikes with the right phase relation-
ship are sufficient. It means that, even in absence of sen-
sory stimulus, a noise with the right phase relationships
may induce a pattern of activity reminiscent of a stored
pattern. Therefore, by adapting the network connectivity
to the phase-coded patterns observed during the learn-
ing mode, the network dynamics builds a representation
of the environment and is able to replay the patterns of
activity when stimulated by sense or by chance.
This mechanism of learning phase-coded patterns of
activity is then a way to adapt the internal connectivity
such that the network dynamics have attractors which
represent the patterns of activity seen during experience
of environment.
13
[1] Nadasdy Z, Hirase H, Czurko A, Csicsvari J, Buzsaki G.
1999. Replay and Time Compression of Recurring Spike
Sequences in the Hippocampus. The Journal of Neuro-
science 19: 9497-9507.
[2] Ji D Wilson MA. 2007. Coordinated memory replay in
the visual cortex and hippocampus during sleep. Nature
Neuroscience 10: 100-107.
[3] Euston DR, Tatsuno M, McNaughton BL. 2007. Fast-
Forward Playback of Recent Memory Sequences in Pre-
frontal Cortex During Sleep. Science 318: 1147-1150.
[4] Carien S. Lansink, Pieter M. Goltstein, Jan V. Lankelma,
Bruce L. McNaughton, Cyriel M. A. Pennartz 2009. Hip-
pocampus Leads Ventral Striatum in Replay of Place-
Reward Information. PLoS Biol 7(8): e1000173.
[5] Schwindel CD, McNaughton BL 2011, Hippocampal-
cortical interactions and the dynamics of memory trace
reactivation, Prog Brain Res. 2011;193:163-77.
[6] Diba K, Buzsaki G. 2007. Forward and reverse hippocam-
pal place-cell sequences during ripples. Nat Neurosci 10:
1241-1242.
[7] Davidson TJ, Kloosterman F, Wilson MA. 2009 Hip-
pocampal Replay of Extended Experience. Neuron 63:
497-507.
[8] Girardeau G, Zugaro M. 2011. Hippocampal ripples and
memory consolidation. Current Opinion in Neurobiology
21: 452-459.
[9] Carr MF, Jadhav SP, Frank LM. 2011. Hippocampal re-
play in the awake state: a potential substrate for memory
consolidation and retrieval. Nat Neurosci 14:147-53.
[10] Pastalkova E, Itskov V, Amarasingham A, Buzsaki G.
2008. Internally Generated Cell Assembly Sequences in
the Rat Hippocampus. Science 321: 1322-1327.
[11] Siegel M, Warden MR, Miller EK. 2009. Phase-dependent
neuronal coding of objects in short-term memory. PNAS
106: 21341-21346.
[12] Montemurro MA, Rasch MJ, Murayama Y, Logothetis
NK, Panzeri S. 2008. Phase-of-Firing Coding of Natural
Visual Stimuli in Primary Visual Cortex. Current Biology
18: 375-380.
[13] Kayser C, Montemurro MA, Nikos K. Logothetis NK,
Panzeri S. 2009. Spike-Phase Coding Boosts and Stabi-
lizes Information Carried by Spatial and Temporal Spike
Patterns. Neuron 61: 597-608.
[14] Masquelier T, Hugues E, Deco G, Thorpe SJ. 2009. Os-
cillations, Phase-of-Firing Coding, and Spike Timing-
Dependent Plasticity: An Efficient Learning Scheme.
The Journal of Neuroscience 29: 13484-13493.
[15] Montemurro MA, Rasch MJ, Murayama Y, Logothetis
NK, Panzeri S. Phase-of-Firing Coding of Natural Visual
Stimuli in Primary Visual Cortex. 2008. Current Biology
18: 375380.
[16] McNaughton BL, Battaglia FP, Jensen O, Edvard I
Moser EI, Moser MB. 2006 Path integration and the neu-
ral basis of the 'cognitive map'. Nature Reviews Neuro-
science 7: 663-678.
[17] O'Keefe J, Recce ML. 1993. Phase relationship between
hippocampal place units and the EEG theta rhythm.
Hippocampus 3:317330.
[18] Lengyel M, Huhn Z, Erdi P. 2005. Computational theo-
ries on the function of theta oscillations. Biological Cy-
bernetics 92: 393-408.
[19] O'Keefe J, Burgess N. 2005. Dual phase and rate cod-
ing in hippocampal place cells: theoretical significance
and relationship to entorhinal grid cells. Hippocampus
15:85366.
[20] Huxter J, Burgess N, O'Keefe J. 2003. Independent rate
and temporal coding in hippocampal pyramidal cells. Na-
ture 425: 828-32.
[21] Burgess N, Barry C, O'Keefe J. 2007. An oscillatory in-
terference model of grid cell firing. Hippocampus 17:801-
12.
[22] Jeewajee A, Barry C, OKeefe J, Burgess N (2008a) Grid
cells and theta as oscillatory interference: electrophys-
iological data from freely moving rats. Hippocampus
18:11751185
[23] Blair HT, Gupta K, Zhang K. 2008. Conversion of a
phase- to a rate-coded position signal by a three stage
model of theta cells, grid cells, and place cells. Hippocam-
pus 18: 12391255.
[24] A. C. Welday, I. G. Shlifer, M. L. Bloom, K. Zhang,
and Hugh T. Blair 2011. Cosine Directional Tuning of
Theta Cell Burst Frequencies: Evidence for Spatial Cod-
ing by Oscillatory Interference The Journal of Neuro-
science, November 9, 2011 31(45):1615716176 1615
[25] Lisa M. Giocomo, May-Britt Moser, Edvard I. Moser
2011, Computational Models of Grid Cells Neuron Vol-
ume 71, Issue 4, Pages 589603
[26] John B. Issa and Kechen Zhang 2012, Universal condi-
tions for exact path integration in neural systems PNAS
2012 109 (17) 6716-6720
[27] Wittenberg GM, Wang SSH. 2006. Malleability of spike-
timing-dependent plasticity at the CA3-CA1 synapse.
Journal of Neuroscience 26:6610-6617.
[28] Geisler C, Robbe D, Zugaro M, Sirota A, Buzsaki
G. 2007. Hippocampal place cell assemblies are speed-
controlled oscillators. PNAS 104:8149-8154.
[29] Amit DJ, Treves A. 1989. Associative memory neural net-
work with low temporal spiking rates. PNAS Biophysics
86: 7871-7875.
[30] Anishchenko A, Treves A. 2006. Autoassociative Mem-
ory Retrieval and Spontaneous Activity Bumps in Small-
World Networks of Integrate and Fire Neurons. Journal
of Physiology-Paris 100: 225-236
Battaglia FP, Treves A. 1998. Stable and Rapid Recur-
rent Processing in Realistic Autoassociative Memories.
Neural Computation 10: 431-450.
[31] Roman M. Borisyuk, Frank C. Hoppensteadt 1998. Mem-
orizing and recalling spatialtemporal patterns in an os-
cillator model of the hippocampus, BioSystems 48, 310
[32] De Almeida L, Idiart M Lisman JE. 2007. Memory re-
trieval time and memory capacity of the CA3 network:
Role of gamma frequency oscillations. Learn Mem 14:
795-806.
[33] Leibold C, Kempter R. 2006. Memory Capacity for Se-
quences in a Recurrent Network with Biological Con-
straints. Neural Computation 18: 904-941.
[34] S Olmi, R Livi, A Politi, A Torcini 2010, Collective os-
cillations in disordered neural networks Phys Rev E 81,
046119
[35] Memmesheimer RM, Timme M. 2006. Designing the dy-
namics of spiking neural networks. Phys Rev Lett 97:
188101.
[36] Scarpetta S, de Candia A, Giacco F. 2010. Storage of
phase-coded patterns via STDP in fully-connected and
sparse network: a study of the network capacity. Fron-
tiers in synaptic neuroscience 2, 1-12.
[37] Hopfield JJ. 1982. Neural networks and physical systems
with emergent collective computational abilities. Proc
NatL Acad Sci USA: 79: 2554-2558.
[38] Hopfield JJ. 1995. Pattern recognition computation us-
ing action potential timing for stimulus representation.
Nature 376: 3336.
[39] Markram H, Lubke J, Frotscher M, Sakmann B. 1997.
Regulation of synaptic efficacy by coincidence of postsy-
naptic APs and EPSPs. Science 275: 213-215.
[40] Feldman DE. 2000. Timing-based LTP and LTD and ver-
tical inputs to layer II/III pyramidal cells in rat barrel
cortex. Neuron 27: 45-56.
[41] Bi GQ, Poo MM. 1998. Precise spike timing determines
the direction and extent of synaptic modifications in cul-
tured hippocampal neurons. J Neurosci 18: 10464:10472.
[42] Bi GQ, Poo MM. 2001. Synaptic modification by corre-
lated activity: Hebb's postulate revisited. Annual Review
Neuroscience 24: 139-166.
[43] Debanne D, Gahwiler BH, Thompson SM. 1998. Event
driven programming Long-term synaptic plasticity be-
tween pairs of individual CA3 pyramidal cells in rat hip-
pocampal slice cultures. J Physiol 507: 237-247.
[44] Shouval HZ, Wang SS, Wittenberg GM. 2010. Spike tim-
ing dependent plasticity: a consequence of more funda-
mental learning rules. Front Comput Neurosci 4: 19.
[45] Graupner M, Brunel N. 2010. Mechanisms of Induction
and Maintenance of Spike-Timing Dependent Plasticity
in Biophysical Synapse Models. Front Comput Neurosci
4: 136.
[46] Markram H, Gerstner W, PerJesper Sjostrom PJ. 2011.
A history of spike-timing-dependent plasticity. Front Syn
Neurosci 3:4.
[47] Sjostrom PJ, Turrigiano G, Nelson SB. 2001. Rate,
Timing, and Cooperativity Jointly Determine Cortical
Synaptic Plasticity. Neuron 32: 1149-1164.
[48] Magee JC, Johnston D. 1997. A synaptically controlled
associative signal for Hebbian plasticity in hippocampal
neurons. Science 275: 209-212.
[49] Yoshioka M, Scarpetta S, Marinaro M. 2007. Spatiotem-
poral learning in analog neural networks using spike-
timing-dependent synaptic plasticity. Phys Rev E 75:
051917.
[50] Scarpetta S, Zhaoping L, Hertz J. 2002. Hebbian Imprint-
ing and Retrieval in Oscillatory Neural Networks. Neural
Computation 14: 2371-96.
[51] Scarpetta S, Zhaoping L, Hertz J. 2001. Spike-Timing-
Dependent Learning for Oscillatory Networks. Advances
in Neural Information Processing Systems 13, MIT Press.
[52] Scarpetta S, Marinaro M., A learning rule for place fields
in a cortical model: Theta phase precession as a network
effect, Hippocampus Volume 15, Issue 7, pages 979989,
2005
[53] Silvia Scarpetta, Masahiko Yoshioka and Maria Mari-
naro, Encoding and replay of dynamic attractors with
multiple frequencies Analysis of a STDP based learn-
ing rule, Dynamic Brain - from Neural Spikes to Behav-
iors Lecture Notes in Computer Science, 2008, Volume
5286/2008
[54] Abarbanel H, Huerta R, Rabinovich MI. 2002. Dynamical
model of long-term synaptic plasticity. PNAS 99: 10132-
14
10137.
[55] Gerstner W, Kistler W. 2002. Spiking neuron models:
single neurons, populations, plasticity. Cambridge Uni-
versity Press, NY.
[56] Gerstner W, Ritz R, van Hemmen JL. 1993. Why spikes?
Hebbian learning and retrieval of time-resolved excitation
patterns. Biological Cybernetics 69: 503-515.
[57] Gerstner W, Kempter R, Van Hemmen L, Wagner H.
1996. A neuronal learning rule for sub-millisecond tem-
poral coding. Nature 383: 7681.
[58] Hasselmo ME. 1993. Acetylcholine and learning in a cor-
tical associative memory. Neural Computation 5: 32-44.
acetylcholine
and memory consolidation. Trend in Cognitive Sciences
3: 351-359.
[59] Hasselmo ME. 1999. Neuromodulation:
[60] Lengyel M, Dayan P. 2007. Uncertainty, phase and oscil-
latory hippocampal recall. Advances in Neural Informa-
tion Processing Systems, MIT Press.
[61] Welinder PE, Burak Y, Fiete IR. 2008. Grid cells: The
position code, neural network models of activity, and the
problem of learning. Hippocampus 18: 1283-300.
[62] E A Zilli, M Hasselmo 2010. Coupled noisy spiking neu-
rons as velocity-controlled oscillators in a model of grid
cell spatial firing. The Journal of Neuroscience, 30(41):
13850-13860.
[63] Legenstein R, Christian Nager C, Maass W. 2005. What
Can a Neuron Learn with Spike-Timing-Dependent Plas-
ticity? Neural Computation 17: 2337-2382.
[64] Song S, Abbot LF. 2001. Cortical Development and
Remapping through Spike Timing-Dependent Plasticity.
Neuron 32: 339350.
[65] Fiete, I.R., Senn, W., Wang, C.Z.H., Hahnloser, R.H.R.
2010. Spike-time-dependent plasticity and heterosynaptic
competition organize networks to produce long scale-free
sequences of neural activity. Neuron 65(4), 56376
[66] Verduzco-Flores SO, Bodner M, Ermentrout B. 2011. A
model for complex sequence learning and reproduction
in neural populations. Journal of computational neuro-
science 2011 Sep 2.
[67] EM Izhikevich 2006. Polychronization: Computation
with spikes. Neural computation Vol. 18, No. 2, Pages
245-282
[68] Scarpetta S, Giacco F, de Candia A. 2011. Storage ca-
pacity of phase-coded patterns in sparse neural networks.
EPL 95: 28006.
[69] Silvia Scarpetta, Antonio de Candia, Ferdinando Giacco,
Dynamics and storage capacity of neural networks with
small-world topology, Frontiers in Artificial Intelligence
and Applications Volume 226, 2011
[70] Zilli EA, Yoshida M, Tahvildari B, Giocomo LM, Has-
selmo ME. 2009. Evaluation of the oscillatory interfer-
ence model of grid cell firing through analysis and mea-
sured period variance of some biological oscillators. PLoS
Comput Biol 5: e1000573.
[71] D'Angelo E, Koekkoek SK, Lombardo P, Solinas S, Ros
E, Garrido J, Schonewille M, De Zeeuw CI. 2009. Tim-
ing in the cerebellum:oscillations and resonance in the
granular layer. Neuroscience 162: 805-815.
[72] D'Angelo E, De Zeeuw CI. 2009 . Timing and plasticity
focus on the granular layer. Trends
in the cerebellum:
Neurosci 32: 30-40.
[73] Buzsaki G, Draguhn A. 2004. Neuronal Oscillations in
Cortical Networks. Science 304: 1926-1929.
[74] Gelperin A. 2006. Olfactory Computations and Network
15
Oscillation. The Journal of Neuroscience 26: 1663-1668.
[75] Lengyel M, Kwag J, Paulsen O, Dayan P. 2005 Matching
storage and recall: hippocampal spike timing-dependent
plasticity and phase response curves. Nature Neurosci.
8(12):1677-83
[76] Gilson M, Masquelier T, Hugues E. 2011 STDP allows
fast rate-modulated coding with Poisson-like spike trains.
PLoS Comput Biol. 7(10):e100223
[77] Tom J. Wills, Colin Lever, Francesca Cacucci, Neil
Burges, John O'Keefe 2005 Attractor Dynamics in the
Hippocampal Representation of the Local Environment
Science. 2005 May 6; 308(5723): 873876
[78] Laura Lee Colgin, Edvard I. Moser and May-Britt
Moser 2008 Understanding memory through hippocam-
pal remapping Trends in Neurosciences Vol.31 No.9
[79] Ringach DL. 2009 Spontaneous and driven cortical activ-
ity: implications for computation. Curr Opin Neurobiol
19: 439-44.
[80] Luczak A, Maclean JN. Default activity patterns at
the neocortical microcircuit level. Front Integr Neurosci.
2012;6:30.
[81] P-M. Lau and G-Q. Bi 2005. Synaptic mechanisms of
persistent reverberatory activity in neuronal networks.
Proc. Nat. Acad. Sci. USA, 102:10333 10338
[82] Han F, Caporale N, Dan Y. 2008. Reverberation of recent
visual experience in spontaneous cortical waves. Neuron
60: 321-327.
[83] Berkes P, Orban G, Lengyel M, Fiser J. 2011. Sponta-
neous cortical activity reveals hallmarks of an optimal
internal model of the environment. Science 331: 83-87.
|
1309.1086 | 1 | 1309 | 2013-09-04T16:10:02 | Characterization of high frequency oscillations and EEG frequency spectra using the damped-oscillator oscillator detector (DOOD) | [
"q-bio.NC"
] | Objective: The surgical resection of brain areas with high rates of visually identified high frequency oscillations (HFOs) on EEG has been correlated with improved seizure control. However, it can be difficult to distinguish normal from pathological HFOs, and the visual detection of HFOs is very time-intensive. An automated algorithm for detecting HFOs and for wide-band spectral analysis is desirable.
Methods: The damped-oscillator oscillator detector (DOOD) is adapted for HFO detection, and tested on recordings from one rat and one human. The rat data consist of recordings from the hippocampus just prior to induction of status epilepticus, and again 6 weeks after induction, after the rat is epileptic. The human data are temporal lobe depth electrode recordings from a patient who underwent pre-surgical evaluation.
Results: Sensitivities and positive predictive values are presented which depend on specifying a threshold value for HFO detection. Wide-band time-frequency and HFO-associated frequency spectra are also presented. In the rat data, four high frequency bands are identified at 80-250 Hz, 250-500 Hz, 600-900 Hz and 1000-3000 Hz. The human data was low-passed filtered at 1000 Hz and showed HFO-associated bands at 15 Hz, 85 Hz, 400 Hz and 700 Hz.
Conclusion: The DOOD algorithm is capable of high resolution time-frequency spectra, and it can be adapted to detect HFOs with high positive predictive value. HFO-associated wide-band data show intricate low-frequency structure.
Significance: DOOD may ease the labor intensity of HFO detection. DOOD wide-band analysis may in future help distinguish normal from pathological HFOs. | q-bio.NC | q-bio | Title: Characterization of high frequency oscillations and EEG frequency spectra using
the damped-oscillator oscillator detector (DOOD)
Authors: David Hsu1; Murielle Hsu1; Heidi L. Grabenstatter2; Gregory A. Worrell3;
Thomas P. Sutula1
Institutions:
1. Department of Neurology, University of Wisconsin, Madison, WI, United States.
2. Department of Pediatrics, University of Colorado, Aurora, CO, United States.
3. Department of Neurology, Mayo Clinic, Rochester, MN, United States.
Corresponding author:
David Hsu
Department of Neurology, Rm 7178
1685 Highland Av
Madison WI 53705-2281
P: 608-265-7951
F: 608-263-0412
E: [email protected]
Funding: NIH R01-25020 and NIH R01-NS63039-01.
Highlights:
A novel method called the damped-oscillator oscillator detector (DOOD) has been
developed that is capable of very high resolution time-frequency analysis.
DOOD can be adapted for the reliable, automated detection of high frequency
oscillations (HFOs) in the electroencephalogram.
HFOs as detected by DOOD are associated with intricate low frequency structure.
Abstract
Objective: The surgical resection of brain areas with high rates of visually identified high
frequency oscillations (HFOs) on EEG has been correlated with improved seizure control.
However, it can be difficult to distinguish normal from pathological HFOs, and the visual
detection of HFOs is very time-intensive. An automated algorithm for detecting HFOs
and for wide-band spectral analysis is desirable.
Methods: The damped-oscillator oscillator detector (DOOD) is adapted for HFO
detection, and tested on recordings from one rat and one human. The rat data consist of
recordings from the hippocampus just prior to induction of status epilepticus, and again 6
weeks after induction, after the rat is epileptic. The human data are temporal lobe depth
electrode recordings from a patient who underwent pre-surgical evaluation.
Results: Sensitivities and positive predictive values are presented which depend on
specifying a threshold value for HFO detection. Wide-band time-frequency and HFO-
associated frequency spectra are also presented. In the rat data, four high frequency bands
are identified at 80-250 Hz, 250-500 Hz, 600-900 Hz and 1000-3000 Hz. The human
data was low-passed filtered at 1000 Hz and showed HFO-associated bands at 15 Hz, 85
Hz, 400 Hz and 700 Hz.
Conclusion: The DOOD algorithm is capable of high resolution time-frequency spectra,
and it can be adapted to detect HFOs with high positive predictive value. HFO-
associated wide-band data show intricate low-frequency structure.
Significance: DOOD may ease the labor intensity of HFO detection. DOOD wide-band
analysis may in future help distinguish normal from pathological HFOs.
Introduction
Detecting transient high frequency oscillations (HFOs) in the human
electroencephalogram (EEG) in the ripple (80-200 Hz) and fast ripple (250-500 Hz)
range (Bragin, Engel et al. 1999) has attracted increasing interest because ripples and fast
ripples are more abundant in seizure onset zones (Staba, Wilson et al. 2002; Worrell,
Parish et al. 2004; Staba, Frighetto et al. 2007; Urrestarazu, Chander et al. 2007; Jacobs,
LeVan et al. 2008; Jacobs, Levan et al. 2009; Khosravani, Mehrotra et al. 2009; Zijlmans,
Jacobs et al. 2009; Bragin, Engel et al. 2010; Jacobs, Zijlmans et al. 2010), and the
resection of brain tissue with high rates of ripple and fast ripple activity is correlated with
better seizure outcomes (Jacobs, Zijlmans et al. 2010; Wu, Sankar et al. 2010). However,
HFOs are also found in normal brain (Buzsaki, Horvath et al. 1992; Ylinen, Bragin et al.
1995; Chrobak and Buzsaki 1996; Draguhn, Traub et al. 2000; Ponomarenko, Korotkova
et al. 2003; Axmacher, Elger et al. 2008), and can be difficult to distinguish from normal
fast brain oscillations (Engel, Bragin et al. 2009; Nagasawa, Juhasz et al. 2011).
Visual inspection remains the gold standard for the detection of HFOs. The
detection of transient oscillatory activity typically involves either band pass or high pass
filtering to remove lower frequency activity. For instance, one may employ an 80 Hz
high pass filter to detect ripples and a 250 Hz high pass filter to detect fast ripples, with
events identified as one or the other if there are at least 4 consecutive oscillations and if
the event is clearly above baseline (Zijlmans, Jacobs et al. 2010). Such visual approaches
are very time-intensive (Urrestarazu, Chander et al. 2007). A certain amount of
subjectivity is also unavoidable.
One early automated approach employs first a band pass filter that passes
frequencies between 100 and 500 Hz, then calculating root mean square (RMS)
amplitudes of sliding 3 ms time windows of EEG data (Staba, Wilson et al. 2002).
Events with peak RMS amplitude greater than 6 standard deviations above the mean and
lasting at least 6 ms in duration are identified as candidate HFOs. A sensitivity of greater
than 84% was reported for this method, but the false positive rate was not reported. The
false positive rate in another automated method, the line-length measure, was reported at
about 80% when compared with expert visual review (Worrell, Gardner et al. 2008), but
sensitivity was not reported. An earlier test of the line length measure showed that it is
superior to the root mean square amplitude measure when tested on EEG data bandpass
filtered between 30 and 85 Hz (Gardner, Worrell et al. 2007). More recently an
unsupervised method for classification of high frequency oscillations has been proposed
(Blanco, Stead et al. 2010).
As reflected by the variety of approaches for detecting HFOs, the definition of an
HFO is somewhat arbitrary, and the various methods that have been applied yield results
with a range of sensitivities and specificities that have implications for distinguishing
HFOs in pathological tissue from similar oscillating activity in normal tissue. An
automated computer algorithm that detects pathological HFOs quantitatively and reliably
would be useful for grading long-term EEG from patients undergoing pre-surgical
localization.
We have recently presented a novel pseudo-wavelet approach for high resolution
time-frequency analysis called the damped-oscillator oscillator detector (DOOD) (Hsu,
Hsu et al. 2010). This method uses a set of mathematical oscillators, i.e., damped
harmonic oscillators, to detect oscillations in EEG. Initial testing suggests that DOOD is
capable of superior time-frequency resolution compared with time-windowed Fast
Fourier Transform (FFT) approaches and with other wavelet and pseudo-wavelet
approaches. Here we show how the DOOD algorithm may be used to detect transient
oscillations such as HFOs. After a transient oscillation is detected, one can also
investigate the wide-band time-frequency structure surrounding that oscillation. We
present illustrative examples from human temporal lobe depth recordings(Worrell,
Gardner et al. 2008), and from in vivo hippocampal rat recordings.
Methods
Rat data: Adult Spraque-Dawley rats were placed in a stereotaxic apparatus and
anesthetized with 2% isoflurane after pretreatment with atropine (2 mg/kg). The area of
incision was injected with 0.5 ml of 0.5% bupivacaine for prolonged local
anesthesia/analgesia and postoperative analgesia was administered once daily
(buprenorphrine 0.01 mg/kg, subcutaneous). Burr holes for the recording probe and
ground screw were placed by conventional surgical techniques. The depth-recording
electrode, a 16-channel silicon probe (NeuroNexus, Inc., Ann Arbor, MI) implanted in
the right hippocampus (3.0 mm posterior, 2.6 mm lateral, 3.1 mm ventral) spanned CA1,
dorsal dentate gyrus, and the hilus at 0.1 mm intervals. Each electrode contact has an
area of
. A skull screw was placed in the left frontal bone to serve as
ground. Dental acrylic was used to secure the electrodes according to standard chronic
methods. At two weeks following surgery for electrode implantation, kainic acid (Tocris
Bioscience, Ellisville, MO) was administered in repeated, low doses (5 mg/kg,
327.0310 mm intraperitoneal) hourly until each rat experiences convulsive status epilepticus for >3 h
(Hellier, Patrylo et al. 1998).
Simultaneous in vivo field potential recordings from 16-channel silicon probes
were obtained in freely-moving rats on a weekly basis for six weeks. One to two minutes
of artifact free EEG was saved from each recording session. Field potentials were
recorded at a sampling frequency of 12,207.03 Hz and filtered on-line between 2 Hz and
6000 Hz using a Tucker Davis Technologies recording system (Alachua, FL). Off-line
conversion of the electrographic data to f32 binary files was accomplished using custom-
written software developed in MATLAB (The MathWorks, Inc., Natick, MA).
The EEG data used for the current study consists of the first 100 seconds from the
first rat at baseline (just prior to kainate-induced status epilepticus), and the last 100
seconds from the same rat 46 days later.
After completion of long-term recordings epileptic rats were perfused with 0.01M
phosphate buffered saline followed by an aqueous solution of aldehyde fixatives (4%
paraformaldehyde-0.5% glutaraldehyde). The brains were removed, post-fixed overnight
in same fixative solution, and sectioned on a Pelco Vibratome 1500 sectioning system
into 100-µm coronal sections. The sections were wet-mounted in 0.9% saline and
imaged with a digital camera (Spot II; Diagnostic Instruments, Sterling Heights, MI) on a
Nikon E600 Eclipse epifluorescent microscope x1-4 planapochromatic objectives in
order to locate and measure the recording tracts. Brightfield images were acquired at an
initial 36-bit tone scale and saved as 16-bit files. The final images were prepared in
Adobe Photoshop 7.0. Adjustments in tone scale, gamma, contrast, and hue and
subsequent sharpening with the unsharp mask algorithm were applied to the entire image.
The 16 recording sites along the probe length were reconstructed and associated with
relevant hippocampal structures using histological measurements and the known inter-
site intervals. Electrodes 2-3 were in stratum oriens, with electrode 1 probably in alveus.
Electrodes 4-5 were in the CA1 pyramidal cell layer. Electrodes 6-7 were in stratum
radiatum. Electrodes 8-9 were in the lacunosum moleculare layer. Electrodes 10-11
were in the molecular layer of the dentate gyrus. Electrodes 12-13 were in the dentate
granule cell layer. Electrodes 14-16 were in the hilus of the dentate gyrus.
Institutional regulations for the care of all animals were strictly adhered to, and
approval of protocols was obtained from the animal care committee. Animals that are
persistently ill after induction of seizures or that suffer frequent, prolonged seizures are
euthanized.
Human depth recordings with microwires. After Institutional Review Board
approval and with informed patient consent, EEG data were collected from a patient with
refractory temporal lobe epilepsy who underwent pre-surgical temporal lobe depth
electrode evaluation (Worrell, Gardner et al. 2008). Thirty minutes of artifact free stages
3 and 4 slow wave sleep were acquired using custom hybrid depth electrodes which
combine both clinical macroelectrodes and special microwires (Adtech Inc.). Each
microwire has a contact area of about
. EEG data were sampled at 5 KHz
and low pass filtered at 1000 Hz (Neuralynx, Inc.). The first 120 seconds from the first 8
microwire channels of this data were subsequently used for analysis.
Visual identification of HFOs. EEG data were high pass filtered at 80 Hz and
visually inspected second-by-second at consecutive one-second intervals for oscillatory
activity. An oscillatory event with at least 3 crests or 3 troughs was marked as a putative
32110 mm HFO if it clearly stood out from baseline. This definition is less restrictive than that used
by most investigators. Our motivation is that we wish to include as many potential
oscillations as possible, and then to test the ability of DOOD to detect both the
ambiguous oscillations and the unequivocal ones depending on a threshold value intrinsic
to the DOOD algorithm. In practice, we found the EEG data from the normal rat to
contain the most challenging visual samples of putative oscillations, as there were many
oscillations that almost blended into the background, when compared both to the sample
from the human subject with epilepsy and to the sample from the same rat 6 weeks after
kainate-induced status epilepticus.
Basic DOOD algorithm. There are several useful variations of the DOOD
algorithm (Hsu, Hsu et al. 2010). Let
denote the voltage measured by EEG
channel k at time t, with k = 1 to
. In the coordinate or X-DOOD variation, the data
driving force from channel k at time t is taken to be
. In the velocity or
V-DOOD variation, the data driving force is taken to be the first time derivative of the
EEG output,
, where the time derivative is performed by finite
difference,
. Here
is the sampling time
related to the sampling rate by
. The mathematical oscillator associated with
the
EEG channel at time t is defined by its coordinate
, velocity
,
frequency
, and frictional damping constant
. The total number of
mathematical oscillators for each EEG channel is denoted
. Its pseudo-wavelet
wavefunction
is given by:
(,)dataxktCN(,)(,)datahktxkt(,)(,)datahktxkt(,)(,)(,)/datadatadataxktxkttxkttt1/Sftthnthk(,,)xknt(,,)xknt()fn()gnfN(,,)knt .
(1)
Taking
and
to be the real and imaginary parts of
, we can
write the coordinate and velocity of the associated mathematical oscillator as:
(2)
(3)
For incrementally small time steps
,
can be easily calculated from
through the recursion relationship:
(4)
For both X-DOOD and V-DOOD, the spectral density is defined as that part of the time
rate of change of the energy at time t due to the data driving force, which is given by:
.
(5)
In practice, we average
over consecutive time windows of time duration
, in
large part to avoid saving enormous files sampled at the EEG sampling time of
. When
0(,,)'(,')exp2()()(')tkntdthktgnifntt(,,)Rknt(,,)Iknt(,,)knt1(,;)(,,)2()Ixkntkntfn()(,,)(,,)(,,)()RIgnxkntkntkntfnt(,,)kntt(,;)knt(,,)(,)exp2()()(,,)kntthktttgnifntknt,(),(,,)(,,)Skfntxknthknt(,,)Skftwintt one wishes to detect events of short duration, one should choose
to be somewhat
smaller than the expected event duration. For the detection of HFOs that last tens of
milliseconds, we choose
ms.
A third variation of the DOOD algorithm is obtained if one squares the data
power before performing the time average:
.
(6)
When Eq. 6 is paired with
, we refer to this variation as V-DOOD-SQR.
The X-DOOD and V-DOOD-SQR variations are best for revealing low frequency time-
frequency structure, while the V-DOOD variation is most useful for detecting high
frequency activity (Hsu, Hsu et al. 2010). For the purpose of detecting HFOs, we will
use V-DOOD only. However, we will also show V-DOOD-SQR time-frequency spectral
densities to demonstrate associated low-frequency structure.
The friction constant
is equal to the half-width at half-maximum of the
spectral density of the mathematical oscillator. Therefore, the relative frequency
resolution of the mathematical oscillator is given by the ratio
. If all the
mathematical oscillators are required to have the same relative frequency resolution, then
must be proportional to
, i.e.,
, where
is a dimensionless
number which is the same for each mathematical oscillator. In addition, in order to
obtain good frequency coverage, the spacing of consecutive oscillators along the
wint5wint22,(),(,;)(,,)Skfntxknthknt(,)(,)datahktxkt()gnthnthn()/()gnfn()gn()fn0()()gngfn0g frequency axis should be no more than
. Therefore,
must have the
form of a geometric series:
,
(7)
where the constant
is a number that is less than or equal to 1. In our experience,
choices of
to 0.10 and
to 1 work well.
Spanning the frequency range from 1 to 6000 Hz with
and
would
require
mathematical oscillators. A larger value of
is permissible (and
computationally cheaper) when one is trying to detect oscillatory activity of shorter
duration, because the spectral widths of brief oscillations are broader than those of
oscillations that persist for many cycles. If we choose
and
,
mathematical oscillators would be required. If one attempts to span the same frequency
range with linearly spaced oscillators with a frequency resolution of 1 Hz, one would
need 6000 oscillators. A geometric frequency series can span the same frequency range
as a linear one but much more efficiently, and with no loss of relative frequency
resolution.
Note that the most natural way of plotting a geometric frequency series is as the
logarithm of the frequency. For convenience, we adopt the base 10 logarithmic scale,
denoted as
where f is the frequency in Hz. A logarithmic frequency axis
allows us to plot wide-band spectra much more conveniently than the more traditional
linear frequency axis. It may be of interest that the cochleas of humans, elephants, cats,
0()()gngfn()fn0(1)1()fngfn00.02g0.500.02g1440fN0g00.10g0.5179fN()LogfLogf mice and a number of other animals that are capable of wide-band auditory processing
also use a logarithmic frequency scale (Greenwood 1990). In fact, the original
motivation for the basic DOOD algorithm came from consideration of a simplified
mathematical model of the human cochlea (DH and MH, unpublished).
HFO detection algorithm. In using DOOD to detect HFOs, we adopt the
following algorithm. First, every EEG channel is individually Z-normalized by
subtracting out the mean and dividing by the standard deviation, with the means and
standard deviations calculated over the entire data sample. Next, for a given sampling
rate
, we take
Hz,
,
, and
. Here
is
the number of mathematical oscillators for each EEG channel, and the frequency of the
oscillator is given by Eq (7). The V-DOOD spectral density
is calculated at
each sampling time with sampling time increment
, and then averaged over
consecutive time windows of duration
ms. Let this time-averaged
be
denoted
. The time resolution of
is thus
ms. The time-
averaged spectral density
is accumulated over consecutive one-second time
intervals. For each one-second time interval and for each individual channel k, the means
and standard deviations of
are calculated with frequencies restricted to a
target range of 80 to 1000 Hz. Data points for
outside of this frequency range
are saved but not used in calculating means and standard deviations. These means and
standard deviations are then used to Z-normalize each one-second segment of
over the entire frequency range,
to
.
Sf(1)1fmax()/2fSffNf0.500.10gfNthn(,,)Skft1/Stf5wint(,,)Skft(,,)Skft(,,)Skft5wint(,,)Skft(,,)Skft(,,)Skft(,,)Skft1 Hzfmaxf We will perform two types of searches on the function
. The first is a
frequency domain search for a given fixed point in time with the search conducted within
the pre-identified frequency range of 80 to 1000 Hz. The purpose of this search is to
identify the largest value of
at a fixed point in time t, denoted
. The
frequency associated with
is denoted
, and a time period of oscillation
is similarly defined. The second type of search is a time domain search,
the purpose of which is to identify onset and end times for the candidate HFO. The time
step for this time domain search is
ms. At every time step of the time domain
search,
and
are calculated and updated if the new value of
is larger
than the prior.
The onset time of a candidate HFO is identified by searching
in the
time domain for the first occurrence where
. The end time of the candidate
HFO is defined as the first occurrence after onset where
, provided a number of
other conditions are met, as follows. First,
must be less than 1 for a time period
that is at least one full oscillation cycle (
) past the last time that
. This
precaution is taken because, in the presence of a well-developed EEG oscillation, the
DOOD spectral density exhibits a pronounced “beating” effect that depends on the details
of the oscillation and whether the oscillation is coupled to other oscillations (Hsu, Hsu et
al. 2010). If the flagged end time is not yet one full cycle past the last time that
, then one keeps stepping through time until one full cycle is reached before
accepting the event as having ended.
(,,)Skft(,,)SkftpeakSpeakSpeakfpeakpeak1/tf5wintpeakSpeakfpeakS(,,)Skftpeak1Speak1SpeakSpeaktpeak1Speak1S Once the candidate HFO has definitively ended, then
is time-averaged
between the onset and end times to produce the doubly time-averaged
. A
frequency domain search is then performed on
to identify its maximum
value, which is denoted
, and is referred to as the oscillation amplitude index for
that candidate HFO. If
does not exceed a threshold
(the oscillation amplitude
index threshold), then the entire event is rejected as a candidate HFO. If
, on
the other hand, then we apply one final test. The purpose of the final test is to exclude
high amplitude events that are non-oscillatory, or that are over-damped. To perform this
test, a frequency domain search is performed on
to identify not only its
maximum value,
, but also its full-width at half-maximum,
. We then require
that
.
If an event meets all of the criteria above for an HFO, then it is accepted as a
candidate HFO, and
is saved as the oscillation amplitude index of that candidate
HFO. The frequency associated with
is denoted
. The oscillation amplitude
index defined in this way has dimensionless units (denoted Z), and represents the number
of standard deviations that
is above the mean.
Note that this algorithm is best suited for identifying the largest amplitude HFO
within a chosen frequency band at any given moment in time. If there is more than one
HFO present at the same time, and if one wishes to identify all HFOs present
(,,)Skft(,,Skft(,,Skft*peakS*peakS0S*peak0SS(,,Skft*peakS*peakW**peakpeakWS*peakS*peakS*peakf*peakS simultaneously, then a more sophisticated algorithm would be needed. For instance, one
may break the target frequency band of interest from 80-1000 Hz into smaller sub-bands,
or one may add additional bands at higher or lower frequency. Then one searches
through each sub-band individually for HFOs in that sub-band. In practice, HFOs with
frequencies greater than 80 Hz are relatively rare and brief events, and it is very
infrequent that more than one HFO is present, per individual channel, at the same time.
Summary of DOOD algorithm. The key definitions of the DOOD HFO detection
algorithm are the V-DOOD spectral density,
, and the oscillation amplitude
index
. The V-DOOD spectral density,
, is defined for every EEG
channel k at every instant in time t. For each candidate HFO, a value for the oscillation
amplitude index
is calculated.
is in essence a time average of
over
the time when a particular HFO is present. A value of
is defined for a given EEG
channel at a given time only if there is an HFO in that channel at that moment in time.
For a candidate HFO to be accepted as an HFO,
must exceed a threshold
, i.e.,
. This threshold may be chosen to be smaller if higher sensitivity is desired,
and larger if higher specificity is desired.
Results
Figure 1 shows a sample of control rat EEG prior to and after bandpass filtering at
80-1000 Hz. There are HFOs apparent after bandpass filtering in electrode channels 2,
14 and 16 (numbered bottom to top) at time t = 4.78 sec. On the right are the V-DOOD
and V-DOOD-SQR time-frequency color contour plots for channel 16. The V-DOOD
(,,)Skft*peakS(,,)Skft*peakS*peakS(,,)Skft*peakS*peakS0S*peak0SS plot is band restricted to the frequency range 80-1000 Hz to show the frequency range
over which we search for an HFO. The V-DOOD-SQR plot is a wide-band color contour
plot which shows the intricate low frequency structure associated with the HFO.
Figure 2 shows another set of HFOs from the control rat EEG which are more
prominent than those of Fig. 1. The HFO in channel 15 has an oscillation amplitude
index of 5.80 standard deviations above the mean. There is again intricate low frequency
structure associated with the HFO. This structure is similar in both channels 1 and 15.
An expanded view of the low frequency structure in channel 15 is shown in Fig.
3a. There is an HFO with
at time 12.24 sec. In addition, there is 750 Hz
activity at 14.75 sec, which does not meet our criteria for an HFO because it has a
bandwidth that is slightly wider than its mean frequency. Note that the 184 Hz HFO and
the 750 Hz activity appear to interrupt the theta oscillation, which resumes after time
. Also note the prominent “beating” effect seen in the theta oscillation. This
beating effect is not an artifact; it is characteristic of well-developed oscillatory rhythms
in the DOOD time-frequency spectra and its detailed analysis can reveal information on
whether the underlying periodicity is sinusoidal or more spike-like, and whether there are
other frequencies to which the dominant frequency is coupled (Hsu, Hsu et al. 2010). In
the time domain, the 750 Hz high frequency activity has a very fast “buzzing” kind of
appearance in all 16 channels. A magnification of the 750 Hz activity, after high pass
filtering at 250 Hz, is shown in Fig. 3b.
Comparison with standard Fast Fourier Transform. Figure 4 shows time-
frequency spectra for the same HFO from channel 15 as in Fig. 2 but now using standard
short-time Fast Fourier Transform (ST-FFT) for time-frequency analysis. In this method,
*peak184 Hzf16 sect each EEG time series is divided into consecutive time windows of duration
. Each
EEG time window of duration
is then zero-padded with a time window also of
duration
. FFT is then performed on each zero-padded time window of duration
. In Fig. 4a,
, while in Fig. 4b,
. The low
frequency structure is much less apparent with ST-FFT as compared to DOOD.
Sensitivity and positive predictive value. A total of 1191 HFOs were visually
identified in the rat data spread across 16 channels, and 1076 HFOs were visually
identified in the human microwire data spread across the 8 channels. In the automated
detection of HFOs, an amplitude index threshold
(in units of standard deviations
above the mean) needs to be specified, above which an event is considered a candidate
HFO. If
is small, then the HFO detection algorithm will be sensitive but not very
specific. If
is large, then the algorithm is more specific but less sensitive. The
specificity of rare events tends to be close to one, even when the positive predictive value
is low. Because HFOs are rare events, we report here the positive predictive value rather
than the specificity. The dependence of sensitivity and positive predictive value of the V-
DOOD HFO detection algorithm is shown in Fig. 5 for both the rat and human data. The
rat data combines 100 seconds of control recordings with 100 seconds of data from the
epileptic rat (6 weeks after kainate-induced status epilepticus). The human data consists
of 120 seconds of continuous recording.
HFO event rate. The control and epileptic rat data are acquired from the same rat.
Is the number of HFO events per unit time different before induced status epilepticus and
6 weeks after? In Fig. 6 we show the HFO event rate for the rat at baseline and 6 weeks
wintwintwint2wint0.084 secwint0.042 secwint0S0S0S after kainate-induced status epilepticus, for a choice of threshold
. The HFO event
rate is much higher for the epileptic rat than for the rat at baseline (2-tailed paired t-test, t
= 0.0002). Note that the event rate in the epileptic rat shows a strongly laminar
distribution, which is not unexpected given the strongly laminar distribution of cell types
in the hippocampus. The EEG channels are spatially separated by 01. mm intervals and
closely approximate the layers of the hippocampus (see the Methods section describing
the rat data). Note that channels 14-16 are all in the hilus of the dentate gyrus and all of
these exhibit high HFO event rates in the epileptic rat. The HFO event rate for the human
microwires is also shown in Fig. 6. The number of HFO events is higher in channels 2, 4,
5 and 7 than in the other channels. The spatial relationship between these channels are
more difficult to determine because the microwires used in these recordings are arranged
as a “spray” that extends from a modified clinical macroelectrode (see the Methods
section on human depth recordings).
Total vs HFO-associated spectral density. The spatial distributions of the V-
DOOD spectral density for the control and epileptic rat are shown in Fig. 7. These
distributions can be calculated in two ways. The total spectral density is calculated by
time-averaging
over all time points in the data file. The HFO-associated spectral
density is calculated by accumulating
only for those times when a candidate
HFO is identified in the frequency range 80-1000 Hz, with
, and then averaging
over those times only. Note again the laminar distribution of the spectral density in all 4
contour plots.
In the total spectral density for the control rat, there are prominent frequency
bands at 7 Hz (the theta oscillation), 10-60 Hz (alpha-beta-gamma), 600 Hz and 2000 Hz.
03S(,,)Skft(,,)Skft03S We have discussed the 600 Hz and 2000 Hz bands earlier (Hsu, Hsu et al. 2010). The
theta and gamma bands tend to co-localize in channels 2, 4, 6, 12, 14 and 16, while the
600 Hz and 2000 Hz bands tend to co-localize in channels 1, 3, 5, 7, 9, 11, 13, and 15.
The HFO-associated spectral density of the control rat shows that the broad 10-60 Hz
band splits into separate alpha, beta and gamma bands. It is also apparent that there are
oscillators in the 100-200 Hz frequency range although they are not as prominent as the
lower frequency oscillators. The 100-200 Hz oscillators appear to co-activate with a
continuum of lower frequency activity spanning the theta, alpha, beta and gamma ranges,
but they do not co-activate with higher frequency activity in the 600 Hz and 2000 Hz
bands.
The total spectral density of the epileptic rat shown demonstrates a marked
degradation of the theta oscillation and a relative increase in the gamma band with an
apparent invasion into lower frequencies. The 600 Hz band is no longer visible. The
HFO-associated spectral density of the epileptic rat shows that there are HFOs of
frequency near 200 Hz which co-localize and co-activate with a continuum of lower
frequency activity spanning the theta, alpha, beta and gamma ranges, but they do not co-
activate with higher frequency activity in the 2000 Hz band. There are no longer clear
divisions of lower frequency activity into separate theta, alpha, beta and gamma bands,
but rather these are fused together into a broader band. Note that the 100-200 Hz band is
invisible in the total spectral density in either the control or epileptic rat. As expected,
the HFO-associated spectral density is much more sensitive to brief, transient oscillations
than the total spectral density.
Figure 8 shows the total and HFO-associated spectral densities for the human
microwire data. The total spectral density exhibits bands at 20 Hz, 60 Hz and 700 Hz.
Channel 5 is unique in having a relative more activity at 20 Hz and 60 Hz and less at 700
Hz. The HFO-associated spectral density shows that channel 5 exhibits peak activity at
15 Hz, 85 Hz and 400 Hz, while channels 2, 4 and 7 exhibit peak activity at 700 Hz. This
distribution suggests that the 85 Hz and 400 HFOs (ripples and fast ripples, respectively)
tend to co-localize spatially and co-activate temporally with lower frequency activity in
the alpha-beta range, but not with higher frequency activity in the 700 Hz range. There is
no activity visible in the 2000 Hz range, but this data was low-pass filtered at 1000 Hz.
Figures 7 and 8 show that the HFO-associated spectral density provides a valuable
adjunct to the frequency analysis of EEG data. It is capable of revealing the frequency
structure of transient oscillations that are not apparent in the total spectral density.
Amplitude and duration of HFOs. Figure 9a shows the number of HFO events vs
HFO amplitude index for the control and epileptic rat. There are a similar number of
HFOs of small amplitude index in the control and epileptic rat, but there are more HFOs
of higher amplitude index in the epileptic rat. Figure 9b shows the number of HFO
events vs HFO duration for a choice of
. From this figure, it is seen that the overall
distribution of HFO durations is similar in the control and epileptic rat. However, Figs.
9c and 9d show that there are more HFOs of higher amplitude in the epileptic rat, and
these have time durations of less than 0.1 sec. Figures 9e and 9f show that the
distributions of HFO amplitude indices vs frequency are similar in the control and
epileptic rat, but there are more HFOs of higher amplitude index in the epileptic rat and
these are mostly in the frequency range of 80-300 Hz (ripple range).
01S A summary of Fig. 9 is that HFOs in the control and epileptic rat span a similar
time duration and a similar frequency range, but there are more HFOs of higher
amplitude index in the epileptic rat, and these tend to be of shorter duration (<0.1 sec)
and lower frequency (80-250 Hz, ripple range).
For the human temporal lobe microwire data, Fig. 10 shows plots of the number
of HFO events vs HFO amplitude index, the distribution of HFO amplitudes vs duration,
and the distribution of HFO amplitudes vs frequency. The presence of very high
amplitude HFOs (amplitude index > 6) suggests the presence of epileptogenic tissue.
However, as compared to the rat hippocampus, the highest HFO amplitudes in this
sample occurs in the frequency range 700-800 Hz. These HFOs have a time duration of
30-60 ms.
Discussion
The DOOD algorithm is capable of detecting very brief, high frequency EEG
oscillations. If high sensitivity is desired, the amplitude index threshold may be set at
1, while if high positive predictive value is desired, it may be set at 3 or higher. We find
that the HFO event rate is markedly increased in the epileptic rat compared to a control
rat. Similarly, the HFO-associated spectral density is much more sensitive to brief,
transient oscillations than the total spectral density.
The DOOD spectral densities shown in Figs. (7) and (8) suggest that high
frequency activity may be classified into 4 types: type 1 encompasses the frequency
range 80-250 Hz (
, i.e., ripples); type 2, 250-500 Hz (
,
fast ripples); type 3, 500-1000 Hz (
, the 600 Hz band), and type 4,
0S1.9 to 2.4Logf2.4 to 2.7Logf2.7 to 3.0Logf 1000-5000 Hz (
, the 2000 Hz band). HFOs in the 600 Hz range and
even into the 2000-2600 Hz range have been described in primary somatosensory cortex
in response to peripheral somatosensory stimulation (Cracco and Cracco 1976; Curio,
Mackert et al. 1994; Gobbele, Buchner et al. 1998; Curio 2000; Sakura, Terada et al.
2009). Based on our own results, we suggest that brain areas other than primary
somatosensory cortex may also be capable of producing high frequency activity of types
3 and 4, and that these events can occur spontaneously. A caveat here is that high
frequency activity in the 2000 Hz band appears to be rather complex and is not strictly
oscillatory (see Fig. 3b).
In agreement with others (Staba, Wilson et al. 2002; Bragin, Engel et al. 2010;
Jacobs, Zijlmans et al. 2010; Jacobs, Zijlmans et al. 2010), we find that HFOs especially
of type 1 are more abundant in the epileptic rat compared to the control rat, and are also
abundant in the temporal lobe of the human subject with temporal lobe epilepsy. These
HFOs can be quantified with the HFO event rate and visualized in the frequency domain
using the HFO-associated spectral density. The HFO-associated spectral density is able
to reveal the frequency structure of associated lower and higher frequency activity as well
as of the HFOs themselves in a way that is not possible with the total spectral density.
HFOs tend to be invisible in the total spectral density because they are so rare and brief.
Both the HFO event rate and the HFO-associated spectral density may be useful clinical
markers of epileptogenic tissue.
It is of interest that HFOs of type 1 tend not to co-localize with HFOs of types 3
and 4, either in space or in time. It may be of interest in the future to investigate whether
types 3 and 4 HFOs predominate in normal brain tissue, while types 1 and 2 predominate
3.0 to 3.7Logf in epileptogenic tissue. If true, then epilepsy surgery should aim to resect brain tissue
where types 1 and 2 HFOs predominate, but spare tissue where types 3 and 4
predominate.
There are many other ways to detect HFOs automatically and quantitatively,
among them the traditional Fast Fourier Transform (FFT), wavelet transforms, other
pseudo-wavelet transforms, and supervised (Firpi, Smart et al. 2007) and unsupervised
(Blanco, Stead et al. 2010) learning computer algorithms. The DOOD algorithm is a
pseudo-wavelet transform that is unique among all wavelet and other pseudo-wavelet
transforms in that it has a simple physical analogy in the form of the damped harmonic
oscillator. This physical analogy allows the DOOD spectral density to be defined in
terms of the time rate of change of the detecting harmonic oscillator that is specifically
due to the driving force represented by the EEG data, and not confounded by any other
contribution to the total energy. As a result, DOOD appears to give superior time-
frequency resolution compared to other wavelet and pseudo-wavelet transforms (Hsu,
Hsu et al. 2010).
How does DOOD compare with FFT in detecting HFOs? FFT algorithms are
unsurpassed when one only wishes to know what frequencies are present in a long stream
of EEG data and when one does not need to know when these frequencies appear and
when they disappear. It is very easy, however, for very brief and rare events such as
HFOs to be lost in an ocean of other EEG activity, if FFT is applied to an entire data file.
To detect HFOs using FFT, it is necessary to break the total EEG data file into shorter
time segments, and then to apply FFT to each segment to determine the frequency
content of each segment (see Fig. 4). However, such short-time FFT (ST-FFT)
algorithms are severely limited by the uncertainty principle, which states that improving
time resolution necessarily degrades frequency resolution and vice versa. For instance, if
the time segments are of duration
, then the frequency resolution becomes
. For the rat data, if we take
, for example, then
.
At this level of resolution, one cannot hope to study delta or theta range activity. If one
tries to improve time resolution by taking
, then the frequency resolution
further degrades to
. The loss of low frequency resolution using ST-FFT
was demonstrated in Fig. 4 which is to be compared to the DOOD results of Fig. 2. Note
that the HFO at t = 12.24 sec appears more smeared out along the frequency axis when
one uses the shorter FFT time window
. In contrast, the effective time
window used for the DOOD calculations in Fig. 2 was much shorter, at
,
and yet the DOOD frequency resolution is better. While DOOD is also subject to the
uncertainty principle (Hsu, Hsu et al. 2010), its manifestation appears to be much less
severe in DOOD than in ST-FFT.
Note that in generating the ST-FFT contours Fig. 4, we actually zero-pad the EEG
data to remove artifactual end-effects that arise from truncating the EEG time series. One
might suppose that such zero-padding improves the frequency resolution of the ST-FFT
time-frequency spectrum thus obtained, but it does not, because zero-padding contains no
information about longer timescale phenomena
We have also been struck by the intricate low frequency structure that
accompanies HFOs, most apparent on V-DOOD-SQR color contour plots. HFOs never
occur in isolation and are always accompanied by activity in other frequency bands,
especially at lower frequency bands. Others have suggested that this low frequency
wint1/winft0.084 secwint12 Hzf0.042 secwint24 Hzf0.042 secwint0.005 secwint structure may help distinguish pathological from normal HFOs (Nagasawa, Juhasz et al.
2011). The remarkable capability of DOOD to time-resolve low frequency structure
makes it ideally suited for this purpose. We conjecture that each HFO may have its own
wide-band “signature,” which may then facilitate its classification into normal vs
pathological. Such wide-band analysis of HFOs may be useful not only for clinical
purposes (in the development of an EEG biomarker of epileptogenic tissue), but also for
basic science purposes, in pointing to a need to explain not only the HFOs themselves,
but also the accompanying lower frequency structure.
Computer learning algorithms such as neural network approaches (Firpi, Smart et
al. 2007) may eventually prove superior to DOOD in detecting HFOs. Supervised
learning algorithms, however, entail a certain degree of subjectivity in determining which
EEG waveforms truly are HFOs and which are not. The DOOD algorithm, furthermore,
is a general purpose time-frequency transform. It can also be used to define
instantaneous phases of oscillation and amplitude-amplitude and phase-amplitude
correlation functions (Hsu, Hsu et al. 2010).
One limitation of the current study is that visual identification of HFOs was
necessary against which the DOOD algorithm can be compared. There is subjectivity in
visual identification that cannot be entirely eliminated. Some of the visually identified
HFOs with smaller DOOD spectral densities may not be deemed true HFOs by other
expert investigators. More stringent standards for HFOs would likely result in the
elimination of HFOs with small DOOD amplitude indices, and thus would likely improve
the apparent sensitivity of the DOOD HFO detection algorithm.
Further, our experience has been that some high frequency EEG oscillations are
not necessarily of high amplitude when visualized in the time domain, even after
appropriate bandpass filtering. They are distinguished from nearby segments of EEG
only in being more nearly sinusoidal. These oscillations, if brief, can be exceedingly
difficult to identify visually. These oscillations may nonetheless show up as prominent
peaks in a DOOD spectrogram because they drive the corresponding resonant DOOD
oscillators of similar frequency in a coherent way. In this case, one may argue that the
DOOD spectral density may be a more objective measure of the presence or absence of
EEG oscillations than visual inspection.
The concept of discrete EEG oscillations that appear and disappear, moreover,
may not do justice to the complexity of high resolution time domain EEG. Sometimes
EEG oscillations have a gradual onset, and sometimes they have a waxing and waning
course. In this case, there is a certain unavoidable degree of arbitrariness when one
defines onset and end times for these oscillations. It may be that, for the purpose of
identifying epileptogenic tissue, these complexities are of minor importance, but it may
also be that, for purposes of understanding the details of brain dynamics, a more nuanced
description of EEG oscillations is needed. The basic DOOD time-frequency spectra may
help to provide such a more nuance description.
Another limitation of our study is that the data sample is quite small, only 120
seconds of human 8-channel microwire data and 200 seconds of 16-channel wide-band
rat local field data. Visual identification of HFOs is very time intensive. It has been
estimated by others that it takes between 3 minutes to 3 hours to mark one minute of one
channel of EEG for the presence of ripples and fast ripples, depending on how active that
channel is (Urrestarazu, Chander et al. 2007). Our experience has been similar. In this
regard, it would be helpful to have a comparison EEG database, with HFOs identified
consensually by multiple experts in the field, on which to test proposed algorithms for
detecting HFOs. Our goals here are limited mainly to demonstrating the DOOD
algorithm as a candidate algorithm and suggesting what to expect for sensitivities and
positive predictive values when using the DOOD algorithm. Future studies will address
much longer EEG recordings, and more widely distributed channels, so that recordings
near known epileptogenic regions may be compared with that near normal brain.
Conclusions
DOOD may be a useful way to quantify HFOs for use in long-term EEG analysis.
In addition, DOOD time-frequency analysis reveals an intricate low frequency structure
that accompanies HFOs that may be useful for wide-band characterization of HFOs. In
future, such wide-band characterization may help differentiate pathological HFOs from
normal brain oscillations.
Acknowledgements: We are grateful to Sue Osting for help with the rat histology.
Funding was provided by NIH R01-25020 and NIH R01-NS63039-01.
References
Blanco J A, Stead M, Krieger A, Viventi J, Marsh WR, Worrell GA, Litt B.
Unsupervised classification of high-frequency oscillations in human neocortical
epilepsy and control patients. J Neurophysiol 2010;104:2900-12.
Bragin A, Engel J, Staba RJ. High-frequency oscillations in epileptic brain. Current
Opinion in Neurology 2010;23:151-6.
Cracco RQ, Cracco JB. Somatosensory evoked potential in man: far field potentials.
Electroencephalogr Clin Neurophysiol 1976;41:460-6.
Curio G. Linking 600-Hz "spikelike" EEG/MEG wavelets ("sigma-bursts") to cellular
substrates: concepts and caveats. J Clin Neurophysiol 2000;17:377-96.
Curio GB, Mackert BM, Burghoff M, Koetitz R, Abraham-Fuchs K, Harer W.
Localization of evoked neuromagnetic 600 Hz activity in the cerebral
somatosensory system. Electroencephalogr Clin Neurophysiol 1994;91:483-7.
Engel J, Bragin A, Staba R, Mody I. High-frequency oscillations: what is normal and
what is not? Epilepsia 2009;50:598-604.
Firpi H, Smart O, Worrell G, Marsh E, Dlugos D, Litt B. High-frequency oscillations
detected in epileptic networks using swarmed neural-network features. Ann
Biomed Eng 2007;35:1573-84.
Gardner AB, Worrell GA, Marsh E, Dlugos D, Litt B. Human and automated detection of
high-frequency oscillations in clinical intracranial EEG recordings. Clin
Neurophysiol 2007:118:1134-43.
Gobbele R, Buchner H, Curio G. High-frequency (600 Hz) SEP activities originating in
the subcortical and cortical human somatosensory system. Electroencephalogr
Clin Neurophysiol 1998:108:182-9.
Greenwood DD. A cochlear frequency-position function for several species--29 years
later. J Acoust Soc Am 1990;87:2592-605.
Hellier JL, Patrylo PR, Buckmaster PS, Dudek FE. Recurrent spontaneous motor seizures
after repeated low-dose systemic treatment with kainate: assessment of a rat
model of temporal lobe epilepsy. Epilepsy Res 1998;31:73-84.
Hsu D, Hsu M, Grabenstatter HL, Worrell GA, Sutula TP. Time-frequency analysis using
damped-oscillator pseudo-wavelets: Application to electrophysiological
recordings. J Neurosci Methods 2010;194:179-92.
Jacobs J, Zijlmans M, Zelmann R, Chatillon CE, Hall J, Olivier A, Dubeau F, Gotman J.
High-frequency electroencephalographic oscillations correlate with outcome of
epilepsy surgery. Ann Neurol 2010;67:209-20.
Jacobs J, Zijlmans M, Zelmann R, Olivier A, Hall J, Gotman J, Dubeau F. Value of
electrical stimulation and high frequency oscillations (80-500 Hz) in identifying
epileptogenic areas during intracranial EEG recordings. Epilepsia 2010;51:573-82.
Nagasawa T, Juhasz C, Rothermel R, Hoechstetter K, Sood S, Asano E. Spontaneous and
visually driven high-frequency oscillations in the occipital cortex: Intracranial
recording in epileptic patients. Hum Brain Mapp. 2011; Epub Mar 25.
Sakura Y, Terada K, Usui K, Baba K, Usui N, Umeoka S, et al. Very high-frequency
oscillations (over 1000 Hz) of somatosensory-evoked potentials directly recorded
from the human brain. J Clin Neurophysiol 2009;26:414-21.
Staba RJ, Wilson CL, Bragin A, Fried I, Engel J. Quantitative analysis of high-frequency
oscillations (80-500 Hz) recorded in human epileptic hippocampus and entorhinal
cortex. J Neurophysiol 2002;88:1743-52.
Urrestarazu E, Chander R, Dubeau F, Gotman J. Interictal high-frequency oscillations
(100-500 Hz) in the intracerebral EEG of epileptic patients. Brain 2007;130(Pt 9):
2354-66.
Worrell GA, Gardner AB, Stead SM, Hu S, Goerss, Cascino GJ, Meyer FB, Marsh R,
Litt B. High-frequency oscillations in human temporal lobe: simultaneous
microwire and clinical macroelectrode recordings. Brain 2008;131(Pt 4):928-37.
Zijlmans M, Jacobs J, Kahn YU, Zelmann R, Dubeau F, Gotman J. Ictal and interictal
high frequency oscillations in patients with focal epilepsy. Clin Neurophysiol
2010;122(4):664-71.
Figure captions
Figure 1. Sample HFOs in control rat at time
. (a) 16 channel EEG from
control rat with no bandpass vs time. Each EEG channel is Z-normalized to itself by
subtracting out the mean and dividing by the standard deviation. The hash marks
between consecutive channels represents 10 standard deviations. (b) Same as (a) but
after bandpass filtering at 80-1000 Hz. (c) V-DOOD time-frequency color contour plot
for channel 16. The frequency axis (vertical scale) is shown in terms of
and it spans the frequency range from 50-1000 Hz (
to 3).
(d) V-DOOD-SQR time-frequency color contour plot for channel 16. The frequency axis
(vertical scale) is shown in terms of
and it spans the frequency range
from 2-1000 Hz (
to 3).
Figure 2. Sample HFOs in control rat at time
. (a) 16 channel EEG
from control rat with no bandpass vs time. Each EEG channel is Z-normalized to itself.
The hash marks between consecutive channels represents 10 standard deviations. (b)
Same as (a) but after bandpass filtering at 80-1000 Hz. (c) V-DOOD-SQR time-
frequency color contour plot for channel 15. The frequency axis (vertical scale) is shown
in terms of
and it spans the frequency range from 2-1000 Hz
(
to 3). (d) Same as (c) but for channel 1. Note that the V-DOOD-SQR plots
of (c) and (d) are not band-passed. The table on the right lists the corresponding
oscillation amplitude indices
for these HFOs in units of standard deviations above
4.78 sectLog([Hz])Logff1.7LogfLog([Hz])Logff0.25Logf12.25 sectLog([Hz])Logff0.25Logf*peakS the mean, along with their associated frequencies
in Hz. Blank spaces in the table
indicate absence of an HFO in the corresponding EEG channel at that moment in time.
An asterisk after a number in the first column indicates that the corresponding time-
domain EEG discharge was visually identified as being an HFO.
Figure 3. (a) Expanded view of V-DOOD-SQR time-frequency structure for channel 15
of the control rat. The frequency axis (vertical scale) is shown in terms of
and it spans the frequency range from 1 to 3162 Hz (
to
3.5). There is an HFO with
at time 12.24 sec, diffuse 750 Hz activity at
14.75 sec, as well as a 7 Hz (
) theta oscillation which is interrupted by high
frequency activity and resumes after time t = 16 secs. The prominent “beating” effect
seen in the theta oscillation is not an artifact; it is characteristic of well-developed
oscillatory rhythms in the DOOD time-frequency spectra (see text). (b) 16 channel EEG
of control rat vs time after bandpass filtering at 80-1000 Hz. In the time domain, the 750
Hz high frequency activity has a very fast “buzzing” kind of appearance in all 16
channels. A magnification of the 750 Hz activity, after high pass filtering at 250 Hz, is
shown in Fig. 3b. Each EEG channel is Z-normalized to itself. The hash marks between
consecutive channels represents 10 standard deviations.
Figure 4. Time-frequency analysis of an HFO using short-time fast Fourier transform
(ST-FFT), frequency vs time. The color intensity represents a Z-score calculated by
*peakfLog([Hz])Logff0Logf*peak184 Hzf0.85Logf subtracting out the mean of the FFT spectral density and dividing by the standard
deviation. This HFO is the same one shown in Figs. (2) and (3), channel 15. (a) ST-FFT
for time windows of duration
. Note the HFO at 184 Hz with several
satellite peaks visible (the satellite peaks are artifact due to the FFT algorithm). (b) ST-
FFT for shorter time windows of duration
. Note the blurring of the
frequency structure with the shorter time window.
Figure 5. Sensitivities and positive predictive values vs threshold amplitude index,
.
On the left are results from 16 channels of rat intrahippocampal data, which includes 100
seconds of control rat data and 100 seconds from epileptic rat. On the right are results
from 8 channels of human microwire data from a depth electrode in the temporal lobe.
These results are calculated from 120 seconds of microwire data.
Figure 6. HFO event rates for rat and human data with amplitude index threshold of
. The number of HFO events per second is plotted against EEG channel for rat (a)
and human microwire (b) data. The rat data comes from a 16-channel shank with
electrode contacts that are 0.1 mm apart. Note the increase in HFO event rates after
epileptogenesis. Also note the strongly laminar structure.
Figure 7. Total vs HFO-associated spectral density for the control and epileptic rat, EEG
channel vs logarithm of the frequency. For the HFO-associated spectral densities, an
amplitude index threshold of
is chosen. The total spectral densities are in the top
row, with HFO-associated spectral densities in the bottom row. The control rat data are
0.084 secwint0.042 secwint0S03S03S on the left, and the epileptic rat on the right. Spectral density intensities are shown on a
color scale representing the Z-score, which is calculated by subtracting out the mean and
dividing by the standard deviation.
Figure 8. Total (left) vs HFO-associated (right) spectral densities for the human
microwire data, EEG channel vs logarithm of the frequency. For the HFO-associated
spectral density, an amplitude index threshold of
is chosen. Spectral density
intensities are shown on a color scale representing the Z-score, which is calculated by
subtracting out the mean and dividing by the standard deviation.
Figure 9. Amplitude, duration and frequency of HFOs in control and epileptic rat. (a)
HFO event rate vs amplitude index (here denoted S), with lowest amplitude index
threshold taken to be
. (b) HFO event rate vs duration, with amplitude index
threshold taken to be
. (c) Scatter plot of HFO amplitude index vs HFO duration for
the control rat. (d) Scatter plot of HFO amplitude index vs HFO duration for the epileptic
rat. (e) Scatter plot of HFO amplitude index vs HFO frequency for the control rat. (f)
Scatter plot of HFO amplitude index vs HFO frequency for the epileptic rat.
Figure 10. Amplitude, duration and frequency of HFOs in human microwire data. (a)
HFO event rate vs amplitude index (here denoted S), with lowest amplitude index
threshold taken to be
. (b) Scatter plot of HFO amplitude index vs HFO duration. (c)
Scatter plot of HFO amplitude index vs HFO frequency.
03S01S01S01S Fig 1
4.54.64.74.84.95.00.51.01.52.02.53.0(d) V-DOOD-SQR, Ch 16t (sec)Log( f [Hz] )0102030404.54.64.74.84.95.0246810121416EEG amplitude (dimensionless units)t (sec)(a) Full band (0-6000 Hz)4.54.64.74.84.95.0246810121416EEG amplitude (dimensionless units)t (sec)(b) Bandpass 80-1000 Hz4.54.64.74.84.95.01.82.02.22.42.62.83.0t (sec)Log ( f [Hz] )04.08.01216(c) V-DOOD, Ch 16 Fig 2
1.67*
5.80*
3.40*
3.52*
1.82*
0.78
2.33*
-
114
184
184
184
184
846
95
-
4.04*
184
-
1.82*
0.52
3.37*
0.73
2.19*
3.02*
-
95
846
184
700
184
184
12.112.212.312.4246810121416(a) Full band (0-6000 Hz)EEG amplitude (dimensionless units)t (sec)12.112.212.312.4246810121416EEG amplitude (dimensionless units)t (sec)(b) Bandpass 80-1000 Hz11.011.512.012.513.00.51.01.52.02.53.0t (sec)Log ( f [Hz] )08.0162432(c) V-DOOD-SQR, Ch 1511.011.512.012.513.00.51.01.52.02.53.0t (sec)Log ( f [Hz] )010203040(d) V-DOOD-SQR, Ch 1*peakS*peakf Fig 3a
111213141516171819200.00.51.01.52.02.53.03.5t (sec)Log( f [Hz] )08.0162432(a) V-DOOD-SQR, channel 15 Fig 3b
14.614.714.814.915.0246810121416EEG amplitude (dimensionless units)t (sec)(b) EEG, high pass filter = 250 Hz Fig 4
11.011.512.012.513.050100150200250300t (sec)f (Hz)00.250.500.751.0(b) ST-FFT, Ch 15, twin = 0.042 sec11.011.512.012.513.050100150200250300(a) ST-FFT, Ch 15, t win = 0.084 secf (Hz)t (sec)00.250.500.751.0 Fig 5
01234560.00.20.40.60.81.0S0 Sens PPVRat, control and epileptic01234560.00.20.40.60.81.0S0 Sens PPVHuman microwires Fig 6
2468101214160.00.20.40.60.81.0Events/secChannel Control rat Epileptic rat(a) Rat HFO event rate, S0 = 3123456780.000.050.100.150.200.250.30Events/secChannel(b) Human microwire HFO event rate, S0 = 3 Fig 7
0.00.51.01.52.02.53.03.5246810121416Log( f [Hz] )Channel01234Total spectral density, control rat0.00.51.01.52.02.53.03.5246810121416ChannelLog( f [Hz] )Total spectral density, epileptic rat012340.00.51.01.52.02.53.03.5246810121416ChannelLog( f [Hz] )HFO-associated spectral density, control rat012340.00.51.01.52.02.53.03.5246810121416HFO-associated spectral density, epileptic ratChannelLog( f [Hz] )01234 Fig 8
0.00.51.01.52.02.53.012345678Total spectral densityChannelLog( f [Hz] )012340.00.51.01.52.02.53.012345678HFO-associated spectral densityChannelLog( f [Hz] )01234 Fig 9
0123456789101101001000 Control rat Epileptic ratEvents/100 secS(a) HFO event rate vs amplitude index, S0 = 11.82.02.22.42.62.83.012345678910(b) HFO amplitude index vs Log(f), S0 = 1Log( f [Hz] )S Control rat Epileptic rat123456789100.010.11HFO duration (sec)S Control rat Epileptic rat(c) HFO amplitude index vs duration0.00.20.40.60.81101001000Events/100 secHFO duration (sec) Control rat Epileptic rat(d) HFO event rate vs duration, S0 = 1 Fig 10a
246810121101001000Events/120 secS(a) HFO event rate vs amplitude index Fig 10b
0.010.11024681012SHFO duration (sec)(b) HFO amplitude index vs duration Fig 10c
1001000024681012Sf (Hz)(c) HFO amplitude index vs frequency |
1610.04255 | 1 | 1610 | 2016-10-13T20:32:37 | The Influence of Streamlined Music on Cognition and Mood | [
"q-bio.NC"
] | Recent advances in sound engineering have led to the development of so-called streamlined music designed to reduce exogenous attention and improve endogenous attention. Although anecdotal reports suggest that streamlined music does indeed improve focus on daily work tasks and may improve mood, the specific influences of streamlined music on cognition and mood have yet to be examined. In this paper, we report the results of a series of online experiments that examined the impact of one form of streamlined music on cognition and mood. The tested form of streamlined music, which was tested primarily by listeners who felt they benefited from this type of music, significantly outperformed plain music on measures of perceived focus, task persistence, precognition, and creative thinking, with borderline effects on mood. In contrast, this same form of streamlined music did not significantly influence measures assessing visual attention, verbal memory, logical thinking, self-efficacy, perceived stress, or self-transcendence. We also found that improvements in perceived focus over a 2-month period were correlated with improvements in emotional state, including mood. Overall the results suggest that at least for individuals who enjoy using streamlined music as a focus tool, streamlined music can have a beneficial impact on cognition without any obvious costs, while at the same time it may potentially boost mood. | q-bio.NC | q-bio | The Influence of Streamlined Music
on Cognition and Mood
Julia A. Mossbridge, M.A., Ph.D., [email protected]
Staff Scientist and Director of the Innovation Lab, Institute of Noetic Sciences
Visiting Scholar, Northwestern University Department of Psychology
Science Director, Focus@Will Labs
I think I should have no other mortal wants, if I could always
have plenty of music. It seems to infuse strength into my
limbs, and ideas into my brain. Life seems to go on without
effort, when I am filled with music.
― Mary Ann Evans (George Eliot), The Mill on the Floss
Abstract1 –Recent advances in sound engineering have
led to the development of so-called "streamlined music"
designed to reduce exogenous attention and improve
endogenous attention. Although anecdotal reports
suggest that streamlined music does indeed improve
focus on daily work tasks and may improve mood, the
specific influences of streamlined music on cognition and
mood have yet to be examined. In this paper, we report
the results of a series of online experiments that
examined the impact of one form of streamlined music
on cognition and mood. The tested form of streamlined
music, which was tested primarily by listeners who felt
they benefited from this type of music, significantly
outperformed plain music on measures of perceived
focus,
task persistence, precognition, and creative
thinking, with borderline effects on mood. In contrast,
this same form of streamlined music did not significantly
influence measures assessing visual attention, verbal
memory, logical thinking, self-efficacy, perceived stress,
or self-transcendence. We also found that improvements
in perceived focus over a 2+month period were
correlated with
in emotional state,
including mood. Overall the results suggest that at least
for individuals who enjoy using streamlined music as a
focus tool, streamlined music can have a beneficial
impact on cognition without any obvious costs, while at
the same time it may potentially boost mood.
Key Words – music, streamlined music, cognition,
precognition, creative thinking, cognitive flow.
improvements
1 See Conflict of Interest and data availability statements at the end of this
article.
BACKGROUND AND MOTIVATION
Music has been used to influence our thoughts and
feelings for millennia. In modern times, as we watch a
movie, are led to think a phrase just spoken is important
because of a shift in the musical score, just as we are led to
believe that the heroine is about to meet her enemy at the
moment the music changes. Aside from the world of film, in
our everyday lives we listen to music at least partially
because we believe it will help us concentrate, get things
done, and shift our moods (for review, see [1]).
Scientific examinations of the influence of music on
cognition and mood are becoming more commonplace,
though they still represent a minority of studies examining
the influence of sensory input on cognition. Two hypotheses
dominate these studies: 1) music will be detrimental to
cognitive tasks to the extent that the listener devotes
attentional resources to processing the music [2], and 2)
music will improve performance on cognitive tasks to the
extent that the listener's state of arousal and mood are
improved by the music [3]. Which is correct? They are not
mutually exclusive hypotheses, and it is likely they are both
correct – music helps boost mood and arousal, but if the
listener devotes attention to the music, any gain in cognition
induced by increased mood and arousal is lost.
the
Overall, the relationship between music and the mind
seems to depend on the task and the music. For example,
high-intensity, fast-rhythm versions of a Mozart Sonata
hindered reading comprehension [4], while a slower version
of the same piece boosted visuo-spatial cognition [3].
Further, familiar music may be more distracting than
unfamiliar music [4], although data supporting this idea are
sparse. A meta-analysis examining
influence of
background music on different kinds of tasks found no effect
overall, but did find positive effects on motor-related tasks
such as athletic skill [5]. The authors of that meta-analysis
asked researchers to examine particular types of music and
their effects on particular types of tasks. This paper and the
experiment it reports is one response to that request.
We examined a type of music called streamlined music,
which is electronically recorded music designed with an aim
of increasing an individual's focus on cognitive tasks by
improving endogenous attention and reducing exogenous
attention, essentially supporting the listener in entering a
state of flow [6]. We define streamlined music as unfamiliar
Cognitive Impact of Streamlined Music -- Mossbridge
1
music to which one or more methods have been applied in
order to reduce the activation of exogenous attentional
mechanisms in the listener. These methods could include
removing frequencies
to be
distracting, or ensuring that the music changes slowly over
time. We tested the hypothesis that streamlined music, when
played in the background while a listener performs cognitive
tasks, produces
(i.e., non-
streamlined) music.
improvements over plain
that have been reported
We predicted that all cognitive measures would be
positively influenced by streamlined music, as would mood.
We were agnostic about whether streamlined music would
influence self-efficacy, perceived stress, or a listener's sense
of self-transcendence, but we
it would be
informative to examine self-reports for each of these factors
as well, as they are related to the emotional state of a
listener. Finally, we hypothesized that improvements in
perceived focus ability would support positive shifts in
emotional states.
thought
METHODS
The aim of this Methods section is to give the reader a
basic sense of what was done in this experiment; for anyone
interested in attempting an exact replication, please contact
the author.
through
http://www.focusatwill.com).
Participants. We solicited participants
the
homepage of a startup that delivers streamlined music for
background listening during computer-based work tasks
(Focus@Will,
Only
participants who wanted to participate were given the link to
the experiment.2 Participants were told that we were
performing a cognitive research experiment consisting of
four testing sessions, and they would receive increasing
numbers of free days of the music service in exchange for
their participation in each of the tests. Note that this
participant payment method eventually
screens out
participants for whom the service is not desirable (as some
of them might not be motivated towards continuing the
testing). Thus the conclusions drawn from the experiment
described here should only be used to help understand the
influence of streamlined music on listeners who enjoy using
streamlined music as a focus tool.
Procedure. All four testing sessions were conducted
online via web software (Word Press and Java), in the homes
of the participants, at the time of their choosing. The delay
between each of the first three testing sessions was 48 hours
to one week. The delay between Testing Session Three and
Testing Session Four was 60-80 days. As with any multi-
stage study, we found
that a decreasing number of
participants performed each successive testing session.
Testing Session One (no background music). Participants
were asked not to listen to any background music during this
2 Note that while consent forms were not used, this offering the link to the
experiment only to participants who pressed a button indicating that they
wanted to participate amounts to implied consent.
the
testing session. We obtained basic demographic information
including gender and age. Participants completed a brief
Big-5 personality trait inventory [7] and several emotional-
state surveys, including the Brief Mood Introspection Survey
or BMIS [8], a general self-efficacy questionnaire [9], a
perceived stress scale questionnaire [10] and a questionnaire
that we had specifically created to assess the experience of
self-transcendence or
the feeling of connection with
something beyond ourselves (available upon request). After
completing
surveys, participants
performed an arrow flanker task [11] and a dimensional
change task [12] to acclimate them to cognitive testing.
Finally, participants were given a one-item self-report
survey to assess their perceived focus on a scale of 1 (low
focus) to 5 (high focus).
emotional-state
the music channels on
Testing Session Two (participant's choice of music or
streamlined music). Half of the participants, randomly
selected, were asked to listen to their own choice of
background music and the other half were asked to listen to
their own choice of streamlined music. Streamlined music
was selected from
the site
http://www.focusatwill.com, and plain music could be
selected from any other source (e.g., radio, internet music
streaming site, digital music files). Participants completed
the same mood, general self-efficacy, perceived stress scale,
and self-transcendence questionnaires that they completed in
Testing Session One. Then they performed the arrow flanker
task and the dimensional change task that they performed in
Testing Session One, followed by a verbal memory task that
was also a test for implicit precognition [13], the Test of
Logical Thinking or TOLT, form A [14], and an alternative
uses creative thinking task [15]. At the end of this testing
session, participants were asked to rank their focus during
the session from 1 to 5.
(no
Testing Session Three (participant's choice of music or
streamlined music). To complete the crossover design, the
half of the participants who were asked to listen to their own
choice of background music in Testing Session Two were
now asked to listen to their own choice of streamlined
music, vice versa. The remainder of this testing session was
identical with Testing Session Two, except we used TOLT
form B [16].
Testing
Session Four
background music).
Participants were asked not to listen to any background
music during this testing session. The purpose of this testing
session was to compare scores on self-rated focus and the
emotional state surveys to baseline scores on these same
measures from Testing Session One.
verification. Because
experiment was
performed online, we could not be sure participants in
Testing Sessions One and Four were actually not listening to
background music, nor could we be sure participants in
Testing Sessions Two and Three were listening to the type
of music that was assigned to them. In an attempt to
determine whether participants followed instructions, all
participants completing both Testing Sessions Two and
Three were asked what they listened to during these
Data
the
Cognitive Impact of Streamlined Music -- Mossbridge
2
RESULTS
Nine hundred and nine participants completed the first
testing session, 157 of these completed Testing Sessions
Two and Three by our deadline, and 50 of those participants
completed Testing Session Four by our deadline. Ten of the
157 participants who completed Testing Sessions Two and
Three were not able to follow instructions about background
music listening and were removed from all analyses. In
addition, participant numbers vary slightly across tasks due
to some participants not completing each task successfully
as a result of technical errors. We note that the bulk of the
participants who completed all four testing sessions were
those who also were randomly assigned to perform Testing
Session Two while listening to streamlined music as
opposed to plain music (40 out of 50 participants), a
significant effect according to an exact binomial test
(probability=0.8, chance=0.5, p<0.00002). We can assume
these participants were largely motivated to continue testing
as a result of their desire to obtain free access to the
streamlined music service, but it is not clear why this should
be more likely to be the case for participants who heard
streamlined music first. We speculate on the interpretation of
this unexpected result in Conclusions.
Visual Attention
We used 30 trials of an arrow-flanker task to assess
visual attention. In this task, participants report as quickly as
possible the direction of a central arrow (left versus right)
that is surrounded by distracting arrows. These arrows can
either be pointing in the same direction as the central arrow
the opposite direction
(congruent) or pointing
(incongruent). The dependent variables are
the mean
response time (for correct trials) and mean proportion
correct on congruent versus incongruent trials.
For the 150 participants who successfully completed the
arrow flanker task in both Testing Sessions Two and Three,
we found no differences between streamlined music and
plain music in either the mean response times or mean
proportion correct on congruent and incongruent trials. As
expected from the original work on a related flanker task
[11], mean response times were slower on incongruent than
on congruent trials, but this was the case regardless of
background music (see Table 1; streamlined music,
t[149]=7.98;
t[149]=8.35;
p<0.000001). There was no significant difference between
any of the four dependent variables in the two conditions (all
p-values > 0.057). The only borderline effect was the
incongruent condition, on which response times were
marginally faster by a mean difference of 20 milliseconds
when the background music was plain music as compared to
streamlined music.
plain music,
p<0.000001;
in
For
sessions. Participants who either did not understand the
question and therefore could not be expected to understand
task instructions, or who responded that they listened to
either no music, their choice of music both times, or
streamlined music both times were removed from the
analysis (10 participants removed in total).
Data analysis: Streamlined versus plain music. To
compare the effects of streamlined music to those of plain
background music, we performed within-participant
comparisons using two-tailed t-tests, except for the data
from the combined verbal memory and implicit precognition
task (see below). Specifically, we compared data between
the two music-listening conditions from participants who
completed both Testing Sessions Two and Three.
the combined verbal memory and
implicit
precognition task, we only used data from Testing Session
Two, which was the first time participants performed this
task. This is because this was a task in which words from a
word list had to be memorized, and therefore we did not
want be concerned about interference from previous word
lists, a concern that would have been introduced if we
considered data from Testing Session Three. While the
combined verbal memory and implicit precognition task was
also performed in Testing Session Three, we ignored these
data, and simply used the task to keep the tasks and timings
between the two testing sessions equivalent.
Data analysis: Independent validation of perceived
focus. We wanted to compare self-ratings of perceived focus
across the two background music conditions (streamlined
music versus plain music), but we were aware of a flaw with
this plan. Specifically, participants reaching completion of
Testing Session Three were likely to be at least partially
motivated by receiving an extension of their free trial at the
streamlined music service, and therefore would likely be
biased toward validating this choice by ranking their
experience of focus higher during the streamlined versus
plain background music sessions. So we used performance
on the TOLT as an objective measure of focus, reasoning
that being more focused should produce higher scores on
this task. Further, participants did not receive feedback on
their performance on the TOLT in either Testing Session
Two or Three, and therefore they could not have used this
score as a way to assess their experience of focus; thus we
could use it as both an objective and independent measure of
focus.
Data analysis: The relationship between changes in
perceived focus and emotional state. To examine whether
changes in perceived focus were correlated with changes in
positive emotional states, we first calculated difference
scores (i.e., Testing Session Four minus Testing Session
One) for each of the four emotional state surveys (BMIS,
PSS, NGSE, and the Self-Transcendence Scale). We then
used linear regression to compare these difference scores to
those calculated from the focus survey in the same two
testing sessions.
Cognitive Impact of Streamlined Music -- Mossbridge
3
Streamlined
Music
Plain Music
RT Cong. RT Inc.
879
(190)
860
(172)
943
(178)
923
(162)
PC Cong.
0.993
(0.02)
0.994
(0.02)
PC Inc.
0.988
(0.03)
0.992
(0.03)
Table 1. Means and standard deviations (in parentheses) of response times
(RT, in ms) on correct trials and proportion correct values (PC) for
congruent (Cong.) and incongruent (Inc.) trials in the arrow-flanker task for
assessing visual attention, for both listening conditions. See text for details.
Task Persistence
To assess task persistence, we used a dimensional change
task consisting of 50 trials. On each trial, just prior to
presenting a cartoon image of an animal, we displayed the
dimension ("shape" or "color") by which the image was
required to be sorted. This task can be used to assess
executive function, and specifically cognitive flexibility
[12]. This is because after a trial on which participants have
sorted the image by one dimension (e.g., "shape"), there is a
tendency to continue sorting according to that dimension
even if the dimension is changed on the next trial (e.g.,
"color"). Participants tend to respond more slowly on these
"dimension-shift" trials than on trials with the same
dimension as the previous trial. The ability to respond
accurately and quickly on dimension-shift trials can be seen
as a measure of high cognitive flexibility.
Alternatively, because the dimension cue is an exogenous
one (a word flashed on the screen prior to each image), and
because an endogenous attentional set towards continuing
the same task is a type of task persistence, one can view less
accurate performance on dimension-shift trials as compared
to other trials as a measure of task persistence. On this view,
the participant is ignoring or not fully processing the
exogenous cue in favor of continuing the previous task. This
interpretation would be contingent on visual attention not
being negatively affected by
the manipulation being
examined (which is the case here, see above). However,
interpreting increased response times on dimension-shift
trials in this same light is not as straightforward as
interpreting reduced accuracy, because response times are
only calculated for correct trials, and therefore on any trials
for which response time is calculated the participants have
already made the switch correctly to the new task. Thus here
we will interpret lesser accuracy on dimension-shift trials as
compared to non-shift trials as an indication of high task
persistence.
for dimension-shift versus non-shift
While listening to both kinds of music, participants were
significantly less accurate and slower when responding to
dimension-shift trials as compared to other trials (proportion
correct
trials:
streamlined music, t[148]=8.06, p<0.000001; plain music,
t[148]=6.03, p<0.000001; response times for dimension-shift
t[148]=5.68,
versus non-shift
p<0.000001; plain music, t[148]=6.98, p<0.000001; see Table
2). However, we found the proportion-correct difference
score (i.e., proportion correct on dimension-shift trials minus
non-shift trials) was significantly higher for streamlined as
compared to plain music (t[148]=2.03, p<0.05, see Figure 1),
while there was no significant difference between the
trials: streamlined music,
response-time difference scores (p>0.278). These results,
when taken together with the novel interpretation of reduced
accuracy on dimension-shift trials, indicate a statistically
significant effect in which streamlined music can be thought
of as boosting task persistence, potentially via reducing
exogenous attention.
Streamlined
Music
Plain Music
RT NonS. RT DimS.
1005
(234)
970
(204)
1058
(234)
1036
(208)
PC NonS.
0.985
(0.03)
0.986
(0.03)
PC DimS.
0.950
(0.06)
0.962
(0.05)
Table 2. Means and standard deviations (in parentheses) of response times
(RT, in ms) on correct trials and proportion correct values (PC) for "non-
shift" trials on which the dimension was not different from the previous trial
(NonS) and "dimension-shift" trials on which the dimension was different
from the previous trial (DimS) in the dimensional change task, for both
listening conditions. See text for details.
Figure 1. Mean task persistence scores in the dimensional change task, as
measured by the difference between proportion correct on trials on which
the dimension was not different from the previous trial and "dimension-
shift" trials on which the dimension was different from the previous trial.
Left blue bar: streamlined music; right red bar: plain music. Error bars
indicate +/- one standard error of the mean (S.E.M.) within participants.
Verbal Memory
In the verbal memory task, participants were shown a list
of 48 nouns that were randomly selected for each participant
from a set of 96 nouns. These were displayed one at a time
in quick succession, and participants were asked
to
memorize the words in this original word list as well as they
could. Then we asked participants to perform a two-
alternative forced-choice word-recognition test, in which
their goal on each trial was to use their mouse to select
which of two words was the one that was in the original
word list. We asked them to perform as quickly and
accurately as possible. For reasons discussed in Methods, we
only analyzed data from this task obtained from Testing
Session Two (N=144 for streamlined music and N=160 for
plain music, all from Testing Session Two; these numbers
are larger than for the other tasks because we were not
constrained to participants who also performed Testing
Session Three). As a result, these comparisons were
necessarily between participant groups, so the statistical
power of these comparisons is lower than in the within-
group analyses.
Cognitive Impact of Streamlined Music -- Mossbridge
4
We found no significant difference between streamlined
music and plain music in accuracy or response times on the
word-recognition test (both p-values>0.206; see Table 3).
These results suggest that streamlined music and plain music
did not differentially affect verbal memory as assessed in
this task.
Streamlined
Music
Plain Music
RT
1547
(282)
1584
(270)
PC
0.797
(0.18)
0.768
(0.21)
Table 3. Means and standard deviations (in parentheses) of response times
(RT, in ms) on correct trials and proportion correct values (PC) in the verbal
memory task, for participants in either listening condition. See text for
details.
Implicit Precognition
these
In addition to being useful for examining verbal memory,
we also used the verbal memory task as a test for implicit
precognition. Implicit precognition is the nonconscious
ability to predict future events that should not be predictable
[13]. After the participant completed an initial word-
recognition test (see Verbal Memory above), 24 words from
the original word list were randomly selected to be
reinforced using training, and these words represented the
"future events" that should not have been predictable on the
initial word-recognition test, as they had not been selected at
the time this test was taken. These words were selected
independently for each participant. The dependent variable
was
that
participants correctly remembered on the initial word test, as
compared to the randomly selected 24 words on the original
list that were not to be trained.
to-be-trained words
the number of
Participants practiced the words selected for training in
two ways. First, from a list of these 24 words they took four
trials to select the words belonging to each of the four
categories the nouns represented (people, animals, food, and
clothing), a task that required them to attend to the trained
words. Second, they saw a picture of each of these 24 words
and they were asked to type the appropriate word under each
picture. They were not allowed to continue the task until
they typed the noun correctly, again requiring them to attend
to each of the trained words. Following this training, they
were administered a second two-alternative forced-choice
word-recognition test on all 48 words in the original word
list. The purpose of this second word-recognition test was to
determine whether the training had been effective. Thus,
participants who recalled fewer trained words than untrained
words on this second test were dismissed from the analysis,
as we assumed they did not pay attention during the practice
portion of the task, or that the training did not work for
them. This data-selection step was pre-planned, as the
motivation for this step drove the decision to include the
second word-recognition test in the first place. We also
introduced a non-planned data-selection step, in which we
ignored data from participants who scored 90% correct or
above on the first word-recognition task, as we reasoned that
the memories of these individuals were likely too good to be
improved by subtle cues from future training, should such
cues exist. After the data from both the streamlined and plain
music groups were processed via these two steps, 97
participants remained in each group.
Based on the remaining data, we found that participants
listening to streamlined music recognized more to-be-trained
words than not-to-be-trained words on the initial word-
recognition test (t[96]=2.41, p<0.018), while this was not the
case for participants listening to plain music (see Table 4
and Figure 2). The
interaction effect was borderline
significant (t[192]=1.82, p<0.07). Note that prior to the second
(not pre-planned) data-selection step, the trend was in the
same direction as after the second data-selection step (to-be-
trained recall vs. not-to be trained recall, streamlined:
t[135]=1.63, p<0.106; plain music: t[142]=-0.51, p>0.609)3.
To-be-
trained
words
18.52
(4)
18.01
(4)
Not-to-be-
trained
words
17.90
(4)
18.11
(4)
Streamlined
Music
Plain Music
Table 4. Means and standard deviations (in parentheses) of the number of
words out of 24 recognized on the initial word-recognition test in the verbal
memory task for the groups of participants remaining after two data-
selection steps, for participants in either listening condition. Columns
indicate recognition of words to be trained in the future (left column) versus
those not to be trained in the future (right column). See text for details.
Figure 2. Mean numbers of correctly remembered words in the implicit
precognition task, for the 24 randomly determined words to be trained in the
future (left column) versus the 24 remaining words that would not be
trained (right column). Blue symbols give means for participants listening
to streamlined music in Testing Session Two, red symbols give means for
participants listening to plain music in Testing Session Two. Error bars
indicate +/- one S.E.M. within participants.
Logical Thinking
logical
To examine
thinking under both
listening
conditions, we used the Test of Logical Thinking (TOLT;
[14,16]), which assesses mathematical and deductive
3 These data are also presented as part of a review of implicit precognition
in a conference proceedings journal that is now in press [13].
Cognitive Impact of Streamlined Music -- Mossbridge
5
reasoning. We found no significant difference between
listening conditions for scores of the 148 participants who
successfully completed this task in both Testing Sessions
Two and Three (t[147]=0.450, p>0.653; see Table 5). This
result suggests that background music had no differential
effect on logical thinking that could be detected with the
TOLT.
personality scores. Examining the creative quality scores of
these participants indicated that streamlined music had a
significant effect on participants with low openness, but the
effect was only borderline significant for participants with
high openness (streamlined vs. plain,
low openness:
t[61]=3.07, p<0.004; high openness: t[70]=1.90, p<0.07).
Streamlined music boosted the mean quality of creative
thinking of participants with low openness above the level of
individuals with high openness listening to plain music
(Figure 4).
Figure 4. Mean quality of creative thinking demonstrated on the alternative
uses tasks, measured by an independent judge blind to the listening
condition, separated between individuals with openness scores less than 4
on a scale of 1 to 5 (left bars) and those with openness scores of 4 or greater
on a scale of 1 to 5 (right bars) based on a median split on the personality
trait of openness. Blue columns on the left give the means for participants
as they listened to streamlined music, red columns on the right give the
means for the same participants as they listened to plain music. Error bars
indicate +/- one S.E.M. within participants.
Cognitive Impact of Streamlined Music -- Mossbridge
6
Streamlined
Music
Plain Music
Number
7.77
(4.85)
7.16
(3.86)
Quality
3.47
(2.05)
2.97
(1.58)
Table 6. Means and standard deviations (in parentheses) of the number and
quality of creative uses in the alternative uses task, for both listening
conditions. Scale for quality was 0-10; number and quality of uses were
ranked by separate independent judges blind to the listening condition. See
text for details.
Figure 3. Mean quality of creative thinking demonstrated on the alternative
uses tasks, measured by an independent judge blind to the listening
condition. Blue left column gives the mean for participants as they listened
to streamlined music, red right column gives the mean for the same
participants as they listened to plain music. Error bars indicate +/- one
S.E.M. within participants.
Streamlined
Music
Plain Music
TOLT score
12.72
(4.14)
12.84
(4.50)
tube
Table 5. Means and standard deviations (in parentheses) of the number of
correct items out of 18 on the TOLT for both listening conditions. See text
for details.
Creative Thinking
To assess creative thinking, we used the alternative uses
task [15], in which participants are asked to list as many
creative uses of mundane household objects as possible. We
used two independent creativity judges to judge the quality
and quantity of these uses (i.e., one judge examined only
quality without regard to quantity, and the other examined
only quantity without regard to quality). These judges were
blind to the listening conditions from which their data were
obtained. For Testing Session Two, the mundane object was
a two-liter bottle and for Testing Session Three, the object
was a cardboard paper-towel tube. We found that the two-
liter bottle used as the mundane object in Testing Session
Two produced more uses, on average, than the cardboard
paper-towel
listening condition),
introducing concerns about counterbalancing. There were 86
people who performed the while listening to streamlined
music in Testing Session Two and 71 who performed this
task while listening to plain music in Testing Session Two,
so streamlined music was over-represented for the more
productive creativity task. To make our analysis more
conservative in light of this fact, we only analyzed data from
the first 71 participants in each listening group in each test
session, to give a total of 142 participants. Note that after
this data elimination the results were essentially the same as
for the original data, which included all participants who
completed both testing sessions.
(regardless of
We found that when participants listened to streamlined
music as compared to plain music they produced marginally
more alternative uses and had a significantly higher quality
of creative uses (number: t[141]=1.78, p<0.078; quality:
t[141]=3.17, p<0.002; see Table 6 and Figure 3). Additionally,
there was a significant and positive linear correlation
between quality scores on
task
performed while listening to plain music with the personality
trait of openness (r=0.173; p<0.05), but this correlation was
not present for the quality scores on the same task while
listening to streamlined music (r=0.033, p>0.705).
the alternative uses
To examine the relationship between the personality trait
of openness and background music more carefully, we
performed a median split on openness for
the 133
individuals for whom we had successfully recorded
Emotional State
state
We obtained complete emotional state assessments from
157 participants who completed both Testing Sessions Two
and Three. The average responses on each of the four
emotional
surveys were more positive when
participants were listening to streamlined music versus when
they were listening to plain music (Table 7), though none of
the emotional state measures showed any significant
difference between listening conditions (all p-values >
0.196). The scale showing the greatest influence from
streamlined music was the Brief Mood Introspection Scale
(BMIS) for pleasant versus unpleasant mood (Figure 5).
Streamlined
Music
Plain Music
BMIS
46.20
(7.26)
45.47
(7.28)
PSS*
2.76
(0.65)
2.77
(0.58)
NGSE
33.46
(5.49)
33.38
(5.97)
S-T
32.47
(5.90)
32.30
(5.73)
Table 7. Means and standard deviations (in parentheses) of responses to the
emotional state surveys. BMIS = Brief Mood Introspection Scale, (pleasant
vs. unpleasant); PSS = Perceived Stress Scale; NGSE = New General Self-
Efficacy Scale; S-T= Self-Transcendence Scale. See text for details. *For
PSS, higher scores are negative (greater perceived stress); for all other
scales, higher scores are positive.
Figure 5. Mean of responses on the Brief Mood Introspection Scale (BMIS)
for participants as they listened to streamlined music (left blue column) or
plain music (right red column). Higher values indicate more pleasant mood.
Error bars indicate +/- one S.E.M. within participants.
Perceived Focus
At the end of each testing session, participants reported
their perceived level of focus during the session. We
recorded 148 of these self-reported focus responses from
participants who performed both Testing Sessions Two and
Three. The responses supported the idea that participants felt
they were more focused when they were listening to
(t[147]=4.55,
streamlined as opposed
p<0.00002; Table 8). As noted in Methods, our process of
obtaining participants was likely to bias our sample towards
those who felt more focused while listening to streamlined
music.
to plain music
To handle these data more conservatively, we examined
whether self-reported focus measures correlated with an
objective and independent measure. Because there were no
differences between listening conditions on the TOLT (see
above), and because this test was the second-to-last test in
in
their self-assessment of focus. Note
each testing session (the last test being the alternative-uses
task) so that it was relatively close in time to the focus
survey, we chose to use scores on the TOLT as an objective
measure with which to examine whether participants were
accurate
that
participants did not receive feedback on their scores for the
TOLT, so any correlation between the TOLT scores and
perceived focus could only be due to focus being correlated
to the TOLT performance itself, not due to participants'
knowledge of their performance on the TOLT. We found
that among all participants who completed Testing Sessions
Two and Three, there was a significant positive correlation
between self-rated focus and scores on the TOLT form used
in Testing Session Three (r=0.313; p<0.0002) but the same
was not true for scores obtained in Testing Session Two
(r=0.036, p>0.667). Thus if we examine only self-rated
focus for Testing Session Three, this more conservative
method again
to
streamlined music perceived themselves as significantly
more focused than participants who listened to plain music
in Testing Session Three (t[146]=2.47, p<0.015; Figure 6).
Note that despite the correlation between perceived focus
and the TOLT in Testing Session Three, there was no
significant difference between TOLT scores in listening
conditions in this testing session.
that participants
listening
reveals
Streamlined
Music
Plain Music
Focus
3.78
(0.80)
3.43
(0.79)
Table 8. Means and standard deviations (in parentheses) of responses to the
focus survey in both Testing Sessions Two and Three; higher values
indicate greater perceived focus. See text for details.
Figure 6. Mean of self-reported focus scores in Testing Session 3 on a scale
of 1 to 5 for participants; these scores were significantly correlated with
scores on the logic test (TOLT). Higher values indicate greater perceived
focus. Left blue column shows the mean for participants listening to
streamlined music and right red column shows the mean for participants
listening to plain music. Error bars indicate +/- one S.E.M. between
participants.
Changes in Self-Reported Focus and Emotional State
We hypothesized that improved focus supports positive
changes in emotional state, and vice versa. To partially test
this hypothesis, we calculated the changes in self-reported
Cognitive Impact of Streamlined Music -- Mossbridge
7
focus scores and emotional state measures between Testing
Sessions One and Four for the 46 participants who
performed Testing Session Four (see Methods). Note that
these two testing sessions were separated by 60-100 days.
We found that for three of the four emotional state measures
(BMIS, PSS,
the Self-Transcendence Scale),
improvements in self-reported focus correlated significantly
with improvements in emotional state (Table 9). This finding
supports our hypothesis that improved focus supports an
improved emotional state and vice versa, though of course
causality has not been established.
and
r=-0.302
p<0.05
Δ S-T
r=0.420
p<0.004
Δ BMIS Δ PSS* Δ NGSE
r=-0.009
r=0.368
p<0.013
p>0.952
correlation
versus Δ focus
Table 9. Correlations and significance levels for change scores derived
from emotional state measures and change scores derived from the focus
survey (Testing Session Four minus Testing Session One). BMIS = Brief
Mood Introspection Scale, (pleasant vs. unpleasant); PSS = Perceived Stress
Scale; NGSE = New General Self-Efficacy Scale; S-T = Self-
Transcendence Scale. *For PSS, higher scores are negative (greater
perceived stress); for all other scales, higher scores are positive. See text for
details.
CONCLUSIONS AND DISCUSSION
The hypothesis that all cognitive tasks and mood would
be positively influenced by streamlined music was partially
supported by these data. Specifically, participants had scores
that were significantly better on four measures while
listening to streamlined, as opposed to plain, music: task
persistence, implicit precognition, creative thinking, and
perceived focus. The most impressive result was the strong
effect on creative thinking, and the most controversial result
was the effect on implicit precognition. We discuss both of
these findings below. There were no differences between
listening conditions in scores on visual attention, verbal
memory, logical thinking. While emotional states were
generally better while listening to streamlined music, with
mood showing the largest effect, these differences were not
statistically significant.
There were no tasks on which plain music produced
significant decrements in performance as compared to
streamlined music, unless the dimensional change task is
interpreted as a cognitive flexibility measure instead of a
task persistence measure. However, in our case, we were
testing the overall hypothesis that streamlined music reduces
exogenous attention and improves endogenous attention.
This hypothesis
the
dimensional change information as an exogenous cue, much
like a smartphone notification that appears while the goal of
proceeding with an ongoing task is being attended via
endogenous attention. Even if one is determined to interpret
the results on the dimensional change task as representing
decreased cognitive flexibility in participants as they listen
to streamlined music, the extent of the effect is small enough
as to be negligible (percent correct on dimension-switch
trials, streamlined music: 95%, plain music: 96%).
interpretation of
instructed our
to an exact binomial
Interestingly, we found that the bulk of the participants
who completed all four testing sessions were those who also
were randomly assigned to perform Testing Session Two
while listening to streamlined music as opposed to plain
music (40 out of 50 participants), a significant effect
according
test (probability=0.8,
chance=0.5, p<0.00002). We can assume these participants
were largely motivated to continue testing as a result of their
desire to obtain free access to the streamlined music service.
This result may suggest that having the opportunity to first
perform cognitive tasks using streamlined music rather than
plain music may have made it easier for these participants to
believe that streamlined music improved their performance
and focus. This could be because these participants could
have noticed that when they performed the same tasks again
but listening to plain music in Testing Session Three, they
did not perform as well as in Testing Session Two.
Meanwhile, participants who listened to plain during Testing
Session Two could easily decide that practice was the reason
that they improved on certain tasks in Testing Session Three,
and might not have attributed their improvement to the fact
that they were listening to streamlined music in Testing
Session Three.
to warrant
intriguing enough
The results from the implicit precognition and creativity
measures are
further
elaboration. As to implicit precognition, this type of
experiment has a rich history that is beyond the scope of this
paper (for review, see [17]). However, for the purposes of
understanding the present experiment, it is important to note
that a previous version of the implicit precognition task used
here has proven to be difficult to replicate [18-19]. The
authors of a recent meta-analysis propose that one reason for
this lack of replication could be that the original version of
this task required lengthy deliberation in order to recall the
words on the original list, and they had found that implicit
precognition tasks requiring much faster responses were
more likely to be replicated [19]. As a result of this
observation, we adjusted the task so that responses had to be
made quickly. But only participants who
to
streamlined music showed the original effect, indicating that
perhaps it is not only response speed that influences whether
the original effect is replicated or not. Along these lines, it is
interesting
that original experiment,
participants did the task only after listening to music that had
some similarities to streamlined music [18]. Thus it is
possible that streamlined music puts people in a state in
which subtle cues from future events can be received,
processed, and/or used more effectively. Replication is
necessary and is underway.
to note
listened
that
in
The influence of streamlined music on creative thinking
was surprisingly strong. We speculate that the reason for this
strong effect is that creative thinking requires the cognitive
resources
to be differentially recruited by
streamlined music. Specifically, streamlined music had a
mild positive effect on task persistence; when participants
listen to streamlined music they were likely less focused on
exogenous cues and more focused on their endogenous
that seem
Cognitive Impact of Streamlined Music -- Mossbridge
8
attentional set. To maintain the process of searching for
creative uses of a mundane task over time undoubtedly
requires endogenous attention. In addition, streamlined
music improved a measure of implicit precognition, an
improvement that can be thought of as the result of better
processing of and access to unconscious information. At the
same time, creative thinking is supported in situations
facilitating
to
unconscious information [20-22], and this may explain the
persistent relationship between creativity and openness [23-
24]. Taken together, it appears there are some similarities
between the mechanisms underlying creativity and implicit
precognition. Finally, streamlined music also improved
perceived focus, an experience that is often necessary for
bursts of creativity, and has been alternately called "flow"
[25].
unconscious
processing
access
and
Overall, the results of this study indicate that there is a
reasonable empirically based argument for using streamlined
music to support focus, creativity, and flow during work.
Although this effect may be specific to those who enjoy
listening to streamlined music as they work, it is possible
that even those who do not enjoy streamlined music could
potentially benefit from it. However, the answer to this
question is in the domain of future examinations of the
influences of streamlined music, and music in general, on
cognition and mood.
ACKNOWLEDGMENTS
Daryl Bem provided ideas related to creating the verbal
memory/precognition task for online use.
REFERENCES
[1] North, A. C., Hargreaves, D. J., & Hargreaves, J. J.
(2004). Uses of music in everyday life. Music
Perception: An Interdisciplinary Journal, 22(1),
41-77.
[2] Lavie, N., Hirst, A., De Fockert, J. W., & Viding,
E. (2004). Load theory of selective attention and
cognitive control. Journal of Experimental
Psychology: General, 133(3), 339-354.
[3] Husain, G., Thompson, W. F., & Schellenberg, E.
G. (2002). Effects of musical tempo and mode on
arousal, mood, and spatial abilities. Music
Perception: An Interdisciplinary Journal, 20(2),
151-171.
[4] Thompson, W. F., Schellenberg, E. G., & Letnic,
A. K. (2012). Fast and loud background music
disrupts reading comprehension. Psychology of
Music, 40(6), 700-708.
[5] Kämpfe, J., Sedlmeier, P., & Renkewitz, F. (2010).
The impact of background music on adult listeners:
A meta-analysis. Psychology of Music,
0305735610376261.
No. 13/843,585.
[6] Henshall, W. R. (2013). U.S. Patent Application
[7] Rammstedt, B., & John, O. P. (2007). Measuring
personality in one minute or less: A 10-item short
version of the Big Five Inventory in English and
German. Journal of Research in Personality, 41(1),
203-212.
[8] Mayer, J. D., & Gaschke, Y. N. (1988). The
[9] Chen, G., Gully, S. M., & Eden, D. (2001).
experience and meta-experience of mood. Journal
of Personality and Social Psychology, 55, 102-111.
Validation of a new general self-efficacy scale.
Organizational research methods, 4(1), 62-83.
[10] Cohen, S. & Williamson, G. (1988) Perceived
Stress in a Probability Sample of the United States.
Spacapan, S. and Oskamp, S. (Eds.) The Social
Psychology of Health. Newbury Park, CA: Sage.
[11] Eriksen, B. A., & Eriksen, C. W. (1974). Effects of
noise letters upon the identification of a target letter
in a nonsearch task. Perception & Psychophysics,
16(1), 143-149.
[12] Zelazo, P. D. (2006). The Dimensional Change
[14] Tobin, K. G., & Capie, W. (1981). The
Card Sort (DCCS): A method of assessing
executive function in children. Nature Protocols –
Electronic Edition, 1(1), 297.
[13] Mossbridge, J. A. (In press). Examining the nature
of retrocausal effects in biology and psychology.
American Institute of Physics Conference
Proceedings.
development and validation of a group test of
logical thinking. Educational and Psychological
Measurement, 41(2), 413-423.
[15] Torrance, E. P. (1972). Predictive validity of the
Torrance Tests of Creative Thinking. The Journal
of Creative Behavior. Vol 6(4), 236-252
[16] Mann, A. E. (1982). The effects of a problem-
solving strategy on the long-term memory of
algorithms. UNF Theses and Dissertations. Paper 16.
Transcendent mind: Rethinking the science of
consciousness. Washington, D.C.: American
Psychological Association Books.
[17] Mossbridge, J. A. & Baruss, I. (2016).
[18] Bem, D. J. (2011). Feeling the future: experimental
evidence for anomalous retroactive influences on
cognition and affect. Journal of Personality and
Social Psychology, 100(3), 407.
[19] Bem, D., Tressoldi, P., Rabeyron, T., & Duggan,
M. (2015). Feeling the future: A meta-analysis of
90 experiments on the anomalous anticipation of
random future events. F1000Research, 4.
[20] Zhong, C. B., Dijksterhuis, A., & Galinsky, A. D.
(2008). The merits of unconscious thought in
creativity. Psychological Science, 19(9), 912-918.
[21] Bowden, E. M., & Beeman, M. J. (1998). Getting
the right idea: Semantic activation in the right
hemisphere may help solve insight problems.
Psychological Science, 9(6), 435-440.
Cognitive Impact of Streamlined Music -- Mossbridge
9
[22] Baird, B., Smallwood, J., Mrazek, M. D., Kam, J.
W., Franklin, M. S., & Schooler, J. W. (2012).
Inspired by distraction mind wandering facilitates
creative incubation. Psychological Science.
[23] Feist, G. J. (1998). A meta-analysis of personality
in scientific and artistic creativity. Personality and
Social Psychology Review, 2(4), 290-309.
[24] Feist, G. J., & Barron, F. X. (2003). Predicting
creativity from early to late adulthood: Intellect,
potential, and personality. Journal of Research in
Personality, 37(2), 62-88.
psychology of discovery and invention. New York:
Harper Collins.
[25] Csikszentmihalyi, M. (1996). Flow and the
CONFLICT OF INTEREST DISCLOSURE: At the time of
conducting and reporting on these experiments, the author
held a contract position as
the Science Director at
Focus@Will, the streamlined music service used in the
experiments.
DATA AVAILABILITY: All raw data are available upon
request, so any individual can verify or contest the results
shown here.
Cognitive Impact of Streamlined Music -- Mossbridge
10
|
1905.10023 | 1 | 1905 | 2019-05-24T04:09:24 | Separation Effect of Early Visual Cortex V1 Under Different Crowding Conditions A TMS Study | [
"q-bio.NC"
] | The visual crowding makes it difficult to identify the patterns in peripheral vision, but the neural mechanism for this phenomenon is still unclear because of different opinions. In order to study the separation effect of V1 under different crowding conditions, single-pulse transcranial magnetic stimulation is applied within the right V1. The experimental design includes two factors: TMS intensity (10%, 65%, and 90% of the phosphene threshold) and crowding (high and low) conditions. The accuracy results show that there is a strong interaction between crowding condition and TMS condition. When the TMS stimulation intensity is lower than the phosphene threshold, more crowding will be perceived under the high crowding condition, and less crowding will be perceived under the low crowding condition. The above results conclude that the high and low crowding condition separate by TMS stimulation. The results support the assumption that the crowding is related to V1 and occurs in the visual coding phase. | q-bio.NC | q-bio | Separation Effect of Early Visual Cortex V1 Under Different
Crowding Conditions：A TMS Study
Xieyi Liu, Junjun Zhang, Ling Li*
Key Laboratory for Neuroinformation of Ministry of Education, School of Life Science
and Technology, University of Electronic Science and Technology of China, Chengdu
610054, P. R. China
* Corresponding author: [email protected]
Abstract: The visual crowding makes it difficult to identify the patterns in peripheral
vision, but the neural mechanism for this phenomenon is still unclear because of different
opinions. In order to study the separation effect of V1 under different crowding conditions,
single-pulse transcranial magnetic stimulation is applied within the right V1. The
experimental design includes two factors: TMS intensity (10%, 65%, and 90% of the
phosphene threshold) and crowding (high and low) conditions. The accuracy results show
that there is a strong interaction between crowding condition and TMS condition. When
the TMS stimulation intensity is lower than the phosphene threshold, more crowding will
be perceived under the high crowding condition, and less crowding will be perceived under
the low crowding condition. The above results conclude that the high and low crowding
condition separate by TMS stimulation. The results support the assumption that the
crowding is related to V1 and occurs in the visual coding phase.
Keywords: visual cortex, visual crowding, phosphene, separation effect, TMS
1 Introduction
When a target is presented with nearby flankers in the peripheral visual field, it
becomes harder to identify, which is referred to as crowding [1]. Crowding is an important
obstacle to object recognition. For ordinary people, peripheral vision is more susceptible
to crowding, while central vision is basically unaffected [2,3]. Studying the neural
mechanism of crowding and related brain regions is helpful to find a breakthrough to solve
diseases such as dyslexia which is related to visual crowding.
At present, the reasons for crowding effect can be classified into two categories in
general [4]: (1) It is considered that the crowding effect occurs in the visual coding phase.
According to this view, extraction or integration of target stimuli has been suppressed or
interfered during the visual coding phase [12,14,22,11]; (2) The target stimuli was not
affected during the visual coding phase, and the participants perform not quite well because
of insufficient attention resolution [8,24-26].
The neural mechanisms involved in the visual crowding are still uncertain, and some
studies supporting that visual crowding effect in the visual coding phase is associated with
V1. Both Blake et al. [12] and Ho et al. [14] found that the crowding decreases the
adaptational aftereffect of the target grating, which indicates that the crowding occurs
before the orientation adaptation, so that they infer the lower visual cortex V1 is the
position of crowding. Millin et al. [15] used fMRI to explore the nerve area related to
crowding. In the state without crowding, when the target and flankers appear at the same
time, the level of blood oxygen (blood oxygen level dependent, BOLD) signal is stronger
than the contrary situation. However due to the suppression between the target and flankers,
the growth of signal reduces when target and flankers appear at the same time instead of
only flankers appear in crowding conditions. They found that this inhibitory effect can be
produced at V1, and the stronger crowding of the stimulus, the stronger suppression of V1.
Chen et al [18]. used ERPs and fMRI to study the neural mechanism of crowding from two
perspectives: time and space. The signal difference between C1 and V1 was related to the
behavioral crowding of participants. The difference between C1 suppression and V1 signal
are stronger when behavioral crowding is stronger. This shows that crowding occurs in V1
and is regulated by attention. Although the studies above all support that V1 is the region
where crowding occurs, the fMRI studies of Arman et al. [16] and Bi et al. [17] did not
find significant crowding effects at V1, instead they found the lowest cortical area
associated with crowding is V2.
At present, the visual crowding is carried out by fMRI and ERP experiments. Those
two experimental methods can only ensure that the changes in signal in V1 occur
simultaneously with the stimulus, but the exact correlation cannot be proved. In order to
further study the correlation between crowding and V1, we performed a transcranial
magnetic stimulation (TMS) experiment. TMS is a technique that produces focal, transient
and fully reversible disruptions in brain activity by delivering strong magnetic pulses to
the cortex that pass through the skull and depolarize the underlying neurons of particular
areas in the brain [13, 10]. This method has been proved to be a useful research approach
to assess the functional and causal role of specific cortical areas in cognition [19]. TMS
experiment can be used to observe changes in results under different crowding conditions
by giving short-term pulse stimulation to specific brain regions. Abrahamyan et al. [20]
found that TMS-induced neuronal activity can be combined with stimulus-induced activity
to enhance visual perception when stimulating the V1 region with intensity below the
phosphene threshold according to TMS experiments. Based on the above conclusion, we
plan to stimulate V1 with different strengths of TMS to study its effects on high and low
crowding conditions. We hypothesized that stimulating the V1 region with a single-pulse
TMS intensity below the phosphene threshold can augment visual perception. Besides, the
visual crowding occurs in the visual coding stage. Under the crowding condition, more
crowding will be felt when the visual perception increases. So that it is more difficult to
integrate and extract the features of stimulate during the visual coding phase, resulting in
the increase of crowding.
In previous studies, it has been found that the appearance of phosphene is related to
V1 and V2. The precise Montreal Neurological Institute (MNI) coordinate is found by the
navigation system of the coil and considered to be the accurate position of V1. Then, the
crowding experiment is started. This method can solve the difficulty of locating V1.
2 Materials and methods
2.1 Participants
In the experiment, we tested a total of 18 participants (7 male). All participants are
right-handed and reported normal or corrected-to-normal vision. Ages range from 19 to 25
years. All participants conform to the following selection requirements: 1) no pregnancy;
2) no metal implants in the body; 3) no history of mental illness. The study was approved
by the UESTC Ethics Board. Written informed consent was obtained from each participant
prior to being tested.
2.2 Stimuli
All the targets and flankers are circular sinusoidal gratings (diameter: 2.12°; mean
luminance: 61.47cd/m2). The background luminance is 61.47cd/m2 [18]. Previous studies
reported that the crowding in the upper field of view is stronger than in the lower field of
view. Therefore, the stimulus was present in the upper left field of view and centered at 8°
eccentricity in the upper left visual quadrant. The orientation of the target was 45°±𝜃,
either left or right rotated. The orientations of the flankers were independently and
randomly selected from 0° to 180° for each trial. The targets and flankers' settings include
high crowding (center distance is 2.33°) and low crowding (center distance is 4.53°).
2.3 Apparatus
TMS pulse stimulation is supported by using Magstim super rapid magnetic
stimulator and air-cooled figure-of-eight coil (Magstim Company Limited, Whiteland,
UK). The coil manually holds the handle horizontally to the right side of the participant
and tangent to the scalp by using the online visualization function of the BrainSight
frameless stereotaxy system (BrainSight Frameless, Rogue Research, Montreal, QC,
Canada).
The stimulation position's MNI coordinates were targeted via the BrainSight
stereotaxic neuronavigator (Rogue Research, Montreal, QC, Canada), equipped with a
Polaris Vicra position sensor system. With the help of Psychtoolbox [5,6], the experimental
program was written by MATLAB (MathWorks) and connected to the TMS instrument to
realize simultaneous triggering of TMS pulses and stimuli. The visual stimulus is displayed
on the computer monitor (refresh rate: 60Hz; resolution: 1024×768) and the background
is gray.
Three TMS stimulation intensities below the phosphenes threshold were set in the
experiment：10%, 65% and 90% of the phosphenes threshold respectively, and were
recorded as 10% TMS, 65% TMS and 90% TMS. The single-pulse TMS stimulation
coincided with the target grating. The three TMS conditions and the sequence of
experiment 1 and experiment 2 were randomly among participants. Because the position
of visual cortex V1 is deep and has great difference among people, it is difficult to locate
it according to the precise coordinates. The Polaris Vicra position sensor system is used
to find the coordinates of that point, which is considered to be an accurate position of V1,
which solves the difficulty of V1 positioning.
3 Design
3.1 Phosphenes threshold test
The position of the coil was fixed according to the MNI coordinates obtained before
with the help of transcranial magnetic stimulation localization cap, and then the stimulation
intensity of TMS was reduced by 2%-3% step size until participants ensure that
phosphenes phenomenon disappeared. The minimum stimulus intensity that produces the
phosphenes reaction was record as phosphenes threshold.
3.2 Orientation discrimination threshold test
experiments 1 and 2. To measure the discrimination threshold, a stimulus (Low or High)
𝜃 was the orientation discrimination threshold (85% correct) for the target in the
was presented for 250ms. The orientation of the target was either 45°+𝜃 or 45°-𝜃.
result. The 𝜃 varied trial by trial and was controlled by the QUEST staircase [29]. We use
Participants were asked to judge the orientation of the target relative to 45° (clockwise or
counterclockwise) and press the left or right button of the keyboard to report the judgment
the same orientation discrimination threshold, which is calculated by the average value of
orientation discrimination threshold in both conditions (Low and High).
3.3 Experiment Procedures
Before we start Section 1 and Section 2, we test all participants to check whether they
can elicit phosphenes while stimulating the right occipital lobe. The participant sat in a
dimly lit room and dark adapted for 5 min and they were asked to relax and look at the
green dot which diameter is 0.5 cm in the center of the black screen. The participant's head
was supported by a sitting posture corrector and the viewing distance is 62 cm from the
monitor. The TMS coil was positioned with the handle oriented to the right side of the
participant in a horizontal direction and is tangent to the scalp. The coil was initially placed
against the back of the participant's head, with the center over an area 3 cm above the inion
and 2 cm right. Single pulses were delivered with intensities reaching 80%-85% of the
stimulator's output and the initial stimulus intensity is determined by the maximum
stimulus intensity that the participant can withstand without causing uncomfortable
reactions. After being stimulated five times in five seconds at the same position, the
participant's will be asked whether they think that there is a change surrounded the green
dot or around the screen, including the appearance of white flash, white dots, white lines,
etc. The coil was moved in steps of 0.5 -- 1 cm and the accurate position will be found which
can induce five times phosphenes reaction [20]. Salminen-Vaparanta et al. [21] found that
V1 and V2 have similar abilities to produce conscious perception that the phosphene
induced by stimulating V1 is brighter than stimulating V2 through TMS and fMRI studies.
Therefore, in present paper, we first find the position of the participants which can produce
phosphene in the right occipital cortex before the experiment, and compared with the
surrounding position to find point that can induce the brightest phosphene phenomenon.
This accurate position is compared with the position nearby, in order to find the point that
can cause the most obvious and brightest phosphenes phenomenon. The navigator and the
fMRI structure image of the participant are used to get the MNI coordinates of the
phosphenes point and recorded it as TMS stimulation coordinates for subsequent
experiments. The coordinates of the phosphene point were obtained by the navigator and
the fMRI structure of the participants, and recorded as the TMS stimulation site of the
experiment.
In the Section 1, the target between flankers are high crowding and the target between
flankers are low crowding in the Section 2. The process of Section 1 and Section 2 are the
same (see Figure 1). Before each trial, there will be a blank screen with time of 750-1250ms
randomly, and then the stimulus will appear 250ms and disappear after. During the whole
process of the experiment, the participant is required to fix their eyes on the black dot at
the center of the screen. There will be a fixed white dashed circle to indicate the position
of the target stimulus. The participant required to pay attention to the white dashed circle,
and judge the direction of the target when it appears. Participants need to judge the
direction of the target stimulus relative to 45° (clockwise or counterclockwise). The
rotation angle 𝜃 for each participant is determined by the orientation discrimination
threshold measured before the experiment. Participants need to press the left or right button
of the keyboard to report the result. After the participant responded, the blank screen before
the next trial will display. During the experiments, the program will record the button status,
reaction times and accuracy of the participant.
Figure 1. Experiment design. The figure on the left shows the screen of each section, and the figure on the
right shows the stimuli are presented in the upper left visual quadrant and the protocol of each section. In
two sections（Section 1: High crowding condition；Section 2: Low crowding condition）, participants are
tested under three TMS condition. Participants need to fixate the black dot in the center of the monitor and
pay attention to upper left field of vision. The grating stimulus is presented for 250ms in upper left field of
vision. During the experiment, a single pulse of TMS is delivered to V1 at the same time as the stimulus
appears.
4 Results
All behavioral results were normally distributed (Kolmogorov -- Smirnov test
values>0.05). Repeated-measures analyses of variance (ANOVAs) is applied to reaction
times (RTs) and accuracy (ACC) results (Bonferroni's corrected). Mauchly's sphericity
test was used for compound symmetry. When the compound symmetry was not satisfied,
Greenhouse -- Geisser correction and post hoc t-test are used. In all tests, an α threshold of
0.05 for assessing statistical significance is consistent.
average orientation discrimination threshold is 6.5125°. The average MNI coordinate of
phosphene position of 18 participants is founded: x=29.9±9.13mm; y=-102.85±8.18 mm;
z=17.29±19.30 mm.
The average phosphene threshold of 18 participants in the experiment is 73% and the
4.1 RTs Results
The 2×3 repeated ANOVA reveal only a main effect of crowding condition
[F(1,17)=21.68, p<0.001] (see Figure 2). There is no main effect of TMS condition
[F(2,16)=0.072, p=0.931]. There has no significant interaction effect between crowding
condition and TMS condition [F(2,16)=0.291, p=0.751]. Under the same crowding condition,
post hoc t-test is used in three TMS conditions (10% TMS, 65% TMS, 90% TMS). There
is significant difference between high and low crowding condition with each TMS
condition [10% TMS, t(17)=2.504, p=0.023; 65% TMS, t(17)=2.346, p=0.031; 90% TMS,
t(17)=3.097, p=0.007].
Figure 2. Mean reaction times (RTs) are shown for three TMS conditions with low and high conditions.
Asterisks mark significant post hoc t-test (p < 0.05)
4.2 Accuracy Result
From the average accuracy of the three TMS conditions, the average accuracy of the
high condition (81.67%) is greater than the average accuracy of the low condition (78.83%)
with 10% TMS condition, and the average accuracy of the high condition is less than the
average accuracy of the low condition with 65% TMS condition (77.39% < 82.28%) and
90% TMS condition (77.5% < 83.11%). The 2×3 ANOVA reveal only a main effect of
crowding condition [F(1,17) = 4.952, p<0.05]. There is no main effect of TMS condition
[F(2.16)=0.100, p=0.906]. There has significant interaction effect between crowding and
TMS condition [F(2,16)=18.393, p<0.001].
Figure 3. Mean accuracy are shown for two crowding conditions with three TMS conditions (10% TMS,
65% TMS, 90% TMS). Asterisks mark significant post hoc t-test (p < 0.05)
Under the same TMS condition, post hoc t-test is used in two crowding conditions
(high, low). The result shows that there has a significant difference in accuracy of two
crowding condition (high, low) with 10% TMS condition [t(17)=2.136, p<0.05], 65% TMS
condition [t(17)=-3.076, p<0.005], 90% TMS condition [t(17)=-3.076, p<0.01].
Under the high crowding condition, Bonferroni corrected is used in three TMS
condition (10% TMS, 65% TMS, 90% TMS) (see Figure 3). There has a significant
performance difference between 10% TMS and 65% TMS [t(17)=2.954, p=0.027], 10%
TMS and 90% TMS [t(17)=3.828, p=0.003]. However, there are no difference between 65%
TMS and 90% TMS [t(17)=-0.079, p=2.814]. Similarly, there has a significant performance
difference between 10% TMS and 65% TMS [t(17)=-2.699, p=0.045], 10% TMS and 90%
TMS [t(17)=-3.341, p=0.012] and no difference between 65% TMS and 90% TMS [t(17)=-
0.531, p=1.806] with low crowding condition. Conclusion is obtained that high TMS
stimulus intensity impact is more obvious then low TMS stimulus intensity on crowding.
5 Discussion
The ANOVA analysis for the reaction times shows that in the case of different
crowding conditions, regardless of the TMS condition, the RTs in the high crowding
condition is significantly smaller than the RTs in the low crowding condition. The
crowding condition has a significant main effect, indicating that crowding phenomenon
exist under different TMS conditions with different intensity. The 2×3 repeated measures
ANOVA analysis of accuracy has a significant main effect of the crowding condition,
indicating that from the situation of three TMS conditions, the overall accuracy in the high
condition is less than the accuracy in the low condition. There is a strong interaction
between crowding condition and TMS condition, means that stimulation on V1 with TMS
inhibited the high condition and promoted the low condition for participants to recognize
the target. Those two effects are significant.
Many previous studies have suggested that the crowding is related to the early visual
cortex and is likely to be related to V1, but these results are previously confirmed by fMRI
experiments [15,18]. The use of gratings to simulate crowding, combining with the TMS
pulse stimulation online, will have an impact on the V1 zone. Previous studies have proved
that the visual sensitivity will improve when stimulating V1 with intensity below the
phosphenes threshold [20]. The crowding effect does not affect the participants' detection
of the target, only affects the recognition of the target. Therefore, the increase of visual
sensitivity will feel more crowding in the condition of high crowding, and the identification
of the target is more difficult. Similarly, less crowding will be felt in the condition of low
crowding and improve the ability to recognize the target, which is accord with our
experimental results [9, 23].
One explanation for the crowding of previous studies is based on long-range
horizontal connections between different populations of neurons, which may be promotion
or suppression [27,28]. Stettler et al. showed that the horizontal connections cover portions
of V1 representing regions of visual space up to eight times larger than classical receptive
fields [7]. In our experiments, the spatial extent of the 8° eccentricity of neurons can
completely cover two gratings. This long-range horizontal connections links neurons with
similar response characteristics, including neurons with similar spatial frequencies in the
experiment [27], which demonstrates that the crowding in the experiment is certain
because of the intrinsic structure of the peripheral visual system. In the case of high
crowding, the establishment of long-range horizontal connections may become easier.
TMS stimulation below the threshold of phosphenes will establish more long-range
horizontal connections, affecting feature extraction, resulting in more crowding.
Conversely, in the case of low crowding, long-range horizontal connections are less stable,
and TMS may break these connections, improve feature extraction, and reduce crowding.
Millin et al. [15] believe that that the main reason for the occurrence of crowding is
the lack of feature integration and segmentation in peripheral vision at the earliest stage of
cortical processing, mainly in V1. The visual system's perception of graphical features is
to integrate certain features in the graphic, and to prevent these features from being
combined with other features. This visual extraction and integration of graphical features
occurs in the visual coding phase. Our experimental results support the conclusion that
visual crowding occurs during the visual coding phase. From the perspective of visual
sensibility and visual feature extraction in the visual coding phase, the improvement of
visual sensitivity will make it easier to integrate the features of the target extracted from
the peripheral visual field with the features of flankers in high crowding condition, which
is difficult to segment. Under the low crowding condition, it is easier to extract features of
the target and the crowding is reduced.
Our experimental results did not rule out the possibility of crowding occurring in
multiple visual processing stages. Because this phenomenon is a complicated process, it
has relationship with the stimulation and experimental methods used in the experiment.
Our work demonstrates the relationship between crowding and V1 and the separation
effect of TMS from different crowding conditions, and is benefit to further clarify the
neural mechanism of visual crowding effect.
Acknowledgement
This work was supported by the National Natural Science Foundation of China under
grants 61773096, 61673087 and 61773092.
References
1. Whitney D, Levi D M. Visual crowding: a fundamental limit on conscious perception
and object recognition. Trends in CognitiveSciences, 2011, 15: 160-168
2. Levi D M. Crowding -- an essential bottleneck for object recognition: a mini-review.
Vision Research, 2008, 48: 635-654
3. Levi D M, Waugh S J. Spatial scale shifts in peripheral vernier acuity. Vision Research,
1994, 34: 2215-2238
4. Fan Z Z, Fang F, Chen J. Neural mechanisms of visual crowding effect. Scientia Sinca
Vitae, 2014, 44: 450-462
5. Brainard D H. The psychophysics toolbox. Spatial Vision,1997, 10: 433-436.
6. Pelli D G. The VideoToolbox software for visual psychophysics: transforming numbers
into movies. Spatial Vision, 1997, 10: 437-442.
7. Stettler DD, Das A, Bennett J, Gilbert CD. Lateral connectivity and contextual
interactions in macaque primary visual cortex. Neuron, 2002, 36:739 -- 750
8. He S, Cavanagh P, Intriligator J. Attentional resolution and the locus of visual awareness.
Nature, 1996, 383: 334-337
9. Watson A B, Pelli D G Quest: a Bayesian adaptive psychometric method. Perception
and Psychophysics, 1983, 33:113-120.
10. Epstein C M, Wassermann E M, Ziemann U. Oxford Handbook of Transcranial
Stimulation. Oxford: Oxford University Press, 2012
11. Nandy A S, Tjan B S. Saccade-confounded image statistics explain visual crowding.
NatureNeuroscience, 2012, 15: 463-469
12. Blake R, Tadin D, Sobel K V, et al. Strength of early visual adaptation depends on
visual awareness. PNAS, 2006, 103: 4783-4788
13. Balconi M. Dorsolateral prefrontal cortex, working memory and episodic memory
processes: insight through transcranial magnetic stimulation techniques. Neuroscience
Bulletin, 2013, 29: 381-389.
14. Ho C, Cheung S H. Crowding by invisible flankers. PLoS ONE, 2011, 6: e28814
15. Millin R, Arman A C, Chung S T, et al. Visual crowding in V1. Cerebral Cortex, 2014,
24(12): 3107-3115
16. Arman A C, Chung S T, Tjan B S. Neural correlates of letter crowding in the periphery.
Journal of Vision, 2006, 6: 804
17. Bi T, Cai P, Zhou T, et al. The effect of crowding on orientation-selective adaptation
in human early visual cortex. Journal of Vision, 2009, 9: 13
18. Chen J, He Y, Zhu Z, et al. Attention-dependent early cortical suppression contributes
to crowding. Journal of Neuroscience, 2014, 34(32):10465-10474.
19. Miniussi C, Rossini P M. Transcranial magnetic stimulation in cognitive rehabilitation.
Neuropsychological Rehabilitation, 2011, 21: 579-601.
20. Abrahamyan A, Clifford C W G, Arabzadeh E, et al. Improving visual sensitivity with
subthreshold transcranial magnetic stimulation. Journal of Neuroscience, 2011, 31(9):
3290-3294.
21. Salminen-Vaparanta N, Vanni S, Noreika V, et al. Subjective characteristics of TMS-
induced phosphenes originating in human V1 and V2. Cerebral Cortex, 2014, 24(10):
2751-2760.
22. Bernard J B, Chung S T. The dependence of crowding on flanker complexity and
target-flanker similarity. Journal of Vision, 2011, 11: 1
23. Pelli D G, Palomares M, Majaj N J. Crowding is unlike ordinary masking:
distinguishing feature integration from detection. Journal of Vision, 2004, 4: 1136-1169
24.
CognitivePsychology, 2001, 43: 171-216
25. Leat S J, Li W, Epp K. Crowding in central and eccentric vision: The effects of contour
interaction and attention. Investigative Ophthalmology and Visual Science, 1999, 40: 504-
512
26. Strasburger H. Unfocused spatial attention underlies the crowding effect in indirect
form vision. Journal of Vision, 2005, 5: 1024-1037
27. Gilbert C D, Wiesel T N. Columnar specificity of intrinsic horizontal and
corticocortical connections in cat visual cortex. Journal of Neuroscience, 1989, 9: 2432 --
2442
28. Fitzpatrick D. Seeing beyond the receptive field in primary visual cortex.
Current Opinion in Neurobiology, 2000, 10: 438 -- 443
Intriligator J, Cavanagh P. The spatial
resolution of visual attention.
|
1105.0866 | 3 | 1105 | 2012-03-12T13:45:18 | A Mathematical Model of Tripartite Synapse: Astrocyte Induced Synaptic Plasticity | [
"q-bio.NC",
"math.DS",
"q-bio.CB",
"q-bio.SC"
] | In this paper we present a biologically detailed mathematical model of tripartite synapses, where astrocytes modulate short-term synaptic plasticity. The model consists of a pre-synaptic bouton, a post-synaptic dendritic spine-head, a synaptic cleft and a peri-synaptic astrocyte controlling Ca2+ dynamics inside the synaptic bouton. This in turn controls glutamate release dynamics in the cleft. As a consequence of this, glutamate concentration in the cleft has been modeled, in which glutamate reuptake by astrocytes has also been incorporated. Finally, dendritic spine-head dynamics has been modeled. As an application, this model clearly shows synaptic potentiation in the hippocampal region, i.e., astrocyte Ca2+ mediates synaptic plasticity, which is in conformity with the majority of the recent findings (Perea & Araque, 2007; Henneberger et al., 2010; Navarrete et al., 2012). | q-bio.NC | q-bio | A Mathematical Model of Tripartite Synapse: Astrocyte Induced
Synaptic Plasticity
Shivendra Tewari†
[email protected]
Systems Science and Informatics Unit
Indian Statistical Institute
8th Mile, Mysore Road
Bangalore 560059, India
Kaushik Majumdar
[email protected]
Systems Science and Informatics Unit
Indian Statistical Institute
8th Mile, Mysore Road
Bangalore 560059, India
In this paper we present a biologically detailed mathematical model of tripartite
synapses, where astrocytes modulate short-term synaptic plasticity. The model
consists of a pre-synaptic bouton, a post-synaptic dendritic spine-head, a synaptic
cleft and a peri-synaptic astrocyte controlling Ca2+ dynamics inside the synaptic
bouton. This in turn controls glutamate release dynamics in the cleft. As a
consequence of this, glutamate concentration in the cleft has been modeled, in which
glutamate reuptake by astrocytes has also been incorporated. Finally, dendritic spine-
head dynamics has been modeled. As an application, this model clearly shows
synaptic potentiation in the hippocampal region, i.e., astrocyte Ca2+ mediates synaptic
plasticity, which is in conformity with the majority of the recent findings (Perea &
Araque, 2007; Henneberger et al., 2010; Navarrete et al., 2012).
1 Introduction
One of the most significant challenges in neuroscience is to identify the cellular and
molecular processes that underlie learning and memory formation (Lynch, 2004).
Cajal originally hypothesized that information storage relies on changes in strength of
synaptic connections between neurons that are active (Cajal, 1913). Hebb supported
this hypothesis and proposed that if two neurons are active at the same time, the
synaptic efficiency of the appropriate synapse will be strengthened (Hebb, 1949).
Synaptic transmission is a dynamic process. Post-synaptic responses wax and wane as
pre-synaptic activity evolves. Forms of synaptic enhancement, such as facilitation,
augmentation, and post-tetanic potentiation, are usually attributed to effects of a
† Shivendra Tewari’s present address is Biotechnology & Bioengineering Center and Department of
Physiology, Medical College of Wisconsin, 8701 Watertown Plank Road, Milwaukee, WI 53226,
USA.
residual elevation in pre-synaptic Ca2+ concentration ([Ca2+]), acting on one or more
molecular targets that appear to be distinct from the secretory trigger responsible for
fast exocytosis and phasic release of transmitter to single action potential (Zucker &
Regehr, 2002). It is now well established that the astrocytic mGluR detects synaptic
activity and responds via activation of the calcium-induced calcium release pathway,
leading to elevated Ca2+ levels. The spread of these levels within micro-domain of one
cell can coordinate the activity of disparate synapses that are associated with the same
micro-domain (Perea & Araque, 2002). The notion of tripartite synapse consisting of
pre-synaptic neuron, post-synaptic neuron and astrocyte has taken a firm root in
experimental (Araque, et al., 1999; Newman, 2003; Perea & Araque, 2007) as well as
theoretical neuroscience (Nadkarni & Jung, 2003; Volman et al., 2007; Nadkarni, et
al., 2008). Astrocytes play crucial roles in the control of Hebbian plasticity (Fellin,
2009).
There is a recent report, that at least in the hippocampus, astrocyte Ca2+ signaling
does not modulate short-term or long-term synaptic plasticity (Agulhon, et al., 2010).
However evidences of astrocytic modulation of synaptic plasticity are more abundant
including in hippocampus (Vernadakis, 1996; Haydon, 2001; Yang et al., 2003;
Andersson, 2010; Henneberger, et al., 2010). Neuronal activities can trigger Ca2+
elevations in astrocytes (Porter & McCarthy, 1996; Fellin, 2009) leading to
concentration increase in adjacent glial cells including astrocytes, which expresses a
variety of receptors (Newman, 2003). These activated receptors increase astrocyte
[Ca2+], and release transmitters, including glutamate, D-serine, ATP (Parpura et al.,
1994; Henneberger et al., 2010) etc. The released gliotransmitters feed -back onto the
pre-synaptic
terminal either
to enhance or
to depress further
release of
neurotransmitter (Newman, 2003; Navarrete & Araque, 2010) including glutamate,
which is mediated by Ca2+ concentration in the pre-synaptic terminal. It is worthy to
note that Ca2+ elevations are both necessary and sufficient to evoke glutamate release
from astrocytes (Haydon, 2001). On the other hand short-term synaptic depression is
caused by depletion of the releasable vesicle pool due to recent release in response to
pre-synaptic action potential (Wu & Borst, 1999). This entire chain of Ca2+ mediated
pre-synaptic activity consisting of both short-term enhancement (STE) and short-term
depression (STD) can be called short-term synaptic plasticity or simply short-term
plasticity (STP).
Synaptic plasticity occurs at many time scales. Usually long-term plasticity (LTP)
happens at a time scale of 30 minutes or more and STP takes less than that (p – 311,
Koch, 1999). Within the ambit of STP, STE has been more widely studied than the
STD. A quantitative definition of STE has been proposed in (Fisher et al., 1997). STE
has been divided into four different temporal regimes, namely fast-decaying
facilitation
(tens of milliseconds), slow-decaying
facilitation
(hundreds of
milliseconds), augmentation (seconds) and post-tetanic potentiation (minutes) (Fisher
et al., 1997).
2
STP is thought to provide a biological mechanism for on-line information
processing in the central nervous system (Fisher et al., 1997) and therefore could be
the key to the formation of working memory and intelligent behavior. A
computational model of how cellular and molecular dynamics give rise to the STP in
the synapses (particularly in the synapses of the hippocampus and the prefrontal
cortex) can be quite useful in understanding intelligent behavior.
In this paper, we present a computational model of astrocyte mediated synaptic
potentiation in a tripartite synapse. The present model is based on experimental work
of Perea & Araque (2007) where they used immature wistar rats for hippocampal slice
preparations. Primarily there are just two models (Nadkarni et al., 2008; Volman et
al., 2007) shedding light over the molecular aspects of astrocyte mediated synaptic
potentiation, where a lot of important details were omitted or were modeled
hypothetically (see Table 1).
Table 1: A Comparison among Nadkarni et al (2008) model, Volman et al (2007) model,
and the proposed model
Volman et al., 2007
Nadkarni et al., 2008
This Paper
No
No
Yes
No
Yes
Yes
Yes / No
Yes / No
Yes / Yes
Yes
Yes
No
Yes
Yes
No
Yes
Yes
Yes / Yes
Yes / No
Yes / No
Yes / Yes
Signaling Processes
Modeled
Bouton Ca2+
Bouton IP3
Synaptic Vesicle /
Glutamate
Astrocytic Ca2+
Astrocytic IP3
Extra-synaptic Vesicle /
Glutamate
Post-Synaptic Current /
Potential
The computational model proposed here makes use of different detailed
biophysical models highlighting specific aspects of astrocyte-neuron signaling. The
following steps have been followed in simulation of our model. (1) Pre-synaptic
action potential train has been generated using the HH model (Hodgkin & Huxley,
1952). (2) Ca2+ concentration elevation in the pre-synaptic bouton incorporating fast
(using single protein properties (Erler et al., 2004)) and slow (using modified Li-
Rinzel model (Li & Rinzel, 1994)) Ca2+ influx. (3) Glutamate release in the synaptic
3
cleft as a two step process (using Bollman et al., (2000) for Ca2+ binding to synaptic
vesicle sensor and, Tsodyks & Markram (1999) for synaptic vesicle fusion and
recycling). (4a) Glutamate modulated enhancement of astrocytic Ca2+ (using astrocyte
specific G-Chi model (De Pitta et al., 2009)). (4b) Glutamate mediated excitatory
post-synaptic current (using Destexhe et al (1999)) and potential (using Tsodyks &
Markram (1997)). (5) Extra-synaptic glutamate elevation is also modeled as a two-
step process (using modified Bertram model (Bertram et al., 1996) to fit Synaptic-
Like Micro-vesicle (SLMV) release probability determined recently (Malarkey &
Parpura, 2011) and, Tsodyks & Markram (1997) for SLMV fusion and recycling).
The motivations and consequences of the specific models chosen have been explained
in appropriate places.
We observed an increase in average neurotransmitter release probability, Pr,
after astrocyte became active (before: 0.25; after: 0.35) which is in close conformity
with the experimental observation (before: 0.24; after: 0.33) of Perea & Araque
(2007). On measuring the windowed average amplitude of the excitatory post-
synaptic current (EPSC) we could observe up to 250% increase from pre-astrocytic
activities to the post-astrocytic activities, which decayed with a time constant of 10 to
12 seconds. This signifies augmentation (Fisher et al., 1997; Koch, 1999).
Figure 1. Information flow from pre-synaptic bouton to post-synaptic dendritic spine-head, as
modulated by an astrocyte. Solid line shows the astrocyte-independent pathway, while, solid-line
combined with dashed line shows the astrocyte-dependent pathway. (1) AP generated at pre-synaptic
axon-hillock. (2) Elevated intracellular [Ca2+] in bouton. (3) Increased [Ca2+] leading to exocytosis of
Glutamate into synaptic cleft. (4a) Synaptic glutamate causes an increase in astrocyt ic [Ca2+]. (4b)
Simultaneously synaptic glutamate can also bind with AMPAR causing an increase in post -synaptic
membrane potential. (5) Increased astrocytic [Ca2+] leads to an elevated glutamate concentration in the
4
extra-synaptic cleft, in a vesicle dependent manner. This extra-synaptic glutamate is free to bind with
extra-synaptic mGluR on the pre-synaptic bouton surface. Glutamate bound to mGluR leads to an
increase in Ca2+ concentration via IP3 dependent pathway. This transient enhancement of bouton [Ca 2+]
forms the basis of improved synaptic efficacy, through an astrocyte -dependent pathway.
2. The Model
In this section, we describe the details of the mathematical model, whose
computational implementation will be presented in the section that immediately
follows. In order to elucidate the major neurophysiological steps in our model we use
the flow chart in Figure 1. The mathematical formulations have been described in the
subsequent subsections.
2.1 Pre-synaptic Action Potential
Action potential (AP) is generated at the axon hillock of the pre-synaptic neuron. In
the cortical neurons there may be eleven or more number of different ion channels
(Lytton & Sejnowski, 1991). The key features of initiation dynamics of cortical
neuron APs are (i) their rapid initiation and (ii) variable onset potential – are outside
the range of behaviors described by the classical Hodgkin-Huxley (HH) theory
(Naundorf et al., 2006). Still the HH paradigm has been used to generate pre-synaptic
AP in computational models (Nadkarni & Jung, 2003; Volman et al., 2007). Since in
this paper our focus is not on the detail of the pre-synaptic AP generation, for the sake
of simplicity here we have followed the HH model for the pre-synaptic regular spikes
and bursts generation.
(1)
where Vpre is the pre-synaptic membrane potential in millivolts, Iapp is the applied
current density, gK, gNa and gL are potassium, sodium and leak conductance
respectively, VK, VNa and VL are potassium, sodium and leak reversal potential
respectively, and x=m (Na+ activation), h (Na+ inactivation) and n (K+ activation).
The detail of the HH model can be found in (Hodgkin & Huxley, 1952). The values of
the different parameters in equation (1) that have been used in this paper are furnished
in the Table 2. αx and 𝛽x for x = m, h and n are defined as
5
43()()()(1)preappKpreKNapreNaLpreLxxdVCIgnVVgmhVVgVVdtdxxxdt0.01(60)0.1(45)70,,0.07exp(),604520exp()1exp()11010707010.125exp(),4exp(),408018exp()110prepreprenmhpreprepreprenmhpreVVVVVVVV Table 2: Parameter values used in the HH model (all are from Hodgkin & Huxley, 1952)
Symbol
Value
36 mS cm-2
120 mS cm-2
0.3 mS cm-2
82 mV
45 mV
59.4 mV
2.2 Bouton Ca2+ Dynamics
The train of AP that has been generated in the axon hillock of the pre-synaptic neuron,
travels all the way down to the axon end feet without degradation and leads to an
increase in cytosolic [Ca2+]. The increase in intracellular [Ca2+] can be attributed to
two components:
i)
ii)
[Ca2+] due to AP, denoted as cfast, and
[Ca2+] due to intracellular stores, cslow.
Because of its rapid kinetics, [Ca2+] due to AP is termed as cfast. Similarly, [Ca2+] due
to intracellular stores is termed as cslow. Total intracellular [Ca2+] denoted as ci
satisfies the following simple equation
(2)
The sensitivity of rapid Ca2+ kinetics over neurotransmitter release is well established
(Schneggenburger & Neher, 2000; Bollman et al. 2000). In immature neurons, the
necessary Ca2+ flux for neurotransmitter release is primarily mediated by N-type Ca2+
channels (Mazzanti & Haydon, 2003; Weber et al. 2010). Also, the contribution of
P/Q- type channels is negligible as compared to N-type channels in immature cells
(Ishikawa et al., 2006). Hence, in this article Ca2+ influx through plasma membrane is
modeled through N-type channels alone. Immature cells have been chosen following
Perea & Araque (2007). The equation governing cfast consists of simple construction-
destruction type formulism and is as follows (Keener & Sneyd, 1998)
(3)
6
KgNagLgKVNaVLVfastslowiifastslowdcdcdccccdtdtdtfastconstructiondestructionCabtnPMCabtnPMleakCabtnCabtndcIAIAJdtzFVzFV
Here,
bouton,
is the surface area of the
is the Ca2+ current through N-type channel,
is the Ca2+ ion valence,
is the Faraday’s constant,
is the volume of
represents the current due to electrogenic plasma-membrane Ca2+
the bouton.
ATPase. This pump is known to extrude excess of Ca2+ out of the cell and it has also
been shown that it regulates excitatory synaptic transmission at CA3-CA1 pyramidal
cell (CA3-CA1) synapse (Jensen et al., 2007). The formulation for this pump uses the
standard Michaelis-Menton (MM) type formulism (Erler et al., 2004; Blackwell,
2005).
is the positive leak from extracellular space into bouton, which makes
sure that MM pump does not decrease cytosolic Ca2+ to 0 (Blackwell, 2005).
The Ca2+ current through the N-type Ca2+ channel is formulated using single protein
level formulation, which is described in detail in (Erler et al. 2004)
Here,
is the N-type channel protein density which determines the number of Ca2+
channels on the membrane of the bouton (
was determined computationally so
that average neurotransmitter release probability lies in the range 0.2–0.3, when
astrocyte is not stimulated, similar to the experiments of Perea & Araque (2007)),
is the reversal potential of Ca2+ ion
is the single N-type channel conductance,
determined by the Nernst equation (Keener & Sneyd, 1998),
(4)
where
is the real gas constant,
is the absolute temperature,
is the extracellular
Ca2+ concentration,
is the total intracellular [Ca2+] at rest. It is assumed that a
single N-type channel consists of two-gates.
denotes the opening probability of a
single gate. A single N-type channel is open only when both the gates are open.
Hence,
is the single channel open probability. The time dependence of the single
channel open probability is governed by an HH-type formulation,
where
is the Boltzmann-function fitted by Ishikawa et al. (2005) to the whole cell
current of an N-type channel,
approaches its asymptotic value
with a time
7
CaIbtnACazFbtnVPMCaIPMleakJ2Single open channel()CaCaCaCapreCaImgVtVCaCaCagCaVextrestilnCaCacRTVzFcRTextcresticCam2CamCaCaCaCammmdmdtCamCamCam constant
. The mathematical expression of other parameters used in equation (3)
is as follows:
Here, vPMCa is the maximum PMCa current density, determined through computer
simulations, so that ci is maintained at its resting concentration. All other parameter
values used for simulation are listed in Table 3.
Table 3: Parameters used for Bouton Ca2+ dynamics
Symbol
Description
Value
96487 C mole-1
8.314 J / K
293.15 K
2
1.24 μm2
0.13 μm3
3.2 μm-2
2.3 pS
125 mV
0.4 μA cm-2
0.1 μM
Reference
Temperature
Araque (2007)
in Perea &
Koester & Sakmann, 2000
Koester & Sakmann, 2000
See text; Page No. 7
Weber et al. 2010
Calculated using equation (4)
See text; Page No. 8
Erler et al. 2004
2.66 x 10-6 ms-1
0.1 μM
See Text; Page No. 7
Erler et al. 2004
2 mM
External [Ca2+] in Perea &
Araque (2007)
-17 mV
Ishikawa et al. 2005
8.4 mV
Ishikawa et al. 2005
Faraday’s constant
Real gas constant
Absolute Temperature
Calcium valence
Surface area of bouton
Volume of bouton
N-type channel density
N-type channel conductance
Reversal potential of Ca2+ ion
Maximum PMCa current
Ca2+ concentration at which vPMCa is
halved
Maximum leak of Ca2+
Intracellular
Resting
concentration
External Ca2+ concentration
Ca2+
Half-activation voltage of N-type
Ca2+ channel
Slope factor of N-type channel
activation
Ratio of ER volume to volume of
Bouton
Maximum IP3 receptor flux
Ca2+ leak rate constant
SERCA maximal pump rate
F
R
T
zCa
Abtn
Vbtn
ρCa
gCa
VCa
vPMCa
KPMCa
vleak
cext
VmCa
kmCa
c1
v1
v2
v3
k3
d1
d2
d3
d5
Time constant of N-type channel
10 ms
Ishikawa et al. 2005
0.185
Shuai & Jung, 2002
30 s-1
0.055 s-1
90 μM s-1
0.1 μM
0.13 μM
SERCA dissociation constant
IP3 dissociation constant
Ca2+
Inhibitory
dissociation
constant
IP3 dissociation constant
Ca2+
dissociation 82.34 nM
Activation
943.4 nM
1.049 μM
8
See text; Page No. 10
See text; Page No. 10
See text; Page No. 10
Jafri & Keizer, 1995
Shuai & Jung, 2002
Shuai & Jung, 2002
Shuai & Jung, 2002
Shuai & Jung, 2002
Cam2iexti22i1 , (), .1exp()/CaCaPMCaPMCaPMleakleakCaPMCammmcIvJvccmcKVVkresticCam
constant
Inhibitory Ca2+ binding constant
Maximum production rate of IP3
Glutamate concentration at which
vg is halved
IP3 degradation constant
0.2 μM s-1
0.062 μM s-1
0.78 nM
Shuai & Jung, 2002
Nadkarni & Jung, 2008
Nadkarni & Jung, 2008
0.14 s-1
Wang et al., 1995
Initial IP3 concentration
160 nM
Wang et al., 1995
a2
vg
kg
p0
The second component of bouton Ca2+, cslow, is the slower component. It is known to
play a crucial role in STP (Emptage et al., 2001). The release of Ca2+ from
endoplasmic reticulum (ER) is mainly controlled by two types of receptors (or Ca2+
channels) i) the inositol (1,4,5)-trisphosphate receptor (IP3R) and ii) the ryanodine
receptor (RyR) (Sneyd & Falcke, 2005). For the sake of simplicity, the flow is
assumed to be through IP3R alone. The IP3 necessary for release of Ca2+ from ER, is
produced when glutamate (agonist) binds with mGluRs (receptor) and causes via G-
protein link to phospholipase C (PLC), the cleavage of phosphotidylinositol (4,5)-
bisphosphate (PIP2) to produce IP3 and diacylglycerol (DAG). We have used the
conventional Li-Rinzel model (L-R model) (Li & Rinzel, 1994) to formulate this
slower Ca2+ signaling process.
There were a few modifications made to the L-R model. The L-R model assumes that,
total intracellular concentration, c0, is conserved and determines the ER Ca2+
concentration, cER, using the following relation
(5)
Such an assumption is not valid in the present model because of the presence of
membrane fluxes, namely ICa and IPMCa. Also, in the L-R model intracellular IP3
concentration, [IP3], is used as a control parameter. To take care of these
“inconveniences” two additional equations governing ER [Ca2+] and [IP3] have been
incorporated in the L-R model. The [IP3] production term was made glutamate
dependent to study the effect of astrocytic Ca2+ over ci (Nadkarni & Jung, 2007). The
mathematical model governing the cslow dynamics is as follows
9
p0iER1.cccc
(6)
Here
denotes Ca2+ flux from ER to the intracellular space through IP3R,
is the Ca2+ flux pumped from the intracellular space into ER,
is the leak
of Ca2+ ions from ER to intracellular space, cER is the ER Ca2+ concentration, c1 is the
ratio of the volume of ER to the volume of bouton, p is the intracellular IP3
concentration, ga is the glutamate in the extra-synaptic cleft, q is the fraction of
activated IP3R. The expressions for the fluxes are
with
. Most of the values of v1,
v2, v3 mentioned in literature are for closed-cell dynamics which is not the case here.
The values of v1, v2, v3 were fixed through extensive simulation runs so that Ca2+
homeostasis is maintained inside the cell and its organelles. Details of parameters are
as listed in Table 3.
2.3 Glutamate release dynamics in bouton
It is now widely accepted that AP waveforms lead to a transient increase in
intracellular [Ca2+] and lead to neurotransmitter release (Bollman et al. 2000; Wang et
al., 2009). However, the study of Ca2+ sensor sensitivity becomes exceedingly
challenging due to small size of nerve terminals (Wang et al., 2009). It is generally
assumed that Ca2+ concentration of at least 100 μM in the terminal is necessary for a
“low-affinity” Ca2+ sensor to activate (Neher, 1998; Nadkarni & Jung, 2008). But,
recent studies performed at giant Calyx of Held terminal have revealed that
intracellular Ca2+ concentration of ~10 μM is sufficient for glutamate release
(Schneggenburger & Neher, 2000; Bollman et al., 2000). The kinetic model
10
slowchanERpumpERleakslowER10.3ag00.30.3ga,1,,(1).pqqdcJJJdtdcdcdtcdtgdpvppdtkgdqqqdtchanJERpumpJERleakJ333chan11iER23iERpump223iERleak12iER,,,JcvmnqccvcJkcJcvcci1222i1i53, ,, qqcpdpmnadacpdcdpd governing the Ca2+ binding to Ca2+ sensor is given by the following equations
(Bollman et al., 2000),
(7)
Where, α and β are the Ca2+ association and dissociation rate constants respectively, γ
and δ are Ca2+ independent isomerisation constants. X is the Ca2+ sensor (of a
synaptic vesicle) with no Ca2+ bound, X(ci)1 is Ca2+ sensor with one Ca2+ bound,
likewise, X(ci)5 is Ca2+ sensor with five Ca2+ bound;
is the isomer of X(ci)5
which is ready for glutamate release. Hippocampal synapses are known as low-
fidelity synapses (Nadkarni & Jung, 2008). We have assumed an active zone
consisting of two-docked synaptic vesicles (Danbolt, 2001; Nikonenko & Skibo,
2006). Since there are few synaptic vesicles the number of vesicles with 5 Ca2+ ions
bound cannot be determined by the average of vesicle pool. Therefore, fraction of
docked vesicles ready to be released fr, has been determined using dynamic Monte-
Carlo simulation (Fall et al., 2002) of kinetic equation (7) and depends on
state.
Apart from evoked release of docked vesicles, spontaneous release of vesicles can
also occur when pre-synaptic membrane is not depolarized. The rate of spontaneous
release depends upon pre-synaptic Ca2+ concentration (Bollman et al., 2000; Emptage
et al., 2001; Schneggenburger & Neher, 2000). The number of vesicles ready to be
released spontaneously, pr, is assumed to be a Poisson process with the following rate,
(8)
The formulation for the rate of spontaneous release is from Nadkarni & Jung (2008).
We have to modify the parameter values (see equation (8)) because as per their choice
of values and system setup, the frequency of spontaneously released vesicles was as
high as 19 per sec (we have determined this through simulation runs of over 10000
trials). However, the experimentally determined frequency of spontaneous vesicle
release in presence of an astrocyte is in between 1 – 3 per sec (Kang et al., 1998).
Thus, we determined the values of a1, a2 and a3 by simulation so that the frequency of
spontaneous vesicle release is between 1 – 3 Hz. The vesicle fusion and recycling
process is governed by the Tsodyks & Markram Model (TMM) (Tsodyks &
11
iiiii5α4α3α2ααγ*i1i2i3i4i5i5β2β3β4β5βδ()()()()()()cccccXXcXcXcXcXcXc*i5()Xc*i5()Xc11ii321exp.accaa Markram, 1997). A slight modification has been made to the TMM to make the
vesicle fusion process pr dependent. The modified TMM is as follows
(9)
where R is the fraction of releasable vesicles inside bouton, E is the fraction of
effective vesicles in the synaptic cleft and I is the fraction of inactive vesicles
undergoing recycling process, fr has the values (0, 0.5, 1) corresponding to the number
of vesicles ready to be released (0, 1, 2), which is determined by the stochastic
simulation of kinetic model in equation (7) or generating a Poisson random variable
with the rate given by the equation (8). τinact and τrec are the time constants of vesicle
inactivation and recovery respectively. Once a vesicle is released whether evoked or
spontaneous the vesicle release process remains inactivated for a period of 6.34 ms
(Dobrunz et al., 1997). The parametric values used for simulation are listed in Table
4.
2.4 Glutamate dynamics in synaptic cleft
Various types of glutamate receptors have been detected pre-synaptically, extra-
synaptically, as well as on glial cells (Danbolt, 2001). Suggesting that, to study
transmission of glutamatergic signals, it is essential to study, how glutamate diffuses
(Danbolt, 2001). However, using Monte Carlo simulation of a central glutamatergic
synapse, in particular a CA3–CA1 synapse, Franks et al., (2002) showed that
glutamatergic signaling is spatially independent at these synapses. The capacity of the
bouton vesicle containing glutamate has been assumed to be 60 mM (Danbolt, 2001).
Since, E gives the effective fraction of vesicles in the cleft the estimated glutamate
concentration in the cleft can be represented mathematically as
(10)
Here g is the glutamate concentration in the synaptic cleft, nv is the number of docked
vesicle, gv is the vesicular glutamate concentration and gc is the rate of glutamate
clearance i.e. re-uptake by neuron or astrocyte (Destexhe et al., 1998). Using this
simple dynamics, we could achieve the estimated range of glutamate concentration
0.24 - 11 mM in cleft (Danbolt, 2001; Franks et al., 2002) and the time course of
glutamate in the cleft is 2 ms (Clements, 1996; Franks et al., 2002). Although similar
equation can be used to model glutamate dynamics at other synapses, however, one
12
rrecrinact,,1,dRIfRdtdEEfRdtIREvvc.dgngEggdt
might have to use different constant values. Thus, the present formulation can be
considered specific to a CA3–CA1 synapse.
Table 4: Parameters used for Glutamate dynamics in bouton and cleft
Description
Ca2+ association rate constant
Ca2+ dissociation rate constant
constant
rate
Isomerization
(forward)
Value
0.3 μM ms-1
3 ms-1
30 ms-1
Reference
Bollman et al. 2000
Bollman et al. 2000
Bollman et al. 2000
Isomerization
(backward)
rate
constant
8 ms-1
Bollman et al. 2000
Vesicle recovery time constant
800 ms
Vesicle inactivation time constant
3 ms
Ca2+ concentration at which is
halved
50 μM
Slope factor of spontaneous release
5 μM
rate
Maximum spontaneous release rate
Number of docked vesicle
0.85 ms-1
2
Glutamate concentration in single
vesicle
60 mM
Glutamate clearance rate constant
10 ms-1
Tsodyks & Markram,
1997
Tsodyks & Markram,
1997
See text; Page No. 11
See text; Page No. 11
See text; Page No. 11
Nikonenko & Skibo,
2006
Montana et al., 2006
Destexhe et al., 1998
Symbol
α
β
γ
δ
τrec
τinac
a1
a2
a3
nv
gv
gc
2.5 Astrocyte Ca2+ dynamics
Porter & McCarthy (1996) showed that glutamate released from the Schaffer
collaterals (SC) leads to an increase in astrocytic Ca2+ via an mGluR pathway.
Recently, De Pitta et al. (2009) proposed a G-ChI model for astrocytic Ca2+
oscillations mediated by mGluR pathway while treating glutamate concentration in
the synaptic cleft as a parameter. They called it G-ChI referring to the dependent
variables and the glutamate concentration parameter used in their model (in their
model G represented glutamate concentration in the synaptic cleft, C represented
astrocytic [Ca2+], h represented the gating variable of IP3R and I represented the
astrocytic [IP3]). We have used the G-ChI model for astrocyte Ca2+ dynamics with an
exception that ‘g’ is a dynamic variable given by equation (10). The G-ChI model
uses the conventional L-R model for astrocytic Ca2+ concentration ca with some
specific terms for intracellular IP3 concentration pa. It incorporates PLC𝛃 and PLC𝛅
(are isoenzymes of the family of phosphoinositide specific PLC) dependent IP3
production. It also incorporates inositol polyphosphate 5-phosphatase (IP-5P) and IP3
3-kinase (IP3-3K) dependent IP3 degradation (for a systematic derivation regarding
the expressions, shown in equation (13), incorporating these effects see De Pitta et al.,
2009). It is a very detailed model based on astrocyte specific experiments (Hofer et
13
al., 2002; Suzuki et al., 2004), model which exhibits IP3 oscillations apart from Ca2+
oscillations. However, the exact significance of IP3 oscillations is yet unknown (De
Pitta et al., 2009). The G-Chi model is a closed-cell model (Keener & Sneyd, 2009)
i.e. without membrane fluxes. In such models ca primarily depends upon two
parameters, i) flux from ER into cytosol and ii) the maximal pumping capacity of the
Sarco-Endoplasmic Reticulum ATPase (SERCA) pump. It is known that IP3Rs are
found in clusters in astrocytes (Holtzclaw et al., 2002). However, the size of the
cluster NIP3 is not known. We have assumed it to be 20 (Shuai & Jung, 2002). We
make use of the stochastic L-R model (Shuai & Jung, 2002). The model can be
represented as follows
(11)
(12)
(13)
Here the first term on the right hand side of equation (11) represents the Ca2+ flux
flowing out from ER to the intracellular space, the second term represents the rate at
which Ca2+ is removed from the intracellular space by SERCA pump and the last term
represents the leak of Ca2+ from ER into the intracellular space. Clearly these terms
are very analogous to the terms involved in production of cslow in equation (6). But
with a major difference, which was mentioned earlier as well, that this model is based
on closed-cell assumption. Under such an assumption, an expression like equation (5)
holds true and can be represented in terms of the astrocyte cell parameters as
(14)
Equation (14) gives us the advantage to represent astrocytic Ca2+ flux terms
completely in terms of cell parameters (compare equation (11) with equation (6)
where a separate differential equation for
is present).
is the maximal rate of
Ca2+ flux from IP3R cluster,
together represent the opening probability of
is the maximal rate of Ca2+ uptake into ER, KER is the affinity of
IP3R cluster.
SERCA pump for intracellular Ca2+.
is the maximal rate of Ca2+ leak from ER. The
14
a2333aaa01,aaERL01,aa22aER11,cdccrmnhcccvrcccdtcKa0.72aδβπaPLCδaδ43aa35aHill,1Hill,Hill,1 Hill,Hill,pRRKDpKdpvvgKCKcKpdtKkvcKpKrpaaaaa1()hhhdhhhGtdt0ER,aER,a1,a01,a.aaccccccccERdcdtacr333amnhERvLr
first two terms on the right hand side of equation (12) incorporate agonist-dependent
and agonist-independent production of IP3 and the last two terms incorporate IP3
degradation by IP3-3K and IP-5P respectively. In equation (13),
is the rate at which
ha,
is the rate at which ha closes and Gh(t) is zero mean, uncorrelated, Gaussian
white-noise term with co-variance function (Shuai & Jung, 2002),
Here, δ(t) is the Dirac-delta function, t and t' are distinct time and
is the spectral density (Coffey et al. 2005). The present model can be classified into
three categories i) amplitude modulated (AM), ii) frequency modulated (FM), and iii)
amplitude and frequency modulated (AFM) modulated (De Pitta et al., 2009). We
have used AFM-encoded astrocytic Ca2+ oscillations as coupling of IP3 metabolism
with calcium-induced calcium release (CICR) does not allow pure AM encoding (De
Pitta et al., 2009). The mathematical expression of other parameters used in equations
(11) and (13) are
is the generic Hill function (De Pitta et al., 2009). Typically, Hill
function is used for reactions whose intermediate steps are unknown (or not
considered) but cooperative behavior is suspected in the reaction (Keener & Sneyd,
1998). Mathematically, it can be said that Hill function is used for reactions whose
reaction velocity curve is not hyperbolic (Keener & Sneyd, 1998). Parametric value of
all the constants is as listed in Table 5.
Table 5: Parameters used for astrocyte Ca2+ dynamics
Description
Maximal IP3R flux
Maximal rate of Ca2+ leak from ER
Total cell free Ca2+ concentration
Ratio of ER volume to cytosol
volume
Value
6 s-1
0.11 s-1
2 μM
0.185
Maximal rate of SERCA uptake
SERCA Ca2+ affinity
IP3 dissociation constant
0.9 μM s-1
0.1 μM
0.13 μM
15
Symbol
rL
c0
vER
KER
d1
Reference
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
ahahaa3aaIP(1)()(')'hhhhhhGtGtttNaa3aaIP(1)hhhhNaa,aa1,aa5a1222aa3Hill(,), Hill(,), Hill,,,.nnnnhhxmpdncdxKxKpdadacpdHill,nxKacr1,ac d2
d3
d5
a2
N
KR
Kp
K𝜋
KPLCδ
kδ
r5pa
v3K
KD
K3
1.049 μM
0.9434 μM
Ca2+
inactivation dissociation constant
IP3 dissociation constant
Ca2+
activation dissociation constant
IP3R binding rate for Ca2+
Inhibition
Number of IP3R in a cluster
20
Glutamate-dependent IP3 production
0.5 μM s-1
Maximal rate of IP3 production by
PLCβ
0.08234 μM
2 s-1
10 μM
1.3 μM
Glutamate affinity of the receptor
Ca2+/PKC-dependent inhibition
factor
Ca2+ affinity of PKC
Glutamate-independent IP3 production
0.05 μM s-1
Maximal rate of IP3 production by
PLCδ
Ca2+ affinity of PLCδ
Inhibition constant of PLCδ activity 1.5 μM
0.6 μM
0.1 μM
IP3 degradation
0.05 s-1
Maximal rate of degradation by IP-
5P
Maximal rate of degradation by
IP3-3K
Ca2+ affinity of IP3-3K
IP3 affinity of IP3-3K
2 μM s-1
0.7 μM
1 μM
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
Shuai & Jung, 2002
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
De Pitta et al. 2009
2.6 Gliotransmitter release dynamics in astrocyte
There is enough evidence that astrocytes actually release gliotransmitters in a Ca2+
dependent manner (Bezzi et al. 2004; Montana et al. 2006; Bowser & Khakh, 2007;
Marchaland et al. 2008; Fellin, 2009). There is again considerable evidence that the
released gliotransmitters modulate synaptic plasticity via extra-synaptic NMDAR
(Parpura et al. 1994; Parpura & Haydon, 2000; Carmignoto & Fellin, 2006; Bergersen
& Gundersen, 2009) and extra-synaptic mGluR (Fiacco & McCarthy, 2004; Perea &
Araque, 2007). But, the exact mechanism by which astrocytes release gliotransmitters
is yet to be determined (Wenker, 2010). However, it is widely agreed upon that
astrocytes release gliotransmitters in a vesicular manner similar to neurons (Bezzi et
al. 2004; Montana et al., 2006; Verkhratsky & Butt, 2007; Marchaland et al. 2008) as
they possess the necessary exocytotic secretory machinery (Parpura & Zorec, 2010).
In 2000, Parpura & Haydon determined Ca2+ dependency of glutamate release from
hippocampal astrocytes and found that the Hill coefficient for glutamate release from
astrocytes was 2.1–2.7 suggesting at least two Ca2+ ions are must for a possible
gliotransmitter release. Recently the probability of vesicular fusion in response to a
mechanical stimulation and the size of readily releasable pool of SLMVs in astrocytes
16
vv
have been determined by Malarkey & Parpura (2011). Based on the observation of
Parpura & Haydon (2000) in this manuscript we have assumed that binding of three
Ca2+ ions leads to a gliotransmitter release. The model governing the gliotransmitter
release site activation is based on Bertram et al. (1996). Our gliotransmitter release
model assumes that three Ca2+ ions must bind with three independent gates or sites (S1
– S3) for a possible gliotransmitter release.
where Cj and Oj are the closing and opening probability of gate S j respectively,
are the opening and closing rates of the gate Sj respectively. The temporal
evolution of the open gate Oj can be expressed as
(15)
As the three sites are physically independent, the fraction of SLMVs ready to be
released can be given as the product of the opening probabilities of the three sites
(16)
The dissociation constants of gates S1 – S3 are 108 nM, 400 nM, and 800 nM. The
time constants for gate closure
are 2.5 s, 1s, and 100 ms. The dissociation
constants and time constants for S1 and S2 are same as in Bertram et al (1996). While,
the dissociation constant and time constant for gate S3 was fixed through computer
simulations to fit the experimentally determined probability of fusogenic (fraction of
readily releasable SLMVs in response to a mechanical stimulation) SLMVs found
recently by Malarkey & Parpura (2011). Once an SLMV is ready to be released its
fusion and recycling process was modeled using TMM. The governing model is as
follows
(17)
17
a, j = 1, 2, 3,jjkjjkcCO and jjkkaa.jjjjjdOkckckOdtar123.fOOO(1/)jkthreshaaaaaraarecthreshaaaaaraainactaaa,,1.dRIccfRdtdEEccfRdtIRE
Here, Ra is the fraction of readily releasable SLMVs inside the astrocyte, Ea is the
fraction of effective SLMVs in the extra-synaptic cleft and Ia is the fraction of
inactive SLMVs undergoing endocytosis or re-acidification process.
is the
is the threshold of astrocytic [Ca2+] necessary for
Heaviside function and
release site activation (Parpura & Haydon, 2000).
and
are the time constants
of inactivation and recovery of SLMVs respectively.
2.7 Glutamate dynamics in extra-synaptic cleft
The glutamate concentration in the extra-synaptic cleft ga, has been modeled in a
similar way to equation (10). This glutamate acts on extra-synaptically located
mGluRs of the pre-synaptic bouton. It is used as an input in the IP3 production term of
equation (6). The SLMVs of the astrocytes are not as tightly packed as of the neurons
(Bezzi et al., 2004). Thus, it is assumed that each SLMV contains 20 mM of
glutamate (Montana et al., 2006). The mathematical equation governing glutamate
dynamics in the extra-synaptic cleft are as follows
(18)
where ga is the glutamate concentration in the extra-synaptic cleft,
represents the
readily releasable pool of SLMVs,
is the glutamate concentration within each
SLMV,
is the clearance rate of glutamate from the cleft due to diffusion and/or re-
uptake by astrocytes.
Table 6: Parameters used for Glutamate dynamics in astrocyte and extra-synaptic cleft
Description
Ca2+ association rate for S1
Value
3.75 x 10-3 μM-1 ms-1
Reference
Bertram et al. 1996
Ca2+ dissociation rate for S1
4 x 10-4 ms-1
Bertram et al. 1996
Ca2+ association rate for S2
2.5 x 10-3 μM-1 ms-1
Bertram et al. 1996
Ca2+ dissociation rate for S2
1 x 10-3 ms-1
Bertram et al. 1996
Ca2+ association rate for S3
1.25 x 10-2 μM-1 ms-1
See text, page no. 17
Ca2+ dissociation rate for S3
10 x 10-3 ms-1
See text, page no. 17
Vesicle recovery time constant
800 ms
Vesicle inactivation time constant
3 ms
Astrocyte response threshold
196.69 nM
SLMV ready to be released
12
Tsodyks & Markram,
1997
Tsodyks & Markram,
1997
Parpura & Haydon,
2000
Malarkey & Parpura,
2011
18
Symbol
threshacainactarecvvcaaaaaa,dgngEggdtvanvagcag1k1k2k2k3k3karecainacthreshacvan
Glutamate concentration
SLMV
in one
20 mM
Montana et al. 2006
Glutamate clearance rate from the
extra-synaptic cleft
10 ms-1
Destexhe et al. 1998
2.8 Dendritic spine-head dynamics
The dendritic spine-head is assumed to be of mushroom type. Its volume is taken to
be 0.9048 μm3 (assuming a spherical spine-head of radius 0.6 μm (Dumitriu et al.,
2010)). The specific capacitance and specific resistance of the spine-head is assumed
to be 1 μF / cm2 and 10000 cm2, respectively. Given the dimension of the spine we
can calculate its actual resistance as
(19)
where Rm is the actual resistance of the spine, RM is the specific resistance of the spine
and Aspine is the area of spine-head. NMDAR (N-methyl D-aspartate receptor) and
AMPAR (α-amino-3-hydroxy-5-methyl-4-isoxazolepropionic acid receptor) are co-
localized at most of the glutamatergic synapses, most of which are found at dendrit ic
spines (Franks et al., 2002). Chen & Diamond (2002) showed that the post-synaptic
NMDAR receives less glutamate during evoked synaptic response, suggesting that
most of the post-synaptic current is contributed by AMPAR, under such conditions.
Also, NMDAR is known to play a crucial role in longer forms of synaptic plasticity,
Long-term Potentiation (LTP) and Long-term Depression (LTD) (Bliss &
Collingridge, 1993; Malenka & Bear, 2004). Hence, in our model of short-term
potentiation the post-synaptic density comprises of AMPAR alone. The post-synaptic
potential change has been modeled using a passive membrane mechanism (Tsodyks &
Markram, 1997)
(20)
where τpost is the post-synaptic membrane time constant,
is the post-synaptic
resting membrane potential, IAMPA is the AMPAR current and is given by the
following expression (Destexhe et al., 1998)
19
vagcagMmspine,RRApostrestpostpostpostmAMPA(),dVVVRIdtrestpostVAMPAAMPAAMPApostAMPA,IgmVV
where gAMPA is the conductance of the AMPAR channel, VAMPA is the reversal
potential of the AMPAR and mAMPA is the gating variable of AMPAR. Although there
exists a more comprehensive 6-state markov model for AMPAR gating (Destexhe et
al., 1998), in our model we have used a simple 2-state model for AMPAR gating. This
two state model has been used keeping in mind it is computationally less expensive,
while retaining most of the important qualitative properties (Destexhe et al., 1998).
Also, it is known that detailed AMPAR mechanisms like desensitization do not play a
role in STP (Zucker & Regehr, 2002). AMPAR gating is governed by the following
HH-type formulism (Destexhe et al., 1998)
(21)
Here,
is the opening rate of the receptor,
is the closing rate of the
receptor and g is the glutamate concentration in the cleft given by equation (10). The
parameter values are as listed in Table 7.
Table 7: List of parameters used for post-synaptic potential generation
Description
Actual resistance of the spine-head
Value
0.79 x 105 M
Reference
Calculated
equation (19)
using
Post-synaptic
potential
resting membrane
-70 mV
Post-synaptic membrane
constant
AMPAR conductance
AMPAR reversal potential
AMPAR forward rate constant
AMPAR backward rate constant
time
50 ms
0.35 nS
0 mV
1.1 μM s-1
190 s-1
Tsodyks & Markram,
1997
Destexhe et al. 1998
Destexhe et al. 1998
Destexhe et al. 1998
Destexhe et al. 1998
20
Symbol
Rm
𝜏post
gAMPA
VAMPA
AMPAAMPAAMPAAMPAAMPA1.dmgmmdtAMPAAMPArestpostVAMPAAMPA
Figure 2. The two types of information processing simulated in this paper. (A) Astrocyte-independent
information processing. (B) Astrocyte -dependent information processing. The input signal is being
amplified by astrocyte-dependent feed-forward and feed-back pathways making up a loop.
2.9 Numerical Implementation
All the computations have been performed using MATLAB. The model equations
were discretized with a temporal precision of ∆t = 0.05 ms. The canonical explicit
Euler method was used to solve the system of twenty-two ordinary differential
equations governing TpS. For the numerical simulation of the noise term, in equation
(13), we have used Box-Muller Algorithm (Fox, 1997) to generate noise-term at each
time-step (∆t). All simulations were performed on a Dell precision 3500 workstation
with Intel Xeon processor with 2.8 GHz processing speed and with 12 GB RAM. The
time taken for model time of 1s (stimulation rate 5 Hz) is approximately 8.5 sec. The
MATLAB script written for the simulation of the model can be requested by email to
any of the authors.
3. Simulation results
How post-synaptic current is being generated with and without the participation of
astrocytic Ca2+ have been shown in this section with extensive numerical simulations
of the model equations presented in the previous section. In the latter case how the
output signal is being amplified through a processing loop, consisting of feed-forward
and feed-back paths, with the help of astrocytic Ca2+ signaling, has been shown in
Figure 2B. Here, we have tried to answer the question, “Does astrocyte play an active
role in modulation of synaptic plasticity?” In order to study the difference in both
types of processing (see Figure 2), first we present the results associated with
astrocyte-independent processing followed by astrocyte-dependent processing.
3.1 Astrocyte-independent Information Processing
In this subsection we simulate the processing elaborated in Figure 2A. We present
results of implementation of the models described in subsections 2.1, 2.2, 2.4 and 2.8
(Figures 3A, 3B, 3C, and 3D respectively).
21
Figure 3. The major variables involved in astrocyte-independent information processing. A. Vpre (mV),
5 Hz input signal generated using HH model, in response to a stimulus of 10 μA per cm2 of frequency 5
Hz and duration 10 ms. B. Ca2+ (nM), fast Ca2+ oscillations in response to the 5 Hz input signal . C.
Synaptic glutamate (mM), elevated glutamate concentration in the synaptic cleft due to exocytosis of
glutamate filled synaptic vesicles from bouton. D. Excitatory post-synaptic potential (EPSP) (mV),
potential change in the membrane of the post-synaptic spine mediated through AMPAR channels.
We used the model described in equation (1) to generate input signal or pre-synaptic
membrane potential. This input signal forms the basis of signal transduction and we
made sure that the tripartite synapse is at rest in its absence. In response to this input
signal, the N-type Ca2+ channels open and bouton Ca2+ starts undergoing very fast
oscillations (see Figure 3B). Note that, here, there is no astrocyte present and hence
there is no contribution of [Ca2+] from intracellular stores. We adjusted the number of
Ca2+ channels on the surface of the bouton (by adjusting ρCa) so that the amplitude of
Ca2+ oscillation is 5 μM i.e., exactly half of the affinity of Ca2+ sensor (β/α, where β
and α are given in Table 4). Doing this we could attain average neurotransmitter
release probability, in the range 0.2–0.3 (see Figure 5), observed experimentally in
absence of astrocyte (Perea & Araque, 2007). Increased bouton [Ca2+] instigates the
process of exocytosis and vesicles release their content (glutamate) in the synaptic
cleft (see Figure 3C). When glutamate concentration rises in the cleft, it binds with
post-synaptic AMPAR, which causes this ligand-gated channel to open. Once opened,
AMPAR causes a change in the post-synaptic potential (see Figure 3D) since this
deflection is positive it has been termed as EPSP. As described in the previous
section, we also keep track of the vesicle recycling process, see equation (9), which is
shown in Figure 4.
22
Figure 4. Fraction of releasable and effective vesicles, in astrocyte-independent information processing,
during an input signal of 5 Hz (see Figure 3A). (A) The fraction of releasable vesicles i.e., ready to be
fused, inside the bouton. (B) The fraction of effective vesicles i.e., fraction of vesicles left in the
synaptic cleft.
In Figure 4 we show the underlying process of vesicle release. In the absence of
astrocyte, it can be observed that nearly 90% of the vesicles are available for release
for most of the time (see Figure 4A). In Figure 4B, we observe that the fraction of
effective vesicles is not as dense as the input signal (see Figure 3A) implying low
probability of vesicle release. In fact, the probability of vesicle release was nearly
0.25 i.e., every fourth input signal is able to release a synaptic vesicle. We next show
Pr i.e., neurotransmitter release probability in absence of astrocyte. Pr has been
calculated as the ratio of the number of successful post-synaptic responses to the
number of pre-synaptic impulses (with a time-window of length 5 seconds).
3.2 Astrocyte-dependent Information Processing
In this subsection we show simulations associated with the biophysical model
governed by equations (1) - (20) i.e., the astrocyte-dependent information processing.
In Figure 6, we give an idea of the processes involved in the loop shown in Figure 2B.
For the simulation of the scheme, shown in Figure 2B, we simultaneously solved
equations (1) - (20). Of particular interest is the astrocyte-dependent feed-forward and
feed-back paths making up a loop (Figure 2B). The same input signal was used in a
feed-back manner into the loop. Using such a feed-forward and feed-back pathway an
input signal can be amplified as per the cognitive process requiring strengthening of
synapses. Ultimately such a process, may, lead to enhanced synaptic efficacy.
23
Figure 5. Probability of neurotransmitter release Pr, without incorporating the feedback loop due to
astrocyte, is computed as the ratio of the number of successful post -synaptic response to the number of
pre-synaptic stimulus (which was 5 Hz for the given simulation) within a time -window of length 5
seconds.
24
Figure 6. The major variables involved in astrocyte -dependent information processing. Here, input
signal is same as in Figure 3A and is omitted. Output F is feeding back into the input A. A. Increased
bouton IP3 concentration in response to elevated extra -synaptic glutamate concentration (see F). B.
Increased IP3 concentration causes the IP3R channels to open and leads to a transient enhancement in
bouton [Ca2+], due to influx of Ca2+ from IP3R (see Ca2+ concentration after 20 seconds). C.
Accumulated bouton [Ca2+] leads to increased transients of glutamate concentration in the synaptic
cleft. D. Transients of glutamate concentration set-off the production of astrocytic IP 3 concentration
through an mGluR dependent pathway. E. Elevated astrocytic IP3 concentration causes the IP3R
channels to open and initiates astrocytic Ca2+ oscillations. F. Astrocytic Ca2+ oscillations instigate the
process of SLMV fusion, which is followed by a raised extra-synaptic glutamate concentration. This
elevated extra-synaptic glutamate concentration forms the basis of bouton IP 3 production shown in A.
All the variables shown in Figure 6 are inter-dependent i.e., variation in one affects
variation in others. When the bouton is fed with an input signal, it shows its response,
in the form of increased cytosolic [Ca2+] (see Figure 6B). This elevated [Ca2+]
exocytose glutamate in the synaptic cleft (see Figure 6C). After being exocytosed,
synaptic glutamate can have either of the two fates (see Figure 2B). It can either bind
with the post-synaptic AMPAR or it can bind with the mGluRs on the surface of the
astrocyte. Once this glutamate binds with mGluR, it instigates the production of
astrocytic IP3 (see Figure 6D) through a G-protein link. During this glutamate spill-
over process astrocytic IP3 concentration goes on appreciating and gradually starts
oscillating (notice after the 20 seconds mark Figure 6D). It can be observed from
Figure 6D and Figure 6E that astrocytic Ca2+ also starts oscillating as soon as
astrocytic IP3 starts oscillating. The biological significance and importance of IP3
oscillation on Ca2+ oscillation is not been fully understood though (De Pitta et al.,
2009). This astrocytic Ca2+ is known to exocytose SLMVs filled with glutamate once
it crosses its threshold value of 196.69 nM (Parpura & Haydon, 2000). Similarly, in
our model whenever astrocytic Ca2+ crosses its threshold value it can spill glutamate
(contained in SLMVs) in the extra-synaptic cleft (see Figure 6F). We have
mathematically modeled this process of astrocytic glutamate release using equations
(15)-(18). Extra-synaptic glutamate binds with extra-synaptic mGluR located on the
surface of the bouton and initiates the production of bouton IP3 (see Figure 6A)
through a G-protein link. It is visible from Figure 6F and Figure 6A that bouton IP3
production starts only when the astrocyte spills glutamate in the extra-synaptic cleft,
reflecting the significance of extra-synaptic glutamate in the model. This bouton IP3 is
free to diffuse inside the cytosol and opens the IP3R on the intracellular stores in a
Ca2+-dependent manner. Transient accumulation of Ca2+ takes place as a result of
opening up of IP3R on the surface of the intracellular store (e.g., see Figure 6B at 20
seconds mark). Flow of Ca2+ through these IP3Rs is a slow process and is known to
play a crucial role in modulating synaptic plasticity and spontaneous vesicle release
(Emptage et al., 2001).
The synaptic vesicle exocytosis from bouton and SLMV release from astrocyte has
been modeled using equations (7) - (9) and equations (15) - (17). Figure 7A and
Figure 7B show the fraction of releasable and effective vesicles respectively during
25
synaptic vesicle recycling process emulated using equations (7) - (9). Figure 7A and
7B are similar to the diagrams in Figure 4, except for the astrocyte-dependent
pathway used here. The SLMV recycling process has been modeled using equation
(17).
Figure 7. Fraction of releasab le and effective vesicles, in astrocyte-dependent information processing,
during an input AP of 5 Hz (see Figure 3A). A. Fraction of releasable vesicles inside the bouton. B.
Fraction of effective vesicles in the synaptic cleft i.e. , fraction of vesicles fused and residual vesicles in
the synaptic cleft. C. Fraction of releasable SLMVs inside the astrocyte . D. Fraction of effective
SLMVs in the extra-synaptic cleft i.e., fraction of SLMV fused and residual SLMV in the extra-
synaptic cleft.
Figure 7C and Figure 7D show the fraction of releasable vesicles in astrocyte and
effective vesicles in extra-synaptic cleft. It can be observed from Figure 7A that
nearly 92% of the releasable (docked) vesicles have been used in astrocyte-dependent
pathway. The fraction of effective vesicles in the synaptic cleft has also considerably
gone-up (compare with Figure 4B). It is because of the transient increase in Ca2+
concentration (see Figure 6B) which improves synaptic vesicle release probability
(see Figure 7). In fact, the average vesicle release probability during this pathway was
nearly 0.35, implying more than one out of three spikes are able to release a synaptic
vesicle. It should be pointed out that the similar amount of enhancement in
neurotransmitter release probability has been observed experimentally as well. Perea
& Araque (2007) reported an increased Pr after astrocyte stimulation (from 0.24 to
0.33). We next show neurotransmitter release probability following the astrocyte-
dependent pathway of
information processing. A
transient
increase
in
neurotransmitter release probability can be observed from Figure 8 in close
correlation with the astrocytic Ca2+ concentration (see Figure 6E). The average
26
release probability under astrocyte-dependent pathway of
neurotransmitter
information processing was 0.338 compared to 0.23 for astrocyte-independent
pathway.
Figure 8. Probability of neurotransmitter release after incorporating the feed-forward and feed-back
loop due to astrocyte, is computed as the ratio of the number of successful post-synaptic response to the
number of pre-synaptic stimulus (which was 5 Hz for the given simulation) within a time -window of
length 5 seconds.
3.3 Comparison between the two-forms of information processing
In this subsection, we have undertaken a comparative study between the two forms of
information processing (see Figure 2A & 2B). We will discuss some of our findings
keeping in mind the recent controversy regarding whether astrocytic [Ca2+]
contributes in synaptic plasticity or not (e.g., Henneberger et al., 2010 vs. Agulhon et
al., 2010).
27
Figure 9. A comparison of the two modes of information processing (see Figure 2) in response to the
same input signal of 5 Hz. Synaptic efficacy is calculated as the windowed-mean of post-synaptic
responses including successes and failures where the window length has been taken to be 5 seconds for
both figures. A. Output signal using astrocyte-independent information processing and B. Output signal
using astrocyte-dependent information processing.
Using their experimental setup Perea & Araque (2007) demonstrated an increase in
synaptic efficacy at single CA3–CA1 synapses during the phase of high astrocytic
[Ca2+] (see Figure 1F of Perea & Araque, 2007). They stimulated the pre-synaptic
neuron and simultaneously increased the astrocytic [Ca2+] through different pathways,
e.g., purinergic receptors (P2Y-R), and recorded the EPSCs. They used caged Ca2+
and used UV-flash to artificially increase astrocytic [Ca2+]. In contrast, in our
mathematical model, we allow an activity-dependent increase in astrocytic IP3
following an AP. As a measure of change in synaptic strength i.e., synaptic efficacy,
Perea & Araque (2007) demonstrated an increase in mean EPSC amplitude when
astrocyte was stimulated. We measured the mean EPSC after every 5 sec. In Figure
9B, the mean EPSCs have been measured relative to the mean EPSC during first 20
28
seconds, because it is the phase during which astrocytic [Ca2+] has not exceeded its
threshold (see Figure 6E). In Figure 9A, the mean EPSCs have been measured relative
to their overall mean. The impact of astrocytic response is clearly visible when we
look at Figure 9A and 9B. In astrocyte-independent information flow, there is not
much deviation (± 20%) from its mean value, while in astrocyte-dependent
information flow there is a transient increase of nearly 80%. This increase is
subsequent to the rise in astrocytic Ca2+ (see Figure 6E) and has decay time constant
(the time necessary to reach 1/e of its initial magnitude (Fisher et al., 1997)) of nearly
10s. This increase in synaptic efficacy falls under short-term-enhancement, in
particular augmentation, given the classification in Koch (1999, p – 311).
Figure 10. Cumulative probability of EPSC amplitude in response to an input signal of 5 Hz.
Astrocyte-dependent curve shifts upwards implying an increased probability of having EPSC amplitude
between 0.5 to 2.5 pA.
29
Figure 11. Cumulative probability distribution of inter -arrival time of EPSP for astrocyte-dependent
and astrocyte-independent information processing. The distribution associated with astrocyte -
dependent process shifts radically to the left suggesting reduced inter -arrival time due to enhanced
synaptic efficacy.
Perea & Araque (2007) also demonstrated an increase in cumulative probability of
EPSC amplitude before (astrocyte-independent) and during (astrocyte-dependent)
astrocyte stimulation (see Figure 1E of Perea & Araque, 2007). Similar to their
experimental observations, we also observed an increase in probability of EPSC
amplitude (see Figure 10). This implies that there are more chances of having EPSC
amplitude between 0.5 to 2.5 pA when astrocyte is present. Apart from an input signal
of 5 Hz we also tested cumulative probability for an input signal of 2 Hz and 10 Hz.
We observed that the astrocyte mediated potentiation (for an input signal of 2 Hz) of
synaptic efficacy becomes more prominent as demonstrated by a significant increase
in cumulative probability of observing EPSC amplitudes between 1.5 to 4.5 pA (see
Figure 1 of the supplementary online material), similar to Figure 10 here. On the other
hand, the contribution of astrocyte mediated potentiation (for an input signal of 10
Hz) of synaptic efficacy becomes less prominent or insignificant when compared with
synaptic efficacy following astrocyte-independent pathway (see Figure 2 of the
supplementary online material). The decrease in astrocyte mediated synaptic
30
potentiation observed with an increase in the frequency of input signal might be due
to the fact that our model has been calibrated for the experiments of Perea & Araque
(2007) where they applied mild pre-synaptic neuron stimulation.
A more comprehensive way of demonstrating synaptic enhancement will be to show
that we have more number of post-synaptic spikes under astrocyte-dependent
processing than astrocyte-independent processing. In Figure 11, we show cumulative
probability distribution for inter-arrival time of post-synaptic potentials. Cumulative
probability graph tells us the probability of having a post-synaptic firing in a time-
interval of length x ms (where x is an arbitrary point on the abscissa in Figure 11).
Obviously the probability of having a post-synaptic spiking will increase as we
increase the length of the time-interval (see Figure 11 where after 4000 ms mark
cumulative probability is 1 under both forms of information processing). It is apparent
from the figure that the probability of having post-synaptic spiking in short time-
intervals has greatly increased in presence of astrocyte (see Figure 11).
Figure 12. Synaptic potency under both forms of information processing (i.e. , astrocyte-independent &
astrocyte-dependent). Synaptic potency is given as a measure of mean EPSC, calculated over a time-
window of 4-sec, excluding failures. Synaptic potency is unchanged in both cases which has also been
31
observed in recent experiments (see Figure 1 of Perea & Araque, 2007); A. mean = -3.21 pA, std =
0.27 pA; B. mean = -3.11 pA, std = 0.24 pA. The two-sample paired t-test helps establish the previous
statement (p = 0.4475).
During this type of astrocyte-induced plasticity, it is known that synaptic potency
remains unchanged (Perea & Araque, 2007). Synaptic potency is given as a measure
of mean post-synaptic response, excluding failures. We calculated the mean of each
successful post-synaptic response in a time-window of 4 sec. It can be observed from
Figure 12 that there is no apparent difference in synaptic potency under both forms of
information processing. This observation was also confirmed statistically using a two-
sample student’s t-test. Synaptic potencies were assumed to be independent normally
distributed random samples. It was tested that both the samples are with equal mean
and equal but unknown variances (null hypothesis), against the alternative that the
means are not equal with 5% significance level. The result returned a p-value of
0.4475 indicating a failure to reject null hypothesis.
Using our simulation, we found that, all these measures (like synaptic efficacy, inter-
arrival time) which are used to demonstrate and establish the effect of astrocyte-
dependent pathway over synaptic plasticity depend primarily on two parameters i)
size of readily releasable pools of SLMVs in astrocyte and ii) rate of IP 3 production
due to pre-synaptic mGluRs. The size of readily releasable pools has recently been
determined using astrocyte cultures (Malarkey & Parpura, 2011). Here we show
change in neurotransmitter release probability for a readily releasable pool ½ (see
Figure 13A) and 1½ (see Figure 13C) in size of readily releasable SLMV pool
determined experimentally (see Figure 13B). Computer simulations shown in Figure
13A–13C reveal the effect of different sizes of readily releasable pool of SLMVs. It is
apparent that for a readily releasable pool of size 6 (i.e., containing 6 SLMVs)
astrocytes do not contribute to enhance pre-synaptic neurotransmitter release
probability. In fact, the average neurotransmitter release probability for readily
releasable pool of size 6 was 0.25, which is similar to the average neurotransmitter
release probability without astrocyte.
Figure 13B is the simulation of the model for default set of parameters listed in Table
2–Table 7. It is apparent from the figure that increase in neurotransmitter release
probability is preceded with increase in astrocytic Ca2+ concentration. In Figure 13C
we again show neurotransmitter release probability but for an increased size of readily
releasable SLMV pool. The effect of increased pool size is apparent from Figure 13C.
The average neurotransmitter release probability in this case was 0.35 . It should be
that coherence between astrocytic [Ca2+] (see Figure 13D–13F) and
noted
neurotransmitter release probability (see Figure 13A–13C) is absent only for
(compare Figure 13A and Figure 13D) which highlights a possible biological
condition under which astrocyte does not contribute to synaptic plasticity. The
average neurotransmitter release probability in three simulations was 0.25, 0.33 and
0.35. There is no considerable difference between the experimentally determined pool
32
va6n
size and a pool size of 18 (i.e. containing 18 SLMVs). However there was
considerable difference in the maximum extra-synaptic glutamate concentration when
latter compared with former (2.56 mM against 1.8 mM; data not shown). It is because
of the negative cooperativity of mGluRs in response to extra-synaptic glutamate
binding which ensures robust response to lower concentration of glutamate and also
ensures insensitivity to higher concentration of glutamate. Thus, extra-synaptic
glutamate is necessary for astrocyte mediated synaptic potentiation but with limited
influence. A more potent contributor to astrocyte mediated synaptic potentiation is the
IP3 production rate by pre-synaptic mGluRs.
Figure 13. Neurotransmitter release probability in response to changing availability of readily
releasable SLMV pool inside astrocyte. A. Neurotransmitter release probability for a readily releasable
SLMV pool of size 6. B. Neurotransmitter release probability for a readily releasable SLMV pool of
size 12. C. Neurotransmitter release probability for a readily releasable SLMV pool of size 18. D – F
Astrocytic Ca2+ concentration corresponding to the three simulations shown from A – C.
The maximum rate of IP3 production vg, by pre-synaptic mGluRs can be expressed in
terms of surface density
of mGluR, let the surface area of bouton exposed to
extra-synaptic glutamate released by astrocyte be S, the Avogadro Number NA, the
volume of bouton Vbtn, and the production rate of IP3 molecule per receptor rp. Then
(22)
Nadkarni & Jung (2007) estimated the maximum rate of IP3 production to be 0.062
nM ms-1 or
molecules/m3ms or
molecules/ms
per unit volume by mGluRs on the surface of the astrocyte. Such an estimate of IP 3
33
mGluRpmGluRgbtnA.rSvVN92330.062106.0231010170.37310
production rate is not known at boutons of CA3 pyramidal neurons. Thus, we used the
IP3 production rate by mGluRs on the pre-synaptic bouton to be same as that
determined by Nadkarni & Jung (2007) i.e., 0.062 nM per ms . Hence, for an average
bouton (of volume 0.13 μm3, Koester & Sakmann (2000)) at hippocampal CA3-CA1
synapse the production rate will be
If we assume
molecules/ms.
i.e. ≈ 0.0055 μm2 (0.31 μm is the radius of
bouton and 0.0028 μm is the strip of bouton exposed to extra-synaptic glutamate) of
bouton is exposed to glutamate released in the extra-synaptic cleft by the astrocyte.
Also if we assume that receptors produce 1 IP3 molecule per ms, then the receptor
density on relevant surface of the bouton is ≈ 0.87 per μm2. This assessment is in
conformity with the experiments as receptor density at synapses is estimated to be
between 200 – 2000 / μm2 (Holmes, 1995) and extra-synaptic receptor density is
known to be 230 times less than the receptor density at the synapse (Nusser et al.,
1995). The exact density of extra-synaptic mGluRs on CA3 pyramidal neurons is not
known. Hence, we simulated the model for a range of possible IP3 production rates
(see Figure 14A – D). The average neurotransmitter release probability for vg = 0.05
nM ms-1 is nearly equal to astrocyte-independent pathway of information processing
(Pr = 0.24 against Pr = 0.23). But as we increased the value of vg the effect of
astrocyte over synaptic plasticity became more prominent. The average
neurotransmitter release probability for vg = 0.1 nM ms-1 and vg = 0.2 nM ms-1 was
0.36 and 0.4 respectively. Please note that Figure 14B is same as Figure 13B and
Figure 8, it has been shown for comparison purposes only. Our simulation reveals that
vg is a critical parameter which can modulate the contribution of astrocyte induced
synaptic plasticity.
Figure 14. Plasticity of Neurotransmitter release probability in response to varying rate of IP 3
production by pre-synaptic group I mGluRs. A. Neurotransmitter release probability for an IP3
production rate of 0.05 μM per sec. B. Neurotransmitter release probability for an IP3 production rate
34
17180.373100.13100.004820.310.0028
of 0.062 μM per sec. C. Neurotransmitter release probability for an IP3 production rate of 0.1 μM per
sec. D. Neurotransmitter release probability for an IP3 production rate of 0.2 μM per sec. Please note
the change in Y-axis bounds for C and D.
4. Conclusion and future directions
There is a debate regarding the mechanism and calcium dependence of
gliotransmission and the role of gliotransmission in synaptic plasticity. Together they
imply that the effect of astrocytic calcium on synaptic plasticity is a controversial
issue. Here we have put together a number of phenomenological and biophysical
models for the processes shown in Figure 2 to simulate the effects on synaptic
strength with and without astrocytic Ca2+. From the computational modeling point of
view this is equivalent to controlling the effect of Ca2+ in astrocytes by genetic
engineering (Agulhon et al., 2010) and by calcium clamp (Henneberger et al., 2010)
in order to study the effects of astrocytic Ca2+ on synaptic plasticity. A better
understanding, through varieties of approaches, of calcium dynamics, signaling and
gliotransmitter release is necessary for settling down the aforementioned debate (Ben
Achour et al., 2010). Here we have taken a computational approach, and concluded
that the astrocytic Ca2+ does contribute to the synaptic augmentation at the time scale
of the order of seconds, for the given mathematical framework.
Here we presented a mathematical model which studies the effect of astrocyte over
the hippocampal CA3CA1 synaptic strength. It is found that given the pathway
(Figure 2B), astrocyte plays a significant role in modulating synaptic information
transfer. It might be possible that under physiological conditions, neurons also exhibit
the two types of information processing: i) astrocyte-independent ii) astrocyte-
dependent. Recent study performed by Di Castro et al. (2011) confirms that astrocytes
are activated under physiological stimulation of neighboring synapses. It is suggested
that neurons process information usually in astrocyte-independent manner unless
there is a need to modify synaptic efficacy according to various plasticity events
taking place in hippocampus (Navarrete & Araque, 2010; Panatier et al., 2011;
Navarrete et al., 2012).
Using our computational model, we identified two important parameters (readily
releasable pool size of SLMVs and maximum rate of IP3 production rate) which affect
astrocyte mediated synaptic potentiation at single CA3CA1 synapse. Our
simulations reveal a possible biological condition under which astrocyte Ca2+
oscillations do not contribute to synaptic potentiation (see Figure 13A). It was found
that maximum rate of IP3 production rate (vg) was a more potent modulator (of the
two parameters) of astrocyte mediated synaptic potentiation. Using equation (22) and
performing simple algebraic calculations we could predict mGluR density on relevant
surface of CA3 pyramidal neuron bouton which is experimentally unknown at
CA3CA1 synapse but was in conformity with experiments from other synapses
(Nusser et al., 1995; Holmes, 2000).
35
It should be pointed out that, it is not possible to conclude and assert that astrocyte
induces a particular type of synaptic plasticity (e.g., augmentation) using only a
temporal model, like the one proposed here, as synaptic plasticity depends on several
spatial constraints. As a future direction, it is proposed to develop a spatio-temporal
model to study the effects of spatial constraints, like release sites, Ca2+ sources etc.,
over modulation of synaptic activity. It is also known that a single hippocampal
astrocyte in CA1 region ensheaths around thousands of synapses (Schipke & Peters,
2009). Thus, it is possible for a single astrocyte to modulate signal processing at
thousands of synapses simultaneously. It has also been shown experimentally that,
astrocytes help to synchronize firings of neurons in CA1 region (Carmignoto &
Fellin, 2006). Hence, it is worthy to study the effects of astrocytes over the networks
of neurons. PortoPazos et al. (2011) recently performed a study where they
highlighted the importance of artificial astrocytes in modulating an artificial neural
network.
The present mathematical model is quite adaptable and can be easily extended to
study longer and other forms of synaptic plasticity (Tewari & Majumdar, 2012).
Another advantage of this model is that it can be extended to astrocyt ic
microdomains, where it is difficult to experimentally manipulate calcium fluctuations.
Simply increasing intracellular calcium is not sufficient for gliotransmitter release, as
evident from conflicting results (Henneberger et al., 2010; Agulhon et al., 2010;
Wenker, 2010). If calcium is required for transmitter release, then it may need to
occur in specific microdomains (Wenker, 2010), which has been over-looked and
needs examination using similar computational modeling approaches among others.
Acknowledgments
The work has been supported by the Department of Science and Technology,
Government of India, grant no. SR/CSI/08/2009. Helpful suggestions from Vladimir
Parpura are being thankfully acknowledged.
References
Agulhon, C., Fiacco, T. A., & McCarthy, K. D. (2010). Hippocampal short - and long-
term plasticity are not modulated by astrocyte Ca2+ signaling. Science. 327, 1250 – 1254.
Andersson, E. H. (2010). Astrocytes impose postburst depression of release probability at
hippocampal glutamate synapses. J. Neurosci. 30(16), 5776 – 5780.
Araque, A., Parapura, V., Sanzgiri, R. P., & Haydon, P. G. (1999). Tripartite synapses:
glia, the unacknowledged partner. Trends Neurosci., 22, 208 – 215.
Ben Achour, S., Pont-Lezica, L., Bechade, C., & Pascual, O. (2010). Is astrocyte calcium
signaling relevant for synaptic plasticity? Neuron Glia Biol., 6(3), 147 – 155.
Blackwell, K. T. (2005). Modeling Calcium Concentration and Biochemical Reactions.
Brains, Minds, and Media. 1, 1 – 27.
36
Bergersen, L. H. & Gundersen, V. (2009). Morphological Evidence For Vesicular
Glutamate Release From Astrocytes. Neuroscience, 158, 260–265.
Bertram, R., Sherman, A. & Stanley, E. F. (1996). Single-Domain/Bound Calcium
Hypothesis of Transmitter Release and Facilitation. J. Neurophys., 75, 1919 – 1931.
Bezzi, P., Gundersen, V., Galbate, J. L., Seifert, G., Steinhauser, C., Pilati, E., & Volterra,
A. (2004). Astrocytes contain a vesicular compartment that is competent for regulated
exocytosis of glutamate. Nature Neurosc., 7, 613 – 620.
Bliss, T. V. P. & Collingridge, G. L. (1993). A synaptic model of memory: long -term
potentiation in the hippocampus. Nature. 361, 31 – 39.
Bollmann, J. H., Sakmann, B., Borst, J. G. G. (2000). Calcium Sensitivity of Glutamate
Release in a Calyx-Type Terminal. Science. 289, 953 – 957.
Bowser, D. N. & Khakh, B. S. (2007). Two forms of single-vesicle astrocyte exocytosis
imaged with total internal reflection fluorescence microscopy. PNAS, 104, 4212–4217.
Cajal, R. Y. (1913). Histologie du systeme nerveux de l’homme et des vertebras.
Maloine, Paris.
Carmignoto, G. & Fellin, T. (2006). Glutamate release from astrocytes as a non-synaptic
mechanism for neuronal synchronization in the hippocampus. Journal of Physiology, 99,
98–102.
Chen, S. & Diamond, J. S. (2002). Synaptically Released Glutamate Activates
Extrasynaptic NMDA Receptors on Cells in the Ganglion Cell Layer of Rat Retina. The
Journal of Neuroscience, 22, 2165–2173.
Clements, J. D. (1996). Transmitter timecourse in the synaptic cleft: its role in synaptic
function. Trends Neurosci. 19, 163–71.
Coffey, W. T., Kalmykov, Yu., P., & Waldron, J. T. (2005). The Langevin Equation:
With Applications To Stochastic Problems In Physics, Chemistry And Electrical
Engineering, 2nd Edition. World Scientific Publishing, Singapore.
Danbolt, N. C. (2001). Glutamate uptake. Progress in Neurobiology. 65, 1–105.
Destexhe, A., Mainen, Z. F., & Sejnowski, T. J. (1998). Kinetic Models of Synaptic
Transmission, in: Methods in Neuronal Modeling, Koch, C. & Segev, I., MIT Press,
Cambridge, 1 – 25.
Di Castro, M. A., Chuquet, J., Liaudet, N., Bhaukaurally, K., Santello, M., Bouvier, D.,
Tiret, P., & Volterra, A. (2011). Local Ca2+ detection and modulation of synaptic release
by astrocytes. Nature Neuroscience 14, 1276–1284.
Dobrunz, L. E., Huang, E. P., & Stevens, C. F. (1997). Very short-term plasticity in
hippocampal synapses. Proc. Natl. Acad. Sci. USA. 94, 14843–14847.
37
Dumitriu, D., Hao, J., Hara, Y., Kaufmann, J., Janssen, W. G. M., Lou, W., Rapp, P. R.,
Morrison, J. H. (2010). Selective Changes in Thin Spine Density and Morphology in
Monkey Prefrontal Cortex Correlate with Aging-Related Cognitive Impairment. Journal
of Neuroscience 30, 7507–7515.
Emptage, N. J., Reid, C. A., & Fine, A. (2001). Calcium Stores in Hippocampal Synaptic
Boutons Mediate Short-Term Plasticity, Store-Operated Ca2+ Entry, and Spontaneous
Transmitter Release. Neuron. 29, 197–208.
Erler, F., Meyer-Hermann, M., Soff, G. (2004). A quantitative model for pre-synaptic free
Ca2+ dynamics during different stimulation protocols. Neurocomputing. 61, 169 – 191.
Fall, C., Marland, E., Wagner, J., & Tyson, J. (2002). Computational Cell Biology,
Springer-Verlag, Newyork.
Fellin, T. (2009). Communication between neurons and astrocytes: relevance to
modulation of synaptic and network activity. J. Neurochem. 108, 533 – 544.
Fiacco, T. A. & McCarthy, K. D. (2004). Intracellular Astrocyte Calcium Waves In Situ
Increase the Frequency of Spontaneous AMPA Receptor Currents in CA1 Pyramidal
Neurons. The Journal of Neuroscience, 24, 722–732.
Fisher, S. A., Fischer, T. M., & Carew, T. J. (1997). Multiple overlapping processes
underlying short-term synaptic enhancement. Trends Neurosci. 20, 170 – 177.
Franks, K. M., Bartol, T. M., & Sejnowski, T. J. (2002). A Monte Carlo Model Reveals
Independent Signaling at Central Glutamatergic Synapses. Biophys. J., 83, 2333–2348.
Fox, R. F. (1997). Stochastic Versions of the Hodgkin-Huxley Equations. Biophy. J., 72,
2068-2074.
Haydon, P. G. (2001). Glia: listening and talking to the synapse. Nature Rev. Neurosci. 2,
185 – 193.
Hebb, D. O. (1949). The organization of behavior. Wiley, New York.
Henneberger, C., Papouin, T., Oliet, S. H. R., & Rusakov, S. A. (2010). Long-term
potentiation depends on release of D-serine from astrocytes. Nature. 463, 232 – 236.
Hodgkin, A. L. & Huxley, A. F. (1952). A Quantitative Description Of Membrane
Current And Its Application To Conduction And Excitation In Nerve. J. Physiol. 117,
500-544.
Höfer, T., Venance, L., & Giaume, C. (2002). Control and plasticity of intercellular
calcium waves in astrocytes: a modeling approach. J. Neurosci. 22, 4850–4859.
Holmes, W.R. (1995). Modeling the Effect of Glutamate Diffusion and Uptake on
NMDA and Non-NMDA Receptor Saturation. Biophys. J. 69, 1734-1747.
38
Holtzclaw, L. A., Pandhit, S., Bare, D. J., Mignery, G. A., & Russell, J. T. (2002).
Astrocytes in adult rat brain express type 2 inositol 1,4,5trisphosphate receptors. Glia
39, 69–84.
Ishikawa, T., Kaneko, M., Shin, H-P & Takahashi, T. (2005). Pre-synaptic N-type and
P/Q-type Ca2+ channels mediating synaptic transmission at the calyx of Held of mice. J
Physiol. 568, 199–209.
Jafri, M. S. & Keizer, J. (1995). On the Roles of Ca2+ Diffusion, Ca2+ Buffers, and the
Endoplasmic Reticulum in IP3-lnduced Ca2+ Waves. Biophys. J., 69, 2139-2153.
Jensen, T. P., Filoteo, A. G., Knopfel, T., & Empson, R. M. (2007). Pre-synaptic plasma
membrane Ca2+ ATPase isoform 2a regulates excitatory synaptic transmission in rat
hippocampal CA3. J Physiol. 579, 85–99.
Keener, J. & Sneyd, J. (1998). Mathematical Physiology. Springer – Verlag, New York.
Koch, C. (1999). Biophysics of computation: information processing in single neurons.
Oxford University Press, New York.
Koester, H. J. & Sakmann, B. (2000). Calcium dynamics associated with action potentials
in single nerve terminals of pyramidal cells in layer 2/3 of the young rat neocortex. J.
Phys., 529, 625 – 646.
Li, Y-X & Rinzel, J. (1994). Equations for IP3 receptor mediated Ca2+ oscillations derived
from a detailed kinetic model: A Hodgkin-Huxley like formulism. J. theor. Biol. 166, 461
– 473.
Lynch, M. A. (2004). Long-term potentiation and memory. Physiol. Rev., 84, 87 – 135.
Lytton, W. W. & Sejnowski, T. J. (1991). Simulations of Cortical Pyramidal Neurons
Synchronized by Inhibitory Interneurons. J. Neurophys., 66, 1059 – 1079.
Malarkey, E.B. & Parpura, V (2011). Temporal characteristics of vesicular fusion in
astrocytes: Examination of synaptobrevin 2-laden vesicles at single vesicle resolution. J.
Physiol. 589, 4271-4300.
Marchaland, J., Cali, C., Voglmaier, S. M., Li, H., Regazzi, R., Edwards, R. H., & Bezzi,
P. (2008). Fast Subplasma Membrane Ca2+ Transients Control Exo-Endocytosis of
Synaptic-Like Microvesicles in Astrocytes. The Journal of Neuroscience, 28, 9122–9132.
Malenka, R. C. & Bear, M. F. (2004). LTP and LTD: An Embarrassment of Riches.
Neuron, 44, 5–21.
Mazzanti, M. & Haydon, P. G. (2003). Astrocytes Selectively Enhance N-Type Calcium
Current In Hippocampal Neurons. Glia. 41, 128–136.
Montana, V., Malarkey, E. B., Verderio, C., Matteoli, M., & Parpura, V. (2006).
Vesicular Transmitter Release From Astrocytes. Glia, 54, 700–715.
39
Nadkarni, S., & Jung, P. (2003). Spontaneous oscillations in dressed neurons: A
mechanism for epilepsy? Phys. Rev. Lett., 91, 268101-1 – 260181-4.
Nadkarni, S. & Jung, P. (2007). Modeling synaptic transmission of the tripartite synapse.
Phys. Biol. 4, 1–9.
Nadkarni, S., Jung, P., & Levine, H. (2008). Astrocytes optimize the synaptic
transmission of information. PLoS Comp. Biol., 4(5), 1 – 11.
Naundorf, B., Wolf, F., & Volgushev, M. (2006). Unique features of action potential
initiation in cortical neurons. Nature, 440, 1060 – 1063.
Navarrete, M. & Araque, A. (2010). Endocannabinoids Potentiate Synaptic Transmission
through Stimulation of Astrocytes. Neuron 68, 113–126.
Navarrete, M., Perea, G., de Sevilla, D. F., GomezGonzalo, M., Nunez, A., Martin, E.
D., & Araque, A. (2012). Astrocytes Mediate In Vivo Cholinergic-Induced Synaptic
Plasticity. PLoS Biol 10, e1001259.
Neher, E. (1998). Vesicle Pools and Ca2+ Microdomains: New Tools for Understanding
Their Roles in Neurotransmitter Release. Neuron. 20, 389–399.
Newman, E. A. (2003). New roles for astrocytes: Regulation of synaptic transmission.
Trends Neurosci., 26(10), 536 – 542.
Nikonenko, A. G. & Skibo, G. G. (2006). Age-Related Changes in Synaptic Vesicle Pools
of Axo-Dendritic Synapses on Hippocampal Ca2+ Pyramidal Neurons in Mice.
Neurophysiology. 38, 407 – 411.
Nusser, Z., Roberts, Z.D.B., Baude, A., Richards, J.G., & Somogyi, P. (1995). Relative
Densities of Synaptic and Extrasynaptic GABAA Receptors on Cerebellar Granule Cells
As Determined by a Quantitative Immunogold Method. Journal of Neuroscience, 15,
2948-2960.
Panatier, A., Vallee, J., Haber, M., Murai, K. K., Lacaille, JC, & Robitaille, R. (2011).
Astrocytes Are Endogenous Regulators of Basal Transmission at Central Synapses. Cell
146, 785–798.
Parpura, V., Basarsky, T. A., Liu, F., Jeftinija, K., Jeftinija, S. & Haydon, P. G. (1994).
Glutamate mediated astrocyte-neuron signaling. Nature, 369, 744–747.
Parpura, V. & Haydon, P. G. (2000). Physiological astrocytic calcium levels stimulate
glutamate release to modulate adjacent neurons. Proc. Natl Acad. Sci., 97, 8629–8634.
Perea, G., & Araque, A. (2002). Communication between astrocytes and neurons: a
complex language. J. Physiol. Paris., 96, 199 – 207.
Perea, G., & Araque, A. (2007). Astrocytes potentiate transmitter release at single
hippocampal synapses. Science. 317, 1083 – 1086.
40
Pittà, M. De, Goldberg, M., Volman, V., Berry, H., & Ben-Jacob, E. (2009). Glutamate
regulation of calcium and IP3 oscillating and pulsating dynamics in astrocytes. J Biol
Phys., 35, 383 – 411.
Porto-Pazos, A. B., Veiguela, N., Mesejo, P., Navarrete, M., Alvarellos, A., Ibanez, O.,
Pazos, A., & Araque, A. (2012). Artificial Astrocytes Improve Neural Network
Performance. PLoS ONE 6, e19109.
Schipke, C.G. & Peters, O. (2009). Glial Influence on Synaptic Transmission, in
Intercellular Communication In The Nervous System, R.C. Malenka, ed., Elsevier, pp.
112 – 120.
Schneggenburger, R. & Neher, E. (2000). Intracellular calcium dependence of transmitter
release rates at a fast central synapse. Nature. 406, 889 – 893.
Sneyd, J. & Falcke, M. (2005). Models ofthe inositol trisphosphate receptor. Progress in
Biophysics and Molecular Biology. 89, 207–245.
Suzuki, Y., Moriyoshi, E., Tsuchiya, D., & Jingami, H. (2004). Negative cooperativity of
glutamate binding in the dimeric metabotropic glutamate receptor subtype I. J. Biol.
Chem. 279, 35526–35534.
Tewari, S. & Majumdar, K. (2012). A Mathematical Model for Astrocytes Mediated LTP
at Single Hippocampal Synapses. Journal of Computational Neuroscience, to appear.
Tsodyks, M. & Markram, H. (1997). The neural code between neocortical pyramidal
neurons depends on neurotransmitter release probability. Proc. Natl. Acad. Sci. USA. 94,
719–723.
Verkhratsky, V. & Butt, V. (2007). Glial Neurobiology: A Textbook. John Wiley & Sons
Ltd.
Vernadakis, A. (1996). Glia-neuron intercommunications and synaptic plasticity. Prog.
Neurobiol. 49, 185 – 214.
Volman, V., Ben-Jacob, E., & Levine, H. (2007). The astrocytes as a gatekeeper of
synpatic information transfer. Neural Comp., 19, 303 – 326.
Weber, A. M., Wong, F. K., Tufford, A. R., Schlichter, L. C., Matveev, V., & Stanley, E.
F. (2010). N-type Ca2+
channels carry the largest current: implications for nanodomains
and transmitter release. Nature Neurosc. 13, 1348 – 1350.
Wang, L-Y, Fedchyshyn, M. J., & Yang, Y-M (2009). Action potential evoked
transmitter release in central synapses: insights from the developing calyx of Held.
Molecular Brain. 2, 1- 11.
Wang, S. S-H, Alousi, A. A., & Thompson, S. H. (1995). The lifetime of inositol 1,4,5
triphosphate in single cells. J. Gen. Physiol. 105, 149–71.
Wenker, I. (2010). An active role for astrocytes in synaptic plasticity? J. Neurophysiol.
104, 1216 – 1218.
41
Wu, L., & Borst, J. G. G. (1999). The reduced release probability of releasable vesicles
during recovery from short-term synaptic depression. Neuron. 23(4), 821 – 832.
Yang, Y., Ge, W., Chen, Y., Zhang, Z., Shen, W., Wu, C., Poo, M., & Duan, S. (2003).
Contribution of astrocytes to hippocampal long-term potentiation through release of D-
serine. PNAS. 100(25), 15194 – 15199.
Zucker, R. S., & Regehr, W. G. (2002). Short-term synaptic plasticity. Annu. Rev.
Physiol., 64, 355 – 405.
42
|
1808.03359 | 1 | 1808 | 2018-08-09T22:03:32 | Collective irregular dynamics in balanced networks of leaky integrate-and-fire neurons | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.AO",
"nlin.CD"
] | We extensively explore networks of weakly unbalanced, leaky integrate-and-fire (LIF) neurons for different coupling strength, connectivity, and by varying the degree of refractoriness, as well as the delay in the spike transmission. We find that the neural network does not only exhibit a microscopic (single-neuron) stochastic-like evolution, but also a collective irregular dynamics (CID). Our analysis is based on the computation of a suitable order parameter, typically used to characterize synchronization phenomena and on a detailed scaling analysis (i.e. simulations of different network sizes). As a result, we can conclude that CID is a true thermodynamic phase, intrinsically different from the standard asynchronous regime. | q-bio.NC | q-bio |
Collective irregular dynamics in balanced networks of leaky integrate-and-fire neurons
Antonio Politi,1 Ekkehard Ullner,1 and Alessandro Torcini2, 3
1Institute for Complex Systems and Mathematical Biology and
Department of Physics (SUPA), Old Aberdeen, Aberdeen AB24 3UE, UK
2Laboratoire de Physique Th´eorique et Mod´elisation, Universit´e de Cergy-Pontoise,
CNRS, UMR 8089, 95302 Cergy-Pontoise cedex, France
3CNR - Consiglio Nazionale delle Ricerche - Istituto dei Sistemi Complessi,
via Madonna del Piano 10, I-50019 Sesto Fiorentino, Italy
(Dated: August 13, 2018)
We extensively explore networks of weakly unbalanced, leaky integrate-and-fire (LIF) neurons
for different coupling strength, connectivity, and by varying the degree of refractoriness, as well
as the delay in the spike transmission. We find that the neural network does not only exhibit a
microscopic (single-neuron) stochastic-like evolution, but also a collective irregular dynamics (CID).
Our analysis is based on the computation of a suitable order parameter, typically used to characterize
synchronization phenomena and on a detailed scaling analysis (i.e. simulations of different network
sizes). As a result, we can conclude that CID is a true thermodynamic phase, intrinsically different
from the standard asynchronous regime.
I.
INTRODUCTION
The use of simple models proves often very helpful to
identify and characterize the mechanisms underlying gen-
eral dynamical phenomena. Computational neuroscience
is a field where this approach is potentially very powerful,
given the myriad of interactions involved in the function-
ing of the mammalian brain [1, 2]. However, setting the
appropriate level of simplicity is not a priori obvious. A
particularly enlightening example is the reproduction of
the background neural activity. Most of the numerical
and theoretical studies are based on the so-called rate
models, where each neuron is characterized by a single
coarse-grained variable representing the strength of the
ongoing activity [3, 4]. However, it is well known that
neurons work by emitting single spikes, so that it is more
natural to represent them as (nonlinear) oscillators. This
is, indeed the philosophy adopted by many studies based
on pulse coupled units, such as leaky integrate-and-fire
(LIF) neurons [5]. Accordingly, a general question arises
as to whether the two approaches are consistent with
one another and, in particular, to what extent spiking
neurons reproduce the scenario observed in rate models
[6, 7].
In this paper we look at the evolution of the so-called
balanced networks, where excitatory and inhibitory in-
teractions compensate each other [8], since the single-
neuron dynamics is rather irregular and reminiscent of
the background neural activity. A detailed theory of this
regime has been developed for rate models in the limit
of large connectivity [9 -- 12], but some theoretical studies
have been also made for spiking LIF neurons in highly
diluited networks [13]. Altogether, the scenario which
emerges from these studies is that of an asynchronous
regime, i.e. a dynamical state characterized by "micro-
scopic" fluctuations, but a steady and constant firing rate
at the macroscopic level (in the thermodynamic limit).
Nevertheless, evidence of irregular collective dynamics
has been recently found in both rate models [14] and
spiking neurons [15], so that a question arises about the
conditions for the emergence of partial synchronization.
In this paper we focus on networks of LIF spiking neu-
rons, as we believe that this is a more realistic and mean-
ingful setup. We extend the preliminary analysis pre-
sented in [15], by including a study of the avalanches and
of the scaling behavior of the fluctuations of the spiking
activity. Furthermore, we explore several variants of the
model to test the robustness of the results of our claims.
More precisely, we refer to a model that has been re-
peatedly investigated in the literature and in particular
in [13, 16] (several parameter values are herein selected
as in Ref. [16]). However, there are important differences
that should be stressed and, which allow us drawing con-
vincing conclusions about the existence of a CID: (i) in-
stead of sparse networks, we consider massive ones (i.e.
the connectivity K is assumed to be proportional to the
network size N ); (ii) the coupling strength is chosen to be
of the order of 1/√K. Furthermore, (iii) the unbalance is
chosen to be of the same order as the statistical fluctua-
tions of the input current, so as to avoid a complete dom-
inance of either excitation, or inhibition. Equipped with
such assumptions, we have simulated different network
sizes, finding that the seemingly CID observed for finite
networks survives in the thermodynamic limit, thereby
validating its identification as of a true thermodynamic
phase.
In section II we introduce the model, briefly review the
numerical scheme adopted for its simulations, and discuss
a problem related to its ill-defined structure. In fact, in
this model strictly simultaneous events can unavoidably
occur, which require an additional protocol to specify
the way they should be treated. This phenomenon is re-
lated to the occurrence of avalanches which we show to
marginally influence the thermodynamic limit.
In sec-
tion III, we deal with the dynamical properties of the
single-neuron dynamics, mostly focusing on the statis-
tics of the inter-spike intervals (ISI) and on the spectral
properties of the spike trains emitted by a single neuron
as well as of the post-synaptic input currents received
by each neuron.
In section IV, we discuss the collec-
tive dynamics, introducing a suitable order parameter to
quantify the degree of synchronization, and characteriz-
ing its stochastic-like behavior by means of the power
spectrum of the global neural activity. The perturbative
approach developed by Brunel in [13] is implemented to
perform a comparison with the numerical observations.
A qualitative agreement is found.
Finally a fractal-dimension analysis is performed,
which confirms the stochastic-like character of the dy-
namics: i.e.
its high-dimensional nature. In section V,
we explore the collective behaviour of the LIF model
when some of the parameters are varied, notably, delay,
refractoriness, connectivity and the presence of external
noise. All of the simulations confirm the robustness of
CID. Section VI contains a summary of the main results
and a brief list of the main open problems.
II. THE MODEL
In this Section we define the model following Ref. [13,
16], equipped with a suitable scaling of some parameters,
in order to preserve the CID observed at finite sizes. Fur-
thermore, we discuss some intrinsic ambiguities present
in the definition of the model that are related to the un-
avoidable occurrence of strictly synchronous events (the
simultaneous arrival of spikes emitted by different presy-
naptic neurons crossing the threshold at the same mo-
ment). The treatment of these events is somehow ar-
bitrary and requires the definition of a specific proto-
col. Furthermore, these synchronous events propagate
in the network as a sequence of events separated by
the propagation-time of the single spikes. The resulting
avalanches are analysed in detail for networks of increas-
ing size.
We consider networks of spiking neurons composed of
N supra-threshold LIF neurons, split in bN excitatory
cells and in (1 − b)N inhibitory ones [13, 16]. The mem-
brane potential Vi of the generic i-th neuron evolves ac-
cording to
τ Vi = R(I0 + Ii) − Vi ,
(1)
where τ = 20 ms is the membrane time constant, RI0 =
24 mV is a constant external DC "current", and RIi cor-
responds to the synaptic current due to the recurrent
connections within the network, namely
RIi = τ Xn
Gij(n)δ(t − tj(n) − τd) ,
(2)
2
where j(n) is the label of the node firing the n-th spike.
The synaptic connections among the neurons are random
without autapses, but with a fixed in-degree K for each
neuron.
In particular, we consider a massive network,
where the in-degree grows proportionally to the size as
K = cN ; the proportionality parameter c is termed
connectivity -- unless stated otherwise we set c = 0.1
throughout the entire paper. The random connections
are embodied in the matrix G, whose elements take the
values Gij = Je (−Ji), if the pre-synaptic neuron j is
excitatory (inhibitory), otherwise Gij = 0. Whenever at
time tj(n) the membrane potential Vj of the j-th neuron
reaches the threshold Vth = 20 mV for the n-th time, two
events are triggered: (i) the membrane potential is reset
to Vr = 10 mV and it is then held fixed for a refractory
period τr = 0.5 ms; (ii) a spike is emitted and received
τd = 0.55 ms later by the post-synaptic cells connected
to neuron j.
In order to maintain a fixed balance between excitation
and inhibition irrespective of the in-degree (and, thereby,
of the system size) we assume that the coupling strength
scales as the inverse of the square root of the in-degree, as
done in most of the literature on the balanced state [9 --
12]. More precisely, we assume Je = Jp1000/K and
Ji = (4 + g1pc/K)Je [15]. With these choices, and for
g1 = 100, we recover the setup studied in [16] for N =
10, 000, K = 1000 and b = 0.8. The coupling strength J
is our main control parameter.
A.
Integration scheme
The model equations (1,2) can be solved by either im-
plementing an event driven integration scheme, such as
described in [17, 18], or a more standard clock-driven
strategy [19]. The former scheme is, in principle, exact;
provided that the integration time step is small enough,
also the latter scheme is sufficiently accurate.
In fact,
the results discussed in this paper have been obtained by
implementing either scheme without any specific prefer-
ence. One exception is the above mentioned presence
of ambiguities in the very definition of the model, which
can be singled out only with reference to the exact event-
driven scheme. In general, two types of event break the
smooth evolution of the membrane potential: (i) a neu-
ron reaches the threshold; (ii) a neuron receives one (or
more) post-synaptic potentials (PSPs), elicited by one or
more pre-synaptic neurons τd = 0.55 ms earlier. In be-
tween these events all neurons evolve as being uncoupled
according Eq. (1) with Ii = 0. In order to evolve the sys-
tem, we first need to identify the next type of event and
thereby update all the membrane potentials {Vi} until
the occurrence of the event itself. This is done by solving
analytically Eq. (1) with Ii = 0. Thereafter, we process
either the threshold passing (i) or the spike-receiving (ii)
event. Each emitted spike is received τd = 0.55 ms later
by K = cN other neurons.
If the sending neuron j is
excitatory, it may happen that the post-synaptic poten-
tials triggers several threshold passings events at exactly
the same time. Since the delay is the same for all synap-
tic connections, in the next round, more than one of the
simultaneous spikes can reach the same neuron.
It is,
therefore, necessary to complement the definition of the
model with a rule to handle perfectly synchronous spikes,
since the outcome depends on the way such events are
treated.
We have decided that the most "neutral" rule consists
in first estimating the net effect of all the PSPs on the ba-
sis of the current value Vi of the membrane potential. Af-
terwards, all neurons which passed threshold are reset to
the common value Vr and the emitted spikes are stored to
be later received by the postsynaptic neurons connected
to the firing ones. We have verified that also in the case
of event driven integration schemes spike avalanches sep-
arated by exactly one time delay can occur. In the next
section we discuss how to quantify their relevance.
B. Characterization of the avalanches
We first monitored the number of simultaneously emit-
ted spikes E and the corresponding density RE per unit
of time; these are shown in Fig. 1 for different system sizes
and two different synaptic coupling values. The density
RE obviously increases with the system size. As shown
in the insets of Fig. 1, all curves collapse on the same
one, once the abscissa has been rescaled as E/√N . This
means that the number of simultaneous events E is of
order O(√N ), while the total number of emitted spikes
is of order O(N ). Therefore, the strictly simultaneous
events become less and less relevant, while approaching
the thermodynamic limit.
The presence of avalanches is a consequence of instan-
taneous synapses and identical time-delays. Avalanches
arise when a spike triggers a cascade of sub-sequent spikes
occurring at times exactly separated by the time delay.
We have monitored the time duration (called length L)
of the avalanches as well as their consistency (size S),
corresponding to the total number of spikes emitted dur-
ing an avalanche. The densities RL and RS per unit of
time of the avalanche length L and size S are reported
in Fig. 2 for various system sizes. There is a clear in-
crease of the length and size of the avalanches with the
system size N and hence (due to the massive coupling)
with the in-degree K = cN . Upon rescaling the RS den-
sities by √N , they collapse onto a same curve (that we
expect to depend on the coupling strength J) as shown
in the inset of Fig. 2 (b). Finally, no simple scaling be-
havior has been found for the avalanche length L. We
tested different scalings assumptions but none of them
yielded a convincing data collapse. We can nevertheless
safely conclude that the length grows even slower than
3
a)
0.01
1
E/√N
10
100
E
b)
0.01
1
E/√N
10
100
E
1000
RE
101
100
10-1
10-2
10-3
10-4
10-5
RE
101
100
10-1
10-2
10-3
10-4
10-5
RE
100
10-2
10-4
1
RE
100
10-2
10-4
1
FIG. 1. Density per unit of time RE of simultaneously emitted
spikes E for four different system sizes: N = 10, 000 (black),
N = 40, 000 (red), N = 160, 000 (green) and N = 640, 000
(blue) and two different coupling J = 0.2mV (a) and 0.5 mV
(b). The time durations of the simulations are set to 100 s.
The insets are based on the same data but the abscissas have
been rescaled as E/√N to visualise the decreasing effect with
the system size.
the size with N . Altogether, our results show that the
avalanches unavoidably appear also in the exact event
driven approach but do not contribute significantly to
the network dynamics in the thermodynamic limit.
III. THE MICROSCOPIC DYNAMICS
The CID is a macroscopically observable phenomenon
originated by an orchestrated interplay of the microscopic
oscillators. Before introducing appropriate indicators to
characterize the collective phenomena (see the following
Section IV), we first shed light on the microscopic dy-
namics. Each oscillator i is characterised by a membrane
potential Vi which evolves continuously in time, but is
affected by discrete events associated with the emission
times {ti(n)} of the single spikes and the corresponding
RL
10-1
10-2
10-3
10-4
10-5
100
a)
101
L
102
100
RS
10-2
10-4
RS
10-410-2
100
b)
10-2
101
100 S/√N
102
103 S
FIG. 2. Density per unit of time of the length (RL, a) and size
(RS, b) of avalanches for synaptic coupling J = 0.5 mV. The
length of the avalanche L has been expressed in multiple of the
time delay. Different system sizes are color coded: N = 10000
(black), N = 40000 (red) and N = 160000 (green). The time
durations of the simulations are fixed to 100 s. The inset of
panel (b) is based on the same data as the main one but the
abscissa has been rescaled as S/√N to visualise the decreasing
effect with the system size.
arrival times. A time trace of the individual membrane
potential V3 is shown in Fig. 3(b) as a red line together
with the mean potential hV i (t) = 1/N Pi Vi(t) (black
line) (here and in the following the symbol h·i stands for
an ensemble average). The membrane potential of a sin-
gle neuron exhibits significantly larger fluctuations than
those exhibited by the mean value hV i with only a limited
correlation among the two observables (see Fig. 3(b)).
The raster plot in the same time interval is depicted in
Fig. 3(c); it clearly reveals irregular population bursts,
whose degree of synchronization can be appreciated by
looking at the global firing activity Fg, i.e. the number
of spikes emitted in a fixed time window per neuron (see
Fig. 3 (d)). This last entity, whose time average corre-
sponds to the average firing rate, reveals clear irregular
oscillations.
As explained in Section II and shown in Fig. 1, simul-
taneous spike events are intrinsic properties of the model
and can lead to the simultaneous arrival of a mixture of
excitatory and inhibitory PSPs. Hence, the net result
p(t) = RIi/τ can be either positive or negative. Fig-
ure 4 shows examples of the post-synaptic input p(t) for a
synaptic coupling J = 0.5 mV in a system of N = 40, 000
neurons. The most probable input corresponds to a sin-
gle either excitatory, or inhibitory PSP. The other dis-
crete p(t) values are related to all possible combinations
of excitatory and inhibitory PSPs.
Additional information on the variability of the spik-
ing activity is contained in the probability distribution
function (PDF) of the ISIs of a single generic neuron.
A few PDFs obtained for N = 10, 000 for different cou-
pling strengths are reported in Fig. 5. For sufficiently
large coupling, the PDF is composed of an exponential
tail characteristic of a Poissonian dynamics plus a peak
at very short ISI, which is the consequence of the occa-
sionally periodic bursting activity of the neurons. These
results clearly indicate that the single neuron dynamics
0
-20
-40
0
20
0
-20
-40
i
102
101
100
6000
<V>(mV)
2000
4000
6000
8000
<V>(mV)
10000
V3(mV)
Fg(Hz)
6020
6040
6060
t(ms)
6100
4
a)
b)
c)
d)
FIG. 3. Characterization of the network dynamics. Mean
potential hV i versus time (a); a zoom in time of the evolutions
is shown in (b) together with the time evolution of the single
membrane potential V3 taken from a generic sample neuron
identified by the index 3 (red). For the same time interval
also the raster plot of the whole activity is reported in (c) as
well as a measure of the global spiking activity Fg in (d). The
dashed red line in (d) corresponds to the average firing rate,
ν 0 = 13.2 Hz. The data refer to N = 40, 000 and J = 0.5
mV, while the time window employed for the measure of Fg
is 0.11 ms.
2
p(t)
0
-2
-4
1000
1100
t(ms)
1200
1300
FIG. 4. Short time sequence of the post-synaptic inputs p(t)
received by a single oscillator for J = 0.5 mV and N = 40, 000.
is increasingly dominated by fluctuations for large J.
There are at least two ways to explore the spectral
properties of the single neurons: by looking at the evolu-
tion of the membrane potential or by recording the spike
events. We have focussed on the latter one, since it allows
for a comparison between the input stimulus and the out-
put response, as well as for the analysis of the collective
spiking activity (discussed in the next Section).
In order to characterize the output signal we counted
the number of emitted spikes within a fixed time win-
dow (we set it equal to 0.11 ms). The resulting signal
is a series of 0s interspersed with a few 1s. The input
is instead determined from the values of p(t) (reported
in Fig. 4) coarse-grained over time bins of 0.11 ms. Fur-
thermore, the power spectrum of the input signal has
been rescaled according to the number of excitatory and
inhibitory connections and their strength, i.e. by the
factor cN (bJ 2
i ), to be comparable with the
spectrum SS(f ) corresponding to the output time series.
e + (b − 1)J 2
5
20
Ss
(Hz)
10
40
Ss
(Hz)
20
1
2
f(kHz)
3
4
0.5
ISI(s)
1
1.5
10
100
1000
f(Hz)
PDF
10-2
10-3
10-4
10-5
10-6
10-7
0
FIG. 5. PDFs of the ISIs for a typical neuron for N = 10, 000
and various coupling strengths, namely J = 0.05 mV (red),
J = 0.1 mV (blue), J = 0.2 mV (magenta) and J = 0.5 mV
(black). The arrow denotes the increase of J.
Figure 6 reports the spectra SS(f ) associated to the in-
put and to the output signal of a single neuron, averaged
over 20 neurons randomly sampled out of N = 40, 000 for
the coupling J = 0.5 mV. At sufficiently high frequencies
(' 500 Hz) the input and output spectra almost coincide
for all system sizes and are basically flat and converge
towards the average firing rate ¯ν0 = 13.2 Hz, as expected
and shown in the inset of Fig. 6.
At lower frequencies, and especially for 50 ≤ f ≤ 500
Hz, the SS(f ) spectra of input and output differ from
one another (the differences persist for N = 160, 000,
where they have reached an asymptotic shape -- data not
shown):
in particular the input spectra exhibit a clear
peak at ≃ 75 Hz, while the output ones reveal just a
shoulder.
In the presence of an asynchronous regime,
input and output spectra should coincide (except for a
scaling factor).
In fact, in a series of recent papers, a
recursive method was developed to generate asymptotic
spectra, exactly by imposing a perfect correspondence
between input and output [20, 21]. The clear difference
shown by our numerical results (Fig. 6) provides a first
indication of a collective dynamics or, otherwise stated,
of nontrivial correlations among the different neurons.
This will be extensively elucidated in the next Sec-
tion, where we computed various indicators, including
the power spectrum of the overall activity, whose spec-
trum is not too different from that of the input (see
Fig. 9).
IV. COLLECTIVE DYNAMICS
FIG. 6. Spike train spectrum Ss of the input (blue and green
dashed lines) and output (black and red solid lines) of a sin-
gle neuron within an ensemble of N = 40, 000 and 160, 000,
respectively, with J = 0.5 mV. The results has been obtained
from time series of 100 s duration and averaged over 20 dif-
ferent neurons. The inset is the linear representation of the
main plot with a focus on high frequencies.
noticeable in the structure of a typical raster plot, which
consists in an irregular alternation of regions of different
density (see Fig. 3(c)). The evolution of hV i provides
a more accurate representation. The fluctuations of hV i
are in fact smaller than those of the individual membrane
potential Vi(t) (see the red line in Fig. 3(b)), but never-
theless definitely appreciable.
On a quantitative level, it is convenient to introduce
the synchronisation measure ρ
2
ρ2 ≡ hV i2 − hV i
hV 2 − V
i
2
,
(3)
where the overbar denotes a time average. Perfectly syn-
chronised neurons behave in exactly the same way, so
that the numerator and the denominator are equal to
one another and ρ = 1. If instead, they are statistically
independent, ρ ≈ 1/√N .
The progression of the running average of ρ can be ap-
preciated in Fig. 7 for J = 0.5 mV, see the middle bunch
of trajectories, labeled by (0.5 , 0.55). The order parame-
ter approaches ρ ≈ 0.35, irrespective of the networks size,
indicating that CID is not a finite-size effect, but survives
in the thermodynamic limit as defined in Ref. [15]. The
jumps observed at early times of the running average are
caused by sudden drops of the mean potential hV i. One
of the strong drops is shown in Fig. 3(a) around t = 9, 500
ms. These rare events appear randomly at all times, but
their effect on the cumulative average obviously decreases
as time progresses.
In this section we discuss the collective dynamics which
emerges from the correlations in the microscopic activity
of the single neurons. A qualitative evidence is already
The coefficient of variation Cv is another measure of
irregularity of the dynamics, based on the fluctuations
of the ISI, rather than of the membrane potential. More
ρ
0.5
0.4
0.3
0.2
0
(0.0 , 0.55)
(0.5 , 0.55)
(0.5 , 0.0)
10
20
t(s)
30
40
50
<Cv>,
Ξ
1
0.1
0.01
0.01
6
0.1
J(mV)
1
FIG. 7. Running average of the order parameter ρ for increas-
ing integration time t, obtained after discarding a transient
of at least 5 s for J = 0.5 mV for different system sizes and
parameter settings labeled by the tuple (τr , τd). The bunch
of curves reported in the middle of the figure corresponds to
the standard setup with delay τd = 0.55 ms and refractoriness
τr = 0.5 ms. The upper family refers to a setup with standard
delay τd = 0.55 ms but without refractoriness, i.e. τr = 0ms.
The lower collection of lines corresponds to no delay τd = 0.0
ms but with refractoriness τr = 0.5 ms. The system sizes are
color coded in ascending order: black, red, green, blue and
orange for N = 10, 000, 20, 000, 40, 000, 80, 000 and 160, 000,
respectively.
precisely, we calculate
hCvi = h
σS
τS i
where σS is the standard deviation of the single-oscillator
ISI, while τS is the corresponding mean ISI.
For J = 0 the single-neuron activity is strictly periodic
and thus Cv is equal to zero. We expect it to increase
when the coupling strength J is switched on. For small J
we can indeed appreciate a power-law growth, hCvi ≈ J α
(see the black squares in Fig. 8) with a value of the rate
α close to 1.6. The growth of hCvi continues for stronger
coupling strenghts, becoming larger than 1, the value ex-
pected for a Poisson statistics. For sufficiently large cou-
pling, we observe bursting dynamics of the neurons, cor-
responding to hCvi > 1. A similar behavior of hCvi with
the coupling strength has been reported for inhibitory
sparse networks of LIF in the absence of delay but for
sufficiently slow synaptic decays [22]. Interestingly, the
hCvi values are substantially independent of the system
size (we have tested values of N up to 640, 000).
The coefficient of variation measures the amplitude of
the fluctuations, but it is insensitive to temporal correla-
tions: hCvi is strictly larger than zero already for a very
regular sequence of ISIs such as a periodic alternation of
two values t1 and t2, with t1 6= t2. In order to have a
more accurate indicator, we have computed the following
diffusion coefficient. Let Tn denote the time of n-th spike
FIG. 8. The mean coefficient of variation of the interspike in-
terval hCvi. The black squares refer to hCvi for N = 10, 000
and the red crosses show the square root of the rescaled diffu-
sion coefficient Ξ. Power law fits J α are denoted with dashed
lines. The black dashed line power law fit matches the lower
part of hCvi with exponent α = 1.62 very well whereas the
blue dashed lines refers to a power law fit with exponent
α = 1.14.
emitted by a given neuron, so that Tn − Tn−1 is the n-th
ISI. Let, then
DS = lim
n→∞
(Tn − nτS)2
n
(4)
be the diffusive coefficient of the process Tn. We finally
define
√DS
τS
,
Ξ ≡
(5)
Ξ is plotted in Fig. 8 (see the red crosses). Above J =
0.1, it basically coincides with Cv, indicating that Tn is
essentially a renewal process. For J ≤ 0.1, Ξ decreases
more slowly (with a rate close to 1.14) than Cv. This
is clearly due to the increasingly periodic character of
the dynamics. The overall scenario is reminiscent of the
phase transition discussed by Ostojic in [16].
The power spectrum of the global activity Sg sheds
light on collective phenomena from yet a different per-
spective. Analogously to the single oscillators, the spike
times have been converted into a single time series count-
ing the number of spikes emitted in each time bin. We
chose the same time bin of 0.11 ms as in the previous
cases. An example of the global field Fg is included in
Fig. 3 in the bottom panel (d); it clearly shows an ir-
regular behavior. The power spectrum of the global ac-
tivity has been divided by N 2 to allow for a meaningful
comparison amongst different system sizes and with the
single-neuron spectrum (Fig. 6). The spectra obtained
for different system sizes are plotted in Fig. 9. For f > 40
Hz, they collapse onto one another, suggesting that the
dynamics remains irregular in the thermodynamic limit,
i.e. that the fluctuations exhibited by the collective vari-
ables are not finite-size effects. It should be stressed that
in the absence of collective effects, i.e. for asynchronous
states, the spectrum of the global activity would be pro-
portional to N rather than to N 2.
Below 40Hz, the spectral amplitude decreases with
the system size, suggesting that the zero-frequency peak
eventually disappears (at least for J = 0.5). Altogether,
the spectral power is mostly concentrated in two fre-
quency ranges: (i) a broad peak around f ≈ 75 Hz, which
corresponds to the peak observed in the single neuron
spectra SS and is presumably related to a time scale of
the order of the membrane time constant τ = 20 msec
and (ii) a peak around f ≈ 1818 Hz (and its multiples),
which corresponds to the inverse of the delay. A compar-
ison with the spectrum of the single neuron activity (see
Fig. 6), reveals that the latter one is characterized by a
much stronger high-frequency component (of white-noise
type) and weaker peaks in correspondence of the inverse
delay time.
We have finally implemented the perturbative ap-
proach developed by Brunel [13], based on the assump-
tion of a sparse coupling. The idea basically consists in
solving a self-consistent noisy Fokker-Planck equation for
the probability density P (v, t) of the membrane potential
v,
τ
∂P
∂t
=
∂
∂v
[(v−µ−µe)P ]+
σ2
2
∂2P
∂v2 +σ0√cτ
∂P
∂v
ζ(t) . (6)
The first term in the r.h.s. is nothing but the determinis-
tic current defined in Eqs. (1,2), with µe = RI0. The sec-
ond, diffusive contribution, accounts for the unavoidable
statistical fluctuations of the input signal arising from
the coupling with the other neurons and its amplitude is
estimated under the assumption of being a Poisson pro-
cess. Finally, the last term is a common noise due to the
fact that different neurons partially share the same input
signals, whenever they share the same afferent neurons.
More precisely, the drift µ is defined as
µ = −√cτ (1 − b)g1J ν(t − τd)
(7)
where ν(t) is the instantaneous firing rate, while an ex-
pression for the diffusion coefficient σ2 can be obtained
by assuming that the spike train follows a Poisson statis-
tics,
σ2 =
bτJ 2
1 − b
ν(t − τd) .
(8)
Finally, σ0 is the value of σ corresponding to the station-
ary value of the firing rate in the asynchronous regime ν0
and J = J√K (see the appendix for a more precise defi-
nition). The power spectrum n(ω)2 of the neural activity
can be determined by solving perturbatively the Fokker-
Planck equation. The technical details are presented in
Appendix A: we practically follow the method introduced
7
3
Sg
(Hz)
2
1
0
100 101 102
f(Hz)
Sg
(Hz)
101
100
10-1
10-2
0
1
2
3
f(kHz)
4
FIG. 9. Power spectra Sg of the global activity for J = 0.5
mV. The spectra reported for different system sizes N =
10, 000, N = 40, 000 and N = 160, 000 are shown in red,
green and blue, respectively. The black line represents the
results from the Brunel's perturbative theory (see Appx. A)
and it is reported only in the main plot and not in the inset.
in [13], the main difference being the numerical strategy
adopted to determine the spectrum. The resulting curve
is shown in Fig. 9 (black line) after converting the angu-
lar frequency ω into the frequency f . The perturbative
approach qualitatively reproduces the shape of the spec-
trum, including the position of the peaks. On the other
hand, the height of the peaks strongly deviates from the
numerical simulations. For large coupling, the agreement
worsens and the perturbative approach fails even in re-
producing qualitatively the spectra at low frequencies, as
shown in [15].
Finally, we have studied the neural activity by im-
plementing nonlinear-dynamics tools, to determine the
(effective) fractal dimension De of the mean potential
hV i (t). Given the sequence of values hV i (tn), obtained
by sampling the original signal every ∆t = 1 ms over 500
s (i.e. resulting in 500, 000 data points), this is embedded
into a space of dimension m, by building vectors of the
type [hV i (tn),hV i (tn+1), . . . ,hV i (tn+m−1)]. The frac-
tal dimension has then been estimated by using a vari-
ant of the nearest-neighbour method recently proposed
in [23]. In particular, Nr reference points are randomly
selected (Nr = 105 in our case), then each reference point
is compared with other n randomly selected points (up
to the number of data points available) determining the
distance εm(k, n) of the k-th neighbour for different val-
ues of m and k. The distance is herein estimated using
the maximum norm. An established theory [24], implies
that for large n,
−
ln n
hln εm(k, n)i ≈ De ,
where the angular brackets denote the average over the
reference points, while De is the information dimension.
10
De
8
6
4
2
0
m=15
m=9
m=6
m=3
0.1
ε
1
FIG. 10. Effective dimension De as a function of the resolu-
tion for J = 0.2 mV and different system sizes: N = 10, 000
(dotted), 40, 000 (dashed), and 160, 000 (solid). The different
groups of curves correspond to different embedding dimen-
sions m.
In practice, the logarithmic derivative of ε varies with n
before eventually converging to its asymptotic value. Ac-
cordingly, it can be interpreted as an effective, resolution-
dependent dimension, which is, in fact, independent of
the order k of the neighbour considered in the simula-
tions. In practice, given ε, the inverse of the logarithmic
derivative is first determined and then plotted versus the
resolution ε. The results are reported in Fig. 10. They
show that, independently of the network size, the effec-
tive dimension increases upon decreasing the resolution.
The stochastic-like nature of the dynamics is further con-
firmed.
V. ROBUSTNESS
In this section we investigate the robustness of CID, by
testing its properties when some of the model parameters
are modified, notably refractoriness, delay, connectivity
and finally after introducing an external noise which acts
independently on each neuron.
We start by showing the dependence of the average fir-
ing rate ν0 and of the coefficient of variation hCvi on the
system size in a setup where either delay or refractoriness
is missing. From the data reported in the Table I, which
refer to J = 0.5mV, we observe that, strange enough, the
firing rate slows down, when the delay is removed. This
is because the absence of delay induces a more homoge-
neous firing activity, which, in turn, is more dominated
by the inhibitory neurons due to the weak unbalance.
This interpretation is confirmed by the lower degree of
synchronization that can be appreciated by looking at
the (0.5,0.0) curves in Figure 7, which refer to different
network sizes. Although ρ decreases, CID is still present
in the absence of delay. In fact ρ is substantially indepen-
8
TABLE I. Mean firing rates ν 0 and mean coefficients of vari-
ation of the ISI hCvi in absence of delay (τd = 0) or no re-
fractoriness (τr = 0) for different system sizes N , for J = 0.5
mV. The last two column reference to the standard setup with
delay τd = 0.55 ms and with refractoriness τr = 0.5 ms.
standard
τr = 0
τd = 0
N
10,000
20,000
40,000
80,000
ν 0 [Hz]
13.8
13.2
12.4
11.9
hCvi ν 0 [Hz]
15.9
1.68
14.3
1.63
13.4
1.58
1.54
13.0
hCvi ν 0 [Hz]
15.3
1.80
14.3
1.67
13.2
1.60
1.55
12.8
hCvi
1.75
1.67
1.59
1.55
dent of N (it actually even slowly increases). Additional
studies of the spectral properties confirm that the collec-
tive dynamics is irregular (data not shown). This is at
variance with the setup studied in [25], a heterogeneous
ensemble of fully coupled, inhibitory, LIF neurons.
In
that context, CID disappears as soon as the delay van-
ishes. It is still to be understood whether the qualitative
difference is due to the heterogeneity (dispersion in the
bare firing rates of the single neurons).
Refractoriness is less relevant. From the data in Ta-
ble I, we see that its absence does neither significantly
modify the firing rate (which naturally increases by a
small amount), nor the degree of irregularity of the sin-
gle neuron. Even though hCvi slowly decreases upon in-
creasing N , it remains substantially larger than 1, the
expected value for a Poisson statistics. As for the collec-
tive dynamics, we notice in Fig. 7 (see the curves labeled
(0.0,0.55)) that synchronization increases upon remov-
ing refractoriness. The convergence is slower than in the
previous case: this is because of the presence of several
sudden burst of synchronizations (see the upward jumps
exhibited by ρ(t)), which require longer time scales for
them to be suitably averaged out.
After having verified that neither delay nor refractori-
ness are necessary ingredient for CID to be observed, we
now explore the role of the connectivity c.
In Fig. 11
we plot three key parameters (the firing rate ν0, Cv and
ρ) as a function of the coupling strength J for different
c values. In panel (a), we see that the firing rate is al-
most independent of c in the small coupling limit, while
it progressively decreases upon increasing c in the strong
coupling regime. This is due to the fact that a strong
connectivity reduces the fluctuations which are known to
be responsible for the larger ν 0 observed for strong cou-
pling [26]. The progressive regularization of the neural
activity is confirmed in panel (b), where we see that hCvi
decreases upon increasing the connectivity.
Quite interesting is the dependence of the order pa-
rameter ρ on c. The clean data collapse for J < 0.4, in-
dicates that up to a 30% connectivity, ρ scales as √c, in
agreement with the perturbative theory developed in [13]
a)
70
ν
0
(Hz)
40
20
0
0
2
ρ/√c
c)
1
0
0
<Cv>
4
b)
3
2
1
0
0
0.2
0.4
J(mV)
0.8
0.2
0.4
J(mV)
0.8
0.2
0.4
J(mV)
0.6
0.8
a)
c)
200
ν
0
(Hz)
100
50
0
0
0.5
ρ
0.3
0.2
0.1
0
0
9
<Cv>
6
b)
4
2
0
0
0.2
0.4
J(mV)
0.8
0.2
0.4
J(mV)
0.8
0.2
0.4
J(mV)
0.6
0.8
FIG. 11. The mean firing rate ν 0, the mean coefficient of
variation in the ISI hCvi and the synchronisation measure ρ
versus the coupling J for different network connectivities c.
The network connectivity c follows in ascending order the di-
rection of the arrow according c = 0.01, 0.05, 0.1, 0.2, 0.4 and
0.6. The usually used c = 0.1 has been singled out by dashed
lines. The system size is N = 40, 000 for all simulations.
and briefly recalled in Appx. A, which predicts a power
spectrum proportional to c. The strong coupling regime
(J > 0.4) seems to be characterized by different scaling
properties, but additional simulations for different net-
work sizes are required to put the statement on a more
firm basis.
So far, our simulations have been performed for a
slight prevalence of the inhibitory activity. In fact (for
N = 104) the ratio between the two coupling strengths
is g ≡ Ji/Je = 5, to be compared with a 1 : 4 ratio of
the two corresponding populations. In order to investi-
gate the role of the degree of unbalance, we have studied
two additional g-values, g = 4, and 5, which, respectively,
correspond to a perfect balance and a stronger prevalence
of inhibition. The results are presented in Fig. 12. The
most important point is that CID is present for both pa-
rameter values, confirming the robustness of this phase.
On a more quantitative level, unsurprisingly, the per-
fectly balanced state is characterized by a much stronger
firing activity.
Finally, we analyse the role of noise, by adding iid
white noise terms ξ(t) to the single neuron dynamics
(Eq. (1)), such that hξ(t + τ )ξ(t)i = 2Dδ(τ ). The de-
pendence of ρ on D is reported in Fig. 13, for two differ-
ent network sizes. The noise tends obviously to decrease
the strength of the collective dynamics, without, how-
ever, killing it. In fact, CID survives even for moderately
strong noise amplitudes, as it is appreciated by seeing
that ρ does not vary significantly upon increasing N .
Altogether, CID is a very robust property, which sur-
vives even when noise is added, the connectivity is de-
creased, the balance is changed, the delay or refractori-
ness removed from the model equations.
FIG. 12. The mean firing rate ν 0, the mean coefficient of
variation of the ISI hCvi and the synchronisation measure ρ
versus the coupling strength J for different balance factors g
for a system size N = 10, 000. The dashed black line reference
to the slight unbalance g = 5 used throughout the paper,
whereas the solid green line shows the situation for stronger
unbalance (g = 6) and the solid red line corresponds to the
perfect balanced setup (g = 4). The blue crosses refer to a
system size N = 40, 000 for a perfectly balanced situation, i.e.
g = 4.
0.4
ρ
0.3
0.2
0.1
0
0
0.5
1
1.5
2
2.5
D
FIG. 13. Order parameter ρ for J = 0.5 mV and different
noise levels D for N = 10, 000 (black circuits) and 40, 000
(green triangles). The solid line is a quadratic fit.
VI. CONCLUSIONS AND OPEN PROBLEMS
In this paper we have presented an extensive analysis
of the collective dynamics emerging in a quasi-balanced
network of LIF neurons. The irregularity of the collective
dynamics is testified not only by the power spectra of the
neural activity but also by a fractal-dimension analysis.
The detailed simulations performed for different param-
eter values confirm that irregular dynamics is very ubiq-
uitous. Several questions are, however, still open. Here
we list the main ones.
(A) To what extent is this irregular dynamics related to
the similar regime observed in globally coupled, hetero-
geneous neurons [23, 25]? In those setups, which are rem-
iniscent of the Kuramoto model, the heterogeneity seems
to be a crucial ingredient, since CID disappears when the
diversity among the neurons is removed. Here, it seems
that the collective, stochastic-like dynamics is the result
of a microscopic pseudo-chaotic evolution, which perco-
lates up to macroscopic scales, as a consequence of the
quasi-balanced regime. Whether this is really the correct
explanation it is however still unclear.
(B) All the models so far explored assume δ-pulses,
but this is obviously an approximation. The limit of in-
finitely narrow PSPs is singular, as shown, for instance,
while investigating the stability of the splay state [27].
Furthermore, we have seen that the presence of strictly δ-
like pulses induces unavoidable synchronous events whose
treatment requires additional ad-hoc hypotheses. It will
therefore, be instructive to explore networks character-
ized by PSPs of finite duration, e.g by considering expo-
nential or α-pulses.
(C) The numerical analysis has revealed that collective
dynamics arises also for a very small coupling strength.
The weak-coupling limit is typically amenable to a per-
turbative treatment. Accordingly, it is plausible that
a model of Kuramoto-Daido phase-oscillators might be
able to reproduce a similar regime and, at the same time,
allow for an analytical treatment.
(D) The only limit where the irregular collective dy-
namics vanishes is that of a sparse network, where
K/N → 0 for N → ∞. However, this statement refers
to random Erdos-R´enyi-type networks. It would be in-
teresting to explore different more elaborated network
structures as well as the role of heterogeneity.
(E) Qualitatively speaking, it looks like some differ-
ences exist between the weak and strong coupling regime.
In the former case, the single neuron spiking activity is
strongly correlated (being far from a renewal process) as
shown in Fig. 8 and the strength of the collective dynam-
ics is reproduced as expected by the perturbative theory
(see the nice overlap among the curves reported in panel
(c) of Fig. 11). In the latter case, the neuronal activity
is very well approximated by a renewal process and, at
the same time, the perturbative theory seems to fail al-
ready for a 1% connectivity. These differences suggest
that at small coupling the neuronal dynamics is mean
driven, i.e.
is dominated by the mean value of the DC
currents, while at large coupling it is fluctuation driven,
i.e. the neurons are in proximity or below the threshold
and the firings are triggered by fluctuations of the input
currents. Similar transitions from mean to fluctuation
driven dynamics has been recently reported in sparse in-
hibitory heterogeneous networks made of LIF neurons in
[22] and composed of realistic models of striatal medium
spiny neurons in [28]. An additional finite-size analysis
is necessary to test whether this is a true transition that
10
persists in the thermodynamic limit, as claimed by Os-
tojic [16] for strongly diluted networks.
(F) In all of our simulations, excitatory and inhibitory
neurons have been assumed to be equal to one another.
This implies that the same combination of excitatory and
inhibitory fields is automatically consistent with the evo-
lution of both types of neurons. This strong limitation
should be lifted before drawing yet more general conclu-
sions about the ubiquity of collective irregular dynamics.
ACKNOWLEDGMENTS
The authors acknowledge N. Brunel, F. Farkhooi, G.
Mato, S. Ostoijc, A. Roxin, and M. di Volo for useful
discussions. One of us (AT) has been supported by the
French government under the Excellence Initiative I-Site
Paris Seine (No ANR-16-IDEX-008) and under the Labex
MME-DII (No ANR-11-LBX-0023-01). The work has
been mainly realized at the Max Planck Institute for the
Physics of Complex Systems (Dresden, Germany) during
the Advanced Study Group 2016/17 "From Microscopic
to Collective Dynamics in Neural Circuits".
Appendix A: Perturbative approach
Starting from the Fokker-Planck equation (6), the fir-
ing rate ν(t) can be also expressed in terms of the prob-
ability current at v = vθ, i.e.
∂P
∂v
(vθ, t) = −
2ν(t)τ
σ2(t)
where P (vθ, t) = 0. Additionally, the probability density
must be continuous at the reset potential vr, where there
is an additional current due to the neurons ending their
refractory period
∂P
∂v
(v+
r ) −
∂P
∂v
(v−
r ) = −
2ν(t − τr)τ
σ2(t)
The list of boundary conditions is completed, by includ-
ing
lim
v→−∞
P (v, t) = 0
and the normalization
Z vθ
−∞
dvP (v, t) + pr(t) = 1
where
pr(t) = Z t
t−τr
duν(u)
is the probability for a neuron to be in the refractory
period at time t.
So long as the fluctuating term in Eq. (6) can be ne-
glected, the dynamics relaxes towards a stationary state
which can be interpreted as an asynchronous regime char-
acterized by a constant current ν0 and constant fluctua-
tions σ0,
µ0 = −cτ (1 − b)g1J ν0
σ2
0 =
ν0 ,
τJ 2
1 − b
(A1)
(A2)
which can be determined self-consistently using the fol-
lowing expression for ν0,
1
ν0
= τr + τ√πZ
Vth −µ0 −µe
σ0
Vr −µ0 −µe
σ0
du eu2
(1 + erf(u))
(A3)
It is now convenient to introduce the following changes
of variables
σ0
2τ ν0
Q =
P
y =
v − µ0 − µe
ν0
n =
ν
ν0 − 1 .
The threshold and reset potentials become,
yθ =
vθ − µ0 − µe
σ0
yr =
vr − µ0 − µe
σ0
while the Fokker-Planck equation can be rewritten as
τ
∂Q
∂t
=
∂
∂y (cid:18)y − n(t − τd)
1 + n(t − τd)
µ0
σ0(cid:19) Q +
∂y2 + √cτ
∂2Q
2
+
∂Q
∂y
ζ(t) , (A4)
∂Q
∂y
(yθ) = −
1 + n(t)
1 + n(t − τd)
and
(y+
∂Q
∂y
1 + n(t − τr)
1 + n(t − τd)
The stationary solution can be expressed as
r ) −
∂Q
∂y
r ) =
(y−
where
Q0(y) = e−y2
Q0(y) = e−y2
F (y)
y > yr
F (yr)
y < yr
F (y) = Z yθ
y
dueu2
.
Finally, from the definition of n it follows that n0 = 0.
As a next step, we linearize the Fokker-Planck equation
around the stationary solution to treat the fluctuations
in a perturbative way. Upon assuming Q = Q0 + q and
neglecting nonlinear terms in q and n,
τ
∂q
∂t
=
+
∂yq
∂y
+√cτ
1
2
∂q
∂y
∂2q
∂y2 − n(t − τd)(cid:18) µ0
σ0
dQ0
dy −
1
2
d2Q0
dy2 (cid:19)
ζ(t) ,
(A5)
accompanied by the boundary conditions
and
11
while the b.c. can be rewritten as
∂q
∂y
(yθ) = −1 − n(t) + n(t − τd)
and
∂q
∂y
(y+
r ) −
∂q
∂y
(y−
r ) = 1 + n(t − τr) − n(t − τd)
Eq. (A5) is a linear Langevin equation operating in an
infinite-dimensional space. The best way to handle it is
to Fourier transform Eq. (A5), introducing q(y, ω) and
n. This way we obtain two first order ode's for each
frequency variables,
dq
dy
du
dy
where
= u
(A6)
= −2y u + 2(iωτ − 1)q + G(n, y)
G(n, y) = e−iωτd n(cid:18) 2µ0
ν0
dQ0
dy −
d2Q0
dy2 (cid:19) − 2√cτ
dQ0
dy
and we have implicitly assumed that the power spectrum
of ζ(t) is flat and equal to 1. The corresponding b.c. write
as
q(yθ, ω) = 0
u(yθ) = n(e−iωτd − 1)
u(y−
r ) = u(y+
r ) − n[e−iωτd − e−iωτr ]
(A7)
(A8)
These equations have been numerically solved by inte-
grating Eq. (A6) backward in y, starting from y = tθ
for each given frequency ω and a tentative value of n(ω),
using Eq. (A7) to select the initial conditions for q and
u. The integration is then stopped at y = yr, where the
right derivative u(y+
r ) is adjusted according to Eq. (A8)
to obtain the left derivative u(y−
r ) and thereby proceed
towards −∞. Only if the initial value of n is correct, q(y)
converges towards zero.
[1] W. Gerstner, W. M. Kistler, R. Naud, and L. Paninski,
Neuronal Dynamics: From Single Neurons To Networks
And Models Of Cognition (Cambridge University Press,
Cambridge, 2014).
[2] P. Dayan, L. Abbott, et al., Journal of Cognitive Neuro-
science 15, 154 (2003).
[3] G. Deco, V. K. Jirsa, P. A. Robinson, M. Breakspear,
and K. Friston, PLoS computational biology 4, e1000092
(2008).
[4] S. Ostojic and N. Brunel, PLoS computational biology
7, e1001056 (2011).
12
[5] A. N. Burkitt, Biological cybernetics 95, 1 (2006).
[6] G. B. Ermentrout and D. H. Terman, in Mathematical
foundations of neuroscience (Springer, 2010), pp. 331 --
367.
[7] E. Montbri´o, D. Paz´o, and A. Roxin, Physical Review X
[17] R. Zillmer, R. Livi, A. Politi, and A. Torcini, Phys. Rev.
E 74, 036203 (2006).
[18] V. Klinshov and V. Nekorkin, Chaos: An Interdisci-
plinary Journal of Nonlinear Science 27, 101105 (2017).
[19] M. Rudolph and A. Destexhe, Neurocomputing 70, 1966
5, 021028 (2015).
(2007).
[8] T. P. Vogels, K. Rajan, and L. F. Abbott, Annu. Rev.
[20] B. Dummer, S. Wieland, and B. Lindner, Frontiers in
Neurosci. 28, 357 (2005).
computational neuroscience 8 (2014).
[9] C. van Vreeswijk and H. Sompolinsky, Science 274, 1724
[21] S. Wieland, D. Bernardi, T. Schwalger, and B. Lindner,
(1996).
Phys. Rev. E 92, 040901 (2015).
[10] A. Renart, J. de la Rocha, P. Bartho, L. Hollender,
N. Parga, A. Reyes, and K. D. Harris, Science 327, 587
(2010).
[22] D. Angulo-Garcia, S. Luccioli, S. Olmi, and A. Torcini,
New Journal of Physics 19, 053011 (2017).
[23] E. Ullner and A. Politi, Physical Review X 6, 011015
[11] A. Litwin-Kumar and B. Doiron, Nat Neurosci 15, 1498
(2016).
(2012).
[24] R. Badii and A. Politi, Journal of Statistical Physics 40,
[12] J. Kadmon and H. Sompolinsky, Phys. Rev. X 5, 041030
725 (1985).
(2015).
[25] S. Luccioli and A. Politi, Phys. Rev. Lett. 105, 158104
[13] N. Brunel, Journal of Computational Neuroscience 8, 183
(2010).
(2000).
[26] F. Mastrogiuseppe and S. Ostojic, PLOS Computational
[14] T. Hayakawa
and T. Fukai,
arXiv
preprint
Biology 13, 1 (2017).
arXiv:1711.09621 (2017).
[27] S. Olmi, A. Politi, and A. Torcini, The Journal of Math-
[15] E. Ullner, A. Politi, and A. Torcini, ArXiv e-prints
ematical Neuroscience 2, 12 (2012).
(2017), 1711.01096.
[28] A. Ponzi and J. R. Wickens, PLOS Computational Biol-
[16] S. Ostojic, Nat Neurosci 17, 594 (2014).
ogy 9, 1 (2013).
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.